Data and Analysis for Clinical and Health Research Clinic Notes (2014)
2014 Dec 11
Jessica Wilson, Endocrinology Fellow
 Retrospective descriptive study about glucagon stimulation test to diagnose growth hormone deficiency
 Indications: concern about pituitary tumor or suspected growth hormone deficiency; pediatric history to be confirmed as an adult; or postpit. surgery
 Data from adults 20082014, n=42; stable endocrinology attending and test used during this period
 Interested in GH peak vs. nadirtopeek glucose, sex, ...
 Sometimes IGF I test used for screening
 GH < 3 is currently suggested for diagnosis; suggestion that higher BMI should have lower cutoff
 Blood glucose and hormones taken q30m for 4h
 Sample includes some diabetics (all controlled)
 Descriptive statistics  boxplots (including quartiles) will be useful; can also consider extended box plots
 May be useful to compute Spearman rank correlation coefficients between all variables and calendar time (use year + fraction of a year)
 Obesity and use of oral estrogen are of interest
 Can use individual Wilcoxon or Spearman test for association with peak GH release
 A multivariable model predicting GH may be of interest if a small number of highly clinically relevant predictors can be prespecified
 Scatterplots are almost always helpful
 Correlation coefficients can be useful for getting a rough ranking of strength of association with GH across multiple patient characteristics
 Sometimes it is interesting to try to correlate lab measurements with things that should not matter, to make sure they don't (e.g., time of day when patient study started)
 Show "spaghetti plots" of timeresponse for glucose and GH; plus individual scatterplots (one for each patient) and hysteresis loops
2014 Dec 4
Angela Joanne Weingarten, pediatric cardiology
Project Description: Diagnostic cardiac catheterization in congenital heart disease is essential for accurately assessing hemodynamics in children and adult cardiac patients. The current standard for pressure measurements in congenital heart disease are fluid filled catheters for transduction of pressure in real time. These catheters in are introduced to the heart using the Seldinger technique in which guide wires are used to introduce catheters to the heart through the neck or groin. In our lab, we use catheters that are 47 french which is an external diameter or 1.2 2.3 mm. While these catheters are good for most pressures measurements, there are certain situations when pressure measurements of smaller or stenotic structures are limited by the thickness of the catheter and its pliability. In adult cath labs, a new technology has been developed and is commonly used to measure pressure gradients in diseased coronary arteries called pressure wires. These wires are similar in size and structure to the introducer wires that we use in our lab to introduce catheters to the heart, but have the ability to transduce pressure. The diameter of the wire is 0.36mm and therefore are able to transverse stenotic coronary arteries. These devices are FDA approved for adult cardiac cath and have been described in the literature for congenital heart disease in children, but the pressure measurements have not been validated for children. The aim of our project is to validate these pressure wires in children and congenital heart disease using a comparison of the standard catheter pressure measurements.
 Question: help with evaluating the power I would need to validate a device used in the cath lab.
 Compute mean absolute disagreement between two measurements
 Make BlandAltman plot (difference vs. average)
 Issue of precision, not power
 Precision (margin of error) gets better as the square root of the number of patients
 To be able to do a sample size calculation requires preliminary data with differences
 Or analyze data sequentially until margin of error is satisfactory
Paula DeWitt, Center for Biomedical Ethics and Society; Madhu Murphy, pediatric cardiac intensive care unit
Email: We are wanting to test the effectiveness of a “journey board” (see attachment) designed to better prepare parents for their child’s stay and eventual discharge from Vanderbilt’s pediatric cardiac intensive care unit. This will entail giving selfadministered surveys with preparedness and satisfaction items to parents of children hospitalized in the pediatric ICU immediately before and immediately after the parent has been exposed to a 15minute educational intervention using the journey board. The intervention will take place in the child’s hospital room (or another room in the unit) and will consist of a clinician walking the parent through the journey board, and answering any questions the parent may have. Immediately before this, a researcher (not the clinician) will approach the parent, explain the study, and ask if the parent would like to participate. If yes, the parent will be given two short (510 minute) selfadministered surveys to complete. He/she will be asked to complete one prior to the intervention and one immediately after the intervention. The data will be used to assess the effectiveness of the journey board in preparing parents, and we would like your advice concerning numbers of parents we will need to interview to obtain statistical significance and statistical techniques to be used.
 Frank's note: Design is confounded with time/fatigue/learning. Also there is little precedent for doing a prepost study with such little time between pre and post. I think you will need to do a randomized study to attribute any effect to the intervention. Randomize 1/2 of families to get the intervention, 1/2 to get the prevailing treatment, and give survey at the "after" time point for both groups.
 Discussed individual patient randomization vs. prepost design (the latter based on calendar time, not pre and post within the same patient)
 Staffing constraints  some patients randomized will not be assessible due to no staff present the day but can likely assume this day of week effect is random
 Sample size could be chosen so there is adequate precision or power for the single most important subscale in the parental stressor scale
 Need a standard deviation of this scale, and a relevant difference on that scale not to miss, or a margin of error in estimating a mean difference between treatment arms
Ryan Delahanty, Epidemiology
 ICD9 codes can be independent or dependent variables
 14,000 patient database; 500k total codes
 Interested in precursor codes
 Example question: using previous ICD9 codes to predict later readmission
 Discussed two data reduction methods: AHRQtype diagnostic groupings of ICD9 codes and projecting the codes onto UHC expected mortality (Dan Byrne)
 May want to consider lab data, number of previous admissions, etc., and don't forget age (spline to account for nonlinearity); watch for importance of "present on admission" ICD9 flags
2014 Nov 20
Tracy Marien  Endourology and Laparoscopic Surgery
My research is in regards to stone composition in the US and the correlation between various compositions and age, gender, and geography.
Large dataset from lab.
 100,000 patients in csv file, one row per patient with a cell that has mixed information for each component per type per patient (type = % vs. absolute)
 Interested to know dominant components, do older patients have different stone types, are there geographical differences?
 Location: city and state, zip code
 Suggest applying for a standard VICTR voucher, time required 35 hours ($2000)
Kate Clouse, VIGH
Question about a power calculation in a study of pregnant women. Want to screen N pregnant women and estimating a strep point prevalence (range expected 0.05  0.25), then want to power to detect differences in detection between 2 assays. Want to be able to detect an absolute difference of 0.1 or more. Two assays are done on the same samples so data are matched. Hope for discrepancy <= 0.05. The culture is considered the gold standard, newer one is a rapid assay.
 Estimate disagreement proportions
 Estimate directional agreement proportions (sensitivity, specificity)
 Oversimplification: what is the sample size N such that one can estimate the probability of disagreement to within a margin of error of +/ 0.05. Answer: N=380. N=1530 pairs to get a margin of error of +/ 0.025.
 To estimate sample size needed to estimate sens and spec need to know the proportion culture positive
 Kappa statistics will also be of interest (chance agreement  corrected proportions, e.g., if two true probabilities are 0.9 and samples were independent, would get 0.81 agreement just by chance)
Meridith Blevins
Reviewing methods of a colleague's paper on sexual behavior in men. Goal is to create a sexual risk (for HIV) score based on 11 behaviours. So far a lot of mysterious univariable analysis has been done. Ordinal predictors currently being treated as polytomous. One general solution is to treat the predictors as polytomous but put a "successive category" penalty on regression coefficients connected to the same question. Hans van Houwelingen had a paper on this in Stat in Med.
 Build a "best" model (e.g. using penalization) then ...
 Do model approximation (sometimes called preconditioning in the literature) where you use backwards stepdown regression predicting the predicted values from the "best" model to develop an easiertouse submodel; how many variables can be dropped before the approximation accuracy (R^2) drops below 0.95?
 Maybe also entertain keeping the variables as linear in the model then at the last step using all the dummy variables for the remaining questions
2014 Nov 13
Schola Nwachukwu  Endocrinology
I and my team will be needing some advice on a research project we are working on. Our project is a biomedical research which entails using electronic medical records with associated genomic date as a tool for discovery in novel diabetes pathways. We are looking at type 2 diabetics with extreme phenotypes of glucose and lipids. We currently have data on triglycerides:HDL ratio which we have plotted into a histogram. We are hoping to get advice on the best sample size of extreme phenotypes to use based on the data we have.
 Need to know minor allele frequencies
 Suggest computing the total sample size n then genotyping the lowest n/2 ratio patients and the highest n/2 ratios; no need to solve for a cutoff
 May need to do an analysis to show that triglycerides and HDL are irrelevant for this purpose once you know the ratio
 Best to plot ratios using a log axis
Erin McGuinn, Matt Semler, General Internal Medicine Division
I will be coming to biostats clinic tomorrow at noon to discuss data analysis for a project I am doing with the VALID database regarding chronic glucocorticoid use and risk of ARDS. Setting is sepsis.
 Look at steroid use preadmission
 Y=ARDS in first 96h; secondary inhospital mortality
 Confounding by indication especially severity of illness
 VALID cohort captures all VUMC ICUs 20062013
 Use a multivariable propensity model as a descriptive tool to understand treatment selection
 Main reason to use propensity score analysis is that the number of potential covariates is too large in relation to the effective sample size
 If <= 15 clinically potentially interesting covariates, ordinary covariate adjustment may be fine; need to include all known reasons that steroids may be used
 n=1000, 410 outcomes so effective sample size is good except for there being 130 on steroids
 Doseresponse curve is of interest; prednisone dose equivalents have been calculated
 Regression spline in cube root of dose eq. may be worth trying
 Steroids may be longterm vs. shortterm; a secondary analysis could include an interaction between dose and type
 If dose eq. distribution varies greatly by short/long, there is great difficulty in figuring out where to put knots on spline functions; ordinary polynomial in cube root of dose eq. has worked well in other situations (quadratic probably, maybe cubic)
 How to handle interplay between ARDS and death? Union the two outcomes? Or create an ordinal outcome scale (0, 1 (ARDS), 2 (death))
2014 Oct 30
Susan Eagle, Anesthesiology
Nick Salterelli, Emergency Medicine
I am working with two emergency physicians to build a clinical study to be performed in the ICU, so it looks like Wednesdays make the most
sense, but any day is fine.
Brief project summary:
* A brief previously studied ultrasound protocol would be conducted at the bedside of patients on ventilation for <48hrs
* The scan would sort patients in one of seven "diagnosis" bins
* The standard of comparison would be the documented chart diagnosis of the cause for their respiratory failure
* The goal would be to calculate sensitivities and specificities for each "diagnosis" bin
The above would be the minimum acceptable analysis. I'd like to make further comparisons, but realize adding additional complexity may be
limited by my ability to acquire an adequate sample size. In order to move forward, I was hoping for some help understanding how much
larger sample sizes would need to be for the following comparisons:
1. The above protocol being completed by two groups of operators, and comparing their performance
2. The above protocol being completed with two different ultrasound devices, and comparing their performance
3. Combining 1 & 2 to compare both different operators and devices
Additional questions:
* Is multiple operators performing the protocol on one patient a valid way to increase sample size for these questions?
I will have funds available through the medical school student research program for statistical support down the line. Additionally, I
plan to apply for a VICTR voucher once our protocol is complete. For now, hoping to get some of these questions answered so I can move forward in
planning for funding/device acquisition/department approval/etc.
2014 Oct 23
Christopher Brown, MD
My project is a randomized trial of once daily versus twicedaily labs for patients who are being actively digressed for congestive heart failure. My question is the best way to randomize the patients given the limitations of The complexity of ordering the labs for the patients. More specifically my question is: can I randomize based on the team they're placed on. Meaning team A does once daily labs and Team B does twicedaily labs and patients are randomly assigned to each team; does that count as randomization or is en bloc sealed envelope a better methodology. Also hoping to determine the number of patients needed to achieve the appropriate power to detect a difference in outcomes.
Email from Robert Greevy:
I often attend the Thursday clinic and I will tomorrow if I'm available. I always ask dozens questions, and I've noted some I would likely ask below.
1) What is your primary outcome?
2) How could the twicedaily impact the outcome differently than oncedaily?
3) What is the expectation of the outcome under the status quo, e.g. median time to some event, 30 day M&M rate, etc?
4) What is the smallest clinically meaningful effect size that you would like to be able to detect for that outcome?
5) If randomizing at the patient level, how well do you estimate the protocols would be adhered to, e.g. would we be lucky if half the patients actually got the number of labs they were assigned to get?
6) If randomizing at the team level, how many teams could be randomized? Are there alternatives to randomizing teams, e.g. randomizing study days such that the team on that particular day will follow a randomly assigned protocol?
7) If randomizing at some sort of cluster level, how well do you estimate the protocols would be adhered to?
8) What are the potential sources of bias to worry about, e.g. team quality, season, week day, etc.?
Your question is essentially how much do I need to randomize, and the answer depends on who you need to persuade and what are the limitations of your setting. Would randomizing groups A and B qualify as having done a randomized study? No, that study design would essentially require just one coin toss. If bias exists between the teams, e.g. one team provides better care, then the study will be biased.
That said, the alternative of randomizing at the patient level may not be preferable. If adherence to the randomized assignment would be poor, the study would require complex analysis and still not be very persuasive. The 2006 SPORT trial is a nice example: http://jama.jamanetwork.com/article.aspx?articleid=204281 and http://onlinelibrary.wiley.com/doi/10.1002/sim.4510/full .
I suspect there may be an inbetween design, such as randomly assigning the intervention based on study day or something in that vein, that would strike the best balance between logistics/adherence and randomization helping to control for unobservable sources of bias.
2014 Oct 16
Mary DeAgostino, MPH student
I am working on my analysis plan for my MPH thesis project surrounding sex differences within the NUSTART study results, and would really appreciate some help in the structure for the analysis.
 Sex differences in body composition following ART initiation in HIVinfected adults. Data is longitudinal with baseline, 6 weeks, 12 weeks. Intervention is nutritional supplement. N=1800, subset=900 with CRP (Creactive Protein, marker for inflammation).
 Outcomes: Fat Mass/Lean Mass, Upper Arm, Leg circumferences, etc. (Body Anthropometrics)
 Exposure is time on ART and grouping by sex?
 Control for age, CD4, etc., is there a difference between men/women in gains following ART?
 Consider CRP in the above question by sex as well.
 Response Feature Analysis: Take a biologically plausible summary of repeated measures (e.g. AUC, slope). This takes you back to 1 observation per patient. For fat mass gain, then AUC could capture the extent to which fat mass increases over time. If you expect a linear relationship between treatment and fat mass gain, then you could get the slope of fat mass for all visits and use this summary measure. If response feature is slope, can take logarithm before modeling.
 Alternatively, some more "fancy" methods take correlation into account. Do both response feature and mixed effects model and investigate that results are the same.
 GEE model is a good method that uses Huber White sandwich estimator to adjust for correlation.
 Repeated measures analysis with random slopes and random intercepts (mixed model).
 For continuous data, generalized least squares takes into account the correlation between repeated measures on the same patient.
 Use of multiple imputation: to retain data because of random missing data, probably wouldn't impute main covariates of interest (i.e. outcome and exposure)
 Effect modifiers: put in an interaction term.
 In limitations, acknowledge differences in completers/noncompleters.
 Make sure that you look at the data to make sure the models aren't "screwy". Try spaghetti plots. Or draw spaghetti plots of subset of patients from percentiles.
 For VICTR, number of hours would be 25 hours.
2014 Oct 9
Matthew Kolek, M.D. Vanderbilt Heart and Vascular Institute
I would like to reserve a spot for biostats clinic Thursday. I have 3 questions concerning my dataset:
1) best analysis
2) interim power calculation
3) quote for stats support for final analysis
I’m doing a VICTRsupported pharmacogenetic study to see if genetic variants modulate how patients with atrial fibrillation respond to betablockers. I’ve studied 31 patients so far and am in the process of asking VICTR for more funds. I would like to present VICTR with an interim power calculation. I would also like to ask VICTR for funding for stats support.
 There are about total 80 subjects. Heart rate will be measured before and after taking atenolol.
 Primary outcome is the hear rate after treatment, which is measured every minutes. Since the total time for each patient is different, can calculate adjusted area under the curve and the take the difference between before and after.
 Imputation is not suggested due to data not missing at random.
 Consider do a longitudinal analysis using generalized least squares using all the data.
 Need to apply VICTR voucher for statistical analysis, suggest $4000 including manuscript preparation
Jo Ellen Wilson, Psychiatry
2014 Sep 18
Jo Ellen Wilson, Psychiatry
Jo Ellenwould like help developing an analysis plan for her VICTR proposal concerning:
Brief Introduction and Background:
Delirium, a syndrome of acute brain dysfunction is routinely screened for and recognized in intensive care unit (ICU) patients. Catatonia, a neuropsychiatric phenomenon, characterized by psychomotor changes, can appear as clinically indistinct from delirium in some patient settings, yet is not routinely screened for in the ICU setting. This study seeks to explore the relationship between delirium and catatonia, the extent to which an overlap syndrome exists, and the extent to which this overlap syndrome is clinically relevant.
Study Aims:
 To determine the degree of overlap between diagnostic criteria for delirium and catatonia in medical and surgical ICU patients. We hypothesize that delirium and catatonia will occur as an overlap syndrome in the critically ill population, such that those who meet delirium criteria will also frequently display signs of catatonia that are missed in routine practice. We will test this hypothesis by screening medical and surgical ICU patients on mechanical ventilation or in shock who have consented to the MINDUSA or MENDSII studies and enrolling them into the DCat (delirium and catatonia) study.
 To determine if patients with delirium manifest more catatonic signs than those without delirium. We hypothesize that patients who are delirious (CAM+) will manifest higher BushFrancis Catatonia Rating Scale scores than those who are not delirious (CAM). To test this hypothesis, we will have each patient enrolled into the DCat study evaluated twice daily while in the ICU and once daily on the general medical ward by a reference rater (performing the CAM) and C&L Consultant Psychiatrist (who will perform the BFCRS).
 To understand the clinical relevance of the coexistence of delirium and catatonia as an overlap syndrome to clinical outcomes including length of stay, survival, and longterm cognitive impairment. We hypothesize that patients with delirium and catatonia (D+C+) will have a longer hospitalization, worse survival, and more severe decrements in longterm cognitive impairment than those without this overlap syndrome. To test this hypothesis, we will monitor length of stay, 3 and 12 month survival, and a battery of neuropsychological tests on all patients enrolled into the DCat study.
 To determine if patients with delirium who are on benzodiazepines experience less catatonic signs than those who are experiencing delirium in the absence of having received benzodiazepines. We hypothesize that the use of benzodiazepines in ICU patients and its differential effect on delirium vs. catatonia (i.e., benzodiazepines are thought to worsen delirium but are used as a treatment for catatonia) will serve as a means of better understanding the differences between and phenomenology of delirium and catatonia. To test this hypothesis, we will study the performance on the CAM and BushFrancis Catatonia Rating Scale scores (see Aim 2) in light of benzodiazepine exposure in each patient and across the population.
 To determine if patients with a hyperactive delirium and catatonia experience an increase in the number of catatonic signs after starting an antipsychotic medication. We hypothesize that the use of antipsychotics in a subset of ICU patients who have a hyperactive delirium and (excited) catatonia will experience a worsening of their catatonic signs as compared to those with hypoactive delirium (i.e., antipsychotics are thought to improve delirium but anecdotally are thought to worsen an excited catatonia). As a complement to Aim 4, this will serve as a means of understanding better the differences between and phenomenology of delirium and catatonia. To test this hypothesis, we will study the performance on the CAM and BushFrancis Catatonia Rating Scale scores (see Aim 2) in light of antipsychotic exposure in each patient and across the population.
 Need to consider censoring due to death
 Would be better to treat D and C always as ordinal severity levels rather than dichotomous
 Can envision this as a scatterplot
 Can phrase one of the questions as "What is the distribution of C for each level of D?", and can address synergy using continuous variable x continuous variable interaction assessment
 Beware of confounding by indication: drug usage depends on severity of symptoms
 Can study drug usage as a function of the joint levels of D and C (i.e., reverse modeling) to understand the impact of drug usage jointly on D and C
 i.e., regression model predicting drug exposure from predrug D and C levels and from their interaction
 What other variables should be in LOS, cognitive outcomes, and survival models? It will probably be important to adjust for prognostic factors (e.g. for survival, the acute physiology score of APACHE II or III); perhaps also consider number of days already in the ICU before the assessment began
 Sample size minimum is 96 (number needed to estimate a single proportion with +/ 0.1 margin of error); will need more for associations; n=200 is perhaps OK
 Pick most important aim to guide sample size
 How to deal with longitudinal data?
 For the purpose of relating C to D sometimes it is OK to pool all available days as if you had more subjects
 For getting a confidence interval for the correlation coefficient (e.g. Spearman rho) can use the cluster bootstrap
 N=200 yields a margin of error of +/ 0.14 in estimating a correlation coefficient
 A summary statistic approach may come into play (e.g., perpatient area under the curve or average response; perpatient slope)
2014 Sep 11
Michael O'Connor and Don Arnold
This project is currently ongoing, but was discussed with Ben Saville during study design, but we will now request VICTR funding for biostats support as we complete the project. We do not have any data yet to discuss, but are study aims are listed at the bottom of this email. Are main statistical questions are as follows:
 Best statistical test for interrater reliability for this dataset (kappa?).
 Is it statistically ok to calculate interrater reliability for the complete AAIRS and for the individual components of the score?
 We’re not doing a sample size calculation, at least for the prevideo time period. Should we do one for the postvideo?
Joseph Conrad, Wright Research Group, Vanderbilt University
 Email: Would like to discuss more statistical analysis of flow cytometry (last discussed on May 22)
2014 Sep 4
Rifat Wahab, Department of Radiology
 Email: Dr. Spalluto and I are currently conducting a survey at the One Hundred Oaks Breast Center. The survey is in regards to whether patients have a gender preference with the person performing their diagnostic mammograms and biopsies. Additionally, the study has been IRB approved. We need assistance from the Biostatistics department on how to best organize our results. Also, could you please help us determine how many surveys would we need to obtain to have statistically significant data?
 Main question: What makes these patients comfortable, i.e. what are patient preferences?
 Surveys filled out by patients before imaging procedure; anonymous
 Started survey last month (on paper); not for women coming for screening exam
 Variance in how survey is requested and how long in waiting room
 Need to get an accurate count of the total number of women who came for the diagnostic test, and compare demographics of those doing the survey with overall demographics of women at the clinic (diagnostics only)
 Regarding sample size, focus on precision. For a yes/no male/female question the margin of error (1/2 the width of the 0.95 confidence interval) goes down as the square root of the number of respondents. For n=96 the margin of error is 0.1 if the true probability is 0.5. For n=200 the margin of error is 0.07. For n=300 it is 0.06. For n=500 it is +/ 0.04.
 When stratifying by race or education the demoninators are lower and hence margin of errors higher
 Margins of errors for measures of association are higher
 One approach to sizing the study is to determine the fraction of subjects that are in the smallest category for which you reallywant to have an estimate of a probability and the make sure there are enough (e.g., 96) in that category
 Or just compute the overall n that will achieve an adequate overall margin of error for estimating the prevalence on one category
 If desire to look at associations between two or more variables, it is best to completely prespecify the associations to be tested, before looking at the data
Jejo Koola, VA Medical Informatics
 Email: I would like assistance with a basic sample size question. Experimental design involves subjects evaluating a scenario and producing a yes/no answer. Outcome measure is % correct for each subject. There will be two cohorts, and each subject will analyze XXX scenarios. I need help figuring out the optimal combination of subjects and scenarios.
 Visualization study for high dimensional data
 Provider decisions using usual approach (having to open up and examine electronic medical record) vs. visualization
 Attendings/residents from various specialities; disease=cirrhosis of liver
 Visualization could provide uncertainty around the risk estimate and the applicability of the risk estimate to the current cohort
 Suggested nailing down a clinical outcome measure before considering power
 Debate in the literature on whether to present uncertainty measures to clinicians
 Categorization of physician predictions doesn't help
 See http://biostat.mc.vanderbilt.edu/wiki/pub/Main/FHHandouts/apha.pdf and Error: can't fetch image from 'http://biostat.mc.vanderbilt.edu/wiki/pub/Main/FHHandouts/Dialog.png': 500 Can't connect to biostat.mc.vanderbilt.edu:80 (Name or service not known)
2014 Aug 28
NO CLINIC: Roundtable Discussion with Dr. Karl Moons
2014 Aug 21
L. Tyson Heller and Todd Rice, Internal Medicine
 We would like to discuss a study on the use of Intraosseous Catheters during code situations for the establishment of central venous access.
2014 Aug 14
Reyna L. Gordon, Otolaryngology & VKC (Peabody)
 Email: I am designing a study for an R03 proposal in which I'd like to use either Structural Equation Modeling or Hierarchical Regression, and I need help conducting a power analysis.
 Correlation between musical rhythm and grammar in kids. Want to explain the mechanism behind this relationship in R03 proposal.
 For the Stat Analysis Plan: Preliminary work will look at validity of grammar measures and how to combine them for one outcome.
 Reference for mediation analysis: MacKinnon, David P., Amanda J. Fairchild, and Matthew S. Fritz. "Mediation analysis." Annual review of psychology 58 (2007): 593.
 For the Sample Size: Instead of calculating for effect size, this study will need to give consideration to the large number of covariates being considered in whatever model (SEM, regression, HLM) in order to prevent overfitting and have reproducibility.
 Why? powering to detect a correlation of 0.7 will give much smaller sample size than is needed to fit these large models
 Discussed limiting sample size based on response type. For a continuous outcome, if the total number of parameters is p (fitted covariates), then the sample size you need for reproducibility is in the range defined by [10*p,20*p].
 Reference table 4.1: Harrell FE. Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis
2014 Aug 7
Jessica Pippen
 I am working on a RCT investigating patient satisfaction with interval vs immediate postpartum IUD insertion. Attached is a draft of the protocol.
2014 July 31
Jo Ellen Wilson, Psychiatry
 I am working on a VICTR application for funds for biostat support, and I'll need to include in my application the proposed length of time for analysis (cost, etc).
 Study on Catatonia, and its association with delirium. Total N=62 patients.
 Original estimate of $5000 for aim 1 and aim 2. Suggest $9000 for all the five aims (need cost share of $3500).
2014 July 24
Daniel Croymans, Vanderbilt Health & Wellness Intern
 We are analyzing health risk assessment data on 20K subjects over multiple collection years. I would like to attend clinic on 7/24 to discuss the following topics.
 errors with regression models * GBM – solved by creating a separate dataset which only included subjects with dependent variable data. It would be nice to not have to make a subset dataset every time. If this is the only possibility, can we go over how to create this automatically in R – currently manually done with sorting in excel:
* gbm_1 < gbm(JobProductivityY11 ~., distribution = "poisson", shrinkage = 0.00001, cv.folds= 10, data = vandy)
Error in checkForRemoteErrors(val) : 8 nodes produced errors; first error: Missing values are not allowed in the response"
 GLM Diagnostics – others listed in attached R code.
 # Print ANODEV table:
Anova(glm_y10g1,type="III")
exp.chi<data.frame(data.frame(Anova(glm_y10g1,type="III"))[,1])
vandy_glm1 < ddply(vandy, vandy$!MedicalCareSickDaysY10)
BMI_calculatedY10, AgeCalcPSY10_AsOf01012013_Y10, GenderPSCodeY10, EthnicGroupAbbrY10, EducationLevelY10, FLSAStatusRecodeY10, SubstanceUseSmokingStatusY10)
rand.chi< data.frame(rbind(replicate(1000, c(data.frame(with(vandy, Anova(glm(sample(MedicalCareSickDaysY10,16302, FALSE) ~ BMI_calculatedY10 + AgeCalcPSY10_AsOf01012013_Y10 + GenderPSCodeY10 + EthnicGroupAbbrY10 + EducationLevelY10 + FLSAStatusRecodeY10 + SubstanceUseSmokingStatusY10, family = poisson),type="III")))[,1]))))
summary(c(rand.chi[1,])>exp.chi[1,])
summary(c(rand.chi[2,])>exp.chi[2,])
summary(c(rand.chi[...,])>exp.chi[...,])
* ERROR:
exp.chi<data.frame(data.frame(Anova(glm_y10g1,type="III"))[,1])
vandy_glm1 < ddply(vandy, vandy$!MedicalCareSickDaysY10)
Error in `[.data.frame`(envir, exprs) : undefined columns selected
* Additional ERROR: related to unequal x and y table.
* Review Spline graphs – how to overlay multiple graphs, and a few general tips for additional 'beautification' of the table :).
plsmo(vandy$!BMI_calculatedY10, vandy$JobProductivityY10, method="supsmu", datadensity=TRUE)
plsmo(vandy$BMI_calculatedY10, vandy$JobLimitationsY10, method="supsmu”)
general advice with generalized linear models: evaluating fit of model (goodness of fit), crossvalidation of the model, graphing the model.
* #GLM model Goetzel replication: http://www.ncbi.nlm.nih.gov/pubmed/20061888
* ## Visitsi = f(BMIi, Agei, Sexi, Race/Ethnicityi, Educationi,, Professioni, Smokingi, Site), where Visitsi is an outcome vari able for an individual i, and f indicates the link function for the model. All dependent variables, except presenteeism, were counts of events.
* #GLM Goetzel Year 10 variables  MedicalCareSickDaysY10, BMI_calculatedY10, AgeCalcPSY10_AsOf01012013_Y10, GenderPSCodeY10, EthnicGroupAbbrY10, EducationLevelY10, FLSAStatusRecodeY10, SubstanceUseSmokingStatusY10
glm_y10g1 < glm(MedicalCareSickDaysY10 ~ BMI_calculatedY10 + AgeCalcPSY10_AsOf01012013_Y10 + GenderPSCodeY10 + EthnicGroupAbbrY10 + EducationLevelY10 + FLSAStatusRecodeY10 + SubstanceUseSmokingStatusY10, family = poisson, data = vandy)
summary(glm_y10g1)
* # GLM model after GBM – separate inquiry from above.
glm_1 < glm(JobProductivityY11 ~ AgeCalcPSY10_AsOf01012013_Y10 + SickDays_HRA_Y6 + SickDays_HRA_Y5 + SickDays_HRA_Y4 + BMIY9 + SickDays_HRA_Y2 + SickDays_HRA_Y3 + BMI_calculatedY10, family = poisson,data = vandy1)
summary(glm_1)
* #GLM model Goodnessoffit testing  if p < 0.05 then model is not a good fit for actual data.
with(glm_1, cbind(res.deviance = deviance, df = df.residual, p = pchisq(deviance, df.residual, lower.tail = FALSE)))
2014 July 10
Paula DeWitt and Jessica Turnbull, Center for Biomedical Ethics and Society
 Email: I would like to sign up for the biostat clinic this Thursday at 12:00 at 2525 West End. One of the physicians in our office (Center for Biomedical Ethics) is a pediatric intensivist at Children’s Hospital (Jessica Turnbull). She emailed me an excel file (attached) with aggregated data from 2 surveys (pre and posttest) conducted with pediatric bedside nurses before and after implementation of a “rounding tool.” Jessica asked me if it was possible to determine whether any of the pre and posttest items are significantly different. 79 nurses took the pretest and 49 nurses took the posttest. The surveys were anonymous, so they are not matched pairs.
 Frank Harrell's Response: In general you can't make conclusions from prepost designs even with perfect matching [due to general time trends, learning, fatigue, and wishful thinking]. Without the matching it's even more challenging. The REDCap Survey system has a way to link up anonymous responses across time, I believe, if set up prospectively. If with the existing data you do not have a way to describe differences between the 49 nurses and the 79  49 = 30 nurses, the survey may need to be redone, unfortunately. We'll know more after the discussion but be prepared for bad news just in case.
 Pediatric ICU palliative care; modeling after a Mt Sinai adult ICU project
 Rounding tool triggers for consideration of palliative care + educational component
 Pre > educational lecture + checklist for consult interventions > Post assessment
 Discussed Hawthorne effect, seasonality, bias due to subjects dropping out and not having a post assessment
 Was anonymous due to nurses' sensitivity to being judged for negative comments
 There may be some nurses who took the posttest who did not take the pretest
 Need to determine how many nurses attended the lecture; think it is less than 1/2 of them
 Attendence of a lecture is perhaps an outcome variable in and of itself
 Can emphasize in an abstract or publication the feasibility argument; use this experience to be ready to do a cardiac study
 Regarding analysis, without the ability to pair responses, would need to treat as unpaired (2sample) but override the sample size in statistical tests to a total of 79 [but also need to note interpretation difficulties]
 In future consider VICTR resources, e.g., to refine future designs
Thomas Reed
 Summer research program with Diabetes, Kidney, GI. Medical student from GWU.
 Interested in SPSS; can cover general questions in clinic that are not SPSSspecific
 Need to present at symposium at end of July
 Looking at SPRINT trial  keep SBP below 140 mmHg vs. below 120 mmHg; does this change outcomes (cardiac events)
 Ambulatory blood pressure measurements at month 27; n=27, only 7 are dippers; monitored for 24h; measured q30m
 Have means separately by sleep and nonsleep intervals
 Interested in nighttime dipping of BP; doesn't happen for some subjects
 Hypothesis is that nondipping is associated with harder to control BP and will need more antihypertensive meds
 Have the raw data
 Simplest analysis is to compare the 24h profile to the current med usage
 Need a continuous measure of dipping; otherwise statistical power will be minimal
 One possibility: get the area under the SBP curve during sleep [normalized for time duration, i.e., mean SBP during sleep] and subtract it from the SBP at some universal reference time (right before sleep? noon?)
 Correlate amount of dipping with amount of meds, show scatterplot
 Need to plot reference SBP vs. AUC, and BlandAltman plot (difference vs. sum of reference SBP and mean sleep SBP)
 How to quantify intensity of antihypertensive med usage?
 May also want to measure SBP control at surrounding visits
2014 July 3
Jennifer CunninghamErves, MMC Department of Surgery and Office of Research
 Email:I would like to reserve a time to attend the Clinical and health research clinic on Thursday. Specifically, I need assistance in calculating the sample size for my current research study.
 Administering a survey on parental willingness to consent adolescent participation in clinical trials of new vaccines/drugs.
 Discussed the issue of responding positively versus actually enrolling child in trial  could result in overestimation. Other bias might result from nonresponse (probability of participating in a trial might be linked to probability of responding to the survey); especially if African Americans are less willing to respond to the survey (as indicated historically).
 Bias should at least be acknowledged.
 Could simulate scenarios for bias from nonresponse to assess the bias effect (ie, sensitivity analysis to assumption of random nonresponse).
 Would like to test the association between race (AA/CA) and willingness to enroll.
 Prior information: Proportion of AA in study: 30%. Proportion of CA parents willing: 50%. Proportion of AA parents willing: 10%.
 May want to do a small pilot study to get a better idea of the proportions needed for sample size and power calculation.
 PS Software available here: http://biostat.mc.vanderbilt.edu/wiki/Main/PowerSampleSize
 For examining factors associated with willingness, use logistic regression modeling. Suggest she choose the predictors that are important apriori (not based on statistical significance). In that case, sample size might be motivated by 1020 events per parameter in the model.
 Give some thought to interaction terms or consider modeling both races separately (though you lose power). Could consider using ridge regression to run a full (overparameterized) model as it shrinks poorly predictive/insignificant parameters. Stata/R would do this kind of regression, but probably not SPSS.
 Consider applying for a CTSA voucher for the analysis portion of the study.
2014 June 19
Tultul Nayyar, Department of Neuroscience & Pharmacology
 Email: I would like to attend the clinic on Thursday (19th June, 2014) to get an estimate for a voucher for biostatistical analysis of my data. I have data from two groups of women on different parameters of depression.
 Want to find a biomarker to distinguish between two groups. Have N=25 patients.
 Estimate $2000 CTSA voucher.
2014 June 12
John M. Flynn, Internal Medicine, Vanderbilt University
 Email: My topic is the study of CD39 mediated regulation of human platelet function and its' influence on cardiovascular disease.
 Identify a certain gene which should be correlated with CD39
 Power is useful in designing study. Since the data is already collected, calculating confidence interval is more meaningful.
 ADPinduced aggregation is the primary endpoint.
 N=220. Frequency of A is 0.16. One way analysis of variance to compare aggregation among GG (0.36), GA (0.48), and AA (0.16). If significant, then compare between two groups (AA vs. GA; AA vs. GG; GA vs. GG) using two sample ttest. Scatter plot to check the distribution of the aggregation. Boxplot is also helpful.
 Need to have standard deviation of aggregation and a meaningful difference to calculate power. Control to case ratio is 0.84/0.16=5.25.
2014 May 22
Joseph Conrad, Wright Research Group, Vanderbilt University
 Email: I’ll be stopping by again with a grad student or two to discuss experimental design and analysis of flow cytometry data in the development of clinical diagnostic devices.
 Discuss flow cytrometry for detecting fluorescence of CD4 receptors
 Want to establish that total fluorescence and cell count have a linear relationship. Collect data from N individuals with different dilutions/pp. Longterm goal is to predict CD4 count from the total fluorescence.
 Keep all data points and use the "linear mixed effects" approach.
 Can calculate a confidence band for the linear relationship in a typical person. Also, a 95% prediction band would give the range for most of the population.
 From this model you can get estimates for experimental variability, population variability, and individual variability.
 If you have data for HIV and HIV+, can test for a difference in slopes of the two groups.
 Another option would be to calculate the slope for each individual patient using simple linear regression. This gives estimates for how density varies across separate patients. Then you have one observation per patient, and you can use traditional tests and simple statistics that do not account for correlation. This approach is called "response feature analysis". It's a way of side stepping this 'nasty' repeated measures issue.
 Key to response feature: you need a biologically sensible summary (here: the slope).
 Some talk of gating which is drawing clusters by hand to determine which data are eligible to calculate population statistics (subjective exclusion of intermediate subpopulations and/or outliers). There are computer algorithms, but they do not always outperform human judgement. Would be ideal to have objective 'gating'.
 What about supervised learning algorithm that does allow some human judgement? It exists.
 What sample size is needed to precisely estimate CD4 count from fluorescence? If only interested in CD4=500? These numbers are commonly used in clinical algorithms. When considering patients below 250, we want to minimize FN at risk of FP.
 How many samples are necessary to ensure that our estimates of quality are precise?
2014 May 15
Joseph Conrad, Wright Research Group, Vanderbilt University
 Email: Applying for VICTRS support to collect biospecimens (blood, saliva, urine, stool) that will be used as reagents in development of diagnostic devices to detect biomarkers of disease. The application requires a sample size justification. Also want to discuss study design and statistical analysis of flow cytrometry experiments.
 Collect data for diagnostic device development in RLS (Zambia). Proposed collection of 250 specimens (1 specimen/visit), need to justify.
 Malaria, CD4, HIV, TB diagnostics would be tested in the same cohort.
 Endpoint is to set up standard curves.
 Start with gold standard count of cells, then put through 2 assays (one SOC; one experimental) and they each yield a count. Wish to demonstrate that the recovery date (count recovered/gold standard count) is noninferior in the experimental condition to the SOC.
 When creating standard curves using prediction models, consider using tenfold crossvalidation to assess the predictive accuracy. Plan to discuss modeling more next week.
2014 May 1
Amanda Salanitro Mixon, Section of Hospital Medicine, Vanderbilt University
 Email: I need help recoding some variables, how to use skewed covariates, and the best multivariable model to choose. I have a dataset describing medication discrepancies in home health patients.
 Discussed possibly restructuring data to one baseline record + one record per med per patient
 Or more general, to allow extension to 1st clinic visit after home visit: yes/no flag for med at discharge, med at home, med at 1st clinic visit
 Poisson and negative binomial models are good candidates. Need an offset for # of meds at hospital discharge.
 Worthwhile to verify that errors of omission operate the same way with respect to covariate effects as errors of commission, dosage, frequency
 Fit two separate models, temporarily assuming linearity for continuous covariate effects, and compare regression coefficients across the two model fits
 Another feature of interest: time until resolution to medication discrepancy
 Could model outcomes for individual medications using GEEtype polytomous logistic regression for Y=omission, commission, other
 would not give credit or penalty for being on lots of meds
 could have as a covariate the number of meds at d/c to test dilution of attention
 Covariate transformations needed are only weekly related to marginal distributions of covariates
2014 April 24
Bindong Liu, Associate Professor of Microbiology and Immunology, Center for AIDS Health Disparities Research, Meharry Medical College
I am writing a grant application. I proposed to test the effect of a new drug in blocking HIV transmission using female tissue samples. Basically, the tissue sample will be drug or mock treated, then HIV will be loaded on the tissue. The amount of HIV pass through the tissue will be measured. By comparing the amount of HIV passing through the tissue between mock and drug treated sample, we will be able to calculate the efficiency of the drug in blocking HIV transmission. It would highly appreciate if you could help me to do a power analysis and sample size estimation.
 Novel drug; nothing known about it in relation to HIV
 Human tissue from surgery, pretreated with drug or mock
 Treated  untreated to measure effectiveness of drug
 Goal: block HIV viral transmission
 4050 samples/year
 Grant application  4 years, around 140 samples
 Can tissue be split to allow each patient to be her own control? Almost always.
 Need to know the transformation of viral load that makes the differences exchangeable (BlandAltman)
 If log transformation is known to yield a Gaussian distribution with constant variance, the log transformation probably works
 Need to find a previous study that quantifies the whole distribution of viral load or provides an estimate of the standard deviation of viral load
 Experiment needs to be similar in the ways that matter, regarding variability (treatment effect not relevant for this)
 Can get an upper bound on the SD of differences in the log; provides conservative power or margin of error estimates
Stephen Clark, Assistant Professor, Department of Neurology, Division of Neurooncology
 Clinical trial on blame tumors with methylating agent before death
 Tumors pretreatment, 46 brains at autopsy
 CIMP methylator phenotype; epigenetics; can change meth. status of DNA
 Global (withinpatient proportion of cells meth.) + CIMP identification
 Every patient is treated; prepost paired design
 Quantity of interest: double difference: difference in prepost differences between normal and tumor tissue
 Genetic heterogeneity across different regions
 Discussed advantages of precisionbased planning (margin of error that is likely to arise from the final estimate of the quantity of interest)
 To do power or precision calculation requires a standard deviation estimate and an estimate of correlation between measurements within the same patient
 If none of those are available, best approach is to say that this is a pilot study with unknown power/precision and the analytic plan is x
 Additional complexities:
 How to handle repeated measurements within patient (regions)
 Need to show differences are exchangeable, e.g., they satisfy the BlandAltman conditions (no correlation between YX and Y+X); sometimes proportions need to be transformed to achieve this
 Different doses are used
 Inherent selection problems associated with autopsy studies
 Recommended VICTR voucher
2014 April 17
Mallory Hacker, PhD, Department of Neurology
 Email: "I am designing a BioVU study to examine the prevalence of a SNP in groups of patients with various movement disorders. I am looking for help selecting the proper control group for these analyses."
 Using bioVU data and patients w/Parkinson's treated with Deep Brain Stimulation, do they have a higher prevalence of this SNP?
 Wants to plan nested case control with Parkinson's patients w and w/o DBS. How select a group of healthy controls? Decision to initiate patient on DBS should not be informed directly by allele (SNP data not available to physician).
 Suggest she analyze w/2 parameter model so that you can separately assess the hetero and homozygous patients. There are 50 w/DBS and SNP info and 200 w/o DBS and SNP info. Could be 700 more than just need sequencing.
 Need a power calculation and statement: "Estimate the odd ratios for DBS related to alleles?"
 Using ICD9 codes, can estimate diagnosis date of Parkinson's, then 912 years later they may receive DBS (<10% population). Can draw KaplanMeier plot of time from diagnosis to DBS (crude analysis). Can adjust for covariates in Cox regression model.
 Only want to include patients with known date of diagnosis. Consider using only patients with some health care interaction prior to diagnosis (to ensure they were not transfers).
 Important to assess dropout following diagnosis. If patients are not engaged in health care system for > 1 year, then censor at last encounter. This assumes noninformative censoring.
 Give consideration to whether death is a competing risk with time to DBS.
 Differences between groups (e.g. Table 1) will inform inclusion criteria, which is not all that different from designing a casecontrol. Important covariates for grant writing might be age, race, and gender.
 When selecting control group, match on time from diagnosis, age, race and gender.
 If you seek VICTR support, go heavy on programmer support and moderate on statistical support.
 Can use PS software to estimate power for survival analysis: http://biostat.mc.vanderbilt.edu/wiki/Main/PowerSampleSize
2014 April 10
Katie Rizzone, MD, Orthopaedics and Rehabilitation
 Would like to discuss project looking at age of sports specialization and how it correlates to risk of injury in young athletes. It would be a retrospective crosssectional design.
 Sports specialization is the age at which an athlete plays one sport exclusively. Research related to injury in kids early age specializing among elite athlete population (D1 students).
 Research questions: "What is the average age of sports specialization? (conditional on being a college athlete)" "Do athletes in major division 1 (Vanderbilt) specialize earlier than athletes midmajor D1 athletes?" "Describe injury history by age of specialization and school."
 Correlate with sport type, age, sex, race, and university (belmont vs. vandy). Detail injury history.
 Would be helpful to better understand the reason for specialization and maybe the amount of time devoted to specialized sport (athletic exposure annually "dose"). The dose will likely modify the effect of age on injury.
 Need to give some thought to how Vanderbilt and Belmont students may be different otherwise (ie potential confounding).
 Plan to distribute REDCap survey, aware of potential for recall bias. This is a descriptive study of student elite athletes.
 Inclusion/exclusion criteria need to be specified, measurement of key outcomes (like injury) will be important, VERY important to maximize response (over 90% at a minimum).
 Survey validation, can distribute survey and analyze results for validity/reliability. Or approach experts for feedback on "face and content validity".
 Plan is to write one manuscript from the results. VICTR study support and/or voucher: 40 hours is a rough estimate for biostatistics support given the complexity of the analysis.
2014 April 3
Lisa Jambusaria MD, Fellow, Division of Female Pelvic Medicine and Reconstructive Surgery, Department of Obstetrics & Gynecology
I am designing a randomized cross over trial to study the potential therapeutic benefit of Montelukast for bladder pain syndrome. I wanted to see if I can meet at a Thursday stats clinic to go over the type of statistics we would require and the cost for funding application.
 Protocol is to be uploaded to the main conference computer by the investigator by going here
 Arithmetic error in dropout calculation; need 56/0.7 not 56 * 1.3
 Wording of difference you want to be able to detect  based on minimum clinically meaningful change
 Better way to justify sample size: state sample size that is affordable and compute power to detect a 30% improvement at that sample size; also quote the precision (margin of error; a function of standard deviation and n)
 Y = 080 symptom scale;
 Statistical methods such as Wilcoxon signed test assume interval scale and smooth distribution of Y
 Applying for VICTR funding (study, not voucher)
 Need to describe randomization process; randomization will be generated by a VICTR biostatistician
 Double check that Invest. Pharmacy is comfortable logging into REDCap
 Need to see if REDCap can add a secondperiod reminder
 Estimated stat time to request: 20 hours ($2000)
Kathleen Weber VUSOM III
Study done by Martin Jordanoff in Radiology. Diagnostic accuracy study.
 Y = binary
 Does it matter whether reading a study outside their area of expertise?
 Timed vs untimed images?
 Attendings vs. residents?
 12 standard images, 14 radiographers; each read twice, total of 336 readings
 Computercontrolled image presentation, diagnosis recorded
 Need to account for intraobserver variability
 Patients independent or radiographers independent?
 Making inference to population of patients or population of radiologists?
 Complete fixed effects analysis would make strong independence assumptions
 Could one use random effects for both radiographers and patients?
 Can adjust for radiographer characteristics as fixed effects
Jessica Pippen
Contraception is an important aspect of postpartum care. Women are often more motivated and receptive during this period to initiating contraception due to immediate access to healthcare and the new responsibility of caring for a newborn. Moreover, this time period presents increased risk for unintended pregnancy. Women may begin ovulation before the return of menses and reinitiation of sexual activity before the 6 week routine postpartum visit, the time at which most forms of contraception are began. Nonetheless, there are still some barriers to implementing contraception within the recommended 3 week period and even by the six week postpartum follow up visit. These barriers vary by population and include inability to obtain prescriptions, failure to return to clinic for postpartum visit, or lack of understanding regarding the safety and side effects of contraception.
The IUD is a known effective long acting reversible form of contraception with a cumulative pregnancy rate of less than 1% within the first year of use. Traditionally IUDs are placed at the 6 week postpartum visit, known as interval insertion. However studies have shown safety and efficacy in placement of IUDs in the immediate postpartum period (within 10 minutes of placental delivery). Although this method ensures that patients initiate contraception within the recommended window, its main disadvantage is higher rates of expulsion (1 – 4.5 % versus 6 – 20 % within the first year)iv. Nonetheless, for some patients in whom transportation, loss of insurance or work scheduling present barriers to interval initiation of contraception, the benefit outweighs the risk of expulsionv.
In this study we hope to compare patient satisfaction along with expulsion, breastfeeding an unintended pregnancy rates in those patients who have immediate versus interval placement of IUDs (Paraguard and Mirena). Our question: Will immediate IUD placement decrease unintended pregnancy rates and provide better patient satisfaction than interval IUD placement?
2014 March 27
Stephanie Fecteau, VKC
 Followup to earlier visit to go over latest analysis
 Longitudinal analysis of cortisol levels
2014 March 20
Nyal Borges, Medicine
 Would like to create a prediction model to estimate probability of discharge (survival) based on condition and comorbidities. Rough rule of thumb is to have 10 events per parameter in the model. We are not concerned about covariate effects, but more interested in the actual prediction performance. Could consider logistic regression.
 Machine learning might be a good setting too  support vector machine, random forest methods, decision tree. These include cross validation which allows for assessing performance of the prediction. Outputs will be predictions.
 Plan is to write 1 manuscript.
 VICTR study support and/or voucher: 40 hours is a rough estimate for biostatistics support given the complexity of the analysis.
Ben Holmes and Nyal Borges, Medicine
 Want VICTR support for project examining outcomes in 260 cardiac arrest patients with statins as main exposure. This is a retrospective cohort study. There is no censoring and minimal missing data. Recommend logistic regression for neurological outcomes score and survival. Potential confounders/moderators include: age, time to intermediate events, dose of anesthesia, indication for taking statins (comorbidities).
 Ridge or lasso regression could be used to shrink parameters in this case where there are many covariates to adjust for.
 Plan is to write 1 manuscript.
 VICTR study support and/or voucher: 40 hours is a rough estimate for biostatistics support given the complexity of the analysis.
2014 Feb 27
Natasha Rushing, Department of Obstetrics & Gynecology, PGY3
 I am planning to attend the clinic on this thursday. I planning for my research and am attaching a copy of the lates IRB protocol draft. The question that I need answered is, how many charts do I need to review in order to power the study. Can you please let me know what information you will need so that I can have it prepared for the thursday session? Just an FYI, there are on average 41004500 deliveries at Vanderbilt per year. 85% of these are term deliveries (national statistic) and 24% of all term deliveries will have chorioamnionitis. The primary fetal outcome will be number of days for admission, and I'd like to be able to detect a difference of 3 days. The primary maternal outcome will be presence of postpartum fever. Not really sure how to address the difference I'd like to power since this is not a continuous variable.
 want to correlate histological diagnosis to clinical diagnosis. Compare between stage I and severe patients
 Use PS program
2014 Feb 20
Yuwei, Statistician
 Ran 250 models and 18 are significant. How should I adjust for multiple comparison?
 If lots of data reference Peter O'Brien manuscript. Predict what groups the persons are in based on their lab values.
 For each marker get Rsquared (or Spearman's rho) and rank each marker. Use bootstrap to get confidence interval of ranking. If marker #20 was the best marker and never worst than the 8th. This gives a perfect multiplicity adjustment. The confidence interval gets wider when you have more markers to compare to each other. Maybe subgroup this analysis by time rather than pool all time points.
 Reference for this method  Frank will check.
Paula DeWitt, PhD, Research Analyst, Center for Biomedical Ethics and Society
 Email: Would like to follow up on information discussed at last clinic.
 Other analysis options for overall improvement (not domain specific).
 Get proportion of questions improved per subject divided by 8 and then calculate the mean proportion and confidence interval.
 Add up all 8 questions with the 15 scale pre and post, then do a prepost Wilcoxon rank sum test.
 If able to run a study with 4x the number of patients, then confidence intervals would be half as wide.
 Confidence intervals are great summary measures because they give information on what is and is not going on (versus pvalues which can only be used to reject the null).
 Wilson score confidence interval is just one interval for a proportion.
2014 Feb 13
Paula DeWitt, PhD, Research Analyst, Center for Biomedical Ethics and Society
 Email: We did a training session/simulation at CELA this week, focused on increasing Nurse Practitioner ability to prepare families for ICU care transitions. As a condition of VICTR funding, we did a pre and posttest survey to see if the training increased the NPs’ ability to prepare families for the various transitions. There were only 4 NPs in the training, so the sample size is very small.... We are planning a grant proposal to do a larger study.
 Paula did a feasibility study collecting prepost data on 4 subjects.
 Suggest looking at the improvement at the questionlevel and summarzing the proportion of subjects who show an improvement from pre to post. If three subjects of four improved, you can report the proportion and 95% confidence interval.
 Here are the 5 possible 95% Wilson score confidence intervals:
 4/4 successes = 100% (51%100%)
 3/4 successes = 75% (30%99%)
 2/4 successes = 50% (15%85%)
 1/4 successes = 25% (1%70%)
 0/4 successes = 0% (0%49%)
 Interpretation for 3/4 successes: Data are consistent with the underlying true chance of improvement being between a 30% and 99%.
 Link to Wilson Score interval in SPSS
 Another option is to test whether the training improved results by more than 50%. It's called a onesided fisher's exact test.
 A second analysis would be to estimate for each person, the proportion of questions they improved upon. However, it gets away from assessing each individual domain. Consider the Wilcoxon signedrank test in this situation.
2014 Feb 6
John Koethe, MD, MSCI, Infectious Diseases
 Email: I have a dataset from a large HIV treatment RCT in Zambia that I would like to use for a secondary analysis of body composition and inflammation. I’d like to apply for a VICTR voucher to support the analysis and wanted to get feedback on the project and a quote.
 RCT just completed in Zambia/Tanzania looking at early mortality (specifically nutrition) in HIV patients with low BMI on treatment.
 1800 patients block randomized to receive plumpy nut regular and plumpy nut with extra nutrients. Supplementation occurred 24 weeks before treatment.
 Result from parent trial were inconclusive; however, they wish to proceed with secondary analyses:
 Reduced CRP is accompanied by increased lean muscle mass and grip strength after 6 weeks of tx/supplementation. N=360 at baseline and N=300 at 6 weeks.
 Will need to give careful consideration for informative missingness (death/LTFU). If you did a complete case analysis, can you justify that the observed effect would have been larger had the CRP for death/LTFU been observed.
 When correlating one change with another it is hard to interpret the causal pathway (it could be circular in reasoning).
 Measure CRP at times 1 and 2 and predict midupper arm circumference at time 3. Then midupper arm circumference is a latent dependent variable.
 May be easier to do baseline analyses, and then a landmark analysis.
 When doing survival analysis, consider including both measurements and not change. Maybe even an interaction term.
 Need to state the question with extreme specificity.
 There are many databases, but they have been cleaned and concatenated for this analysis.
 VICTR study support and/or voucher: 45 hours is a rough estimate for biostatistics support given the complexity of the analysis.
2014 Jan 30
Sylvie A. Akohoue, MMC
 Topics: Missing data, Nonparametric tests with adjusted variables, and interpretation of results.
 Discussed validation of scales to measure food intake and consumption behaviors.
Maxim Terekhou, Anethesiology
 Email: Population 1) We want to demonstrate difference in behavioral risk for LGBT vs nonLGBT patients and STD. 2) Approximately 220 employees have registered for samesex domestic partner benefits; Reported incidence of smoking in LGBT patients approaches 50%, whereas it approaches 25% in the general population. We therefore anticipate that with 220 patients in the LGBT group, at a 2sided alpha of 0.05, we will have 90% power to demonstrate a difference of X, between LGBT and nonLGBT groups.
 Would like to estimate probability of smoking (or STD) while adjusting for covariates (e.g. alcohol consumption).
 Need to understand how people are enrolled and identified as LGBT... (self report or somehow this is recorded via insurance).
 Is it fair to treat LGBT as a homogenous group?
 Not clear what the control group would consist of. Representativeness of sample needs to be established.
2014 Jan 23
Laura Edwards, MPH student
 I am looking at 1 year longitudinal health systems data measured each month in 10 districts in Mozambique. I want to confirm which statistical test to run and need help knowing how to best display this information graphically, because right now my graphs are separated by district and look really busy.
 health indicators (targets within a health system), and how close districts get to the target
 time series of proportions (% of target)
 some districts have incomplete reporting
 recommend time series analysis (e.g. logitlinear model of time)
 for multivariate outcomes, might generate a score or fit a series of models to avoid having to deal with complex analysis
 spline models
Some recommended references:

2014 Jan 9
Matthew Taussig, medical student
 I have a few questions regarding statistical analysis of a case control study that I inherited in the Schoenecker lab.
We are looking at levels of hypertension in children with slipped capital femoral epiphysis and tibia vara (orthopedic problems in children associated with obesity) compared to age and sex matched controls from an obesity clinic without orthopedic problems. I have done frequency matching and
1) want to make sure I did that correctly, and 2) I want to know which analyses are most appropriate
working with Dr. Jonathan Wanderer on updating the PACU pain score analysis project and I have a couple of questions regarding the modeling methodology that was used.
 Compare incidence of hypertension in two disease group. Both diseases are associated with obesity. So include obesity controls and compare two disease groups to the control.
 Primary outcome: hypertension. Have continuous BP (average of three BP) and calculated percentile based on age and gender. Should use original BP
 Suggest work with Ben Saville and Meng Xu through Pediatric collaboration
 Since the sample size is large enough, can include all the patients and fit a linear model: Blood Pressure ~ age + gender + groups, or logistic regression: hypertension ~ age + gender + groups
Maxim Terekhov, Center for Human Genetics Research
 I am working with Dr. Jonathan Wanderer on updating the PACU pain score analysis project and I have a couple of questions regarding the modeling methodology that was used.
 Update models with more pts/more variables.
Older Notes
Imputation recommendations for ongoing clinical trial about to close. Request Brian Shepherd be there. Will be inviting other faculty.
I am currently working on the randomization excel sheet to support the randomization component to a redcap registration form.
Yuxi