< < 
2013 December 11
Catherine Bulka, Anesthesiology
 Wants to demonstrate/examine group balance wrt baseline characteristics after matching using propensity scores.
 Calculated propensity scores for receiving regional anesthesia during surgery, as opposed to general anesthesia, based on several factors that anesthesiologists in the department deemed important in their decision making process, such as the age of the patient, duration of the surgical procedure, etc. I then matched patients who received regional anesthesia to those who received general anesthesia by their propensity score.
 consider calculating the "standardized differences" between the groups and plotting it. Something like this: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3472075/figure/fig01/
 Computing confidence limits for effects are probably in order, as opposed to hypothesis testing. It's not just large samples where hypothesis testing presents interpretation
problems; we generally like interval estimation.
 In building the propensity score it is appropriate to not be parsimonious (i.e., it is not approriate to remove variables from the model), and to allow effects of continuous variables to be nonlinear.
 There is controversy over matching vs. covariate adjustment for logit propensity score. Matching usually results in some patients being discarded, which is problematic.
 research question is to compare hospital length of stay in those who got regional anesthesia vs. regional. They used prop. score matching to match.
 Wants to demonstrate group balance wrt several variables.
 Instead of using propensity scores to select the patients, you could take all patients who met inclusion criteria, but use weighted propensity score analysis
 Catherine did try the standardized differences that Meng suggested, but the problem is the surgical duration variable still looks like it is too different. However, it turns out that the standard deviations she used were from the selected patients, while the paper recommends using the ones from the original group of patients from which the two groups were selected. She will reevaluate the quantities and reassess whether the groups are balanced.
 If the groups still don't look like they are comparable, could look at some of the methods Robert Greevy has written for matching in observational studies.
2013 December 4
Jo Ellen Wilson, Psychiatry fellow in psychosomatic medicine
 Needs AP and quote for VICTR proposal.
 Not a reliable way of distinguishing delirium from catatonia; plans to assess patients in two ongoing delirium studies for catatonia to see if there is a clinically relevant overlap between definitions
 Main study personnel will administer main delirium assessment (CAMICU  yes, no or unable to assess); Jo Ellen will administer, within 2 hours, delirium subtype screening (hyperactive, hypoactive, mixed or no delirium) and catatonia screening (scoring: first 14 items are administered; if patient meets two criteria, considered to have catatonia and an additional 9 items are done to assess severity)
 Delirium subtype screening and catatonia screening both look at previous 24 hours, whereas CAMICU is immediate, current assessment
 Technically, patients can't be considered catatonic if delirious
 Suggest doing entire BuschFrancis assessment even on patients who don't meet catatonia criteria on items 114
 Plan to look at individual assessments (two per day in ICU, one per day in wards), not full days or summary measures
 Treatments for delirium and catatonia are sometimes completely opposite  e.g., benzodiazepines avoided in delirious patients but recommended for catatonia, and vice versa for antipsychotics
 Aims 12 focusing on inhospital outcomes (agreement between delirium and catatonia assessments); aim 3 focuses on hospital LOS, longterm outcomes, and might require more complex modeling to account for possible relationship between delirium/catatonia diagnosis and treatment
 Aims 4 & 5 deal with medications  parent databases collect 24h totals, but Jo Ellen could get exact times and doses from StarPanel later
 Suggest splitting into two projects: aims 1 & 2 (perhaps 4 & 5, depending on what exactly is needed  Jo Ellen will discuss with mentor Wes Ely), then aim 3  longterm and clinical outcomes
 For sample size calculations, will need to think about specific differences hoped for (differences in proportion or test scores, for example); Jennifer will look at parent study enrollment numbers for average # enrolled per month at VUMC
 For aims 1 & 2, we estimate 50 hours (likely using kappa statistics for delirium vs catatonia agreement and model of catatonia score ~ CAM [yes/no/UTA] with repeated measures for delirium vs catatonia scores); if aims 4 & 5 are added, more time will be needed. Aim 3 will be addressed at a later date.
2013 November 20
Justin Gregg, Chang
 previously approved study with VICTR funding that may need some additional statistical support ($1100 left from VICTR grant)
 looking at time to recurrence/progression of nonmuscular invasive bladder disease
 retrospective data collected 20022011
 could possibly fit under collaboration with anesthesiology  suggest talking to anesthesiology department first
 suggest estimate of 60 hours to account for data management and analysis if VICTR grant is needed
Mayur Patel
 Prospero082713.pdf: Mayur's file
 systematic review. Want to know whether any metaanalytical technique is possible considering it is largely based on prepost retrospective caseseries data (mostly exposure only, partial cohort type studies; full 2x2 rarely available) and where the event rate is nearly always zero.
 Our current thought is that there is no analysis possible here – but wanted to confirm that, as I've read about continuity correction of 0.5 for zeroevent studies, but that seems more applicable for nonrepeated measure type samples where controls/treatment groups are distinct.
 http://www.prismastatement.org/statement.htm
 He provided more detail on the studies included in the systematic review. Our recommendation was that a metaanalysis would not be appropriate but directed him to prismastatement.org
2013 October 30
Jackie Shuplock, Pediatric Cardiology
 Prospective cohort (20072013) of cardiac surgery patients. Some received dex, others did not. Interested in examining the association of dex use with cardiac arrhythmias in the postop period.
 Needed input on how best to define the model of interest. Currently using a variable selection process based on univariate analyses. Concerned about differing significance with the inclusion/exclusion of particular covariates. Recommended using clinical knowledge and information from literature to guide the selection of covariates.
 Four anesthesiologists were involved in the surgeries (any need to adjust for anesthesiologist?) and there were a total of 783 arrhythmias recorded.
 Dex was rarely used 20072008. We recommended limiting the scope of data to 2008  2013 to avoid any kind of time bias. We also recommended adjusting for year of surgery in the model in case there were any changes in surgical protocol across time.
2013 Rocktober 21
Francheska Desravines, Meharry
From last clinic:
 study: The risk of early discharge following pediatric cardiac catheterizations in infants and young children.
 aim is to find relative risk of a minor or major complication occurring after 6hrs post cardiac catheterization procedure.
 .xlsx file saved as csv on ClinicsData
 Data are retrospective chart review on kids 04 with cardiac cath between July 2007 and 2012
 Want estimate of proportion of events with confidence interval
 To identify important factors, can fit a multivariable model with factors such as age, diagnosis (one v. two ventricle), source of pulmonary
blood flow
 Data on children over 4 would be important.
 There were 24 patients with major complications that would require major intervention
 For a 15 kg child, the estimated modelbased prob of major complication is 0.026 (95% CI 0.0150.046). logistic regression with weight
 Can see Vincent Agboto, who is a statistician at Meharry.
A new dataset is on ClinicsData.
 Describe the outcome (major complications) against some of the variables graphically, as in boxplots
<>
> binconf(x = sum(card$majorYN == "Major complication"), n = nrow(card),
+ method = "wilson")
PointEst Lower Upper
0.03259452 0.02217373 0.04767391
 The (unadjusted) estimated probability of a major complication is 0.03 (95% CI 0.02, 0.05).
 Recommend a forest plot showing the modelbased probability of major complications with CIs for each combination of ventricles and blood source.
Catherine Bulka, Anesthesiology
Questions:
 Looking to identify risk factors for developing pnemonia.
 In her poisson regression, she used the binary outcome of whether the patient developed pnemonia
 Discussed best way to measure outcome.
 Have been using a Poisson regression model with a log link function and persontime offset. The persontime was calculated as the time the patient was at risk for developing postop pneumonia (i.e. #days between date of surgery and pneumonia onset, death, or hospital discharge). I am now questioning whether another type of model would have been more appropriate. I was trying to avoid survival analyses because in this context, it seemed like rate ratios made more sense as opposed to hazard ratios. I also wanted to account for time at risk in the model, since some patients were lost to followup, which is why I avoided logistic regression. Is Poisson regression the appropriate choice for my data?
 Had been using a backwards elimination approach to select the final model, but there are a lot of issues with using statistical significance to identify which variables to keep in the model. Is there a better method? I have about 20 covariates that have been identified as potential confounders/effect modifiers.
 A reviewer suggested using bootstrapping for model selection. While I am able to create bootstrapped samples with 1,000 replications and replacement in SAS, I’m not sure how to use these samples for variable selection.
 We discussed data reduction techniques as an alternative to backward elimination. You could use a propensity score.
 Cluster analysis.
2013 September 25
Heather Jackson, Anesthesiology, and Kyla Stripling, Neurosurgery
 Planning small study (30 pts) examining a pain consult prior to surgery: "Does a preoperative pain management consult effect postoperative lumbar surgical outcomes?"
 Research protocol attached
 Contact Anesthesiology admin to see whether you have access to the existing collaboration with Department of Biostatistics (Jon Schildcrout and Matt Shottwell)
 Recommend trying to get a mentor who has some research experience
 One of the main outcomes is the ODI, a disability index.
 May need to consider the sample size you would need to have power to show the differences you are interested in showing.
 To calculate the required sample size, consider the smallest clinically meaningful difference in ODI and also the distribution of the outcome, including standard deviation.
Jun Dai, Epidemiology
 Jun is responding to reviewer comments to her manuscript.
 She has a survival model with covariates for the amount of food eaten. There are three types of fruit, and three corresponding variables.
 The original research question was about the effect of total fruit consumption on survival. However, the reviewers asked for analysis which considered the amount of fruit separately by type of fruit. Then the reviewer wanted a test of whether the effects of the different fruits are the same.
 We recommended that she respectfully explain that while his questions are interesting, they are beyond the scope of the paper.
 Additionally, the test the reviewer recommended is a test for interaction, and doesn't address his own question. :[
 We recommend that she give some descriptive summaries of graphs showing the types of fruits in her data to answer this reviewer's question without deviating from the paper's original focus.
Vaibhav Janve, Institute of Imaging Science
* Hakmook Kang works with imaging data and with Vanderbilt University Institute of Imaging Science, and we recommend he contact him.
2013 September 18
2013 September 11
Scott McLaurin, Rehab Services
 Does physical/occupational therapy affect hospital length of stay?
 Stroke patients
 Patients were assigned a frequency of visits. The patients were assigned to be in the "meet frequency" group by a systematic algorithm. The patients who were in the "experimental" group were ensured to get their recommended therapy visits. The other group got standard of care.
 Want direction for designing a new study
 Could store the data in redcap
 Could think about using a randomized list of group assignments. There is a mechanism for this in redcap.
 Need to think about factors that would influence length of stay. You would need to collect these quantities and can adjust for these using a multivariable analysis.
 We think their project might come under Dan Byrne and Hank Dominico's project.
 Could go to redcap clinic to help ensure you set up the data the best way. Discussed data quality checks you can build into your database. See https://www.mc.vanderbilt.edu/gcm/rate/index.php/course/view/id/0168
 If you are not covered under Dan's project, and funds are available, can use BCC (Biostatistics Collaboration). Contact Rameela.
Trisha Pasricha, medical student; mentor Clements.
 study in GI Laproscopic Surgery
 in the experimental design stage
 looking at patient scores on the Beck Depression Inventory before bariatric surgery and scores 1 year following surgery to assess any correlation between bariatric surgery and depression.
 The Beck Depression Inventory is a 21 question survey that results in a numerical score of 063, with higher scores indicating more severe depression and lower scores indicating minimal depression.
 have several questions about the optimal experimental design (debating between 2 different designs in particular) and how to power the project accordingly.
 Plan to readminister Beck one year after surgery.
 Population who have premeasurment is people who have bariatric surgery.
 Discussed problems with inference based on prepost designs.
 Need to consider the best control population to make the inferences that are of interest. Could consider a general patient group that had any surgery or a specific type of surgery, or a group of obese patients that did not have bariatric surgery. Would want to ensure that patients who did not get the surgery are not systematically different from the patients who did have the surgery. Examples of this would include making the control group consist of the people who were ineligible for the surgery because they were unable to follow the pretreatment regimen.
 Discussed importance of minimizing bias that could be caused by people not coming to the oneyear followup visit. This could bias the results towards patients who are already are not too depressed to come to the followup visit. Making the followup assessment webbased would mitigate this, but you would also want the pre and post surgery assessments to use the same modality.
 Also discussed looking at the correlation between amount of weight lost and depression
 Group wants to consider whether they should select patients to include in the study based on degree of pretreatment measurement. The concern is that there might not be enough patients with lower depression. If you include all patients, then you would get the maximum possible number of patients with low depression, and there wouldn't be value in excluding patients with higher depression.
 Discussed considering confounding factors like social support. These could be controlled for in a multivariable analysis. The degree of complexity of the relationship (and thus the model required) would inform the sample size required.
 We would need estimates of the standard deviation of the BDI score and its distribution to calculate the number of patients required.
2013 August 28
Angela Joshi, resident, VA
 observational study of patients who undergo coronary angiogram via the radial artery to determine factors that can predispose to radial artery thrombosis following angiography.
 wants to calculate number of patients needed
 Planning a prospective observational study
 Have a list of factors that they want to identify the association with having thrombosis
 Have a limited period of time over which to collect patients
 Estimate around 1200 patients could be collected
 Estimate about 2 percent of patients will have an event, for about 24 events
 We advised that with this number of events, you will not have much power to learn much
 Recommended that they rank the variables in order of their research priorities. We will calculate power based on their top priority variables.
2013 August 21
Francheska Desravines, med student at Meharry
 study: The risk of early discharge following pediatric cardiac catheterizations in infants and young children.
 aim is to find relative risk of a minor or major complication occurring after 6hrs post cardiac catheterization procedure.
 .xlsx file saved as csv on ClinicsData
 Data are retrospective chart review on kids 04 with cardiac cath between July 2007 and 2012
 Want estimate of proportion of events with confidence interval
 To identify important factors, can fit a multivariable model with factors such as age, diagnosis (one v. two ventricle), source of pulmonary blood flow
 Data on children over 4 would be important.
 There were 24 patients with major complications that would require major intervention
 For a 15 kg child, the estimated modelbased prob of major complication is 0.026 (95% CI 0.0150.046). logistic regression with weight
 Can see Vincent Agboto, who is a statistician at Meharry.
JoAnn Alvarez, for Center for Surgical Quality Outcomes Research
 Outcome is workup quality:
None Incomplete Complete
6002 1991 1248
 Plan on using prop odds
 In this case, I think we can assume exchangeable cov structure.
 County info: 1867 unique counties in the data, with 9241 patient records. Here is the distribution of number of observations for each county:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 40 43 44 45 46 47 48 49 53 55 58 59 60 64 71 102 139 161
671 368 212 149 73 59 49 36 30 24 18 15 15 11 15 10 12 1 9 8 5 5 4 8 1 3 4 1 3 3 5 6 3 1 1 2 1 3 1 1 1 1 1 3 2 3 1 1 1 1 1 1 1 1 1 1
So there were 671 counties that had only one record in the data, 368 that had 2 obs, and the county that had the most representation in the data had 161 observations.
 Could use GEE or a random intercept for county to account for practice variation by location.
 Let m be the number of observations per cluster (subject is the cluster in many longitudinal studies), and n is the number of clusters (subjects). So m is the number of obs per county, and n is the number of counties.
 GEE
 Efficiency issues
 Choice of weight mat: In this case, I think we can assume the data have an exchangeable cov structure. Loss of efficiency should be very very small (Liang and Zeger 1986) with use of independence working covar when truth is exchangeable. Or we could use exchangeable weight mat.
 correlation of covariates within subjects: when the covars are betweensubject (do not vary within a person (county here)), negligible efficiency loss. Actually, would need to check this.
 cluster sizes: we have small m (number in each cluster), which should mean small loss of efficiency (see slide 47 of mod 4)
 If we want to do gee with nonindependence weighting, could use geeglm or geese (see slide 87 of mod 4). Can use rob cov in Frank's rms package. This is GEE with independence weight matrix / working covariance.
 Gee requires large sample, but we have that. (n>>m Number of different counties is much bigger than the number of obs per county)
 GEE requires MCAR.
 Disadvantage of using gee would be that you can't use LRT since you are not specifying the full likelihood. Can use wald tests.
 Performance with sparse clusters? Is this separate from efficiency issues? In this case, we do have sparse clusters and many clusters of size 1. The issue is variability in the cluster sizes. Jon thinks this is a validity issue.
 GLMM (generalized linear mixed effects model)
 GLMM requires MAR.
 Can implement via clmm in ordinal package. logit link gives prop odds.
 With glmm, since it's a conditional model with a nonidentity link, we cannot get populationlevel contrasts. The contrasts are subjectspecific, and I believe they approximate the poplevel values.
 The marginal parameters are smaller in magnitude than the conditional parameters. Can think of the conditional params as "inflated" marginal params?
 There is actually a formula for the relationship between the conditional and marginal parameters in terms of the variance of the random effect (Zeger 1988).
 Example of interpretation of subjectspecific effects "exp β represents the ratio of the expected odds of respiratory infection for an individual with vitamin A deficiency to that same individual (with same covariate values but) replete with vitamin A"
 Can check PO assumption using chi square lrt test (difference in deviance) between the prop odds model and the cumulative logits model.
2013 July 31, August 7 and 14
no clients
2013 July 17
David Osborn, Urology
 We estimate that this would take about $3500 for a VICTR biostat support request
 Retrospective study looking at urinary incontinence in patients with traumatic brain injury.
 Want to estimate the incidence of incontinence in this group
 Outcomes are death, hospital los, 3 outcomes
Breanna Michaels and Jennifer Morse
 *Data attached below*
 looking at patients who received non depolarizing neuromuscular blocks during surgery and their incidence of post operative residual curarization (PORC) (means leftover block) in the PACU.
 Main outcome is train of four, and whether it is greater than 0.9, which is the designation for PORC. Train of four is a ratio of two weakess measures.
 Recent studies have shown that patients with PORC have a higher risk of respiratory and nonrespiratory post operative complications.
 want to know the proportion of patients with PORC, as well as the factors that affect the occurrence of PORC.
 Factors may include gender, type/dose of nondepolarizing block, or antibiotics given during surgery.
 Discussed whether to use the train of four ratio or the dichotomization of PORC, which is greater than 0.9,
 Want to see if there are systematic differences between those who had outcomes measured and those who did not.
 Could fit a regression model to identify the important factors.
 Recommend getting a confidence interval for the proportion with PORC
 Try plotting a histogram of the outcome variable
 Discussed events per variable. If you use a logistic regression, dichotomizing the outcome, use the number in the smaller group to determine the number of variables you can fit in your model. A rule of thumb is to have at least ten events per variable (parameter in model).
Roop Gill
 We estimate that the first manuscript, which will encompass two outcomes DVT and seroma will require $4000.
 Tummy tucks
 Have a tenyear cohort of patients
 Want to see if skinonly abdominoplasty
 Want to calculate the power they have for their questions with the data they have.
 One of the main outcomes of interest is deep vein thrombosis, which is rare, so you will need a very large number of patients
2013 July 10
Bill Wester, ID/VIGH
 Bill first consulted the group on 6/19; he would like to continue discussion his proposed study "LongTerm Renal Outcomes Among HIV1 Infected and cARTTreated Adults in South Africa" during biostatistics clinic.
 Need an estimate of hours/cost for VICTR biostatistics support.
 Last time we postponed the estimate to wait for more information on the condition of the data.
 Now we have more information on the completeness of the data. Dr. Wester is going to look through the data dictionary and the data itself to assess how clean it is.
 Dr. Wester has decided to concentrate on Aims 13.
 We estimate that this will require 100 hours of support.
 Additional data
 1. How many total patients in the cohort? 2500
 2. How many (what %) have baseline urinary protein dipstick results? 2000
 3. How many (what %) have longitudinal urinary protein dipstick results/measurements; and at what frequency (I suspect they are irregularly done; but can we see how many (what %) have 1 year post ART and ? 2 year post ART values recorded/captured? 75% have a second measure
 4. How many (what %) have baseline creatinine results? The majority.
 5. How many (what %) have longitudinal creatinine results/measurements; and at what frequency (I suspect they are irregularly done; but can we see how many (what %) have 1 year post ART and ? 2 year post ART values recorded/captured? 75%
 6. How many were ARTnaïve? (total numbers); as it seems best to look at our main outcomes (specific aims per our protocol) among patients who initiated ART (and were ARTnaïve at the time of entering the longitudinal cohort)? Almost 95% are HAART naive at baseline.
 Aims: Compare rates of proteinuria
 Has dipstick measurements on about 80% of patients.
 An alternative to a longitudinal model would be using a summary measure such as slope for each patient
Raj Keriwala, Fellow, Pulmonary and Critical Care
 data from EDEN and FACTT trials, part of the ARDSnet.
 ARDS Acute respiratory distress syndrome
 Low title volume mechanical ventilation is the only therapy that has been shown to be helpful for ARDS
 Want to look at vasopressor use over time and compare by cohort
 retrospective study looking at the use of vasoactive agents in these cohorts and associated characteristics for any identifiable relationships.
 will be bringing the full datasets in csv format with me tomorrow
 What will be the best way to organize this previously collected data to facilitate analysis?
 Discussed developing a set of inclusion criteria for the combined data to get a comparable group
 Allow for time to implement a policy change by excluding a period of time, say, 3 months, immediately after a change
 Recommend scripting data manipulation to ensure reproducibility
 Will need to make sure the appropriate variables have the same variable names, the same variable types, and same variable categories/levels/formats.
 Our department supports R. There is limited support for Stata. Few department members would be knowledgeable in SPSS.
 Discussed whether Dr. Keriwala has access to our dept. through an existing collaboration.
 Based on the data collected, what types of relationships can we look at in both cohorts and compare?
 Is REDCap the right software to use for this type of project? We think it will not be worthwhile to put all the existing data in redcap.
Jennifer Morse, clinical trials specialist
 research study on PACU delirium prevalence
 How do I appropriately deal with the missing pain score values?
 RASS Score is ordinal data that is broken down into 2 groups (hyper and hypo active) based on clinical significance. Is this appropriate?
 Would a chisquare analysis be appropriate? Or an odds ratio? Or a logit analysis to compare odds ratio? Are there any other recommendations for performing the analysis?
 Is there a relationship between pain score and RASS scores (>=1 vs <=0)?
 Is there a relationship between 1st column (did pt wake up delirious) and gender?
 Is there a relationship between 1st column (did pt wake up delirious ) and pain?
 Is there a relationship between RASS scores (>=1 vs <=0)? and gender
 Is there a relationship between CAM positive and gender?
 Is there a relationship between CAM positive and RASS ((>=1 vs <=0) and gender?
 Discussed the characteristics of the outcome.
 We think it is a bad idea to combine negative scores with zero scores, since both negative and positive scores are "bad." Only two of ~400 are positive. We requested Jennifer bring us the frequencies of the different levels of this variable to aid in the decision of how to model it.
 We recommend the investigator come to clinic so we can get more information about the clinical hypotheses behind the analyses requested
 There is a Monday Anesthesiology studio at 4 that may be a helpful resource.
 Part of the complexity of this analysis is that the CAM and RASS are not simply ordinal.
2013 July 3
John Koethe, ID/VIGH
 John emailed the group on 6/20; he will briefly present some data collected in Zambia. He has pre and posttreatment serum levels of several inflammation biomarkers from a cohort of 20 malnourished, HIVinfected adults starting antiretroviral therapy. He would like to get a quote for VICTR biostatistics support.
 We estimated ~ 40 hours for this project.
2013 June 19
Bill Wester, ID/VIGH
 Bill emailed the group on 6/17; he would like to discuss his proposed study "LongTerm Renal Outcomes Among HIV1 Infected and cARTTreated Adults in South Africa" during biostatistics clinic.
 Aim 1: Determine prevalence and incidence of proteinuria among HIVinfected adults in the pop. Proteinuria is a risk factor.
 Aim 2: Compare rates and time to development of ESKRD (CKD4) among cases and controls in the cohort, where case/control is having proteinuria.
 Aim 3: Compare rates of allcause mortality between cases and controls.
 Aim 4: Estimate the impact of of cART on preventing adverse renal outcomes.
 Proteinuria isdefined as having a urinary dipstick of 1+ or greater. The possible values are 0, 1+, 2+, and 3+.
 Recommend avoiding framing study as "case/control" and instead preserving the information in the dipstick result.
 Could use creatinine values to impute missing dipstick results. There may be few instances of missing dipstick but nonmissing creatinine.
 One of the main objectives is looking at the timedependent relationship between proteinuria and eGFR level/development of an eGFR event.
 Ideally we would want to use time to event, but the outcome is not regularly measured. This would be an argument for treating this as a longitudinal study.
 We will estimate the hours required after we have more information about the condition of the data.
 Existing cohort size ~ 5000.
2013 June 12
Catherine Bulka, Shane
 sample size calculation for a study on the prevalence of redgreen color vision deficiency
 The proportion of colorblindness in men in the US is about 7%
 The study is ongoing and is based on a convenience sample of over 300 patients, many of whom are females.
 Want to estimate the proportion of male physicians at Vanderbilt with color blindness and the proportion among other male providers at Vanderbilt.
 Can estimate the difference in proportions between the physicians and other providers. Can see if the CI for proportions in each group includes the value for the general population (7%).
 Can power based on width of confidence interval.
 Recommend a better sampling scheme. Could use a simple random sample.
 Want to ensure high response rate.
 PS software does power and sample size calculations and is provided for free here: http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/PowerSampleSize.
 To compare the two proportions, you need many more subjects. Can increase power by making the ratio of physicians to other providers closer to 1:1.
 Just to estimate the proportion within +3% if the true value is 7%, you would need about 278. For +5%, you would need 101.
 Online calculator is here: http://epitools.ausvet.com.au/content.php?page=1Proportion
2013 June 5
No client investigators.
2013 May 22
Kirk Kleinfield
 Wants to identify predictors of epilepsy for an abstract
 Recommend he focus on variables that he has nonmissing data for and that are a priori identified by experts.
Kurt Niesner, VA GRECC, neurosurgery
 retrospective study of 58 patients who received a fiesta MRI
 Exposure is compression: no compression, possible compression, and definite compression, and also whether or not the patient's symptoms matched the presentation
 Wants to evaluatie the efficacy of preoperative MRI in predicting postoperative outcome in patients with Trigeminal Neuralgia undergoing Microvascular Decompression surgery.
 This study may be able to evaluate the association between compression, symptomaticness, and outcome.
 comparing the presence and locus of trigeminal nerve compression during preop MRI to early (1st followup visit) and late outcome (1 to 5 years postsurgery).
 Early outcome is improvement in pain within one month, measured on a scale of 13 inclusive. Late outcome is Brief Pain inventory, which is on a scale of 0150 inclusive.
 Patients are almost all in severe pain before the surgery.
 We recommend considering whether there are important factors that would likely influence your outcomes.
 Could run a proportional odds model with the short term pain improvement as the outcome with the compression and symptomaticness as covariates, as well as their interaction.
 The mentor could apply for a VICTR voucher for biostatistics support
 discussed the need for a letter of support for amount above $2000.
 We estimate that this project would require about $4000.
 The long term outcome is problematic because of the proportion of missingness.
 To estimate an adequate sample size for this type of study, you need to consider the smallest effect that you consider important.
2013 May 15
Roshi Markley
 Needs a quote for VICTR biostat support
 had a design studio with Frank Harrell and Ayumi Shintami who have helped with the design of the study.
 Looking at hospital free days following assessment of severe aortic stenosis. Not sure about distribution  similar outcome, delirium/comafree days, is quite bimodal, but not sure if hospitalfree days over one year will be similar.
 Want to identify what factors are associated with lower hospital free days and develop a multivariable model to predict individuals that are at risk of poor outcome, including therapy of TAVR, surgery (gold standard) or no therapy.
 About 30% mortality rate expected during first year (50% among patients with no therapy)
 Patients mainly in 80s and 90s
 Secondary outcomes: oneyear mortality, major vascular complications (eg bleeding)
 Planning to submit manuscript
 Estimate 70 hours for work, manuscript prep and anticipated revisions
2013 April 17
Jeremy S. Pollock, Internal Medicine, Wes Ely, Mentor
 Has VICTR biostats request
 aim of this retrospective study is to determine the incidence and risk factors of delirium in survivors of cardiac arrest treated in the CVIUC at Vanderbilt University. We hypothesize that delirium, as defined by the CAMICU, will be present in more than 25% of patients and that use of benzodiazepines, age, shock, and time to return of spontaneous circulation will be associated with delirium.
 Retrospective data have been collected on all survivors of cardiac arrest who were been treated with therapeutic hypothermia at Vanderbilt since 2008. We will use this database to assess the incidence of delirium and it’s associated risk factors in postcardiac arrest patients who survive to hospital discharge.
 Statistical analysis will include determining the incidence of delirium in the cohort who survive to rewarming. Baseline demographics will be compared between patients who experience delirium and those who do not. Multivariate regression analysis will be undertaken to determine the association between delirium and predefined risk factors (age, use of benzodiazepines, shock, comorbidities, time to rosc, PEA/asystole) and any differences in baseline demographics with pvalues<0.10 (assuming the incidence of delirium is high enough to provide adequate power to include these).
 Want to estimate the incidence of delirium after the first 24 hours of therapeutic hypothermia. Also want to estimate time in coma and number of coma free (alive) days. Also looking at length of ICU stay. Want to identify risk factors for (first transition to) delirium (in logistic regression).
 Want to address probably all of these research questions in one manuscript.
 Could address delirium using time to delirium or using a Markov transition model.
 Would like to request Jennifer Thompson because of her expertise. This would need to be approved by Frank.
 We estimate this would require about 50 hours of biostatistics support. We discussed the cost sharing information.
2013 April 3
Harry Wright, Otolaryngology
 Estimate 35 hours of analysis for purposes of VICTR biostatistics support application.
 See DataTransmissionProcedures
 Conducting a retrospective chart review for 71 patients who had surgery to remove a skullbased lesion. The surgery places the facial nerve at risk. All these patients had to have a repair procedure.
 The main question involves the impact of the time between the surgery and the repair on recovery of function. The functionality is measured on a 16 scale. Of interest is whether the patient can voluntarily close his eye (3 of 6). All of these patients have values of 3, 4, 5, or 6.
 Want to consider patient's age, surgeon,
 The time between the surgeries is thought to be random, which is good for making inference on the impact of time on the outcome.
 Proportional odds would be a good choice.
 You don't really have power to adjust for surgeon id with fixed effects.
 Would want to look at the distribution of the outcome: frequencies of each 3, 4, 5, and 6.
 If you do think that surgeon id is an important factor on the outcome, you could model this appropriately using random effects (a random intercept for surgeon id). The purpose in your research objective is not to estimate an effect for a particular surgeon, but to appropriately account for it in the model so that your inference about the variable of interest, time, is correct. However, estimating the variance of the random effects may be hard with only five surgeons.
 Another way to account for this extra variation due to surgeon is to adjust for the surgeon's case mix.
 May want to adjust for the severity of each case.
Catherine Bulka
 project on PACU length of stay and regional anesthesia.
 based on clinic comments decided to use a stratified Cox regression model with sandwich estimator (because some patients in the dataset had more than 1 surgery).
 Outcome of interest is time (in minutes) to successful discharge from the PACU.
 Predictors are regional anesthesia (yes/no), ASA Class, duration of surgery, and patient’s age based on previous studies that have looked at PACU length of stay.
 also matched patients on CCS Grouper (based on CPT codes) so that is how the Cox model is stratified.
 I tested the predictors in the model to see if they met the proportional hazards assumption using 3 methods: looking at them graphically, looking at the Schoenfeld residuals, and including timedependent versions of the covariates in the model. For all covariates (regional anesthesia, ASA Class, duration of surgery, and patient’s age), the proportional hazards assumption was violated.
 Because I already am stratifying by CCS grouper, I think my best option may be to use an extended Cox model rather than stratifying on even more variables, but I’m unsure of how to interpret the hazard ratios if I use this approach.
 We think the departure from the proportional hazards assumption may not be too serious. Whether the assumption holds would not impact the interpretation of the results but rather whether the method (Cox regression) is valid.
 If the Schoenfeld residuals trend toward zero, you need to use an accelerated failure time model.
 The Cox model tends to fit better for chronic conditions, where the impact of the exposure is more constant over time.
2013 March 20
Catherine Bulka
 working on a project looking at hypoxemia in children during surgery & 30day mortality.
 includes all children who had some type of anesthetic care between 2000 and 2011, so there are some patients who had multiple surgeries.
 I’m planning on creating a model with a binary outcome for 30day mortality and using oxygen saturation (a continuous variable) as a predictor along with other predictors like ASA class, age, race, sex, procedure type, preexisting conditions, etc.
 Because there are repeated subjects, I was thinking of using a GEE model, but I’m unsure of what kind of correlation structure to use.
 is using a GEE model the best approach
 if I do use a GEE model, how do I figure out what kind of correlation structure to use?
 If data are large enough, should be less critical in choice but recommend Independent correlation structure, especially since some covariates will change over time (from repeated measures).
Roop Gill, Marcia Spear, Plastic Surgery
 Cranial facial data base (10 yrs). Data is stored in REDCap
 Age at which head was fixed, type of suture, race.
 Outcome: reoperations: redone, voids, voids/reshaping, minor and whether there is an optimal age at which to do the initial surgery
 Need for initial surgery diagnosed anywhere from birth to 6 months of age.
 Abstract deadline April 3. Applying for VICTR voucher to get the analysis done in time to submit the abstract.
 Question 1: Is there a reop or not. Question 2: Does age at initial operation correlate with type of reoperation.required
 Might be interested in the age at which the reop takes place
 For abstract, look at descriptive statistics * VICTR estimate 60 hours
2013 March 13
Jayant Bagai
 Needs estimate for VICTR
 Health outcomes research study of veterans at the Nashville VA hospital.
 The objective is to compare death, MI, stroke, repeat revascularization and major bleeding in 4 cohorts of patients startified by a combination of arterial access site (radial vs. femoral artery) and type of anticoagulant used for coronary intervention (heparin vs. bivalirudin). One of the key comparisons the incremental benefit of a radialangiomax strategy in reducing bleeding compared with a femoralangiomax strategy .
 The 4 cohorts are:
 femoral + heparin n=293
 femoral + angiomax n= 255
 radial + heparin n = 469
 radial + angiomax n= 489
 We know from previously published data that
 Radial access reduces bleeding compared with femoral access and
 Femoral angiomax reduces bleeding compared with femoral heparin use.
 What is not known is if the combination of radial and angiomax is superior to femoral angiomax and if so by what magnitude and if that translates into other benefits such as reduction in mortality.
 The following analysis need to be performed
 Comparison of baseline variables in 4 cohorts
 Comparison of outcomes between cohorts in composite and individual outcomes: length of stay, inhospital death, MI
 Multivariable logisticregression models to identify predictors of major bleeding.
 Cox regression models to determine association between bleeding and death, MACE (major adverse cardiovascular events composite of death, MI, stroke and unplanned/urgent revascularization), length of stay and readmissions. * Need to consider other systematic differences between the four cohorts.
 May want to control for number of stents placed
 Data includes repeats of individuals
 May be underpowered to detect differences in overall survival.
 We estimate that this analysis would require about 60 hours of statistics work.
 Dr. Bagai also has a study involving a device used to apply pressure for venus closure. They want to compare three devices and manual pressure on bleeding, patient satisfaction, cost. They need to know the sample size required to provide adequate power. If the main outcome is bleeding (y/n), then the sample size will depend on the number of bleeds
Alison Kemph, Hearing and Speech
 looking at contamination in ear molds in children. What is the percent of ear molds with bacteria? Need to know the sample size required to estimate the proportion. If the true proportion is about 0.5, would need about 43 patients to estimate the proportion to precision of +0.15. If the true proportion is 0.7, you would only need about 36 patients to get the same precision.
 May want to record the date of the culture and the child's age.
 We estimate that this project would take about 40 hours of work.
 If the scope of the manuscript/project is limited to only estimating the proportion with confidence interval, it can be done in 20 hours.
 See DataTransmissionProcedures
Dupree Hatch, Neonatology
 Studying adverse events in intubations
 Describe the rate of adverse events
 Have new standardization procedure to implement
 Need power to detect difference between proportions. For 80% power to detect a reduction of 50% or greater if the baseline rate is 30% with 95% confidence, you would need 134 patients per time period.
2013 March 6
Jennifer Morse, Clinical Trials Specialist, Periop Clinical Research Inst.
 Survey questions sent to Vanderbilt and MTSA and want to compare the results. (Vanderbilt=74, MTSA=58) 88 and 63% response rates.
 Want to see if there is a difference in how the two groups answered the questions.
 Survey of nurse anesthesists about opinions on classes.
 Students were also asked to identify the five most important aspects out of 22 and least important aspects.
 Plan on making a basic bar graph for each question but am interested if there are further suggestions for organizing the results, or statistical tests that should be performed.
 Interested in analysis techniques for displaying results for the ranking question: Participants were asked to pick their top 5 and bottom 5. How can this data be presented collectively? Most common answer? Weighted results?
 A good way to display the individual questions would be a dot plot. You could, for the graph, dicotomize all 22 questions into either highly or critically important and give one "line" with two dots giving this percent, one each for CRNAs and SRNAs. Or you could give the means
 Can sort the items in the chart by either the CRNA's score (proportion or mean), or the SRNA's score, or sort by the difference between the two scores.
 Could do Wilcoxon RankSum Test for each individual test to check for a difference between groups (for one time point).
 Consider: What information do you have on characteristics of survey responders vs. nonresponders? What evidence do you have that the response rates were the same for Vandy and MTSA?
2013 February 27
Eileen Duggan, General Surgery
 Laproscopic vs. open pyloromyotomies in terms of cost, LOS, complications, etc
 So far have done univariate analyses (ttests or Wilcoxon, Fisher's)
 Strongly recommend multivariate analyses (linear or logistic regression) rather than univariate; may need to transform variables (LOS) to fit model assumptions
 Create list of main hypotheses (research questions) and confounders before doing model fitting, working with mentors and looking at literature to get a plan together
 283 patients total, 127 vs. 156 (ten different surgeons; surgery type is nearly always determined by surgeon more than other factors)
 Might use surgeon as a random effect in mixed effects model  no estimates, but adjust for surgeon
Catherine Bulka, Anesthesiology
 help analyzing length of stay in the PACU (outcome) and regional vs. any other type (general) anesthesia (predictor) on a large dataset. We have identified potential confounders such as surgical procedure, ASA Class, age, etc. We wanted to try to match patients on CPT code to account for the different surgical procedures. The only way I know how to do a matched analysis is using conditional logistic regression, but the outcome for this analysis is continuous. Is there a way to do a matched analysis using linear regression?
 LOS has a strange distribution that is best handled by the Cox proportional hazards model. And if you have more than, say, 1% of the patients die (which rightcensors their LOS) you can censor them on the day of death. That way you don't give credit for a short LOS for those who died. The outcome then becomes "time until successful discharge".
 With the Cox model you can stratify on CPT code group with no matching necessary. It would be nice to have at least 200 subjects per stratum if possible, so probably need fairly broad groups of CPT codes.
 There are 198,712 surgical cases and 3,300 unique CPT codes. Will group the CPT codes into broader categories.
 Looking at all surgeries at VUMC since 2008. Hypothesis is that patients receiving regional anesthesia will have shorter PACU stay than those getting general.
 Patients can be included for multiple surgeries, so need to account for withinpatient correlation.
 First step is to check death rate  if more than 12%, need to do Cox regression, but harder to control for repeated measures. Poisson regression is another possibility (count number of minutes), possibly including indicator for death in the model but would need to interact that term with everything else.
 Lots of potential confounders to consider (BMI, surgery type and length, ASA class).
 Regional vs. general anesthesia is very often decided by type of procedure  not sure you'll be able to tell which is the real cause of any difference. Maybe focus on surgeries where patients have a good chance of getting either type.
 Only ~20,000 out of 198,000 surgeries had regional anesthesia.
 Check with anesthesiology collaborators  JonathanSchildcrout, DavidAfshartous, MattShotwell
2013 February 20
Yuri van der Heijden, Division of Infectious Diseases. Mentor: Tim Sterling
 K08 submitted, score 'in range' for resubmission in May
 Looking for support from BCC.
 Drug resistance in TB patients (specifically, fluoroquinolone) in cohort in South Africa. Aim 1: describing yearly incidence of fluoroquinolone resistance 20072011. Aim 2: Describe risk factors for FQ resistance, primary of those is HIV status. Aim 3: Death/culture conversion as outcomes.
 Bryan Shepherd has been involved in pregrant preparation. Based on that, It does not appear that he needs VICTR funds for pregrant submission work. We recommend that he first talk with Bryan as to how much would be required from VICTR to perform the analysis. [BRYAN: I have discussed the analyses with Yuri. As part of his Kaward he proposes learning how to do these analyses and implementing them with my supervision. Therefore, I will have protected time (%effort) on his Kaward, and he will only need 100 hours of support from the Biostatistics Core, which will be used in case Yuri needs some additional help with specific parts of his proposal. This is, of course, a rough estimate, but is the amount budgeted into his application.]
2013 February 13
Michael Dewan, Neurosurgery. PI: J Mocco
 developing a RCT comparing seizure frequency and clinical outcome in patients with subarachnoid hemorrhage who are given levetiracetam (treatment) or no drug (control). Plan to observe for 30 days after discharge.
 Wants to discuss power calculations, randomization, and a couple other topics
 Currently, prophylactic meds given if patient has history of seizure; otherwise, just physician preference
 Between 820% patients experience seizures after subarachnoid hemorrhage, typically in the first few days. Up to 1/3 of these are before they get to the hospital.
 Plan to randomize patients with SAH who have not yet had a seizure (documented absence/no history of seizure). Patients with previous history or who seized in the field will be excluded.
 Interested in incidence of seizure and modified Rankin scale (6point scale of ability to perform daily functions) at discharge and 30 days.
 At VUMC, see approximately 200 patients/year with SAH; estimate about 10% of patients have seizures.
 Ideally want to balance groups in terms of HuntHess score (15); suggested minimization randomization using a web program (JoAnn has experience with this using VUMC investigational pharmacy). Stratified randomization also an option, as is basic randomization.
 Do not currently plan to blind patients or nurses, but don't expect any differences in monitoring. Recommend some sort of placebo and blinding if possible. VICTR may be a good resource for this, maybe using IV formulation rather than pill.
 Recommend applying for 20hour VICTR voucher for initial study planning, sample size, analysis plan.
Joyce CheungFlynn, Surgery
 Wants to submit manuscript to Journal of Thoracic Surgery, which requires signature from statistician signing off on statistical methods. Recommended VICTR voucher for someone to look over data, rerun numbers and review manuscript; statistician could be coauthor.
Mick Edmonds, PMI
 Has data and would like to answer several sensitive questions. Suggested speaking with CQS statisticians, who have expertise in specific types of data/questions. JoAnn can send email to set him up with a contact (maybe WilliamWu ?) and get accurate quote for number of hours involved. In the meantime, plans to apply for 60hour VICTR voucher to get started.
2013 February 6
Catherine Bulka, Anesthesiology 
< < 
* NSQIP data  looking for association between postanesthesia drug and postop pneumonia * ~ 1500 patients with 1600 observations. 65 got pneumonia; * Interested in estimating relative risk of pneumonia; how to handle modeling with repeated records for patients * 43 had 2 surgeries * Potentially use frailty model in a survival analysis rather than Poisson. * Some concern regarding who is given the postop drug and who is not given it.
Lawrence House, Anesthesiology
* Ability of cardiac troponin's to ID those with postop MIs. * Objective 1: ID variables which are associated with a postop cardiac troponin ordered * Objective 2: How do cardiac troponin values correlate with MIs, etc. * Does not include cardiac surgeries but does include vascular surgeries. Suggest removing vascular surgeries. * Impact is from a resource uttilization point of view. * There are duplicates in the data. Recommend using the sandwich estimator for SEs in modeling.
William Sullivan,, 4th year medical students
* Database of medical student notes. * Compare last year's and this year's 2nd year medical student notes * Interrater reliability between four coders who will be coding the 'sophistication' of the notes. * Coding anakyzes language using rating from 0  2 that assesses arguments and language of arguments on semantic complexity * Will not be able to assess whether the diagnosis is correct * Coders will be trained. * Need sample size requirements; may be driven by time/resource limitations. Need to deidentify the notes so need some estimate on a sample size. * There will be multiple notes from the same student but all from different patients. * Suggest pulling a small sample and looking at interrater reliability for estimates on how long it will take and how different the raters are. Rather than use kappa could use intraclass correlation (ICC). * Ultimately looking for publication but also impacting training policy in the SOM.
2013 January 30
Andre Marshall, General Surgery
 Planning on a logistic regression model
 Outcome is 30 day readmission after apendectomy for acute apendicitis
 1800 pediatric patients, 103 of which had readmissions within 30 days
 Wants to identify risk factors for readmission. Specifically insurance status.
 Perforation status is already known to be a factor.
 Also wants to describe the group of patients who are readmitted
 Has thirty days follow up for all patients, and is also interested in time to readmission.
 Insurance is a substitute for socio economic status. categories are are public insurance and private insurance
 could have interaction between perforation, length of stay, and insurance status.
 Planning on adjusting for ethnicity. This may wash out the effect of the insurance because they are often highly correlated.
 We don't recommend computing power after you already have your data. The width of your confidence intervals will give you an idea of the precision your data are giving.
David Young, Resident, Psychiatry Department
 mentor is Peter Martin
 Study on detox med in psych hospital
 Testing hypothesis that response to detox protocol is associated with dx
 Response is scored as drowsy, irritable, etc. Patients will get several observations over time, and the mode will be taken as the response.
 possible diagnoses are depression, PTSD, dementia, about 5.
 Hope to use the patient's response to detox protocol to determine the diagnosis.
 Can make a frequency table of the combinations of diagnoses and responses.
 Currently have 100 patients meting inclusion criteria with existing data
 Can assess correlation without adjusting for other factors using a chi square test.
 Each person would only be counted once in the table and analysis
 You can group different outcomes to avoid cells (combinations) with very few (less than 5) observations, but agree upon the grouping with your colleagues on one set of grouping before running the tests
2013 January 23
Steven Goudy, Otolaryngology
 Working on basic science R01 due Feb. 5  mice with knockout genes, looking to determine temporal relationship between gene expression and bone/blood vessel development in upper jaw
 Frank recommended twoway ANOVA design with time and treatment (and time*treatment interaction), measuring bone density, protein expression, etc as outcomes
 One potential complication is multiple comparisons  want to study up to 25 genes, and calculating sample size for this is complex
 Emphasize in grant that each mouse is only studied under one condition  no need to account for repeated measurements
 For grant purposes, can use boilerplate language to specify BCC for stats help
 Also look at VANGARD clinic in cancer center for genomics assistance
Damon Michaels & Ricky Sierra, Anesthesiology
 Looking at pain control using nerve blocks after knee replacement; sciatic nerve splits in two above knee, and if it is blocked before this split, can't assess nerve injury
 New treatment: tibial nerve block; everyone gets femoral nerve block
 Outcomes: pain control (patient rating) plus use of opioids, measured in PACU and then every six hours for first 24h
 Previous studies have showed difference in pain scores and ~10mg morphine equivalents over 24h (SD ~15mg)
 Main outcome for power calculation: total amount of opioids given in first 24 hours (measured in morphine equivalents); however, using this might result in reduced power for ordinal outcome of pain rating scale
 Preliminary power analyses in clinic  probably 50 patients per group for 90% power, 35 patients will get about 76% power
 Email Matt Shotwell (matt.shotwell@vanderbilt.edu) protocol, set up appointment for further discussion
2013 January 16
Scott Zuckerman, neurosurgery
 project polling expert neurosurgeons on how they would treat recurrent aneurysms at 1 year with Dr. J Mocco.
 N=40 physicians to be surveyed about when and how they would retreat a brain aneurysm that was treated one year (more or less) prior, and that had certain characteristics (total of 400 permutations of 6 fundamental aneurysm characteristics); how best to construct questionnaire (with all 400 questions, some sort of latinrectanglelike design, etc.).
 would like quote
2013 January 2
Yaa KumahCrystal
 Has questions about stata.
 has longitudinal data on measurements taken before and after an intervention.
 the times are different for each patient
 the origin is the time of intervention.
 your data need to be formatted with one observation on each row, with a variable for id, time/date, and the response variable.
 Main first question is how to reshape wide to long in stata.
 The problem may lie in the variable names. For stata's reshape function, it depends on the number of the time being on the end of the variable name. Make your variable name in that form first.
 Then use the reshape long function.
Joshua Squiers
Over a 6year period, Vanderbilt Medical Center developed
multidisciplinary ICU teams to provide expanded coverage to five of
their adult tertiary care ICUs, including the Surgical Intensive Care
Unit, the Neurosurgical Intensive Care Unit, the Medical Intensive
Care Unit, the Cardiovascular Intensive Care Unit, and the Burn Unit.
Currently, within this model one MD intensivist partners with a number
of ACNPs to provide care for a significantly larger number of ICU
patients than the MD could usually provide alone. This medical team's
core consists of ACNP intensivists and MDs intensivists who provide
billable medical services for ICU patients, in conjunction with other
ancillary services.
This project is a descriptive observational study assessing the
association of NP care on critical care patient outcomes. This study
will utilize the Vanderbilt ICU database to provide data on a variety
of outcome measures, including mortality and inpatient healthcare
associated quality indicators. The ICU database contains records from
more 150,000 ICU patient visits since 2005, and contains descriptive,
morbidity, mortality, and quality data. In conjunction with the ICU
database, the Social Security Administration's Death Master File will
be searched to determine 30 day and 1 year mortality.
This study will utilize a pretest/posttest design to compare outcome
and quality indicators in each of the above ICUs. Each ICU initiated
their NP teams at different times during the past six years. The
initiation time point will serve as the demarcation for
pre/posttesting in each of the ICUs. There will be a six month
washout period following the team initiation to control for the
Hawthorne effect, and allow time for team establishment. Pretest and
posttest analysis will contain all of the patients for that given ICU
1 year prior to the initiation of the NP/MD teams, and following the
washout period, will include 1 year of posttest data. All data will
be deidentified prior to final analysis.
Data from the following ICUs will be collected, utilizing the
following team initiation dates.
1.Cardiovascular ICU (January 2007)
2.Surgical ICU (January 2010)
3.Medical ICU (October 2010)
4.Burn ICU (June 2010)
5.Neuro ICU (August 2009)
Catherine Bulka, Anesthesiology
 Has a rejected manuscript that needs to be improved
 Has data from a prospective group that was matched to a retrospective group
 Research question is whether continuous monitoring of hemoglobin during surgery reduces the number and amount of transfusions.
 Belief is that without monitoring, surgeons tend to overprescribe transfusions
 type of surgery is elective orthopedic.
 since the prospective, controlled trial was not blinded, they later decided to also include retrospective data.
 One strategy to address this concern could be to compare the rate of transfusion between the retrospective group and the control group. Give the odds ratio for receiving a transfusion with a confidence interval.
 We recommend the main analysis be only on the prospective patients.
 For your main analysis, instead of providing a p value from a Fisher's exact test, give an odds ratio with confidence interval.
 Reviewers are also concerned about heterogeneity caused by different types of surgeries. One way to address this is point to the fact that the groups were randomized and also show the breakdown of types of surgeries by group. (The distribution of types of surgeries don't appear to differ across groups.)
 The reviewers are asking for p values for the baseline comparisons. This would require a statistical argument that the two group populations are by definition the same, so testing the hypothesis would not make sense, plus some reference. If you do end up doing statistical tests, use wilcoxon rank sum, kruskal wallis, or chi square tests.
 The original randomization was done by surgery room, not based on surgeon or patient.
2012 December 19
Kelly Green, Cardiology
 Developing registry (~130 records) for two types of TAVR devices (aortic valve replacements); TAVR is fairly new  approximately 50006000 cases performed so far in US
 Noticed large difference between percentage of white vs. AfricanAmerican patients in registry
 Screened 350 patients for procedure, only 15 were AfricanAmerican, and only two taken for procedure
 Prevalence of aortic stenosis among AfricanAmerican population (or other minorities) is unknown
 Looking to use BioVU, CMS (Medicare) data to compile registry  disease is almost exclusively in elderly patients; however, limited by diagnostic codes' accuracy and fact that population is preselected
 One option could be to collaborate with Jackson heart study (~Framingham, recreated in Jackson, MS); might be better estimate of true prevalence, but this procedure isn't performed in Jackson
 Main questions: What is the prevalence of aortic stenosis among AfricanAmericans? Is there a racial disparity in how TAVR technology is applied, when adjusting for severity of disease, SES, etc?
 These two questions require different data, different studies. #1 requires communitybased survey.
 National Cardiovascular Data Registry (NCDR) is in development, but not mature enough to use yet; Medicare Current Beneficiary Survey may be a good bet (age is available), but need enrollment file, ICD9s, procedures  need to get as close to entire population as possible, not just sick patients. Dave Penson at VUMC might be able to help with this.
 Any major database like Medicare will take lots of time/power to work with
 Applying for VICTR studio  no estimate of statistician time needed for that
2012 December 12
Scott Zuckerman (working with J Mocco); Cerebrovascular Surgery
Consultants: Sharon Phillips, Frank Harrell
 Brain aneurysm surgical approaches
 Coils can impact into the aneurysm. Is retreatment needed? Open surgery needed?
 Survey design  30 people at a national meeting; hypothetical patients
 May be 30 most experienced in the country; need to achieve nearly 100% response rate for the survey to be useful for the intended purpose
 Can put confidence intervals around estimates
 Think about either taking random sample of all possible permutations of patient characteristics or generate systematic combinations of char.
 Each surgeon could get a different set of patient char. combinations
 BUT Some combinations may never occur in nature; can use sampling from real cases to populate survey (a different random sample for each surgeon respondent)
2012 December 5
Alex Jahangir, Ortho/Trauma
 Planning to submit for VICTR RFA
 Looking at nontrauma joint replacement patients; ortho has highest incidence of rapid response calls, and 31% are from arthoplasties  perhaps due to age and comorbidities of patients?
 VUMC has funded an NP to comanage  round on patients, manage meds, troubleshoot, facilitate discharge process; want to look at outcomes for a year before vs. a year after NP starts
 Death, codes, PEs/DVTs, rapid responses, UTIs, length of stay, 30day readmission, etc all considered potential outcomes along with costs
 Collaborating with anesthesiology, ICU, finance
 Suggest applying for minivoucher from VICTR to help with grant preparation (Li Wang can instantly approve for <5 hours)
 Eventually looking for manuscript; will need some sort of help with data cleaning (bioinformatics? biostats CSAs? student for data entry?)
 Start prospective cohort a few months after NP starts to accommodate transition time
 Account for double replacements or patients with >1 surgery during time frame?
Kaushik Mukherjee & Kendell Sowards, Trauma and Surgical Critical Care
 Changes to a VICTR project  suggest talking to Li for specifics on biostats involvement with RFA
 Three main aims: 1) does insulin resistance differ between patients who have diagnosed diabetes, occult diabetes, and stress hyperglycemia; 2) does insulin resistance differ between patients with and without ventilatorassisted pneumonia; 3) does knowing amount of insulin resistance help predict development of VAP over and above other diagnostic factors (xray, cultures, etc)
 Recommend timevarying Cox model for time to VAP, adjusting for varying insulin drip amount or M (multiplier); other factors could also vary over time  medications, nutrition, etc, along with baseline factors (age)
2012 November 21
Chad R. Ritch, Fellow, Department of Urologic Surgery
 Proposal development for a pilot study: need help for statistical analysis section
 A pilot study of preoperative enteral supplementation (nutrient shake, similar to Ensure) before Radical Cystectomy (RC) surgery
 Randomized ~150 cancer patients into 2 groups (intervention vs. control; all patients will get shake for 46 weeks after surgery; patients who previously got chemo excluded), primary outcome: number of complications per patient
 Population averages about 25% patients with complications from surgery (about half of these have >1 complication); anticipate that intervention will decrease by 10%
 Complications measured using Clavien scale
 Potentially use albumin levels on day of surgery as primary outcome: main interest is complication rate or count, but with low N and low complication rate in the population, unlikely to have sufficient power to see a difference in dichotomous/fewcategory outcome; albumin is independent predictor of complications in previous studies
 Should also collect data on feasibility for future grant application  how many shakes were drunk, etc (working with nutrition center on this)
 Other potential secondary outcomes: time to first complication, length of stay after surgery, change in BMI or body composition/muscle mass (using dexa scan in nutrition core)
 Albumin measurements: baseline, day of surgery (~46 weeks after baseline), day of hospital discharge postsurgery, 6 weeks and 90 days after surgery
 Need to find mean and SD of albumin measurements in control population (mean around 33.5?); LeenaChoi can help with sample size
 Also planning a weekly phone assessment of food intake from nutritionist
 VICTR RFA deadline December 15
2012 November 14
Kaushik Mukherjee, MD, Division of Trauma & Surgical Critical Care
 Has data from 20052011 on insulin resistance and blood glucose control for intubated patients. Stressinduced insulin resistance (IR) increases infection, including ventilator associated pneumonia (VAP). All vent patients are on an insulin drip.
 Protocol for monitoring blood glucose (BG) has been in place since about 2002. There is a formula used to determine the insulin drip rate (IDR) = [BG(mg.dl)60]xM. M is amount of adjustment for the drip to maintain a BG level between 80 and 110 (now using 100130). When M is high, insulin resistance is increasing, so IDR goes up. The IDR is the best measure of IR.
 The data consists of 314 patients with VAP and no other infection and 4098 patients with no infection. VAP is defined using CDC criteria, a patient has to be on the vent for 2 days before it is considered VAP.
 The number of VAP infections peaks at about 5 days. After 10 days there is little information, so we would be looking at the time they were intubated to 10 days out.
 They would like to know if there is a difference in IR between patients who are diagnosed with VAP and those who are not. Suggestions were to look at the change in IR and BG for all patients and to look at time to infection. They would need to adjust for nutrition,(i.e., glose drip, tube or TPN), mechanism of injury, age, comorbidities (a score used in trauma can be used for this).
 Would like a voucher for biostats support, we estimated ~ 80 hours for this project.
2012 November 7
Moises Huaman, Fellow, Infectious Disease
 Has a case control study with 4 groups. Cases are extrapulmonary TB, 11 patients
 Three sets of controls: pulmonary TB, latent TB, and no TB, ~20 per group
 Want to look at association between groups and vitamin D levels in blood.
 Have done kruskal wallis and pairwise wilcoxon tests
 Want to account for covariates, including season of vitamin D assessment, US/foreign born, age, sex, race, ethnicity
 Recommend against excluding variables based on pvalues
 How to analyze residuals in stata?
 Could use proportional odds ordinal logistic regression with vitamin D as the dependent variable. This is an extension of the kruskal wallis test. You don't want to group the outcome, vitamin D.
 Also want to model tuberculosis and assess the impact of vitamin D, adjusting for other factors. You will have to prioritize the other covariates in this model. The number of parameters you can put in the model depends on the number in the smaller group.
 Could use a propensity score to adjust for the propensity of TB adjusting for many other factors
 Would like a victr voucher for biostatistics support. We estimate this will require about $3000.
 Do not require the propensity score to be linear in the model.
Emily Reinke, Sports Medicine
 http://biostat.mc.vanderbilt.edu/wiki/pub/Main/ClinStat/bbr.pdf Has information about measuring agreement
 Question about preferred method of establishing agreement. ICC or BlandAltman? Frank prefers mean absolute discrepancy (within or betweenraters). Zhouwen has programmed R functions for this.
 Patients who have undergone reconstruction are xrayed to measure the space between the femur and the tibia of each knee on both the lateral and medial side and compare the reconstructed knee to the normal knee. We have imaged ~260 people of 420 people so far. The knee images are measured unpaired and blinded to outcome scores. Telltale hardware is hidden, but drilled tunnels cannot be hidden so blinding to a reconstruction is not complete. 28 pairs have been measured, 12 right knees additionally have been measured.
 Initially, we had intended to use a single person to do the measurements. For various reasons we want to add a second reader. Before we did so, we wanted to establish that we had reliability between the two. Both readers measured 10 publically available images from the osteoarthritis initiative and then several months later did it again to determine the inter and intra rater reliability of the measurement method with these two readers. We are using Bland Altman to assess reliability. Note, previous evaluation of the reliability of the method has been performed using ICCs.
 We had an unexpected finding. In our best case scenario, which came from the comparison of the experienced measurer with herself, the limits of agreement for the measurement[noise] and the greatest anticipated difference in the measurements between knees[signal] are roughly equivalent. In all other cases the limits of agreement are greater than the greatest anticipated difference in the measurements between knees. However, we are talking about tenths of millimeters in difference, and despite extremely similar bone orientation on the xray images this is why I thought the publically available images could act as surrogates for the study images the dissimilarity between the older arthritic knees in the publically available images and the young athletic ACL reconstructed knees in our study images are believed to be potentially sufficient to explain what we are seeing.
 Is it permissible to take 10 pairs of images chosen at random from the 260 study images, have the two readers each measure them twice, perform BlandAltman analysis on the measurements, and if the signal to noise ratio is reasonable, measure the whole group as if the 10 were never taken out?
 We think using some of the actual data to assess the raters' reliability is fine, since it will be only a small subset and the raters are not likely to remember the images.
 Recommend using more than 10, maybe 20.
 Mean absolute discrepancy will be in the same units that you are working with, so that it is more interpretable. Measures degree of disagreement. You can get a confidence interval using bootstrap.
 Bland Altman will show you the amount of discrepancy by the size of the measurement. It lets discrepancies in different directions cancel each other out.
2012 October 31
Frank Virgin, Otolaryngology
 Working on a cystic fibrosisrelated questionnaire study and needs help making sure that it is set up in a way that will allow analysis. Additionally, for the purposes of funding, I want to get a sense for the amount of statistical help I will need to analyze the results.
 Discussed importance of getting a high response rate
 Went over codings of particular questions, and advantages of truly continuous visual analog scales as implemented in REDCap survey
 Bring back a revised questionnaire to discuss at clinic
 VICTR $2000 voucher is likely to be adequate for analyzing the data
Curtis Baysinger, Anesthesiology
 Dr. Baysinger is assessing agreement between two devices that assess effectiveness of a nerve block in cessarian delivery.
 We encouraged Dr. Baysinger to contact Jonathan and Matt to arrange statistical support through an existing collaboration
2012 October 24
Kaushik Mukherjee, Surgery, Division of Trauma and Surgical Critical
 attempting to determine if there is a temporal correlation between increased insulin resistance (as manifested by an increased insulin infusion rate) and the diagnosis of ventilator associated pneumonia in critically ill trauma patients.
 Plans to apply for VICTR biostats support
2012 October 17
Irving Basanez, surgery resident
 Looking at association of hearing loss and school performance
 Inclusion: 1112 years old, 492 total
 Hearing measurement: each child was presented with a tone of 1000, 2000, 4000 Hz, and they were sounded with a volume of 30 dB. If they failed to hear a certain frequency, the volume would be increased to 35, 40, ... 80 dB. After the threshold is established, the next frequency was tried with 30dB. Resulting volume is the lowest volume in either ear.
 Measurement of success in school: 4 exams (language, math, ...). Total score: # of correct questions divided by the # of total questions for all 4 tests together.
 Patient characteristics: age, sex, grade, ear exam.
 Analysis:
 Linear (or proportional odds, depending on the distribution of the outcome) regression. Outcome: total test score. Covariates: age, grade, volume (at which the sound is heard). Might want to think about interaction term: age*volume (to allow the effect of hearing loss to be different for different age groups). Another possible interaction term: hearing exam * ear exam.
 We recommend to choose the covariates a priori.
 Check correlation of [age and grade] and [hearing exam and ear exam] , for example redundancy analysis.
 Possible nonlinear effect of hearing loss on test performance (might want to include hearing as a nonlinear effect)
2012 October 10
Sunya Sweeny, orthodontic resident
 In digital teeth models, a bite registration helps show where the teeth fit together.
 Wants to find the ideal material for bite registration. The comparison gold standard is the plaster model.
 The articulation is going to be measured by measuring the distance between the top and bottom teeth in several different places. This will be done on the plaster model and with each material.
 She is planning on using only one person's measurements. The variability comes from doing the measurements. * possible model: distance from physical ~ location + material + time OR
 distance from physical ~material + time (with a separate model for each location)
 Planning on doing about 2530 measurements with each material.
 Wants help with design and sample size
 Topic is comparing the accuracy of interocclusal record materials in articulating digital dental models.
 Associated literature has shown that laser scanned digital models are dimensionally accurate representations of plaster models. While plaster models can be easily and accurately hand articulated, articulating digital models is less accurate and time consuming, especially when anually or visually articulating them. An accurate interocclusal record could make the articulating process more efficient. No studies have looked at the accuracy different bite registration materials in mounting digital models.
 Tentative materials and methods: 1 typodont mounted in maximum intercuspation on an articulator, 5 different bite registration materials.
 Methods: Place 6 vertical interarch markers on each maxillary and mandibular arch (on the first molars, first premolars, and laterals). Control measurements between the interarch markers will be made using digital calipers on the typodont. Typodont will be scanned using Ortho Insight 3D laser scanner to create digital models. Bite registrations will be made on the articulated typodont using the five different types of bite registration materials. (Need to determine sample size of how many of each type of bite registration material needed). Bite registrations will be scanned using Ortho Insight 3D laser scanner to create digital bite registrations. Digital models will be articulated using the digital bite registrations. Experimental measurements will be made between the digital interarch markers using Ortho Insight 3D model viewing software. Results will compare the differences in the interarch measurements from the control group (physical typodont) and experimental groups (digital models articulated digital bite registrations).
Sarah Hill, nursing doctoral student
 Have an existing collaboration with Pediatric Department. The statistician is Ben Saville: b.saville@vanderbilt.edu
 possible strategy: one model each for the number utilized during the operation and one model for number utilized post operatively.
 number of units post op ~ weigh + post op lab values
 note that the outcome, number of units, is count data. If using a regression model, appropriate options would be poisson regression or proportional odds ordinal logistic regression.
 project for my DNP studies at VUSN.
 retrospective chart review from 1/1/2010 12/31/2011 for patients who have undergone craniofacial reconstruction for the treatment of craniosynostosis. I am specifically looking at the blood utilization with this encounter.
 Research questions:
 Is there a relationship between the patient's weight & total # RBC transfusions
 Is there a relationship between the patient's weight & total # FFP transfusions
 Is there a relationship between intraop Amicar use and # of transfusions given intraoperatively
 Is there a relationship between intraop Amicar use and # of transfusions given postoperatively
 Is there a relationship between the postop PCV and RBC transfusion administration at 6, 12, 18, 24, 36, & 48 hours postoperatively
 Is there a relationship between the postop INR and FFP transfusion administration at 6, 12, 18, 24, 36, & 48 hours postoperatively
 Is there a relationship between the total number of transfusions and length of PICU stay
 Is there a relationship between the EBL (estimated blood loss) and amount of RBC transfusion intraoperatively
 Is there a relationship between the EBL (estimated blood loss) and amount of FFP transfusion intraoperatively
 Frank: "Please convert ages to all be in months and take away the "m" after the numbers."
2012 October 3
Thomas Abramo
 Using near infrared to measure oxygen saturation in brain tissue. Interest in seeing if these measurements are predictive of a CT scan result.
 The measurements are taken over time, and the hypothesis is that a big difference in the two hemispheres or overall low values are predictive of pathological CT results.
 One idea is to have blinded clinicians rate the near infrared graphs as pathological (yes/no), and then look at the predictive ability of the ratings of the CT result.
 Another approach is to get a summary measure. The summary measure could be the variance in each side.
2012 September 26
Emily Bullington, Pharmacy
 Project is “Preventing and managing refeeding syndrome in the acutely ill: the importance of electrolyte replenishment.”
Gopi Shah, ENT research fellow
 doing a project looking at the prevalence of hearing loss in the high schools here in Nashville.
 would like to screen kids and then give them a survey about their perception of hearing loss.
 question is how to figure out the number needed in order to give the study power.
 Frank: "This is not a power problem but a precision (margin of error) problem. My approach to this is to solve for N such that the 95% confidence interval for a probability is +z when the probability is at the worst case (0.5). z might be 0.1. Statisticians at that clinic can help and can discuss this further to make sure what are your needs."
2012 September 19
Juliana Kyle and Susan Hamblin, Pharmacy
 Protocol for open fractures (bone pierces skin)
 Main compliance issue is antibiotic duration. There is also an antibiotic rotation. Also a physician could be noncompliant by not prescribing any antibiotics at all. Right now the group is interested in a composite yes/no overall compliant.
 Want to assess compliance and its relation to outcomes in terms in drug resistance
 Primary outcomes are infection and resistant infections before discharge.
 Secondary endpoints are mortality, length of stay, icu length of stay.
 Have about 200 patients. Rate of infection may be around 50%. Resistant infection is about 2070% of those, based on literature.
 Factors that may affect infection: length of stay
 Analysis strategy: logistic regression with types of compliance, patient factors (diabetes), physician factors, length of stay
 Limitations: have know way of knowing each patient's prior exposure to resistant organisms
 May want to consider other sites besides the original site.
 It would be useful to collect time to infection, but need to consider the uniformity of assessment. May use time to first positive culture.
 Many issues, better to have a statistician on board. For purposes of a VICTR voucher, we estimate that this would take roughly 50 hours.
Suseela Somarajan, General surgery
 This is a study on finding the amplitude of gastric slow waves measured in human with different BMI before and after food.
 For each BMI category, I have only 3 samples, have 9 total subjects.
 Question: Is there any statistically significant difference in amplitude before and after food in each BMI category  performed Student’s ttest on the data (statistical significance set at p < 0.05). Since the sample size is very small the results are misleading. I need to show the data and discuss it.
 Gastric slow waves are measured serially. At each time point there are 16 measurements. They take averages at each time point. Before the meal, there are eight observations, and eight observations after the meal.
 For the question about the amplitude of the waves, you should adjust for BMI. Since there are many time points within each person, the data are correlated.
 Since the sample size is small, focus on graphical presentation. Could plot each person's trend, with time on x amplitude on y, and show BMI with color. Indicate the time of the meal with a vertical line.
 If you collect more patients, focus on modeling, maybe with the difference in pre post activity as the dependent variable. Could also use longitudinal data, but would need to account for correlation within patients, maybe with random effects.
2012 September 12
Jenny Rothchild, Urology fellow
 Recommended that she plot histograms of each count or continuous outcome (in aggregate) to assess the distribution
 Recommend controlling for physician variables or physician in a model
 May need help in analysis, so recommend considering applying statistical support through a VICTR voucher
 Results of the histograms will help us to recommend models.
Eric Thomassee, Cardiology fellow
 Would like feedback on the datapoints prior to data collection. The project involves acute pulmonary emboli and all assoicated therapies (surgical, percutaneous, and medical).
 We looked at his redcap database and made recommendations.
 We emphasized collecting actual dates rather than, for example, length of stay or three month mortality (yes/no).
 Found some instances where radio buttons were more appropriate than check boxes, when the categories were mutually exclusive.
2012 August 15
Daniel Muñoz, Fellow, Cardiovascular Medicine
Background: Patients who are admitted to ED with chest pain are usually given EKG, troponin test and assessed for a cardiovascular risk score. If all these are within acceptable ranges, they are often admitted for stress test or observed in ED, both of which use lots of resources. Rates of MI/CV death within a month, given the patient has acceptable EKG/troponin, are extremely low. Planned "intervention" is discharge to outpatient with stress test within 72 hours, vs. "control" of typical ED/hospital observation. Planned outcome: major CV event within 30 days (CV [or not clearly nonCV] death, MI). Event rate is hypothesized to be roughly the same for both groups at about 1%. Will also collect resource data, like costs.
Potential issues:
 Very low event rate > very large sample size required. Possibly include hospitalization for chest pain and/or revascularization as part of "CV event" outcome.
 Compliance/patient selection  patients would be responsible for getting their stress tests. Could select patients based on insurance or other factors, but that would limit generalizability and possibly create selection bias.
 Followup: Planning on SSDI for death, phone followup for other event(s).
 Secondary outcomes: repeat ED visits
My specific question: what might be the approximate target sample size for a potential trial of the following design?
 1:1 randomization of low risk patients into one of two treatment arms (ED eval vs Outpatient eval) in which we aim to test a hypothesis about noninferiority of outpatient eval as compared with current standard treatment (ED eval)
 Primary endpoint is a dichotomous composite endpoint (Death/MI or no Death/MI at 30 days) with an anticipated event rate of approximately 1% in each arm
 Level of significance 0.05
 Desired statistical power = 0.90
To use the most commonly used approach (relative efficacy using odds ratios) can be used as follows. The standard error of the log odds ratio is the square root of twice 1/[np(1p)] where p = 0.01. The upper confidence limit of the log odds ratio will be log observed odds ratio + 1.96 times this standard error. Setting 1.96 se to log(1.25) if the multiplicative margin of error is allowed to be 1.25 yields n = 2*((1.96/log(1.25))^2)/(p*(1p)) = 15,586 patients in each group for a total of over 31,000 patients. The large n is the result of the tiny incidence rate.
< > p < .01 2*((1.96/log(1.25))^2)/(p*(1p))
< >
2012 August 08
General Surgery, Pediatric Surgery: Brian Bridges, Michael Northrop
 Examine the effect of new laboratory procedure on the outcome of pediatric ECMO
 Control: patients in the previous year; Treatment: patients in the current year receiving the new procedure (N=80 each group)
 The new procedure is to monitor the activity of heparin
 Outcome: survival to hospital discharge, time to bleed, (heparin and other parameters)

2012 July 18
General Surgery, Pediatric Surgery: Andre Marshall
 Congenital Diaphramatic Hernia
 Does placing chesttube during repair improve mortality at 6 months.
 Does not have date that mortality was assessed.
 Recommend collecting date of death, or at least verifying that mortality assessments are made after 6 months (or 30 days).
 Apply for VICTR support in the amount of $3500
2012 July 11
MPB: Kate Ellacott
 Main analysis: the effect of weight loss on a biomarker.
 Population: patients who had a gastric bypass surgery.
 Main question: would like to get power calculation of how many subjects are needed to detect a difference in biomarker level before and after surgery.
 Recommendations:
 define a clinically meaningful difference in the biomarker from the previous literature (instead of putting a difference that you observed in the previous study)
 main analysis should be based on the same test that the power calculation was done (probably paired ttest)
 if the meaningful difference is not know, might be a good idea to estimate possible number of subjects that can be collected and calculate power for difference that can be detected in that many subjects
MPB: Ember Sympson
 Outcome: post liver volume
 regression covariates: planned volume, day from surgery,
 get estimate and confidence intervals for coefficients for each of the covariates
Anesthesiology, Paul W. Hannam
 Looking at how well some three scores predict certain outcome
 Outcome: morbidity defined as renal failure, organ failure, or death
 Recommendation:
 logistic regression with outcome: 1  had an outcome within 30 days, 0 didn't have an outcome within 30 days
 run separate regression for each score
 covariates: ????
 compute predicted probabilities of outcome, and Brier index
2012 June 20
General/Pediatric Surgery: Andre Marshall, Martin Blakely
 Looking at adverse event rates among appendectomy patients from two separate trials  difference between two treatments?
 Interested in funnel/forest plots, getting overall ORs and AEspecific rates
 Planning to do formal metaanalysis? Need input from CTSA statisticians in terms of number of hours
 Would a mixed effects model adjusting for trial be sufficient, rather than formal metaanalysis?
Plastic Surgery: Joshua Anthony
 Data on hand surgery calls from 111 hospitals in TN capable of handling hand emergencies
 Do availability for hand call, other outcomes differ for medically underserved areas vs. others, or teaching vs. nonteaching hospitals?
 For outcome of hand surgery availability, recommend logistic regression model: hand call = medically underserved + teaching hospital + urban area (all yes/no variables); similar for other outcomes like hand surgeon availability
 Very sparse data for level 2/3 trauma centers (1 and 2 hospitals respectively); recommend combining with level 4 trauma centers to compare level 1 (surgeon should be available 24/7) vs. everyone else
 20 hours of VICTR help is probably sufficient for a few logistic models, explanations, manuscript edits (emphasized that clean data is extremely helpful in keeping time down!)
2012 April 25
Anesthesiology : Damon Michaels, Jennifer Morse, Lesley Lirette
 General:
 Observational study
 2 months
 two groups: residents, attending physician performing regional anesthesia (nerve block)
 group sizes: 131 (attending physician) 150 (residents)
 there are repeated measures: there are only (6 residents+one fellow), and 7 attendings. And for each procedure different attendings supervised different residents.
 it was not recorded what resident/attending did what procedure.
 Outcome: delay = time when the patient actually arrives into OR minus time when the OR is ready for the patient. Originally wanted to dichotomise the outcome into 5 minutes or less and more than five minutes.
 Hypothesis: there is no delay
 Suggestions:
 keep the outcome continuous
 if possible, recover the information what resident/attending did what procedure, and put everything into one model
 or leave it as is and state in the limitations the fact that there are repeated measurements
 Recommend you look at the data graphically. You can learn both about whether there was a delay and about the distribution of the times. Do two boxplots (one for residents and one for attendings) overlayed with the actual data.
Medical student group: Monika Jering
 General:
 looking at the adult APGAR (score based on blood loss) per minute
 general question: if a drastic change in the score affects mortality (morbidity)
 Population: 400,000 admissions.
 Outcome: death or complication within 30 days
 Hypothesis: A "sharp" drop in the score is associated with higher risk of death (or complications)
 Suggestions:
 define a drop as a maximum drop within one minute period
 For simplicity: use first admission for each patient
 adjust for baseline, age, length of surgery, comorbidity index, type of surgery, and other mortality
2012 April 11
Medical student group: Beth Greer, Michael Maggart, Neelam Patel, Calvin Sheng, Tenisha James, Cooper Lloyd
 Children are classified as exceptionally behaviorally inhibitted or uninhibited, as assessed by a questionnaire
 Children are given an MRI to measure the amygdala size in volume by summing cross sections. There is specific protocol. The person measuring the images will be blinded to the behavioral status.
 Could do a separate study to assess intrarater reliability (have the person assessing the images rate some of the images twice in a blinded fashion).
 May be a benefit to including patients who are more in the middle of the spectrum
 There is probably a selection bias in the sample of patients who participate in the study, due to the mechanism of accrual.
 There are other assessments given to exclude patients with other disorders, including autism.
 The group wants to know about sample size
 They will supplement their data with data from a separate database. Recommend that a random procedure be used to draw a sample from this database
 Research question is whether there is a difference in amygdala volume in the two groups.
 To find the sample size needed, would need to think about what the smallest clinically meaningful difference in amygdala volume would be and also the standard deviation of the amygdala size in the control group.
 Another way to look at this research question is to use the inhibition score without categorizing, and testing for association with amygdala size. Depending on the way the score is assessed, this could have the advantage of preserving statistical power.
 Can use the PS software by Bill Dupont. Free download.
 Ideally would want to use the amygdala as the outcome variable and adjust for the total brain size as well as inhibition score.
 The amygdala size is hypothesized to influence the behavior score. If you keep the behavior as two groups (inhibited and uninhibited), the appropriate analysis would be a logistic regression predicting the odds of being in the inhibited group.
2012 April 4
John Newman, Voranaddha Vacharathit SOM
 Research question: Does animation improves comprehension.
 Hypothesis: Animation improves learning
 Outcome:
 We recommend: linear regression with the posttest score as outcome and four covariates: pretest score, animation (1=yes, 0=no), style (1=visual, 0=experimental), interaction of animation with style (animation*style). The interaction term allows to model different effect size for different styles .
Katie Guess, Stessie Dort, Fayrisa Greenwald, VUSM
 Population: trauma coordinators in trauma ICU. The researcher are planning to send a survey to 219 trauma centers.
 Research question: Whether trauma coordinators assess and educate about acute and post traumatic stress disorder.
 Concerned about: low response rate, the lower response rate the higher the bias might be.
 Recommendation: set questions in a way that you don't have to categorise. RedCAP allows to have a scroll bar, where you can mark a percent between 0 and 100. Continuous measurements have more power.
 Output: descriptive statistics for each question of the survey. For yes/no or categorical questions we report to report proportion (out of the whole population in each region), for ordinal or continuous variables we recommend to report medians and inter quartile range. To compare responses between regions we recommend Pearson chisquare test for binary or categorical variables, and Wilcoxon Rank Sum test for ordinal or continuous variables.
Rich Latuska, VUSM
 Study question: Do patients requiring multiple specialties for care (ortho, neuro, etc) have better/worse comprehension of their injuries and satisfaction with their care?
 Likely prospective study enrolling trauma patients, with survey given at first clinic visit and followup (36 months)
 To determine required sample size, talk with mentor about what detectable difference would be important (5% difference in patients satisfied vs. 25%...) and how many patients it's reasonable to survey over available time period
 Keep instrument mostly the same at both time points, likely with some additional questions at longterm followup
2012 March 28
Kevin Carr, Candidate for MD degree
* See attached files
2012 March 21
Michelle Huber, Pharmacy
 Project is titled: Delirium and pain in the post operative cardiac surgery patient: a retrospective review. IRB: 111619, PI: Chad Wagner, MD.
 Have 1 month of patient information for my project with hopes to receive 4 additional months for a total of 5 months. I have already data collected (everything is in RedCap) for that first month of data. As a resident I am required to present my research project at a meeting in April. I will not be able to data collect the rest of the patients in time for this presentation but I do want to analyze what I have for this meeting.
 How to analyze this data for one month? (we can do something very simple, keeping in mind analysis will be on a larger scale once I have all data)
 The cost to do a simple analysis for the one month of data so that I can have something to present in April?
 For the whole project of 5 months, how to analyze this data? Need to send an estimate for VICTR funding.
 Already have data for 84 patients.
 Retrospective chart review of patients that had cardio surgery and were transferred to CVICU.
 In first 48 hours after surgery. Looking at pain and CAM scores (delirious vs. not.). Have data on all drugs.
 Want to see whether uncontrolled pain is a predictor for delirium.
 One thing to account for is people who wake up from surgery already delirious. In this case, it will be impossible to assess pain.
 Is there a possibility of working with Jennifer and Amy, who have experience in working with delirium. This would need to be billed through the BCC. Contact Ayumi Shintani to discuss this possibility. ayumi.shintani@vanderbilt.edu
 Need to define uncontrolled pain. Could leave this variable as ordinal.
 Probably for the April 20 presentation, given the time constraints and current number of patients, probably a more simple analysis would be appropriate. Expect ~1828% events, which would limit the complexity of the analysis or model. Use graphs to display the data.
 We estimate that this project would take about 45 hours over the next year or so with some effort within the next month for the presentation in the end in April, so an estimate of the cost is $4500.
Stan Pelosi, ENT
 Descriptive study of pediatric patients with auditory neuropathy who get cochlear implants (inner ear implant for deafness)
 Auditory neuropathy involves a functioning cochlea, but they are still unable to hear.
 Risk factors for auditory neuropathy from literature:
 Initial goal is to describe the patient population.
 Need help in interpreting a previous analysis
 The output in stata is a an estimate of the proportions of some variables in the data and also a test of whether the proportions were equal. For example, they tested whether the proportion of premature patients was equal to the proportion of nonpremature patients.
2012 March 14
Shilo Anders
 Wants to address reviewer comments:
There should be analysis for trend used for the analysis rather
than multiple tests with Bonferroni adjustments.
The authors state that less time was devoted to instability and error
detection. However, the total number of interventions was
significantly larger in the 2007 observation period. Thus the total
number of instability and error detection events may have remained
constant, while the team significantly expanded their role in other
areas. The nature of the interventions that increased in frequency in
2007 might well reflect improved acceptance of the teleICU system by
the bedside providers, with better collaboration and expanded requests
for assistance. I believe that the manuscript could be improved if
the authors expanded their analysis of the nature and possible causes
of the observed distributional changes in teleICU interventions.
The minor revision required is to provide statistical evidence on data
reliability and data validity.
2012 March 7
Dave Janz, Critical Care
 We are designing a trial of using acetaminophen to treat sepsis. Before conducting a large trial looking at clinical outcomes, we are going to do a smaller one just looking at changes in markers of disease after 2 days of acetaminophen therapy compared to prior to starting therapy. My specific question is in regards to how to power the study to be able to detect these changes. We will be studying isoprostane levels and the mean level in this population is 65.7 pg/ml, SD 67.4 (isoprostanes are not normally distributed in humans). We would like to be able to detect a change of 20 pg/ml in those receiving acetaminophen and also to be able to detect a difference of 20 pg/ml compared to a group that receives no acetaminophen.
 We are planning a unrelated study trying to determine the reason why patients in the ICU who receive red blood cell transfusions are more likely to die than similar patients who are not transfused. Specifically, it will be an observational study that will measure 21 markers in the blood of transfused patients immediately before transfusion, along with 4 hours after and 24 hours after. There will be no control group. These markers are also not normally distributed in humans and the marker with the largest standard deviation is IL8 with a mean of 457.6, SD 7641.4. What I would like to be able to detect a change of half a standard deviation in this group. Do I need to account for multiplicity?
 Recommend getting sample sizes for the different possibilities of the standard deviations of the difference in prepost drug outcome. 65 may be the upper bound on the standard deviation.
Tom Golper, Nephrology, Department of Medicine.
 Need a randomization scheme. Two groups. The experimental condition is the vehicle
 Recommend a randomized permuted block design. with varying block sizes.
 Want to do an interim analysis based on the outcome during which a decision will be made to terminate early for efficacy or futility. We recommend from the description of the need for an interim analysis, that an interim analysis based on the outcome is not needed. We recommend that after, say, 40 endpoints are observed, the standard deviation be reassessed, and the sample size be recalculated.
 Account for the number of patients who die.
 Must account for the baseline values in the analysis, since someone with a very low level to begin with has less room to decrease.
 They assume that about one fourth of the patients will have filters that last 70 hours, when the patients will automatically have their filter changed. Thus, there will not be an observed failure time for those patients.
 There is a censoring issue, both from lasting more than 70 hours, and also people dying.
 Want to show that the syringe is not inferior to the infusion.
 JoAnn Alvarez will help with the permuted blocks.
2012 February 29
Steven White, Emergency Medicine
 Study of an informatics tool to automate access to the state of tennessee prescription monitoring program (PMP).
 The tool is much faster than the existing tool
 The primary study objective is to determine whether automated access results in a higher proportion of ED patients being screened for controlled substance misuse.
 Additional objectives will be whether increased screening results in a decrease of opiate prescriptions at discharge and whether there are any PMP report variables which influence decision to prescribe opiate at discharge.
 The study population involves all patients discharged from the ED over a twomonth period, with tool ON/OFF at two week intervals. During tool OFF periods, clinicians still have access to the PMP using the web browser.
 The following data is collected from each PMP automated query: number of prescriptions for controlled substances previous 12 months, number of different prescribers, number of different pharmacies, number of opiate prescriptions, number of days since most recent prescription, pain score at triage, oneway encrypted case number, oneway encrypted MRN, oneway encrypted user racfid, data viewed flag, file opened flag. For periods during which the automated tool is OFF, we will obtain only counts of queries without specific patient data.
 From ED discharge instruction summary (generated from discharge instruction writer), for each patient discharged during both tool ON and OFF intervals, we will have flags for patientvolunteered opiate use at home, flag for opiate prescription at discharge, name and quantity of product prescribed, oneway encrypted case number which will match up with encrypted case number from automated tool query, encrypted racfid of discharge attending.
 Could analyze this at the doctor level, and have repeated measures. For example, each row in the data could correspond to one shift. You would collect the number of patients the doctor accessed in each of the two systems, and the number of opiate prescriptions the physician prescribed in that interval.
 We have an existing collaboration with the Emergency Department. See Cathy Jenkins. We think Jonathan Schildcrout works with DBMI.
Andre Marshall, General Surgery
 See attached document
 Wants to study readmission in pediatric surgery with retrospective data. Want to see which diagnoses and procedures predict readmission within 30 days.
 A big question is whether the readmission was caused by the procedure or not. An algorithm could exist that payers use. If you could use that same algorithm in your study, it would be much more objective than having a group in your team decide each one.
 Jonathan Schildcrout is working on outcomes research.
 Recommend putting patients that did not get readmitted.
 If you can't enter all the "control" data into redcap, you need to randomly select them.
 If you go through VICTR, we estimate that it will take about 40 hours, or about 4000 dollars.
2012 February 22
Mei Liu, Biomedical Informatics
 Retrospective study. Outcome: Abnormal lab test. Exposure: Certain medication
 Planning to do propensity score matching to compare each medication group to a control group.
 Recommended to talk to Michael Matheny (maybe to Jonathan as well)
2012 February 15
Michelle Collins, School of Nursing, and Sarah Star, OB Anesthesiology
 Planning a retrospective chart review of patients who got nitrous oxide in labor.
 What patient factors influence success of the the nitrous oxide.
 Outcome is whether the patient got an epidural.
 Have about 350 patients over about one year period
 They think parity, use of oxytocin induction, oxytocin augmentation, length of labor, provider type (midwife, ob, combo). May want to adjust for time since nitrous oxide was made available in the hospital.
 Nitrous oxide relieves anxiety. Some women may want nitrous oxide but not an epidural because they can more effectively push, and because they have more control on the amount
 If you include patients that did not get nitrous oxide in the first place, it may be important to include a propensity score for using nitrous oxide.
 Can email Jonathan Schildcrout to see if there is already a collaboration in place. jonathan.schildcrout@vanderbilt.edu
2012 February 8
Pat Keegan, Urologic Oncology
 Planning a trial to investigate differences in using staples vs. ties to close blood vessels in prostatectomy.
 The outcomes they want to compare by are surgical margins and continence.
 The investigator had already made some sample size calculations in PS and wanted to review them with us.
 We recommended accounting for attrition in the sample size
 We recommended collecting the actual measurement of the margins rather than only whether they were positive, if feasible.
 Recommend collecting the patientreported number of pads used per day rather than only collecting three levels: 01, 25, 5+. The number of pads can always be grouped later if necessary, but if you only collect the three groups, there will be no way to obtain the actual value.
2012 February 1
Blake Hooper (Anesthesiology)
 Evaluate two new cath methods compared to gold standard; cardiac ouput as outcome
 15 subjects, A was performed on all 15 subjects, B was performed on 5 subjects, gold standard from all the subjects
 measurements of different methods were made at the same occasions for each subjects; the number of measurements ranged from 8 to 12 for the subjects.
 Expect the diff between methods and gold standards change with time; nonmonotonic trend is expected.
 Mixed effects model
 Apply for a Voucher $2000 (http://www.mc.vanderbilt.edu/victr/pub/message.html?message_id=139)
2012 January 25
Lucy He (Neurosurgery)
 Looking for risk factors for Grade 1 vs. Grade 2 meningioma (diagnosed via radiograph)
 Outcome is Grade 1 vs. Grade 2 tumor; currently have 128 patients, 94 Grade 1 and 34 Grade 2
 Limits degrees of freedom possible for well fit multivariable regression
 A priori chosen potential risk factors include edema (four categories), necrosis (yes/no), and location (four quadrants)
 Eventual goal is prediction model
 For now, getting more data probably not practical
 Suggested VICTR application; either logistic regression (possibly using just edema and necrosis  34 df max), or possibly using propensity scores or other data reduction techniques
2012 January 18
Todd Morgan
 Looking at patients with metastatic kidney cancer who have undergone nephrectomy.
 Want to see if the tumor pathology data predicts survival. Also lab values. The want to adjust for type of chemotherapy.
 They have run a Cox regression
 They want to create a nomogram
 Want to look at the predictive ability of the model. Suggest internal validation using bootstrap.
 Also want to validate a prior model that only uses preoperative data.
 Have 63 events out of 88 records.
 Suggest adding the postop data to the preop model and assessing the contribution of the additional variables.
 Discussed ways to account for chemotherapy in the model yet adequately capture the difference between different chemotherapy regimens. All of the current therapies are similarly not very effective. They aren't interested in quantifying this effect; they want to control for it.
 In validation, also need to validate the model selection process.
 Need to check the prop hazards assumptions.
 Recommend at least 40 hours to complete this project.
2012 January 11
Stephen Kappa
 Looking at cost in cystectomy in bladder cancer, comparing laproscopic vs. traditional surgery.
 Recommended that all dollar amounts are adjusted to one standard across years using the consumer price index.
 Talked about approaches for getting biostat support: Check with Tatsuki Koyama to see if this project is covered under existing grants. Another option is a VICTR grant.
Eric Gehrie, pathology resident
 Looking at abo incompatibility in heart transplant
 Hypothesis is that poor outcomes (death within a month) are correlated with amount of incompatible antigen in the transplant heart.
 Have data from 18 cadavers.
 Has tested amount of antigen in the hearts using two ways: one is tissue staining and categorizing into 3 levels by a pathologist; the other is scoring by a machine the percent positive.
 Wants to see whether the amount of antigen is associated with factors like gender, BMI, and time between death and autopsy.
 Recommend displaying all the data graphically.
 To look at the variability, make a strip plot of the percent positive variable. This will show how variable the data is compared with the median.
2012 January 4
Kassatihun GebreAmlack, School of Medicine and Department of Anesthiology
 Want to understand characteristics associated with transfer to ICU.
 Outcome is transfer to ICU for any surgical procedure (nonthorasic).
 763 required transfer from floor. 63,000 did not require transfer from floor.
 Predictors of interest include: past medical history (diabetes, cad, cholesterol, renal function), demographics, surgical procedure (length, medicine, heart rate, rescue).
 Consider logistic regression model conditioning on all covariates to model probability of transfer to ICU  single (1 predictor) and multivariable (>1 predictor)
 Surgeries with longer hospital stays may see more transfer to ICU, will include length of stay.
 Five years of data, have protocols for transfer to ICU changed? Consider including date of procedure to adjust for this effect.
 Could consider a combined endpoint of death/transfer to ICU to identify patients at risk.
2011 November 30
Rebecca Snyder, Department of Surgery; Martin Blakely is the mentor
 Wants to estimate the agreement between three raters of an ordinal scale. There are four levels.
 Also want to know if the agreement varies for different types of surgery.
 Recommend using a weighted kappa for the measurement of agreement.
 For the different types of surgery, they could get separate kappas for the different surgery types (depending on sample size within the types) or do a logistic regression predicting log odds of at least one disagreement between the three raters, including type of surgery as a predictor.
 Wants to know the sample size required. Recommend estimating sample size based on the precision of the kappa statistic.
 Estimate that a $2500 VICTR Biostatistical support request should be sufficient
2011 November 16
Robert Kelly, General Surgery, Bariatric DIvision
 Wants to look at agreement between biopsies taken from different sections of the liver in surgery patients.
 Each of the biopsies get an ordinal score with 6 levels (majority are in 4 categories)
 Two sections will be taken from each patient.
 We recommend that the evaluator of the specimens be completely blinded to the location of the specimen and also the identity of the patient.
 Wants a refined estimate of number of samples they need.
 Think about what difference in measurement would be clinically significant.
 Could calculate sample size required to estimate a weighted Kappa statistic with a 95% confidence interval of a certain width.
2011 November 9
Nick Burjek, Ryan Hollenbeck
 Retrospective study of 150 patients
 Outcome variable is good vs. bad neurological outcomes: Outcome comes in a 15 ordinal scale: CPC scale at ICU discharge
 Want to assess prognostic value of bispectral index, which measures brain function and also amount of sedation
 They have many patients who are on both fentanyl and versed
 One option for the measurement of bis is time under level of 40.
 Probably need to incorporate an interaction between bis and sedation requirement.
 Discussed different ways to summarize the predictor variables.
 Possible summaries could be slope, intercept (where the started), final level, area under the curve, minimum, time less than 40. (Some of these could be used for either sedation of bis.)
 Encourage to look at plots of raw bis and sedation data over time for each patient.
 Can evaluate the effects of many variables in one regression model
 Want to apply for VICTR statistics support. We estimate that $3500. OR, they may have access to Jonathan's group through an existing collaboration.
Khensani Marolen, Anesthesiology
 Prospective study on patients going in for surgery. Collecting data on all drugs the patient is on.
 Two sources of data: one the patient enters at home, and the other is at a clinic visit
 160 patients have both sources of data.
 Want to test the hypothesis that the patiententered data is more accurate and has more detail and more drugs. We have data that can answer how the responses agree and which source has more drugs, on average
 Right now excluding OTC and vitamins.
 Wilcoxon signed rank test and confidence interval
 Be sure to include limitations of this analysis: assumption that all drugs listed are actually taken by the patient
 Can interpret point estimate with confidence interval as the difference in medians between the two groups
 Could also check the actual concordance between the lists
 The patient selfentry device also collects info on comorbidities
 Want to increase enrolment midstudy by offering a gift card
2011 Roctober 26
Elizabeth Card, Anesthesiology
 Want to look at the longterm effects of hypoxia.
 Have two groups of about 650.
 Designing a followup study to a randomized trial in which all patients were were monitored constantly for blood oxygen saturation levels, and those in the treatment arm had the monitor data monitored in real time
 One of the aims is to look at the hypoxia that was detected by the stealth monitor, versus the the hypoxia which was detected by the standard of care, which disturbs the patient.
 Discussed whether the outcome of interest is the sum of time in a hypoxic state.
 One main outcome is whether the patient was readmitted to the hospital for any reason within 30 days.
 One other confounder that could be adjusted for or used as an offset or as a predictor is the length of time that they were on the stealth monitor. We think that the differences in the time on the monitor is independent of the patient's health status.
 Length of stay should be controlled for in models with hypoxia as an outcome. Also amount of blood loss and sedation medicine. (And everything else that is known to affect readmission for the readmission model)
 One way to measure the hypoxia is the average amount below the lower limit for health. This is the area under the limit divided by the time monitored.
 Look at the correlation of the nurses' oxygen saturation measures with the stealth monitor's measures.
 Could also consider the minimum oxygen saturation reached during the observation period. This could be used as a predictor for the patient health outcomes (like readmission) along with the average amount below the lower limit described above.
 Could look at the average O2 itself (and variance) rather than looking at it only when it is below 90%, which is the definition of hypoxia.
2011 Roctober 19
Cathy Jenkins, Emergency Medicine
 Designing a study to compare nurses' assessment of vital signs with a robot's assessment of vital signs on healthy volunteers.
 Can the robot be compared to just one nurse? If not, how would it be analyzed?
 Change wording of aims from accuracy to agreement.
 What patient factors influence the robot's agreement with the nurse's measurements?
 Discussed ways to measure agreement. +
2011 October 5
Khensani Marolen, Anesthesiology Consultants: Meridith Blevins, Cathy Jenkins, Svetlana Eden, Pingsheng Wu, Steve Ampah, Students
 Risk assessment tool  hospital based versus home based. Two reviewers (anethesiologist) review medical history by tool and nurse practitioner, then score. 160 patients with two scores.
 16 ordinal (1=healthy to 6=extremely sick).
 Want to assess the differences between two reviewers in one modality
 suggest (weighted) Kappa  agreement between reviewers for a given patient  report confidence intervals {http://www2.sas.com/proceedings/sugi22/STATS/PAPER295.PDF}
 paired Wilcoxon sign rank  test for difference in distribution of scores between reviewers
< > library(irr) kappa2(r1[,c("ASA_Score_Experimental","ASA_Score_Control")], weight="equal")
< >
2011 Sep 28
Paul Moore, Susan Hamblin, Pharmacy Consultants: Frank Harrell, Svetlana Eden, Mario Davidson, JoAnn Alvarez
 Reducing atrial arrhythmias esp. fibrillation, in trauma patients; role of antioxidant therapy (AO)
 Are applying for VICTR voucher
 Mainly interested in arrhythmias during trauma ICU stay
 7d AO protocol; included if receive >= 72h on the protocol
 Patients must survive 3d and be in the ICU at least 3d to be in study
 AKI would have also stopped the protocol
 How many patients starting the protocol who survived 3d did not finish >=3 3d of protocol
 Find out how many patients had creatinine < 2.5 at baseline who developed > 2.5 by day 3 that led to interruption of the protocol
 Retrospective cohort study
 TRACKS database (trauma)
 Goal: approximate a randomized clinical trial
 Time zero (randomization time in RCT) patients start
 All patients included in analysis
 Patients are similar between 2 treatment groups or you know the measurements needed to adjust for to make them similar
 Survey 10 experts not involved in the study, have them list all the reasons for starting such a protocol
 They must be blinded to the database content
 Verify that all the variables so named are available from the database or you have variables highly correlated with the needed variables
 Verify that 100% of persons who qualified for getting the protocol actually got it or that the reasons for not doing it are captured with available variables (or if reasons were random)
 Can't have comorbidities being the reason for not putting someone in the protocol
 Need to be accurate in identifying arrhythmia/nonarrhythmia
 Random sample of 100 patients  read charts to check concordance with ICD9determined presence of arrythmia
 Recommended analysis: multiple regression (logistic regression model if outcome is yes/no)
 Adjust for potential prognostic variables, especially those in the list assembled from expert opinion
 There is some value in using time until event (censored at discharge from ICU and possible death) because this recognizes that a patient in the ICU a short time has less opportunity for the arrhythmia
 Do we want to define time zero as the end of day 3?
 Assume that arrhythmia that happened day 1  day 3 should be ignored
 How to account for arrhythmias within first 3d?
 Verify that arrhythmia surveillance/coding is constant over the years (also the use of APACHE etc.)
 A good way to justify the sample size is to estimate the number of qualifying patients on each of the two treatment arms, and the overall proportion with arrhythmia, then compute the margin of error (e.g., halfwidth of the 0.95 confidence interval for the difference in two proportions; foldchange margin of error in estimating a hazard ratio)
 Suggest applying for regular VICTR voucher, for $4000 of biostatistics assistance (home dept. will need to pay $1000)
2011 September 21
Bret Alvis, Anesthesiology
 Research question: whether preoperative use of ketorolac prolongs healing of ankle fracture
 Hypothesis: preoperative use of ketorolac DOES NOT prolong bone healing
 Data collected:
 recovery within one month (yes/no)
 drug use (yes/no) * recommended: to use PS
Chad Wagner, CV ICU
 Retrospective study. Would like to look at association of time to delirium and 1. pain, 2. different medications.
 Concerns:
 pain scale is subjective and depends on many characteristics (age gender) it's important to adjust for these characteristics.
2011 September 14
Michelle Collins, SON
 Among women with a positive pap smear result, is positivity of biopsy different by hormone usage?
 Exclusion: Menopausal, Pregnant
 Inclusion: PAP smear with LOW or HIGH grade result
 Outcome: Biospy with NEGATIVE, LOW or HIGH grade result
 Predictor: Hormone Group (Progesterone (IUD or injection), Control)
 Four possible research questions (need to clarify interest in 'overcall' versus 'false positive'):
 Is there an association between progesterone condition and LOW or worse grade biopsy for women who screen with LOW or worse grade pap smear?
 Is there an association between progesterone condition and HIGH grade biopsy for women who screen with LOW or worse grade pap smear?
 Is there an association between progesterone condition and LOW or worse grade biopsy for women who screen with HIGH grade pap smear?
 Is there an association between progesterone condition and HIGH grade biopsy for women who screen with HIGH grade pap smear?
 Null hypothesis: no difference in biopsy result for different hormone groups
 test this hypothesis with a chisquare test
 model the odds of positive biopsy for each hormone group using logistic regression; can adjust for potential confounding (e.g. age, progesterone exposure)
 this requires dichotomizing biopsy result (LOW and HIGH) vs negative
 consider modeling the proportional odds of positive biopsy for each hormone group using ordinal logistic regression
 this preserves the ordinal nature of the outcome variable
 consider an interaction term with PAP smear type and hormone group
2011 August 31
Jim Phillips, Anesthesiology
 Wants to look at sedative for pain management and safe for sickle cell patients with acute pain crisis in ED
 Wants to see the blood's effect on 5 biomarkers
 30% reduction in pain is considered significant
 Safety outcomes
 Wants to design a single arm unblinded pilot study
 Advised that he might have biostats support through an existing collaboration plan
 Wants to know sample size
 3 to 4 patients may be available for consenting every week. Study needs to be over by next summer.
 Pain management is main outcome. It will be measured three times for the patient.
 Could you obtain a control group?
 Recommended checking pain distribution in other studies.
Wonder Drake, Medicine
 Want to study immune response in patients with a lung disease compared to healthy controls.
 Needs to know sample size for grant. Has preliminary data.
 Main outcomes are measured by flow cytometry and western blot.
 The important numbers from the previous data are the standard deviation of the outcome.
 We recommended looking at the PS software and showed some scenarios.
 If there are more complicated relationships you need to use a model for, you will probably need even more patients than the number calculated by PS for a ttest.
2011 August 24
Jeff Waldman, anesthesiology
 Research question: whether the use of PVP (peripheral venous pressure), "low PVP technique", instead of CVP (central venous pressure), "low CVP technique", will reduce blood loss and transfusion requirement in hepatic resection.
 Details: PVP and CVP are correlated and PVP is normally 2 mm Hg higher than CVP
 Hypothesis: "Low PVP technique" is not inferior to "low CVP technique"
 Design: double blinded randomised trial. Randomise patients to those whose doctor uses CVP to manage blood loss vs those who uses PVP to manage blood loss.
 Outcome:
 blood loss, measured by the volume of blood in a canister measured by 50 cc
 plus some eye estimate of number of sponges and how blood is in them
 Concerns:
 Because PVP and CVP don't agree (2 mm Hg difference) the group with PVP may show less blood loss.
 Outcome measure (estimate by eye) is a subjective measure
 Question:
 how exactly PVP and CVP are related (slope and intercept, mean difference and mean absolute difference)
 Suggestions
 Report CVP or PVP, revealing which one was reported, let medical team make their decisions on an adhoc basis
 Report CVP or predicted CVP and stratify outcomes by whether real CVP was used
 Consider VICTR studio
2011 August 17
Patrick Norris, Surgery
 Wants to apply for biostatistics help through VICTR for writing a grant.
 Consult on statistical methods and sample size for an R21 submission.
 Glutamine supplement in trauma patients vs isonitogenous placebo.
 Primary outcome is improved glucose metabolism and stress induce IR
 Want help with identifying inclusion criteria and sample size
 May want to select a group of patients for inclusion criteria that the biggest benefit of the intervention is expected
 One outcome could be the amount of time within range of glucose
 One suggestion is also to apply for a different grant for a one arm observational study to look at glutamine levels and demonstrate feasibility.
 We estimate that this work could be accomplished in $2000.
2011 August 9
Thanh Nguyen, Pediatric Anesthesiology
 Ben Saville was present and we discussed that he or Jonathan Schildcrout may be resources from our department.
 Looking at emergence delerium after tonsilectomy in children. Wants to look at a combo of two drugs
 Wants to design a randomized trial to test this, and wants help in sample size.
 The main outcome is yes/no whether they have emergence delerium, which comes from the PAED scale, which ranges from 0 to 20 and defines emergence delerium if the scale is greater than 10.
 We strongly recommend letting the original PAED scale, which contains more information and will give greater power.
 The control group will get a different drug.
 The literature reports incidence between 10 and 80 percent, but Nguyen estimates that it may be about 40%. The smallest reduction that he considers clinically important is 10%
 We recommend that he check the literature to find out the mean and variance in the actual PAED, and also how high the scale ranges.
 We showed him PS software and showed him several scenarios.
 May want to control for myaps scale, which measured the preop anxiety.
 Planning to set up a redcap database.
 Discussed stratifying by agitation to make sure the two groups are balanced with respect to agitation (or other important factors).
 Discussed interrater reliability of the PAED scale, and recommended he check the literature for whether there has already been a study on the interraterreliability of this scale.
 Discussed pain as an outcome, maybe measured by amount of pain medicine given.
Nick Ettinger, Pediatrics
 Wants to see if admission times tend to cluster around the times of shift changes.
 Has data on the 7000 admissions during one year.
 Start by graphing
 Need to control for when the patients are arriving in the ED.
 Could adjust for the number of patients at risk to be admitted at each time, by considering it an offset term.
 Ben will work with Dr. Ettinger further.
2011 July 27
Raafia Muhammad, Cardiology
 Discussed how to record survival data: one column for the date, and one for status. The status can be recurred, dead, or neither. If the person is neither, the date column will record their last known date that they were alive. If the person has recurred, the date will contain their date of recurrence.
 Her main outcome is time to recurrence of atrial fibrillation after ablation procedure. The patients are not considered eligible to recur until after three months after the procedure. After three months after the ablation procedure, if they are still experiencing afib, they are considered as having a recurrence.
 For the time to event, you will also need the beginning date. This will be the date of the ablation.
 She also wants to compare those who do and do not have family history of afib among those who have pharmalogical therapy.
2011 July 20
Kevin Sexton, Plastic Surgery
 Working with Dr. Thayer
 Wanting to submit a Defense grant that is due 8/25/2011.
 Wanting Biostats support to (1) review/refine the grant before submission (ie, sample size calculation and statistical analysis plan); (2) determine the details needed for a BCC "contract"; and (2) the actual statistical analysis once the grant is funded.
 Feel a $2,000 Voucher would be sufficient to review/refine the grant and determine the specifics for the BCC "contract".
 The BCC "contract" will cover the actual statistical analysis once the grant is funded.
Chelsey Smith, Anesthesiology Summer Intern
 Health literacy and surgery outcomes
 Outcomes: length of stay; ER repeats (30 days from hospital discharge)
 Predictors: measures of literacy
 Specific aim: determine if there are associations between literacy measures and outcomes
 Data: all different surgeries; past 10 years
 Important: many different types of surgeries  may want to "adjust" for type of surgery or look at only a specific group of surgeries
 Important: possible multiple surgeries per person in the retrospective data collection timeframe, which could be correlated with each other  may need to select only one surgery for each patient (depends on the overall proportion of patients who have more than 1 surgery)
 Important: Elective surgery vs surgeries after being admitted to the hospital
 Important: will be difficult to tease out coming back because of surgical complications vs coming back because of health literacy issues  really want to look at those who returned because of health literacy issues
 In such a large group of data (N ~ 20,000) these numbers may even out
 Would need to be stated as a limitation of the study
Shilpa Mokshagundam, Anesthesiology Summer Intern
 Effects of methylphenidate (primary component of ADHD medicine) on time of emergence from anesthesia
 Population: children on ADHD drugs
 Four groups of interest: No ADHD, No medication / ADHD, No Medication / ADHD, Medication w/out methylphenidate / ADHD, Medication w/ methylphenidate
 May want to create more than 1 variable to capture these four groups  ie, (1) ADHD as No / Yes and (2) Medication as No Medication / Medication w/out methyl / Medication w/ methyl
 Some children have had multiple surgeries  may have been on ADHD medication for some surgeries, not on medication for other surgeries
 Data: 60 kids (w/ ADHD), but ~50 different kind of surgeries; ton of kids without ADHD
 To adjust for different kind of surgeries: incorporate measures of the different phases of anesthesia  specifically, measures of induction and maintenance (third phase of "emergence" is the outcome).
 Also: weight, age, current ADHD dosage (or the fact that the child was taking the medication within a specific time frame of surgery)
Karen Kagha, Anesthesiology Summer Intern
 Time frame: Oct 2004 to March 2011
 Noncardiac procedures
 Time frame covers a period when there were no alerts to clinician to administer betablockers; a period when there was a popup alert; and a period when there was a hard stop alert (clinician has to either give betablocker or give reason why they weren't administered).
 Popups appear during surgery (ie, interoperatively)
 Were patient outcomes different between three different periods
 Thought is that clinicians' compliance should have improved over the three time frames, so patients' outcomes should have improved
 Betablockers given to patients who have specific clinical characteristics
 Important: need to know (in detail) whether each patient continued to receive the betablocker postoperatively if that's when the patient outcomes are measured
 Thought: exclude emergency surgeries
2011 July 13
Stephen Kappa, Medical Student working in Urology
 Prospective, randomized, trial comparing stapler vs. ligature during cystectomy surgery for preventing blood loss. Has data from 80 patients. The main outcomes are amount of blood loss and operative time, total device cost, number of additional staples (in ligature group). The number of staples is a cost issue.
 A clinically meaningful reduction in blood loss is around 300 mL.
 Another study: prospectively collected radical cystectomy database of ~1000 patients over 10 years. The goal is to look at the overall survival and cancer survival. Want to look at neoadjuvant chemotherapy. They want to look a the frequency of use of chemotherapy in the database over time.
 Want to know why patients who did not get the neoadjuvant therapy didn't get it. Their question is about how to categorize the reasons why people didn't get the chemo.
 Question about how to graphically display the usage over time: Error bars suggested to show variation.
2011 June 29
Brannon Mangus, ENT
* Research question: compare two types of surgery: 1) replacing STAPES with a prosthesis or 2) combining STAPES with a prosthesis part using laser. Each surgery was performed on a different population. Type one was performed on 600 patients. Type two was on about 100 patients. We have to compare the outcome and cost.
 Outcome is the difference between hearing level before and after surgery: less or equal to 5 or greater than 5 decibels.
 We suggest logistic regression model adjusting for age, gender, race, and the type of surgery.
 Criticism of this approach: two groups may have different baseline readings (which is not available for the first group). Suggested solution: use current data (second group) only in a regression model with outcome of actual decibels difference between baseline and aftersurgery (continuous) adjusted for age, gender, and the baseline reading. If baseline reading isn't significant, we'll feel "better" not having baseline reading in the main model. * Cost analysis. All costs should adjusted to 2010 dollars. Outcome: average procedural cost. Analyse the difference between groups using Wilcoxon Rank Sum test.
2011 June 1
Jesse Ehrenfeld, Damon Michael, Elizabeth Card, Anesthesiology
* First question about how to handle missing data in outcome. Had a nonrandomized intervention trial. * Groups assigned sequentially  those who received standard of care group and then those who received intervention * Outcome was nerve damage after surgery. Outcome was assessed via 30day followup attained from phone calls. * About 20% of intervention subjects could not be reached for followup. * Given that it was outcome data missing, suggested describing differences in those losttofollowup and those with complete information. * Final analysis would need to be on complete cases. * Second question related to vital signs in patients. They have a device that measures the vitals continuously in addition to the values the nurses are taking. They have ~1300 patients and would like to compare the two values.
2011 May 18
Damon Michaels, Elizabeth Lee, Anesthesiology
 Post operation pain assessment in autistic children who have trouble communicating
 Want to evaluate a validated tool that assesses pain and discomfort in children who have trouble communicating. This includes parental input. This will be the intervention.
 The children are all undergoing similar dental rehab surgeries
 Evaluation using a parent questionnaire
 Outcomes include pain medicine use, parents survey on satisfaction, length of stay
 The researchers feel that a 2530% increase would be clinically significant
 Recommend controlling for for body mass in the model for amount of pain medicine.
 Recommended randomization rather than prepost, but this seems hard logistically.
 May need to control for medicine received pre and during surgery.
 Control for autism severity
 Recommend at least 30 per group, the more the better.
 Could come back to clinic after around twenty patients have data to recalculate the sample size based on the better standard deviation data.
2011 May 11
Koffi Kla, Stuart Mcgrane, Anesthesiology
 Hypothesize that nurses in PACU are less experienced in and less likely to have specific training in critical events as compared to nurses in ICU.
 Will have a lecture in and simulation of a critical event
 Is doing a pre and post intervention survey in redcap
 We recommend talking with redcap personnel and making sure that the pre and post surveys can be linked (can match an individual's first response with her second response, although the responses are still anonymous.)
 Recommend using a visual analog scale (VAS) for overall comfort level in critical situations for both pre and post. That way you can use a regression on the post comfort level, using the pre comfort level as a predictor.
 If there is an important binary outcome (yes/no), a yes/no answer choice is appropriate. You can use logistic regression for these.
 Discussed strategies for increasing recruitment.
 May want to ask each respondent if she is a charge nurse.
 Interested in comparing the survey result before taking training and after training. The training is focused on educating nursing staff to respond to critically emergent events.
2011 May 4
Tim Geiger, Colorectal Surgery
 Tim is interested in investigating the association between average intraop temperature and 1) surgical site infection and 2) length of stay. There are 296 patients and 20% have infection and were primarily discharged within 4 days. Other functional forms of temperature are of interest, but a simplified analysis plan is desired due to presentation deadline (May 14). It is our recommendation that this is doable in 1020 hours, but Jeffrey needs to be contacted to determine if personnel are available.
2011 April 27
Mike Stoker, Neurosurgery
 Looking at the association between cerebrospinal fluid leakage and complete reconstruction of suboccipital cranial defects. The main outcome is leakage (Yes/No). There are 100 subjects with one record per subject, and there are 17 events. Reconstruction is a categorical variable Yes/No.
 Recommendation: to adjust for other variables including previous MVD, age, sex, type of closure (complete, partial, incomplete), and what it was closed with. To avoid overfitting, use propensity score data reduction. To account for surgeon (there are two surgeons), include it as a random effect or to acknowledge in the propensity score.
2011 April 13
Kelli Rumbaugh, Pharmacy
 Have pre/post data on hemoglobin, hct, for 26 patients on xigris  severe symptom = bleeding.
 Consider multivariable linear regression if potential confounding exists (e.g. APACHE score).
< > Pre_Hgb < c(10.4,9.7,11.5,13.7,10.6,11.5,12.1,9.3,10.2,9.6,12.5,9.6,6.5,9.8,7.9,8.5,14.5,9.8,7.1,10.2,9.4,13.3,14.8,11.3,8.7,8.7) Post_Hgb < c(6.8,9.2,10.3,9.9,8.5,9.6,8.7,9.2,8.5,7,10.3,8.6,9.7,7.1,8.8,10.5,9,9.6,11.3,8.3,10,8.4,10.6,9.9,8.1,9.1)
t.test(Pre_Hgb,Post_Hgb, paired=TRUE) diff < (Pre_Hgb  Post_Hgb) mean(diff)
wilcox.test(Pre_Hgb,Post_Hgb, paired=TRUE,conf.int=TRUE) # results are consistent with ttest
# Null hypthesis: No difference from pre to post xygris # There is sufficient evidence to reject the null hyptohesis of no difference in Hgb from pre to post tx (p=0.008). # The mean difference from pre to post tx is 1.32 dg/ml (95% CI: 0.37, 2.62).
apache < c(25,27,28,16,27,30,21,30,23,21,32,26,25,30,24,24,25,20,27,26,31,27,32,29,32,27) plot(Pre_Hgb,Post_Hgb) plot(apache,diff)
library(Design) d < datadist(Pre_Hgb,Post_Hgb,apache) options(datadist="d") # No d > no summary, plot without giving all details
f < ols(Post_Hgb ~ rcs(Pre_Hgb,3) + apache, x=TRUE) anova(f) summary(f,Pre_Hgb=c(9,10)) # Pre_Hgb and Post_Hgb have a nonlinear association. # The effect of Pre_Hgb on Post_Hgb adjusting for APACHE score no longer achieves statistical significance (p=0.0581). # However, absence of evidence is NOT evidence of absence, so best to quote the effect size (95% CI). # For an individual with Pre_Hgb=9 versus an individual with Pre_Hgb=10, the Post_hgb difference adjusting for APACHE # is 0.37 (0.74,0). # For an individual with Pre_Hgb=5 versus an individual with Pre_Hgb=6, the Post_hgb difference adjusting for APACHE # is 0.51 (0.97,0.05).
par(mfrow=c(2,1)) plot(f)
< >
2011 April 6
Matt Landman, Surgery
 Is interested in rate of organ donation designation over time. Has countylevel data.
 Poisson regression with a random effects model.
 Or poisson regression with county and time as a fixed effect and an interaction between county and time.
 Include countylevel covariates, like median income, percent income, etc.
2011 March 9
Joyce CheungFlynn, Surgery
 The project was designed to study the expression level of the HSP27 protein in human saphenous vein remnants obtained from bypass surgery and to determine if HSP27 represents a new vascular biomarker for the metabolic syndrome.
 Advised that they did not have enough data to run a multivariable regression (n = 11). Recommended scatterplots of the continuous variables by HSP27, and strip plots separately for dichotomous variables, possibly showing another dichotomous variable using color.
 Excel file: ExcelFile(CheungFlynn_BiostatClinic0309111.xls)
 For a separate study, they are applying for VICTR support. One criticism from the VICTR prereview was that the tests mentioned were incorrect. We reviewed the stats section and improved the wording and made some small tweaks. We also discussed whether a biostatistician will be working on this project. Dr. CheungFlynn was planning on doing the analysis herself. I advised to either list herself and give some qualifications for this type of analysis or amend her request to include biostatistical support.
Padmini Komalavilas, Surgery
 We did some sample size and power analysis in PS software and discussed that these calculations are based on the assumption that the data will come from independent units and that if they get samples from the same patients they will need more statistical methods to analyze it.
Mary Williamson, ENT
 Is looking at patients with down syndrome who have tonsilectomy they are getting chart reviews. They have 120 records, and so far 15 of them have required a second surgery. They are interested in finding variables that are associated with requiring a second surgery. She is working on a conference presentation at the end of April. We advise that her data would need a lot of work to be analyzable and also that the VICTR statisticians may not be able to accommodate this deadline. We recommended narrowing down the candidate predictors to a handful to avoid overfitting. http://biostat.mc.vanderbilt.edu/wiki/Main/DanielByrne. We estimate that this request can be fulfilled with $2000.
2011 March 2
Ted Towse, Radiology
 Looking at matched casecontrol study of ALS patients
 Two time points: time 0 and 6 months later
 VICTR review
2011 Jan 12
Kathy Edwards: NIH Multicenter vaccine trial  Dichotomization issue
< > set.seed(1) mu < 100 sigma < 20 cutoff < 125 true.prob < 1  pnorm(cutoff, mu, sigma) # .106 xlim < c(0, round(2*true.prob, 2)) nsim < 10000 par(mfrow=c(4,2)) for(n in c(100,200,1000,5000)) { p1 < p2 < double(nsim) for(i in 1:nsim) { y < rnorm(n, mu, sigma) p1[i] < 1  pnorm(cutoff, mean(y), sd(y)) p2[i] < mean(y > cutoff) } rmse1 < sqrt(mean((p1  true.prob)^2)) rmse2 < sqrt(mean((p2  true.prob)^2)) eff < (rmse1/rmse2)^2 w < paste('n=', n, ' RMSE=', round(rmse1,4), sep='') hist(p1, xlab='MLE of Exceedance Prob', main=w, nclass=50, xlim=xlim) abline(v=true.prob, col='red') w < paste('n=', n, ' RMSE=', round(rmse2,4), ' Efficiency=', round(eff,2), sep='') hist(p2, xlab='Proportion', main=w, nclass=50, xlim=xlim) abline(v=true.prob, col='red') }
< >
2010 Oct 6
Laura Chang Kit, Urologic Surgery
Is working on two projects. The first concerns a surgery where an artificial sphincter is placed on the urethra for patients with urinary incontinence. The artificial sphincters are measured in circumference, and available in 0.5mm increments. The size used is the smallest size that is larger than the patient's urethra sphincter. The surgery reduces incontinence, but Dr. Chen hypothesizes that the difference in size between the patient's sphincter and the artificial sphincter is related to the amount of postsurgery incontinence. The incontinence is measured as the number of pads the patient uses per day. This is reported by the patient before surgery and at a fixed time point after surgery. We suggested a regression model with the difference in pad use (prepost) as the outcome, and the difference in sphincter size as the predictor while controlling for the presurgery number of pads and some other factors hypothesized to be relevant. We recommend she apply for a $3000 voucher for biostatistics support for an abstract and manuscript and advise that she will need a letter from her department.
Her other project involves a mesh device which is surgically placed around the urethra in women with incontinence. In some patients, the mesh can erode into the urethra or the vagina. She hypothesizes that there are certain risk factors for the erosion, such as higher BMI and smoking, and wants to do a retrospective chart review to identify these risk factors. The problem lies in getting a group of comparable cases and controls. Vanderbilt has many patients who have had this surgery here, but a very small percent of them (about 2 out of 400) have the event. Vanderbilt serves as a referral center for patients in the region who have had surgery elsewhere, but have had erosion and need treatment. She has 36 records of such patients, but there are no controls to go with them. We discussed why the Vanderbilt control patients may not be comparable to these 36 referred with events from various regional hospitals.
2010 Sept 29
Don Arnold, Pediatrics
Rondi Kauffmann, Surgical ICU
Wants to look at the influence of nutrition has on patient outcomes in the surgical icu. Specifically, the time at which nutrition is given is of interest. We discussed controlling for the amount of food and whether the nutrition was given intravenously or through the stomach. She will need to control for how well the patient is doing upon admittance. This is retrospective chart review. We advised to use the time at which nutrition was given rather than grouping early and late nutrition.
2010 Sept 1
Robert Mercie and Fernando Orvalle, Neurosurgery
Retrospective chart review of about 180 records. Evaluating 14 variables' predictive ability for a binary outcome, hemorage during surgery. The main causal outcomes that they have in mind are size of the lesion and the amount of metabolic agent. There are severely limited by their number of events, which is 6. The size of the lesion is correlated with the amount of metabolic agent.
Rachel Idowu, Gen Surgery
Rachel has worked with Meridith to decide on a sampling scheme for sampling hospitals.
2010 Aug 18
Stuart Reynolds Urologic Surgery
Has data from a survey about pelvic symptoms in adult women. Wants to find out which pediatric urologic symptoms are associated with these outcomes. Has about 600 observations. There were four symptoms, each of which are likert scales from never to frequent. One option would be to make one binary outcome based on whether at least one of these was not "never." This outcome could be modeled using logistic regression. We advised that his number of predictors will be limited to the minimum of the number of events and number of nonevents. Also, you could fit a separate ordinal logistic regression model for each of the adult symptoms.
We estimated that this project would require about 20 hours of statistical support and recommended applying for funding through victr.
Osgood, Sexton, Hocking, Surgery
Has redcap data on viability of vein grafts. They have about three different outcomes, but one of them only has 15 observations. We recommended only looking at descriptives for that outcome. We estimated that this project would require 3040 hours, and recommended that the group apply for victr funding.
John Koethe, Medicine
Has about observations on about 800 individuals of cd4 counts over time. They want to look at the association between BMI and change in cd4 counts over time. We recommended using a regression model rather than categorizing and using a Kruskall Wallis test. Got in touch with Cathy, who works with their department.
2010 June 2
Igal Breitman, Surgery
Has repeated measurements from surgery patients. There are two groups. One got a dietary treatment and the other didn't. We recommended a linear mixed model with a random intercept for the ID variable. Igal is interested in finding if there is evidence for an association between the treatment and three continuous outcome variables: glucose, insulin, and and c peptides.
Rondi Kauffmann, Surgery
Rondi has data with multiple observations per subjects. She wants to assess whether a hormone, estradiol, can predict mortality in patients. Sharon suggested a survival model.
2010 Jan 27
Brad Lindell, Med Student
Discussed a study that will be comparing children who are cancer survivors. One group receives traditional followup while the other group receives a more intensive followup. They would like to determine if the children in the more intensive followup group has a higher understanding of their diagnosis. He can have 200 patients in each group. We performed a calculation in PS to determine the detectable alternative.
Rachel Idowu, Surgical Resident
Performed a survey for trauma. This looked at their understanding of trauma and where they felt their understanding was. She is going to bring the data back next Wednesday and we will look at some logistic regression models. See GenClinicAnalyses#Rachel_Idowu_Surgery
2009 Nov 18
Ken Monahan, Cardiology
Measurement: pulmonary vascular resistance. Gold standard: cardiac catherization. Unit of measurement: WOODS, approximate range: .3 to 20. Normal value is about 1.5. Previous not invasive methods: take ultrasound, then run a regression with gold standard as an outcome and ultrasound as a covariate, then use coefficients to find WOODS given the ultrasound. What is a reasonable sample size. Suggestion by Jeffrey Blume: organize a pilot study of 2030 patients. Need to know the variability of WOODS and ultrasound, withinsubject correlation between two measurements, clinically useful range of the difference. Then estimate the variance of the difference of two measurements and use that estimate to project your sample size based on a confidence interval of this difference. After looking at BlandAltman plot you may see different variability depending on the value of the measurement (ignore this sentence for sample size calculations. )
2009 Nov 11
Dan Barocas, Urology
 Prostate cancer and predicting upgrading
Fenna Phibbs, Neurology
 Studying Deep Brain Stimulation in patients with Parkinson's
Marcus Dortch, Pharmacy Trauma
2009 Nov 04
Marcus Dortch, Chris Jones, Surgical Critical Care
Marcus' clinic implemented an antibiotic rotation system, and they want to look at the effect of this system on instance of MDR, drugresistant infection. We recommended some changes to his data setup to include the population at each quarter and the number of resistant infections at each quarter. We recommend that he return to clinic with this data setup, and we will fit a loglinear regression model with negative binomial distribution. One of the important covariates will be whether the quarter was before or after the implementation of the antibiotic rotation system. Another approach would be segmented regression.
2009 Oct 28
Sanjay Athavale, otolaryngology
Had an article with reviewer comments. We recommended some more explanations on some graphs. We also recommended using a logistic regression model instead of several chisquare tests. For a proposal for a new study, we recommended a statistical justification of the sample size.
October 21, 2009
Rondi Kauffman & Rachel Hayes from Surgery
 *Cohort
 patients in ICU. Patient are checked for blood glucose level and the amount of insulin given to them is adjusted by a protocol. The protocol gives a "multiplier". Based on this "multiplier", the insulin level is changed. Aim: assess trend in the "multiplier" right before the hypoglycemic event. Recommendations: 1. Start from spaghetti plot. Can try timevariant proportional hazard model, linear mixed effect model (since we have repeated measurements).
October 21, 2009
Yukiko Ued from Surgery
 *Cohort
 Obese patients having a gastric bypass. Measurements: leptin, 8isoP before and after the surgery. Interested in correlation b/w leptin and 8isoP. Recommendation: Spearman's correlation coefficient for correlation. For looking at the difference b/w before and after surgery use Wilcoxon Signed Rank test.
October 14, 2009
Ian Thompson, Urology
*Prospective Study, 17 patients in each arm, main outcome is blood loss, but about 30% get a transfusion. Need to adjust for amount of fluids received and baseline hematocrit. Goal is to compare two devices used in surgery in terms of blood loss. We have informative censoring. Use a ttest or regression comparing amount amount of blood lost during surgery. Use the amount collected "in the bucket" as the outcome. Overall estimated difference in blood loss.
September 23, 2009
CJ Stimson, MD student
 Had survival data for patients who had radical cystectomy.
 Wanted to test difference in hazard functions for male and female patients, since a difference was found in another study.
Rachel Hayes, Informatics
 Wants to find the best way to determine a cut point incorporating positive predictive value.
September 16, 2009
Jonathan Forbes, Neurosurgery, PGNY4
 plotted data from the study on MRI characteristics of cerebellar neoplasms * R code for plots #Clear existing data and graphics rm(list=ls()) graphics.off() #Load Hmisc library library(Hmisc) #Read Data data=read.csv('DATA_WHPEDS_FOSSA_NEOPLASM_FORBESJ1_20090916122420.CSV') #Setting Labels
label(data$mrn)="Medical Record Number" label(data$criterion_1)="Criterion 1: Diffusion Restriction" label(data$criterion_2)="Criterion 2: T2 Hyperintesity" label(data$criterion_3)="Criterion 3: Laterality" label(data$criterion_4)="Criterion 4: Tumor Exit" label(data$dwi_cgm)="Relative DWI Value of Cerebellar Grey Matter" label(data$dwi_cnp)="Relative DWI Value of Cerebellar Neoplasm" label(data$t2_hi_tumor)="Relative T2 Hyperintensity of Tumor" label(data$final_path)="Final Pathology" #Setting Units
#Setting Factors(will create new variable for factors) data$criterion_1.factor = factor(data$criterion_1,levels=c("2","0","1")) data$criterion_2.factor = factor(data$criterion_2,levels=c("1","0")) data$criterion_3.factor = factor(data$criterion_3,levels=c("1","0")) data$criterion_4.factor = factor(data$criterion_4,levels=c("0","1")) data$final_path.factor = factor(data$final_path,levels=c("0","1","2","3","4","5","6"))
levels(data$criterion_1.factor)=c("DWI Hyperintense","DWI Isointense","DWI Hypointense") levels(data$criterion_2.factor)=c("T2 Isointense","T2 Hypointense") levels(data$criterion_3.factor)=c("Hemispheric","Midline/Indeterminate") levels(data$criterion_4.factor)=c("No tumor exit from Luschka/Magendie","Tumor exit from Luschka/Magendie") levels(data$final_path.factor)=c("Ependymoma","JPA","Medulloblastoma","Other","Other","Other","Other") names(data) = gsub("_", "", names(data))
#pdf("dwi.pdf", height=8, width=8) color="gray55" jpeg("dwi.jpeg", width=8, height=8, units="in", quality=100, res=600) boxplot(data$dwicnp ~ data$finalpath.factor, outline=FALSE, border="black", ylab="Relative DiffusionWeighted Intensity (DWI)") stripchart(data$dwicnp ~ data$finalpath.factor, add=TRUE, method="jitter", jitter=.1, vertical=TRUE, pch=19, cex=.8) abline(h=c(1, 1.35), lty=c(2, 3), col=color) legend("topleft", legend=c("Relative DWI of Cerebellar White Matter", "Relative DWI of Cerebellar Grey Matter"), lty=c(2,3), col=color) dev.off()
 Second study
 study on trigeminal neuralgia looking at outcomes when the by the replacement of the bone from the craniotomy
 craniotomy for pain relief by moving the blood vessel off the trigeminal nerve
September 9, 2009
Sylvie Akohoue, gastroenterology
 We look into risk factors colon cancer in morbidly obese patients.
 We collect hormones, biomarkers, weight.
 Design: 4 groups, with 810 subjects in each group, 1  control, 2  diet, 3  exercise, 4  diet + exercise. They are followed for 12 weeks. The data collected at the baseline and at the end of the study. For groups 2 and 4, the weight data is collected each week.
 Analysis
 linear regression; outcome 12week reading, covariates  baseline reading + treatment group (four levels: see the design). (Find out how SPSS defines reference groups). Analysis of the effect of physical exercise to weight change (for groups 2, 4): look at the average weight for every week and see what group decreases more.
September 2, 2009
Bill Heerman, med peds resident
 Working on understanding analysis in a paper about a plavix study. The study looked at duration of plavix therapy. We went over some of the concepts: splines, competing risks.
Marc Bennett, Alejandro Rivas, ENT
 Working on setting up database.
 Looking at ear surgery patients. Want to track information such as complications. One goal is to be able to compare outcomes with other institutions.
 TIme variables? They need to follow longitudinally over time.
 Current software? Underlying structure?
 Pros and cons of using RedCap. Contact Janey Wang janey.wang@vanderbilt.edu about getting current data imported.
 Pros: Free. Easy to build. Easily exported to different statistical software.
 Cons: Not good at handling longitudinal data?
 Microsoft Access would be an alternative.
 Ways to directly access the data from star panel?
August 26, 2009
Steve Deppen, Thoracic Surgery
 BMI vs. resource use in lung cancer surgical resection
 Research by others has viciously dichotomized BMI
 relationship may be nonmonotonic; need a smooth nonlinear relationship
 Relating complications to obesity is also of interest but the multiplicity of complications is a problem
 May be helpful to score complications against some other outcome as has been done by G Marshall, F Grover, W Henderson, K Hammermeister in the cardiothoracic surgery literature
 Dataset has height and weight so can do a statistical test of the adequacy of the BMI formula in predicting risk, e.g., try adding log(weight) to a model that has log(BMI) in it to check if the coefficients of log(weight) and log(height) have a ratio of 1 : (2)
July 15, 2009
Raphael See and Lisa Mendes, Division of Cardiovascular Medicine
 End stage renal disease; nuclear perfusion imaging: dobutamine stress thought to be less effective in ESRD
 Goal is detecting/quantifying coronary artery disease (CAD)
 Head to head assessment of CAD detection ability with 2 diagnostic modalities
 Prevalence of CAD around 0.2
 Can take as the goal to estimate the difference of diagnostic accuracy of the two methods or alternatively to estimate the probability that method 1 is more accurate than method 2
 Nuclear perfusion imaging provides segmental perfusion assesssment (17 segments)
 Stress echo uses wall motion abnormalities in segments (but different segments)
 Ischemia, infarct, LV function can all be quantified
 Referral bias: negative tests less likely to lead to cardiac cath
 Would CT angiography be an alternative? Calcification is a problem.
 A hybrid design would be to always send to cath every patient who is positive on at least one of the two tests
 If use a CAD severity index (ordinal scale with 10 or so levels) can correlate each noninvasive test index with this severity scale and assess the difference in two rank correlations (correlated correlations; could use the bootstrap to get the final confidence interval)
May 27, 2009
Kathy Hartmann, Epidemiology/Ob/Gyn
 Question concerning grant.
Apr 29, 2009
Tom Compton, Biomedical Informatics
 Studying accuracy of nurses recording of glucose measurements compared to what's recorded on the glucometer.
 Are the nurses recording accurately?
 Now studying "multiple levels of inaccuracy"
 25 values per day per patient, about 1900 patients in one ICU, approx 90,000 values
 Is there a relationship between errors and blood glucose variability?
 May want to consider looking at length of stay, where/when in the hospital are these errors occurring, range of errors.
 Concern: What if, for example, it's a trauma patient with high variability initially and all of the errors occurred after they stabaliized?
 Recommended requesting a VICTR voucher.
Apr 22, 2009
John Wood and Brian Burkey, Otolaryngology
 110 patients, retrospective study, all possible selected from the late 90's on.
 Only 2 cases with hand function deficits
 Looking to create a predictive model of morbidities
 3 tests: Could create at 2x2 tables, look at Kappa statistics
 Created 3 2x2 tables, doppler vs allens; doppler vs surgical allens; allens vs surgical allens.
 Is somebody with diabetes/coronary artery disease/etc more likely to lose their flap?
 Recommend getting a voucher from CTSA for more statistical help.
Apr 8, 2009
CathyJenkins, Pediatric Emergency Medicine
 Measurements of severity of asthma attacks
 about 90 patients
Patrick Norris Phd, Trauma
 New way of measuring bone density, doesn't require a DEXA scan, just a CT
 Some populations commonly get these scans: older women, people on chronic steroids, etc.
 Have DNA measurements of a group of these people
 Defense department is interested in this study; finding young men getting stress fractures; not all men are getting these. Is there a genetic explanation?
 Continuous outcome.
 Sharon recommends looking at Ordinary Least Squares Model and/or Proportional Odds.
 Primary analysis: use OLS, but for "clinical interpretation" use PO.
Feb 22, 2006
Patrick, Trauma
 Relationship between intracranial pressure(ICP) and reduced heart rate(HR) variability
 Look for noninvasive monitoring of brain trauma patients
 less variability with ICP  due to low dynamic range?
 N=146 dinstinct patients, with 4~5 days of data (every 5 minutes), high mortality patient population
 avoid uncoupling?
 regress current on previous time point, prediction as a function of time lag, different behavior between live or die. Is it due to difference in HRV or sampling? Informative missing?
 determine predictors of completeness
 take out any indices related to completeness
 wider sampling intervals, use 1 hour interval instead of 5 minutes
 two completeness indices: % complete and longest gap, use multivariable logistic model to find predictors of completeness
 therapeutic scoring system, how intensively the patients were treated
 repeat completeness analysis with and without zeros
 Identify patients with zeros and ECG measures, and see what zeros are doing
 Graphs: show some individual patient data
Mar 1, 2006
Mark Kelley and Brian Gray, surgical science
 Q: relationship between time of tissue out of body and immunohemo characterization of tissue cells, zenograph of tumor
 Q: quantify chemical stains  different tumors have different stain patterns
 How to test for trend
 grow tumor in mice, take biopsy, divided into samples, let them sit outside for a while, do hemochemical stain
 validatione: a pathologist resident redo a subsample, blinded
 How to organize the data
 three outcomes: score, intensity and pencent of positive cells on each sample at each time; 7 tumors, 4 time points, 3 replicates, 8 stains
 sources of variations: tumor type, samples, time, stains
 nested block design:
 estimate trend within the same tumor type and stain method
 deal with the 3 outcomes separately
 A: repeated ANOVA, test for time effect
Muyibat Adelani, Vascular Surgery
2 May 2007
Greg Polkowski, Orthopedic Surgery
Sample size for paired ttest using estimated standard deviation of withincadaver differences from existing data (n=5). Power of 0.9 was used, sample was estimated to be 15 pairs.
< > # peak load x < c(100, 200, 300, 0, 100) # differences resected vs. control sd(x) # 182 # energy to fracture x < c(3.7, 0.1, 2.7, 0.2, 1.6) sd(x) # 1.67 delta < .15*mean(c(20.1, 3.2, 4.5, 20, 22.4)) # 2.1 used 1.5
< >
23 May 2007
Wes Ely and Delirium Assessment study group
 *Goal
 Development of a score for delirium assessment (severity vs accuracy).
 *CAMICU
 based on 4 features. 1,2,3 or 1,2,4 ==> delirium. Derive a point system, assign points to features, validate the scoring system. Clinicians can do at bedside. Feature 1: 0/1, feature 2: 010, feature 3: 05, feature 4: 5 to 4.
 *Questions
 Is it for diagnosis or risk stratification? Which patient population? What clinical outcomes  mortality, long time cognitive function, differential response to treatment? Do features have equal weights? *Current databases: VALID Database: 131 patients, daily assessment of ICU for each of the features, followup to max 14 days, about 5.5 days per patient, all ICU patients, Delirium can be on and off over time. MINd/MEND: 210 patients, with short term cognitive outcome, up to 21 days Delirium/Mortality(JAMA): n=275, with six month survival data *Outcomes:death, ICU LOS, Hospital LOS, long term CI
 *A
 For risk stratification, the scoring system should be outcome specific. Look at the raw profiles of each feature. Look at how big changes in profiles relate to the outcome. Modelling individual features in the regression models, but this is not the same as directly modelling delirium, still maybe useful. Come up a theoretical framework first instead of emprically deriving it. Go back and rescore each of the 4 features, then you have to go through the estimation and validation for each one. Itemresponse, split sample, psychometics method. When a score is derived, add other features deparately into the model, this is to test the adequacy of other components or information loss.
28 Jan 2009
Parker Gregg, Medical Student
 Doing a pilot study in Guatemala to study two different methods of treatment compared to both each other and a SOC
 "Treatments" are different methods of educating patients with STDs to educate their partners
 Initially considered doing a crossover study with three clinics  two change treatments and a third just does SOC
 Goal of study is to increase the number of people who need to be treated for STDs by telling patients to tell their partners. SOC of does not include this extra step.
 Recommended to ignore clinic effect since it is such a short study over the summer  do one treatment in each clinic. There would need to be a good washout period in between treatments, and there is not enough time to do so. Also, all three treatments would be needed at all three clinics to consider this a true crossover study.
 Need to answer the question "What to measure?" Just a global number showing a patient increase or track individual patients and who they recommend to come back.
 Can you get information on patient volumes from the summer before to compare?
 Recommended giving color coded cards (color based on the clinic) to count referrals.
 Send survey to MarioDavidson once completed.
18Feb2009
Christina Edwards, Surgery
 Studying how often people go to the ICU after a Whipple procedure
 Consider a logistic model with restricted cubic splines for OR Time, estimated blood loss.
 Include other variables such as age, may want to use splines for that variable as well.
 Examine the ROC curve for the logistic model as a measure of classification ability.
 Examine pseudoRSquare and other statistics as well.
 Fit and examine various models, CAD seemed to be an important predictor.
04Mar2009
Doug Atkinson, Pediatric Critical Care
 20 patients, 4 measurements each about 4 hours apart, 1 missing point
 comparing devices used to get measurements of blood oxygen saturation
 What statistical tools to use?
 Is a sample size of 20 enough? Go ahead and do the analysis, see if there is anything conclusive. If the confidence intervals are too wide, may need to add more patients.
 Make a scatter plot of all 80 points across time.
 Several scatter plot suggestions: Measurement 1 vs Truth, Measurement 2 vs Truth, Measurement 1 vs Measurement 2
 Kappa statistic  for first aim
 Can fit a regression model but need to take into account that data within each patient is correlated.
 See spreadsheet tips at the bottom of this page.
 For statistical help, may apply for a voucher through the CTSA.
 Recommend getting STATA to do statistics
 Check with department head to see if you have a collaboration plan.
 Don't worry about Bonferroni adjustments when doing power and sample size analysis.
 Do false discovery rate analysis at the end.
 Do one Ftest for the ANOVA
11Mar2009
Dr. Meghan Lemke, Pulmonary and Critical Care
 Has a poster presentation this weekend, question about graphs.
 If you can get the raw data, email JoAnnAlvarez and she can make some graphs
 For pvalues, use Wilcoxon SignedRank Test
Dr. Thomas Pluim, Pediatric Critical Care
 CaseControl study
 Recommend propensity scores, can be used in both matching and in analysis
Joyce CheungFlynn, Surgery
 Wants to submit manuscript to Journal of Thoracic Surgery, which requires signature from statistician signing off on statistical methods. Recommended VICTR voucher for someone to look over data, rerun numbers and review manuscript; statistician could be coauthor.
