Biostatistics applications in surgery, anesthesiology, and emergency and critical care medicine Clinic Notes (2014)
2014 December 31
Sarah Greenberg, Research Coordinator and Health Policy Fellow
2014 November 26
Dr. Hernandez, Jennifer Morse
 Comparing the difference in time between two different endotracheal devices.
 Patients in operation room. 12/15 succeeded in A, 6/13 in B. Failure was claimed when more than 10 minutes was taken.
 Could do twostep analysis. 1. examine binary success using chisquare test or logistic regression. 2. compare time difference among success patients using linear regression.
 Could use all the data (including censored data) and do survival analysis (logrank test, or cox proportional hazard model).
 data from multiple sites. Note the differences between sites.
 The total time can be broken time into three phases. Could analyze the three periods separately.
2014 November 19
Michael Kenes and Joanna Stollings, MICU Pharmacy
 This study seeks to analyze the natural history of delirium. In order to achieve this, the study will divide patients into two cohorts: those with a continued stop of sedation after spontaneous awakening trial (SAT), and those with sedation restarted after SAT. Based on these two cohorts, we will characterize the time to resolution of delirium after SAT and the time needed to remain deliriumfree for 48 hours. Additionally, we will assess the time until reappearance of delirium once sedation is restarted in those patients who become deliriumfree following SAT.
 Given the outcomes and potentially complex data management, we estimate 120 hours for a VICTR voucher for this project.
Silky Chotai, Spine Center
 I am working on a spine research project and planning to apply for the VICTR grant.
 Implemented an IT protocol in July 2014 where each part of a standardofcare protocol is checked off in the EHR system. Want to see if compliance rates and outcomes improved after this IT system was initiated.
 Part I: Compare compliance rate for standard of care before and after IT system. This lets us determine whether there was any actual improvement in adherence to standard of care before and after IT system was implemented.
 Part II: Compare outcomes before and after IT system, while adjusting for potential confounders (severity of injury, age, etc  determine based on clinical knowledge). Hypothesis is that protocolizing standard of care will improve patient outcomes.
 Suggest classifying each piece of the standard of care protocol into three groups: fully compliant with best practice; received best possible care for that patient, even if not ideal standard of care; received care that did not match ideal care for no recorded reason.
 We estimate 95 hours for this project.
2014 November 12
Wes Self, Tyler Barrett; Emergency Medicine
Examining associations of 2 hour rate control in the Emergency Department with various genetic variants.
2014 November 5
Mary Van Meter, Medical student
I am a 3rd year medical student in the planning phase of a research project that will look at the cost of sterilizing surgical trays within different specialties of gynecology. This project is not really going to be focusing on patient outcomes, and I wasn't sure if I should plan to attend a clinic on Monday because it focuses on cost outcomes of various surgeries, or if I should attend on Wednesday as it involves surgery, or if it even matters at all. I would like to attend the week of November 37 if possible.
Kyla Terhune, Department of Surgery
Have not been to a biostats clinic before so new to this, but would
like to bring two very simple projects to the clinic:
Both on surgical education:
1) A completed project in which we assessed 50 of our interns
(assessment was done by both residents and attendings). Wanted to
review the assessments and query the best way to compare the
assessment done by residents compared to that by attendings. (I can
send data in the morning for review)
 comparing knot tying and suturing skills between two groups ("novices" and R2's)
 interested in evaluating interrater reliability (use a Kappa statistic, need to decide if a weighted Kappa is necessary)
 encouraging using two raters rather than three, better interpretability
 if any variables are continuous consider intraclass correlations (should focus on Kappa, Likert variables)
 Need to consider:
 how to define a clinically significant difference
 cautioned on the possibility of high variability in the small data set for the second group (n=8)
 consider randomization within the population of the "novices"
 estimate 40 hours of statistical support
2)Review the study design for an upcoming project that we intend to
complete on video assessment of interns and basic technical skills.
We have submitted the IRB for approval, and have the basic study plan
in mind but wanted to review the study design prior to beginning, and
would potentially apply for a VICTR grant with this one.
 also consider using spaghetti plots to compare scores between raters
 estimate 60 hours of statistical support
2014 October 29
Jonathan Schildcrout and Yaping Shi
2014 October 22
Jennifer Morse, Perioperative Clinical Research Institute

We created an educational module for 7 fellows who answered daily questions to prepare them for their board exam. We would like to compare their daily quiz participation and scores with their final exam scores. Data to be collected includes:
Quiz Data
Number of questions attempted
Number of questions answered correctly
Above data broken down by question category
MCCKAP Data: scores broken down by question category
Fellowship Data
Number of hours worked during the specified time period
Number of procedures performed
Number of attendances to lectures
Number of attendances to evaluations

Basic plan: write up educational intervention for critical care/anesthesia fellows (daily question emailed to all fellows). Compare pre and postintervention board exam scores, overall and by subscores. Suggested basic descriptive statistics for pre and postintervention groups (N=7 fellows in current data, with another 14 fellows undergoing intervention now), stripcharts and Wilcoxon test to compare pre and postintervention scores.
2014 October 1
Craig Sheedy, Emergency Medicine
 This is a followup visit to the one made September 10. Craig is working with Don Arnold in Pediatric Emergency Medicine as his mentor on a study looking at whether passive oxygenation positively influences O2 sat levels during intubation. They have retrospective data on 44 subjects who were intubated without passive oxygenation. Because this practice is now standard of care at Vanderbilt, a randomized trial cannot be used to study their question of interest.
 Retrospective data is limited to 44 patients, with main outcome of interest = lowest O2 sat during time it takes to intubate. Outcome is not normally distributed, so for actual analysis, something like Wilcoxon test is more appropriate, but using PS's ttest calculations and retrospective data, we roughly estimate 80% power to detect a difference of 15% saturation with 1:1 ratio (new vs retrospective data), and about 13% with 2:1 ratio.
 Using lowest saturation during intubation attempt as the outcome, adjusting for age and possibly race and gender, we estimate about 40 hours for analysis.
2014 September 24
Justin Godown, Pediatric Cardiology
 Email: I have an analysis to be performed involving a multivariable logistic regression looking at risk factors for antibody development after pediatric heart transplantation. I was planning to apply for a VICTR voucher for this project. Do I need to attend a clinic to discuss this prior to applying? How many hours do you usually estimate for a project like this? It should be fairly straightforward.
 Suggest limiting followup to first five years after transplant, limiting patient population to those transplanted >=5 years ago; this deals with issue of very different followup times (patients in database transplanted from 1987  2014). Possible secondary analysis looking at development of antibodies by one year after followup.
 Covariates include ischemic time, pretransplant antibody presence, etc. Anticipate ~50 events, so can include only 5 parameters in model.
 Goal is abstract with a possible manuscript; Justin says data is very clean (stored in REDCap). Estimate 40 hours for manuscript.
2014 September 17
Travis Ladner and Eric Wise, Surgery
 Genetic research in cardiovascular disease using BioVU
 100 patients, 20 vasospasms (on initial screen), 40 possible SNPs; thinking of ~40 logistic regression models (vasospasm = SNP + covariates)
 recommend either Tuesday biostat clinic or VANGUARD clinic in PRB
Justin Godown, Pediatric Cardiology
 The project is looking at strain (a measurement by echocardiogram) in patients after heart transplant. We want to compare the measurement to a group of normal controls at different time points after transplant. We also want to see how this measurement changes in the setting of rejection or coronary disease.
 Transplant patients get echos twice a week right after transplant, spaced out to every three months or so eventually
 Hypothesis is that rejection and coronary disease might be able to be detected earlier via a change in strain measured on echocardiogram
 "Normals" will be patients who are referred to clinic for murmurs, etc, but have normal echocardiogram; will be matched by age and gender at each time point
 Question 1: compare strain values in transplant patients to "normals" at transplant, 1 month, 1/3/5 years after transplant; probably estimate 40 hours for this portion
 Question 2: predict rejection by previous echocardiogram values, time between previous echo and rejection, and interaction term between them for effect modification (logistic regression with repeated measures  use HuberWhite sandwich estimation using patient as cluster); estimate about 100 hours for this portion
 Check with Frank Harrell, Yanna Song and/or Chris Slaughter for possible collaboration plan; otherwise, plans to apply for VICTR
 Strongly suggest using REDCap for data collection, since it will make data management/cleaning/analysis much easier
2014 September 10
Craig Sheedy, Emergency Medicine
 "We are trying to start a project in the pediatrics ED and would like to calculate sample sizes needed for the study."
 Craig is working with Don Arnold in Pediatric Emergency Medicine as his mentor on a study looking at whether passive oxygenation positively influences O2 sat levels during intubation. They have retrospective data on 45 subjects who were intubated without passive oxygenation. Because this practice is now standard of care at Vanderbilt, a randomized trial cannot be used to study their question of interest.
 We used PS to look at various sample sizes and different ways of approaching the study to determine whether the study will have sufficient power to detect a difference based on the number of subjects that will be feasible to recruit within a year.
 Craig will discuss with his mentor and both will return to clinic for further evaluation after discussing the results we saw today.
2014 September 3
Silky Chotai, Vanderbilt Spine Center
 I am a research fellow at the Vanderbilt Spine center, we are working on a grant proposal. I have some questions regarding the biostatistics.
 Primary aim is predictors of patientcentered outcomes, specifically pain (continuous score from validated scale). However, different surgery types get different pain scales, so cannot include all surgery types in the same model for pain (six types of pain scores). (EQ5D, a QOL score, is used across surgery types.)
 Secondary aim is to use surgical patients in the spine registry to look for predictors of high direct/indirect/total costs for patients with various surgery types. Start by looking at distribution of cost outcomes/residuals  linear regression might be appropriate, but need to look at distribution to see whether transformation or other model type is necessary.
 Tertiary outcome: identifying cost "outliers"  factors which determine patients who have unusually high costs.
 Some patients are included multiple times due to revision surgeries, and all are followed up for pain scores at multiple time points, so repeated measures are important.
 Strongly advise against using univariate analyses to select potential predictors; this will lead to potentially misleading and/or nonreproducible results. Instead, use clinical knowledge to select potential risk factors and interactions/effect modifications to include in models.
 Plan is to apply for VICTR voucher.
 For future projects with smaller sample sizes, discussed data reduction techniques like propensity scores, prioritizing degrees of freedom, etc.
 Mentor is unable to come on Wednesdays due to OR schedule, so recommend returning on a different day and having him available for discussion.
 Possibly split into two separate projects, since all the above would easily run >200 hours. Planning to use REDCap for data collection.
2014 August 27
L. Tyson Heller, Jennifer Green, Dr. Rice, not present, Med/Peds
 Under the umbrella of improving IV access on the general medicine floors in general, we have a proposal for a simultaneous study on the placement of ultrasoundguided peripheral IVs placed by Medicine Housestaff.
 Design for testing intraosseous catheters against central lines. the catheters can be put in much more quickly than the central lines
 Discussed whether/how to randomize in previous clinic
 For patients who have codes, there is about 510%
 Increase in return of spontaneous circulation (y/n)... event note is sent to redcap
 cerebral
 Questions about designing a study.
 How is time off of iv access measured?
 other study: would training in ultrasound _ decrease the time off
 Clinically, being without iv access for more than 4 hours is unacceptable.
 They are trying to decrease the time between being without access until IV therapy places the IV.
 Timepoints: time of loosing access, IV consult request, time IV placed by IV therapy.
 Could there be additional variablility due to the prioritizing for urgent patients.
Douglas Conway, Vanderbilt Institute for Clinical and Translational Research
 I would like to attend the Wednesday, 8/27/14 biostats clinic regarding an upcoming RCT. We need some power calculations done to determine the study size needed. We have collected pilot data from around ~70 individuals through a survey. One of the outcomes of the study will hopefully be a significant change in quality of life (QoL), measured by a 29 item instrument/questionnaire taken at the beginning of the trial and several more times throughout. We want to know how many people we will need to enroll to hopefully show significant, powered stats of change. The 29 item instrument is within a larger set of questions that makes up our pilot data. That raw data has been attached as well as a scoring guide generated by the institution that created the instrument. I look forward to meeting with you all tomorrow.
2014 August 20
Ahilan Sivaganesan, MD  Neurosurgery
 Needs help with database design, data cleaning, and possible extraction of data from StarPanel /Wiz.
Ashly Westrick, MPH  Neurosurgery
 She is working on a VICTR application for funds for biostat support for a retrospective abusive head trauma study, and I'll need to include in my application the proposed length of time for analysis (cost, etc).
 Wants to describe the population. Model with disposition as the outcome, i.e. home with mother/father, home with other family, rehab. Model with death as the outcome (has about 25 deaths).
 Approximate 60 hours for analysis and manuscript.
2014 July 30
Luke Krispinsky, Pediatric Critical Care Fellow
 Needs an estimate of time needed for VICTR application.
 Main outcome is endothelial function postbypass vs exposures including age, weight, baseline endothelial function, potentially sedation (though sedation is complex because different drugs are used)
 Secondary analyses: correlate endothelial function with additional outcomes (peak lactate, fluid, vasoactive ionotrope score) and correlations between endothelial function and biochemical qualities
 Eventually look at mortality, time on MV, time to ICU and hospital discharge, but not in the scope of the first manuscript
 Check into potential collaboration with cardiothoracic surgery (Frank or Hui?); if VICTR voucher is submitted, estimate max of 100 hours for above
Malena Outhay, SOM
 Gave intervention in trauma education to both medical personnel and laypeople in Mozambique, along with pre and postintervention tests; currently have 88 subjects, roughly 50/50 medical vs laypeople
 Main question: are pre and postintervention test scores different, and does that difference depend on medical vs layperson (may be other potential confounders as well, but incomplete data on these)
 Test scores could be 0100; actual data ranges from about 1390% and is pretty normally distributed
 Potential collaboration with IGH  check with Meridith to see if this project is covered
 Recommend paired ttest for first question, then linear regression: posttest = pretest + group, or possibly posttest = pretest * group (interaction + main effects) if hypothesis is that regression slopes will be different for medical personnel vs laypeople
 Adjusting for additional confounders difficult due to incomplete data
2014 July 23
Calvin Gruss, Department of Anesthesiology
 Calvin is asking for help in analyzing data from an anesthesiology project comparing the accuracy of a neck circumference estimation with a true neck circumference.
 Two main questions: is "gold standard" neck circumference measurement repeatable, and is a new method (via digital photo) as reliable as gold standard?
 Recommend BlandAltman test for both, but not in Excel addon package; look for resources like SPSS (possibly contact Jonathan Schildcrout or Matt Shotwell for direction, since falls under anesthesia collaboration)
 Possible reference
Austin Adams, ENT Surgery
 Followup questions from last week's clinic. Helped with PS calculations.
2014 July 16
Kelvin Moses, Urologic Surgery
 wants preliminary results for a grant submission.
 requesting data from Southern Community Cohort Study (SCCS). Needs power analysis and statistical plan for the data request.
 applying for VICTR biostats support for funding for this prelim project. Needs estimate.
Alex Seelochan, Anesthesia
 Original email: My name is Alex Seelochan, and I am currently affiliated with the summer anesthesia program at Vanderbilt. I wanted to request whether I may come to the Wednesday class of the Bio statistics Lab to revise Fisher's Exact Test. Specifically, I do have data to consider. I have attached the following table for your reference. My mentor (Dr. Thomas Austin) and I are trying to do the appropriate analysis and extrapolate patient sample needed for significance. Moreover, I have been using PS.
 There are two groups, intubated at 20 and 30 centimeters of water; main question is whether postintubation pressures are different between the two groups. Original plan was to collapse into "in acceptable range" vs. not; we instead recommend keeping these values continuous, graphing data (boxplot + stripchart), and doing a ttest or, more likely, a Wilcoxon rank sum test to compare the two groups.
 Recommend using SPSS for graphs and analysis rather than Excel.
 PS can calculate number of patients needed in each treatment group to see a difference of __ assuming one exists.
Austin Adams, ENT Surgery
 Planning to apply for VICTR voucher  will eventually need estimate of hours
 Plans prospective RCT comparing two types of intubation, with several outcomes of interest; will collect data via REDCap
 Will use PS to calculate sample size on primary outcomes: aspiration (yes/no) and patient satisfaction postop; need pilot data/estimates for both quantities (eg, what % do we expect to aspirate in each group, or how satisfied are patients under usual care and how much of a difference would be meaningful  need measures of variability, like standard deviation, in addition to estimates)
 Also need to figure out how to measure patient satisfaction  visual analog scale, Likert scale, simple satisfied vs. not satisfied? This will affect power calculations
 Use REDCap to full advantage  take advantage of numeric fields/ranges, dropdowns, etc (this will maximize stats support by minimizing data cleaning time)
 Talk to Matt Shotwell and Jonathan Schildcrout (PhD biostatisticians) about possible collaboration plan with anesthesiology
2014 July 9
Catherine Bulka, Anesthesiology
 General question: propofol dosing required for loss of consciousness has been shown to differ by race, but providers are often not considering race in choosing doses; wondering whether this will be improved by educational intervention for providers
 Concern is that studies which suggest difference between races may not be strong and/or generalizable
 Also, given VUMC patient population, unlikely that we'd be able to see any differences between races other than white vs. AfricanAmerican
 Alternate research question: assuming everyone starts at same loading dose (by weight), do different races require different maintenance doses? Recommend mixed effects approach to account for differences among providers.
Anji Wall, General Surgery/Bioethics
 Original email: "I am a general surgery resident, with a PhD in bioethics, and am planning to start a project assessing the common ethical issues discussed in MMI conferences. I have a paper survey tool and a coding guide, which I am planning to use for data collection... I would like assistance with determining sample size, format for data collection and the type of analysis to conduct. I do not have research funding or a formal research mentor but will attempt to get funding through VICTR if this is something that you all think would be warranted."
 Recommend collecting data in REDCap  no need for numeric vs. character coding, etc
 Main goal is to determine how best to educate surgeons on clinical ethics topics; main question: are ethical issues discussed equally often in morbidity vs mortality cases?
 For descriptive purposes, plan to start with 100 cases, then use that as pilot data or proceed with morbidity vs mortality comparisons from there
2014 June 18
Jennifer Morse and Emmanuel Okenye, Perioperative Clinical Research Institute
* "I am assisting one of our summer students with a research study. We are planning on attending the Wednesday clinic together to get some advice on
performing an analysis.
The investigator is trying to determine if there is a statistical difference between the time to successful intubation between 2 devices when used on
mannequins. There are two sites (Us and San Antonio). At each site, there were 5 anesthetists who each performed 6 trials with each device.
In my initial analysis, it was determined that the total time to success was not normally distributed. The two devices appear to result in different
lengths of time but I am unsure what test to use. MannU? Can that account for the multiple trials per individual? In addition, there was a large
difference between the two sites due to a different mannequin and different experience levels of the anesthetists (This demographic data was not
captured).
* Data is skewed, but not terribly, so suggest a linear model with sandwich estimation to adjust for withinsubject correlation (checked model diagnostics in clinic). Example code:
library(rms)
## Model without interaction, since interaction may be underpowered with 10 subjects
mod1.ols < ols(Total.time ~ Site + device, data = mydata, x = TRUE, y = TRUE)
## fit original model, without accounting for withinsubject correlation
mod1.robcov < robcov(mod1.ols, cluster = mydata$Subject)
## use HuberWhite sandwich estimation to account for withinsubject correlation
mod1.robcov ## get coefficients, pvalues
dd < datadist(mydata); options(datadist = 'dd') ## needed to get predicted values
plot(Predict(mod1.robcov))
## Plot predicted times for each site, device held at mode of other variable (eg, predicted times for Vanderbilt and SA held at most frequently tested device)
## Calculate predicted values for all combinations of site, device, save as data set to use in additional plots
pred.data < Predict(mod1.robcov, Site = c('Vanderbilt', 'San Antonio'), device = c('LMA', 'igel'))
## Repeat above for interaction model as sensitivity analysis, replacing "+" with "*" in ols() call
## Repeat also for time for first step in process (similar outcome distribution)
2014 June 11
Stuart Ross, Anesthesia
 "I'm in the beginning stages of a project with the anesthesia department and I wanted to get some ideas about how to best collect information. Later I'll be getting data from charts here at Vanderbilt, but the goal is to compare patients here with those elsewhere. The gist of it is comparing how patients from various contries/ regions differ from those here. What I want to do is take information from however many sources I find and organize it in a way that makes it easy to use/ search/ etc etc. Maybe a simple excel spreadsheet will do, but I wanted to make sure there wasn't an easy way of doing this that I might not be aware of."
2014 June 4
Tom O'Lynnger, Neurosurgery
 "I’m interested in attending the Wednesday biostats clinic to discuss a project I am conducting about outcomes in pediatric traumatic brain injury after ICU protocol implementation. The main analysis is an ordered logistic regression involving Glasgow Outcome Scale and a second ordered logistic regression involving discharge disposition. I’d also like to predict favorable discharge disposition using logistic regression. I have 129 total patients and have already done an analysis myself (I’m an MPH student in addition to being a resident in neurosurgery) that I believe is accurate but would like to confer with an expert. If possible it’d be great to go over the analysis during the session, though if not, I’d plan on getting VICTR support."
 We went over Tom's analysis and made some suggestions, including: describing continuous variables with medians and IQRs instead of means and SDs; doing Wilcoxon tests rather than ttests for descriptive statistics/table 1; removing sex and race from the model to avoid overfitting; combining the single patient discharged to acute care with the patients discharged to rehab; making sure that Stata is coding the outcome variable as expected; producing a boxplot of raw data for before/after and favorable/unfavorable discharge disposition vs GCS.
2014 May 21
Heidi Smith and Natalie Jacobowski, Psychiatry and Anesthesiology
 Study to describe pediatric delirium in ICU.
 Want to study relationship between diagnoses of delirium and physicians' use of certain descriptors.
 Intensivists' description of patients who were diagnosed as having delirium.
 They have developed a list of areas: agitation,
 One factor is whether delirium is mentioned in the problem list and in the plan. There is a daily physician note.
 Some medications could contribute to delirium. This could be reflected in nurse notes or medication record.
 They have three observation times for each patient. The day prior, the day of, and the day after the day delirium was diagnosed.
 Discussed need for comparison with patients who were not diagnosed with delirium for the inferences they are interested in.
 Could select controls based on matching on important patient factors.
 It is important to consider which day's observation to use for patients who didn't have a diagnosis of delirium. One option is to consider the
 You could also potentially consider including all observations for the patients.
 They are considering using VICTR for biostatistical support.
 We think Jennifer Thompson would be well suited for this project and should give the time estimate.
Ben Mackowiak, Neonatology
 Acidosis and pulmonary hypertension in neonates.
 Has experiments on piglets whose pulmonary vessels were exposed to acid in three doses until a certain pH is reached.
 They have a machine that reads the percent dilation
 Discussed problems with analysis on percent change. An alternate way to control for the initial size is to use a regression model controlling for the initial size.
 You can estimate the (absolute) mean difference between the initial and final data. You could do a paired ttest between the baseline and result after the first application of acid.
 A good approach would be a mixed effects regression model with a fixed effect for dose and a random intercept for subject (pig vessels).
 Should plot the trajectories and see how linear they are.
2014 May 7
Shreyas Joshi, Urology
 Appling for a VICTR grant and would appreciate assistance powering our study and determining the most appropriate data analysis plan for the study.
 Overall, 49 patients died.
 Looking to correlate preoperative sarcopenia with postoperative outcomes in patients undergoing surgery for Renal Cell Carcinoma (RCC).
 We are using a program that determines the skeletal muscle index on preoperative CT scans to obtain our "preoperative sarcopenia index" variable. Sarcopenia is lack of muscle mass. It is a newer measure of nutrition.
 It may be nice to have some or all of the scans rescored to see how reliable it is. You can look at the agreement in the scores. Or, if there is already a study published,
 Have 250 preop ct scans. All patients who get the surgery should have the scan. Whether they have the scans would maybe depend on the referral patterns.
 So far, we have overall and diseasespecific survival data, and we are working on gathering 30/90 day complication rates and hospitalfreedays.
 We hope to be able to power the study for survival (overall or diseasespecific) in order to move forward with data analysis.
 They are applying for biostatistics support, and we estimate that the project will require 50 hours of statistician work for this project and manuscript.
2014 April 30
Calvin Gruss, Anesthesiology
 studying the effects of acute hypoxia longitudinally in ~100 healthy subjects. Study has three phases, subjects had hypoxia first, then had carboxyhemoglobin/methemoglobin, at last they had hypoxia+elevated carboxy/methemoglobin. Each subject had two runs. The # of measurements for each subjects is from 2026. Interested in assessing the relationship between % carboxyhemoglobin and hemoglobin concentration in the blood.
 suggest using mixed effect model. Additional stat help can be get from Dr.Schildcrout or Dr.Shotwell or VICTR biostat support.
2014 April 2
Christy Goben, PICU
 Needs quote for VICTR biostat support; study in pediatric critical care with a focus on sedation trends in the PICU over the last decade with delirium impact
 No delirium screening prior to 20082009 (PCAM introduced); now done on children 5yo+
 Delirium reeducation done in 2011, so plan to compare three time periods (no screening, postscreening education, postreeducation)
 Expect overall use of sedation to decrease over time with possible exception of Precedex/dexmedetomidine
 Pulling data from StarPanel, ICU only (not after transfer to floor); suggest collecting in longitudinal format rather than summary if feasible
 Likely to have multiple ICU stays for some patients; will have identifier, so can control for withinpatient correlation
 Plan to describe delirium prevalence, use/dose of several sedatives and antipsychotics (primary goal), and secondarily, correlate/model these vs outcomes (ICU LOS, hospital LOS, time on vent, mortality)
 For outcomes, need to collect potential confounders as feasible (eg, use of pressors, sepsis, diagnosis/procedure codes, SOI, etc  whatever seems reasonable)
 Suggest very preliminary estimate of 120 hours, pending discussion on informatics/pulling data
2014 March 26
Luke Krispinsky, PICU
 Pediatric critical care fellow looking at endothelial dysfunction before and after cardiopulmonary bypass in infants (012m) undergoing repairs of congenital heart defects using iontopheresis and monitor that can quantify distal perfusion made by Perimed
 Since this is a new machine, suggest taking multiple measurements on same patient/same time to gauge reproducibility  may be restricted by cost ($13/probe)
 Mainly interested in a) describing change in endothelial function b) seeing how change is associated with outcomes (eg, ICU LOS)
 Exposures: difference between baseline and lowest (postsurgery) measurement, difference between postsurgery and 24h measurement, and AUC
 Outcomes: ICU (average 314 days, depending on type of surgery and other variables) and hospital LOS (varies widely), ianotrope score (need for BP meds), fluid requirements, vent LOS, mortality (<10%)  within defined study period, like 30 days?
 Next steps: define exposure(s) and outcomes(s) of primary interest, get estimates of variability on those outcomes from literature, think about potential confounders like severity of illness
2014 March 12
Christy Goben, PICU
 Needs quote for VICTR biostat support.
 Study in pediatric critical care with a focus on sedation trends in the PICU over the last decade with delirium impact
 RESCHEDULED
Sarah Scott, MD candidate, Director of Pharmacy Shade Tree Clinic
 Needs a quote for VICTR biostat support.
 Small study using a cohort of pediatric critical care patients. Her topic is the association of acute kidney injury and mortality in children on ECMO.
 RESCHEDULED
Angela MaxwellHorn, Developmental Medicine
 Briefly, I am doing a training on developmental screening for
pediatric residents when they rotate through my department
(developmental medicine). I am going to examine the wellchild checks
that they do in their continuity clinic both before and after the
training to see if their practice changes in how they screen and
refer. Another aspect is that I am also contacting their preceptors
in the clinic to see if they think the resident does a better job of screening after the training. Currently, there are 26 pediatric interns. I want to know how many
patient charts I need to look at before and after the training to make
any results significant. Additionally, we are thinking of breaking up
the next intern class into two groups...one group would get the in
person training by me, and the other group would watch a video
recording online. I am concerned, however, that this will
considerably lower the power of my study.
 Recommended: sampling all relevant charts for each intern from month prior to and month after training; maybe stratified chi square test for # charts with appropriate screening. First step: how many interns are currently doing appropriate screening?
 Angela will talk to preceptor and finalize outcome variable, get initial idea of how many people are doing screening correctly pretraining.
2014 March 5
Kendell Sowards, Instructor in Surgery
 Has requested feedback on a power calculation.
Eileen Duggan, Pediatric Surgery
 Questions regarding how to treat missing race and insurance data in large dataset; how to build best model for overall adverse event binary outcome (specifically how to adjust for hospital clustering, how to treat race and insurance variables (i.variable or different variables for each race/insurance status), and choosing the best model); and working with an interaction term in this model.
 We recommended adjusting for hospital using a random effect in her logistic regression model and that the best way to build her model was through prespecifying predictors to include based on literature and clinical knowledge.
 Her data has missing data at random so recommended that she impute data to avoid biased estimates.
 We also discussed how best to deal with and present terms with interactions  always report together, never just report a main effect.
 Finally, we discussed different ways of coding categorical variables such as race. She had seen in the literature that sometimes it is recorded as a multilevel single variable and other times it is recorded as separate indicator variables. Much depends on her question and discussed the different interpretations of the different ways of coding the variables.
2014 February 19
Aaron Benson, Urology, mentor/PI Nicole Miller
 multivariable logistic regression predicting sepsis
 an important independent variable (preoperative nephrostomy tube) is omitted(?) because there are no observations of the dependent variable (sepsis) in patients with the preoperative nephrostomy tube (n = 67) and 9 observations of sepsis in patients without a preoperative nephrostomy tube (n = 152). For this test, the results state that the preoperative nephrostomy tube "predicts failure perfectly". My questions are: how might I explain this to reviewers of our manuscript and is there another test that I should use?
 Having a nephrostomy tube shouldn't have impacted the length of times.
 Addressed whether there is a time after which everyone gets nephrostomy. The use has increased. This would be something to discuss in the discussion section.
Here are some followup questions I sent Dr. Benson, along with his answers:
 What is the purpose of the model? Is it to make predictions for patients based on their characteristics? Or to identify the important predictors of sepsis?
 The study is a retrospective analysis of our percutaneous nephrolithotomy (PCNL) experience. Most of these patients have access to the kidney obtained as part of the PCNL (i.e., no preexisting nephrostomy tube). Other patients have a preexisting nephrostomy tube placed ahead of time  either because of their history of recurrent UTI, pyelonephritis, high infection risk features, etc. or because they presented acutely and had the nephrostomy placed at that time. We are basically trying to determine whether patients with a nephrostomy tube prior to PCNL are less likely to develop postPCNL sepsis. We are not necessarily trying to identify predictors of postPCNL sepsis (already lots of data), but rather whether prePCNL nephrostomy tube (with renal urine culture and specific antibiotics) may be protective against postPCNL sepsis. After two manuscript reviews, our journal reviewers are recommending multivariate analysis to make sure that the differences in sepsis rates is not due to other factors.
 When you say the variable (nephrostomy tube) is omitted, do you mean that your group decided to exclude the variable (nephrostomy tube) from the model? If so, is that because of the result you are getting?
 No, I mean that STATA itself is showing the word "omitted" in the row for PCN (nephrostomy tube) where the data should be. We are not omitting the data on PCN because that's the focus of the study.
 I'm unsure of what you mean by "For this test, the results state that the preoperative nephrostomy tube "predicts failure perfectly"." Which test? Is it part of the automatic regression output? And where is the "predicts failure perfectly" coming from? The output in stata?
 By "this test", I mean logistic regression. The phrase PCN "predicts failure perfectly" is from the STATA output just above/below the results table it produces.
 When you say "no observations" of sepsis in patients with the tube, you mean all of the patients with the tube were known to not have sepsis, right? You don't mean whether they had sepsis is unknown/missing?
 Correct, in the group of patients who had a nephrostomy tube prior to PCNL (n = 67), there were no sepsis cases. In the group the did not have a nephrostomy tube prior to PCNL (n=152), there were 9 sepsis cases. There is not any unknown/missing data for whether patients developed postPCNL sepsis.
 How many other variables did you have in the model, and how many were considered?
 Unfortunately, I left my jumpdrive at home today or I would have already sent you the STATA file. But, off the top of my head, there are probably 1012 other variables.
 We recommended using exact logistic regression. In stata you would use exlogistic.
 Here is a web page explaining this issue and why you need exact logistic regression: http://www.ats.ucla.edu/stat/stata/dae/exlogit.htm. You should be able to try this (you don't need the [fw=] option) in stata and use this to explain your analysis in the methods.
 We also discussed that the model would be really overfit using 10 variables with only 9 events. We think one variable would be appropriate, but it would also be okay to use 2 variables, maybe nephr. tube and operating time.
Here is some R code:
counts < matrix(c(143, 67, 9, 0),
nrow = 2,
byrow = TRUE,
dimnames = list(c("No sepsis", "Sepsis"),
c("No tube", "Tube")))
prop.test(counts)
binconf(x = 0, n = 67, method = "all")
2014 February 12
Imani Brown, IGH, MPH candidate
 Looking at preliminary data from an intervention in HIV+ people in Mozambique. She is interested in assessing what factors are associated with receipt of 9 different messages included in the intervention.
 Each message is defined as having been 'received' or 'not received' so we suggested logistic regression. We also recommended ranking the predictors of interest so that depending on how many events she has, she can fit a model with the proper number of parameters based on the 10:1 or 20:1 ratio.
 For those messages with very few events, we recommended descriptive tables and graphs as opposed to tests of association or models.
2014 January 22
Stephen Humble, Mayur Patel and Patrick Norris, Trauma
 Planning to perform noninferiority power calculations applied to paired observations of heart rate variability.
 Want to describe heart rate variables to hopefully determine a norm for ICU population
 Suggest spaghetti plots to describe data for now
Heather Kistka, Neurological Surgery
 Planning to submit VICTR voucher
 Compared VUMC residency applications in 2007 (N = 148) vs 2012 (N = 191) to determine whether "misrepresentation" has increased among applicants based on Pubmed searches for publications
 Problems: 1) application changed in that window; 2) various kinds of misrepresentation (existence, author order, peer reviewed vs not); 3) if misrepresentation had multiple types (eg, not peer reviewed and changed author order), only "worst" was recorded
 Analyses planned: lots of descriptives by year (number/types of misrepresentation, demographics, etc), plus logistic regression model among 2012 applicants looking at risk factors misrepresentation vs no misrepresentation (ie, "red flags" indicating that application should be closely investigated)
 Logistic model: 89 df (AOA membership, grad degree, board scores [nonlinear?], gender, top 20/nontop 20/foreign med school, # works on CV), works with 84 events in 2012
 Suggest secondary analysis using a "scale of misrepresentation" as outcome (flagrant vs something getting put in the wrong section) in proportional odds model  lots of ways to look at this (done only on applicants with misrepresentation, looking at "badness" of misrepresentation, or look at worst misrepresentation per applicant...)
 For above, possibly create weighted score, along lines of DNE*3 + FAO*2 + (MAO + NPR + OPO)*1, for outcome per applicant; if distribution is wacky (probably will be), perhaps take out applicants who didn't misrepresent anything and reduce number of covariates to account for lower N
 Clinic estimate is 40 hours without secondary analysis, 60 hours with secondary analysis
2014 January 8
Catherine Bulka, Anesthesiology
 Working on a project analyzing geographic variation in hospital billing practices. I have a dataset of hospitals and what they charge for certain orthopedic surgeries. I have aggregated the hospitals to the corebased statistical area level (these are geographic areas designated by the Office of Management and Budget that are based around an urban center of at least 10,000 people and any adjacent areas that are socioeconomically tied to the urban center by commuting). Rather than look at differences in hospital billing practices nationwide since there are so many confounders, I decided to aggregate the data and look at the amount of variation in billing practices within each corebased statistical area because I’m assuming that the socioeconomics/cost of living/overall health of the patient population/any other potential confounders are likely pretty similar within these areas.
 I’ve calculated the means and standard deviations in what the hospitals in each area bill for the same procedures, but I’m not sure what the best way is to compare these. The data are not normally distributed, so I’m not sure that standard deviation is even the best way to represent variations in the amount billed. Further, some areas have many more hospitals than others – from 2 in one area to 105 in another, which I think should be taken into account. I thought about using ANOVA, but I’m not so much interested in the mean amount billed by the hospitals in each area, since certain areas of the country (California, NYC, Florida) are known to charge more for certain procedures than other areas for economic reasons.
 How can I best compare the amount of dispersion between many groups (there are > 500 areas that I’d like to compare), while addressing differences in sample size? * Have calculated the coefficient of variation (standard deviation/mean * 100) for each core based statistical area, although I am not sure if that's the best metric to show variation. Also not sure how to compare these areas with hypothesis testing. * After discussing her project, we suggested she explore a linear mixed effect model as well as further explore some of the geographic representations she had started with somehow including some aspect of the variation of charges by region in addition to reporting mean charges by hospital within a region.
Michael DeLisi, Biomedical Engineering
* Michael has a project comparing how his intervention to imageguided surgery for minimally invasive eye surgery helps in time to reaching the desired target and the ability to hit the desired target.
* The study uses 4 skulls with different targets per eye. In each skull, one eye is operated on using standard imageguided methods; the other eye uses the enhancement to the standard methods. Sixteen surgeons were tested, each operating on each of the skulls. The order of skulls for each surgeon was the same but the order of methods of surgery was randomized.
* Currently, he has tested for differences using ttests and Ftests.
* Our recommendation was to use linear and logistic mixed effects models to account for the correlation among surgeon, including method of surgery, skull, and eye (?) as covariates in the model with surgeon as the random effect.
Older Notes
I am looking for guidance on how to proceed with performing a validation of the bedside swallow screening used for acute stroke patients in the ED, neuro ICU, and the neuro care unit.
The project is development of risk prediction models for placement of a ventricular assist device vs medical management with outcomes of survival to transplant and 1 year post transplant survival in pediatric patients. Considering using propensity matching due to variability within groups.
I am planning a clinical study and would like my sample size calculations to be reviewed by a biostatistician before I submit for VICTR funding.
The study is a randomized, doubleblind, experiment in human volunteers examining the effects of a drug commonly given to liver failure patients on oral glucose tolerance.
Pooja Santapuram