Biostatistics applications in surgery, anesthesiology, and emergency and critical care medicine Clinic Notes (2015)
2015 December 16
Jason Singer, Medical student
 "I have a few quick questions for biostats clinic related to how to treat data from Likert scales. For example, if the average choice is somewhere between agree and strongly agree or if the average choice is between
usually and always, how do I report a mean and standard deviation?"
 Due to limitations with REDCap on tablets, their survey had to be restructured from using VAS's to categorized answers. Because of this, the type of analysis required to answer their questions of interest has changed.
 Doing a chisquare test will determine if there is a difference in distribution of responses between inpatient and outpatient groups. If there is a significant difference, then differences in proportions could be done in which a specific cutoff is determined.
 Another alternative is to fit a proportional odds model with the responses as the outcome and group (inpatient versus outpatient) as the main covariate. The advantage to this approach is that confounders can be included in the model. The proportional odds assumption needs to be tested once the model is fit.
2015 December 9
Mike LeCompte, Surgery and Critical Care
 "I am trying to design a study on surgical resident education techniques and wanted to get some input on my study design and setting up my statistical methods."
 We would recommend doing the prevideo at the first junior surgical experience to then compare to the postvideo after all surgeries are complete. We would also recommend having the same number of assistant surgeon experiences between the two groups if the prevideo is made at the time we recommend.
 We recommended that he determine what would be clinically meaningful differences in the different outcome measures to help determine what power he would have given his fixed sample size.
 We think it will require about 40 hours of statistical support to do the analysis and help in manuscript preparation.
Tommy An, Medical Student
 Question about Stata output for logistic regression
2015 November 25
Renee Hill, Physical Medicine & Rehab
 Postdoc in Center for Integrative Medicine; interested in seeing whether an 8week intervention results in reduced emotional distress, and whether this reduction is different depending on level of selfcompassion. No control group; all patients received the intervention, and have pre and postintervention distress scores.
 For main question (reduction in scores), could do a paired ttest or (more likely) a Wilcoxon signed rank test, if data is not normally distributed, comparing pre and postintervention scores.
 To see whether the association of pre and post scores differs based on level of selfcompassion (continuous value), could do a multivariable regression model: postintervention score = preintervention score + selfcompassion score + (prescore * selfcompassion). Can also add potential confounders (age, gender, etc) to this model. Linear regression is most common, but need to check distribution of the outcome first  if the outcome is not normally distributed, this model may not be appropriate or reliable. Also check model assumptions, such as residual vs. fitted plots.
 Secondary question: do postintervention scores stay stable in the months after the intervention, or rebound? Data available on cohorts from 2012 through 2015, all with different followup times. Could do a regression model with followup (current) score as the outcome, months since completion of intervention as the main exposure, and adjusting for other confounders (anger score, age, etc).
 Suggested looking into a VICTR voucher (check Starbrite for more info).
2015 November 18
Michael Ghiam, medical student  canceled
 "Hello, I am a third year medical student working on a epidemiology study using Stata to analyze my data. I am currently working on writing code for my study and I am running into some roadblocks. I was wondering if I could come in on Wednesday or Thursday and get some pointers about which codes would be best to use and if I’m analyzing my data the right way. Please let me know if this is at all possible and what I should provide you with in advance."
Tommy An, medical student
 "I have attempted a logistic regression in STATA to predict whether a patient has MRSA or MSSA musculoskeletal infection based on presentation data from the emergency department. I need some advice to see if I'm on the right track with my statistical analysis."
 Logistic regression seems fine. Make sure to note that results will only be generalizable to patients who actually did develop MRSA or MSSA, not patients who developed neither.
Jamie Robinson, Surgery
 " I am trying to do a Kmeans analysis and having some difficulty with figuring out if what I am getting is what it should look like. "
 Suggest looking at varclus() function in Hmisc package to cluster variables instead of patients
 Also could create principal components (using princomp()) before clustering to reduce data to two dimensions before using kmeans
2015 November 11
Nick Kramer
 "...here is our project and question: literature review of weight bearing after posterior acetabular fracture. The data available is relatively limited looking at the question we are asking so we are attempting to combine several studies to see if there is a trend for benefits of early vs late weight bearing. We have several questions regarding the best way to do this, if it is possible at all."
 Clinic statisticians will email the department to ask if anyone has further expertise in this area. This study presents a challenge because there is no common exposure in each study included in the review; rather, each study is a cohort of either early or late walking patients. Raw data/SDs are also available on very few of the studies.
Michael Benvenuti, Orthopedic surgery
 "I am working on a pilot study to determine the effect of antibiotics on length of stay and culture sensitivity in pediatric musculoskeletal infection. i have done some preliminary analysis using Stata and have a few questions about how to continue."
 We discussed logistic regression for the main outcome: positive culture = covariates. The number of covariates that can be reliably put in the model is roughly equal to [minimum of positive/negative cultures] / 1015. So, if the split is 50/50 and there are 120 patients, 60 "events" / 10 = 6 possible covariates in the model.
 For length of stay, we suggested using a Cox model looking at time to hospital discharge, rather than using LOS as a continuous outcome. (LOS generally has a distribution that is difficult to model.) All covariates for this model are measured at ED presentation.
 About 30 patients are included in the cohort but have no culture measurement. It is important to look at whether these 30 patients are different from the other 120  was no culture done because they were sicker/less sick, younger, etc.
2015 November 4
Courtney Baker
 "In terms of the project, it is a collection of data on intraoperative coagulation factors and transfusion data for pediatric scoliosis patients over the last 3 years. I have asked a number of questions around "what determines/predicts intraoperative blood loss in these cases?" I have done rudimentary (and most likely not entirely correct) statistical analysis between the coagulation factors and the transfusion results. What is needed is a rigorous discussion about employing multivariant statistical analysis on the associations I see. The goal of this data set is to publish some new observations and associations in order to develop a quality improvement protocol AND a more rigorous research project into one or more of these specific associations."
 Recommend doing multivariable regression models instead of univariate analyses: eg, "end of surgery platelets = baseline platelets + other baseline variables". Specific regression type depends on distribution of outcome; linear regression is appropriate for truly continuous outcomes with wide enough range (eg, 0100) as long as assumptions are met (eg, residuals are normally distributed). Logistic regression is appropriate for dichotomous outcomes, and something like proportional odds logistic regression would be appropriate for integer values with small ranges (eg, 03).
 Suggested spaghetti plots to describe fibrinogen loss during surgery, one line per patient, with reference line at time of fibrinogen intervention.
 Models could be fit using restricted cubic splines for continuous covariates, such as baseline fibrinogen.
 Discussed multiple comparisons, which GraphPad uses by default.
 Use nonparametric tests of association (such as Wilcoxon rank sum test/MannWhitney) unless it is known that variable is normally distributed (and even then, nonparametric is a safe choice).
 Possibly talk with Shirley Liu about doing regression analyses due to ortho collaboration; otherwise apply for 90hour VICTR voucher.
2015 October 28
Emily Buttigieg, Medical student
 "I am a medical student (VMS III) working on a research project in the Pediatric Surgery department. My project involves measuring body composition using tissue resistance and reactance measurements and comparing it to standard measurements, BMI, weight and height. I have collected my data and am inquiring as to the best software and approach to analyzing my data. I was hoping to attend next Wednesday’s clinic. Thanks in advance for your help. "
 About 30 patients total with repeated measurements. Suggested using Spearman correlations on Zscores from device and BMI calculations, using a) only one measurement per patient and b) all measurements per patient. BlandAltman plots might also be helpful, and scatterplots of raw data will be very useful.
2015 October 21
Mary Bayham, Global Health

We seek to describe the burden of fever, diarrhea, and respiratory illness among children aged 659 months in Zambézia Province, Mozambique as well as predictors (individual and system level) of health care utilization for these children. The goal of this thesis is to identify significant predictors (individual and system level) of healthcare utilization for children under five with fever, diarrhea, and cough. These findings could inform future planning, policy and interventions in reducing under five morbidity and mortality in Mozambique.
 Dataset includes 3,800 families; 2,700 children < 5yearsold; and 14 districts (3 of which were oversampled).
 Suggest reporting descriptive statistics for 3 individual districts and comparing to all other districts combined.
 Suggest utilizing a multivariable logistic regression model for healthcare utilization  initially without any weighting and once more weighting for district  and comparing results.
Erin Hamilton, Global Health
 Goal to assess impact of a nutrition education intervention in children.
 Outcome is Zscore BMI adjusted for age measured at 4 time points. Dataset includes 151 children with Zscore BMI measured at least twice.
 Planning to use multilevel mixed effects linear regression model adjusted for gender.
 Suggest utilizing paired Wilcoxon test for change in Zscore BMI between time points (pre vs. post) and boxandwhisker plot of median and IQR at each time point.
2015 October 14
Rachel Hayes, Surgical Sciences
 "I am struggling to interpret an interrupted time series using logistic regression and proc glm in R. I’ve attached a deidentified data set (date is shifted) and some R code."
 Difference in interpretation between lrm() and glm() models is due to differences in default anova() tests  anova.rms() uses added last tests, while anova.default() uses sequential tests (eg, time after adjusting for only variables ahead of it in model formula).
2015 October 7
Thomas An, Medical student
 "I am in Dr. Schoenecker’s lab studying musculoskeletal infection. I have outcomes data and numerous variables for patients with musculoskeletal infection and am hoping to set up a multivariable analysis to predict which variables are most predictive of outcomes."
 We discussed linear regression and diagnostics to use to evaluate whether assumptions for linear regression have been met (residual/predicted plots). We also discussed potential alternatives if the assumptions for linear regression are met including transformations, ordinal regression, or negative binomial regression.
Teerayut Tangpaitoon, Department of Urology
 "Our study is about evaluate outcome of Holmium laser enucleation prostate surgery (HoLEP) compare to HoLEP with concurrent Cystolitholapaxy in same setting(retrospective)."
 The two groups are completely confounded by presence of bladder stones. Those who received the additional therapy were patients with bladder stones.
 We discussed ttests versus Wilcoxon Rank Sum tests. We also discussed using linear regression to adjust for potential confounders.
 We recommended the UCLA website for help in how to perform the analyses. ( http://www.ats.ucla.edu/stat/spss/ )
2015 September 30
Kelly Maguigan, Critical Care Pharmacy
 Project involving enteral intolerance in spinal injury patients who are on concurrent vasopressors.
 Planning to request VICTR voucher for analysis.
 Retrospective chart review of 80100 patients from TRACS registry who were on pressors + enteral feeding for at least one hour.
 Primary analysis: risk factors either count of enteral intolerance during hospital stay (Poisson or negative binomial model?), and/or daily yes vs. no enteral intolerance outcome, using lagged covariates.
 Strongly suggest using REDCap for data collection  easier to build databases, especially with assistance from REDCap clinics, and built to be easy for statisticians to export data on the back end.
 Secondary outcomes (ICU/hospital LOS, etc) are mostly descriptive.
 For VICTR voucher, we believe this project will fit within the voucher time frame (90100 hours).
Mike Benvenuti, medical student
 I have been working with some retrospective data and am not sure how best to represent the effects of antibiotics on culture yield (odds ratio, negative/positive predictive value…) and would like some quick input. I also have another data set and I would like to show that patients have a ddimer above the clinical threshold following joint replacement and am again not sure how to best show that.
 ~200 patients. Primary relationship of interest: time of first antibiotic use vs. hospital LOS and/or possible secondary outcome: extension of antibiotic prescription posthospital stay
 Issues: immortal time bias  patients who get antibiotics later are, by necessity, in the hospital longer. Could use timevarying Cox model to address this? Looking at secondary, posthospital outcome would help with this issue but could be a more blunt measure, and may not be that helpful depending on how many patients had antibiotic prescriptions extended.
 Analysis will be complicated either way  suggest talking to MarioDavidson and/or research immersion course instructors about how to get stats support.
 Definitely need to adjust for confounders (severity of infection, etc).
2015 September 23
Don Arnold, Pediatric Emergency Medicine
 Dr. Arnold is applying for VICTR biostatistics support for writing the analysis plan and sample size justification for his grant proposal. I estimate that this will take approximately 20 hours of support.
 This is a clusterrandomized trial and will have correlation due to providers on multiple patients, so we recommend an analysis that will account for that, such as GEE, mixed effects models.
 "We propose a 2arm cluster (by clinician) RCT of the Asthma Prediction Rule (APR) in 3 children's hospitals to determine if implementation of the APR electronically will result in a decrease of the "unnecessary hospitalization rate" for children with acute asthma exacerbations from 23% to about 19%. My specific question is around the power calculations I'm doing. For an effect size this small I'm calculating that I need a couple thousand clinicians in each arm, whereas we will likely have at most 60 in each arm. I'm looking to alternative designs or methods to analyze the data."
Richard Lesperance, Surgery
 Want to use an ICC to measure interobserver reliability. We recommend giving a confidence interval along with the point estimate.
 A useful measure for one of the main analyses could be the mean difference (with confidence interval) in length between the anterior and lateral chest wall.
 "1) For measurements of chest wall thickness among trauma patients, should the mean measurements have standard deviation or 95% CI calculated for them? 2) We have 4 observers performing measurements on a total of 450 CT scans. How do we calculate / report interobserver reliability? (Cohens’s kappa?) Should all 4 observers rate the same 5 or 10 scans to obtain the measure?"
 "Detailed background: We are reviewing about 450 CT scans of patients who had a prehospital intervention: needle decompression of a tension pneumothorax. This is a traumatic condition where pressure builds up inside the chest (but outside of the lungs), and can kill a patient unless the pressure is drained. Commonly, paramedics are taught to insert a needle through the chest wall to drain the pressure – but there are concerns that the needles may not be long enough. Previously published literature suggests that this procedure may be ineffective in many people due to chest wall thickness – there are several published studies using cadavers, and CT scans of healthy volunteers, that show quite often the chest wall is thicker than the needle size commonly used – if the needle can’t reach the pleural space, the excess pressure cannot be drained.
We are looking at those 450 CT scans and measuring chest wall thickness in 2 locations – the front (where the procedure is traditionally taught, just below the clavicle) and also the side (which is an alternate location that some EMS
providers are taught). We are measuring both sides (4 measurements total per CT scan).
Additionally, we are trying to measure the distance from the chest wall to the nearest cardiac structure (if this is too close, maybe the needle can injure something if we just use a longer needle!)
Besides the measurements (which other people have done, just not on actual trauma patients), we also theorize that many of these procedures are being performed unnecessarily. This condition is difficult to study because it is
difficult to diagnose prehospital, and can be lethal if not effectively treated. The only marker that a patient both needed the needle decompression, and that is was successful, is the subjective EMS report that there was a “whoosh”
of air when inserting the needle and the patient “got better”.
Further, we have noticed many patients who received a needle prehospital, but when they get to the CT scanner, they have no pneumothorax (air in the pleural space at all) – which means by definition, not only did they not need
the procedure at all, but the procedure was completely ineffective. So we will be left with a % of patients who we can prove never had a pneumothorax and even though they had needle decompression attempted, it definitely failed.
We realize that this number is the
minimum; in effect these are the patients we can PROVE had both inappropriate and ineffective treatment. The true number is much higher but unknowable.
Statistical questions:
1) Some previous literature reports the mean thickness and standard deviation – others report the mean and a 95% Confidence Interval. Is there a benefit to one way or another?
2) We have 4 people reviewing CT scans. What test should be used to measure interobserver reliability? Should all 4 observers rate the same scans to obtain the measure? How many samples – 5, 10, more?"
2015 September 16
Jamie Robinson, Biomedical Informatics NLM Fellow
 "My topic is Surgical Resection for CPAMs. In particular, I believe I need help with regression analysis."
 We discussed specifics of R code, how to fit splines to a continuous variable and how to report the resulting estimates.
Thomas An, Medical student
 "I am trying to compare hospital outcomes for MRSA vs. MSSA pediatric infection. I have done some analysis in GraphPad that I am not sure is correct."
 We discussed nonparametric tests to use (Wilcoxon Rank Sum test, KruskalWallis) as opposed to parametric tests. We also discussed whether a regression that adjusts for potential confounders is appropriate.
2015 September 9
Isa WismannHorther and Katie Ryan, OB/GYN
 The project is a retrospective case control study looking at healthcare system based factors (postpartum counseling on birth control etc.) and their affect on short interval pregnancies. We wanted to meet with you all before we finished our IRB to make sure our methods were using the best possible statistical model of analysis. We were thinking about creating a scale to make a composite score which would put a numerical value to the descriptive data we're looking at (if birth control counseling was documented in the chart before discharge, did they pick a preferred method etc.) We were curious how many charts we would need to look through to power the study.
 We recommended that the group write out their specific aims in order to keep them as focused as possible.
 We also recommended that they check out whether a VICTR design studio would be appropriate for more intense statistical support. We also recommended that they check if there is a collaboration between our department and OBGYN.
 We discussed what would be an appropriate outcome. We advised against using a scoring system. Rather, we suggested they consider either looking at the elements of counseling and treat it as an ordinal outcome (number of elements covered in counseling) or as a binary outcome (any counseling versus none).
2015 August 26
Carissa Cascio, Psychiatry
 "If possible, we'd like to review a recent VICTR submission and the prereview feedback we received regarding our power analyses, assessment of normality, and imputation."
 Addressed three critiques from VICTR application:
 Power calculations: initially used difference seen in pilot data; suggest adding power analyses using the smallest effect size that would be clinically meaningful.
 Assessments of normality (for ttests): Suggest using nonparametric versions (Wilcoxon) instead of assessing normality. Loss of power is minimal if assumptions are met, and benefit is large if assumptions are not met.
 Multiple imputation: Attrition is expected to be a major issue, but not sure if data will be missing at random (an assumption of multiple imputation). Suggested doing complete case analyses as primary, and doing multiply imputed analyses as a secondary analysis. Specify beforehand what variables will be used for imputation, and describe the differences between patients with and without missing data.
 Possibly helpful paper for types of missing data for missing imputation: http://www.sciencedirect.com/science/article/pii/S0895435606001971
2015 August 19
Maya Yiadom, Emergency Medicine
 Looking at differences in screening for EKG vs. stemi incidence for multiple emergency departments for pilot data for a grant submission. Each center has its own screening criteria, and rates of stemis vary across centers.
 Suggested simple descriptive analysis: table with one row per center describing incidence, sensitivity, specificity, NPV and PPV with confidence intervals. Think about describing demographics as well.
Jamie Robinson, General Surgery
 Wants to look at the difference in costs using PHIS on performing appendectomy + 30 days post op before and after the intervention of a clinical practice guideline.
 Didn't come to clinic.
Brian Long, Surgery
 Looking at risk of cancer recurrence in pediatric neuroblastoma patients and whether surgery is helpful in highrisk patients. At the very beginning of the process (database design, etc for a retrospective multicenter study), but looking ahead to requesting a VICTR voucher for statistical analysis. Analysis will likely involve multiple survival models and competing risks, so we recommend the 90hour VICTR voucher level to accommodate all analyses and data management. Also suggested attending the REDCap clinic for help with database design, and gave some data collection suggestions.
2015 August 12
Brian Long, Surgery
 Planning a pilot retrospective chart review
 Brought a proposed variable list for data collection and proposed outcome measures for a retrospective study looking at outcomes after an operation.
 Research questions of interest are: Which patients with highrisk disease will benefit from surgical resection of the primary tumor? Does complete resection of the primary tumor benefit patients with highrisk disease? Do patients in whom a complete resection is unlikely to be accomplished benefit from partial resection of their tumor?
 Discussed data needed for survival analysis: cause of death, last date of follow up, etc.
 Discussed types of survival events
 We discussed applying for VICTR biostatistics support, but we did not estimate the time required yet.
2015 Aug 5
Nick Carter, Surgery resident
 The sample size estimation is completed using the Kappa statistic. With a total measurements of 48, it provides at least 80% power to detect a Kappa = 0.8 with a twosided type I error = 5%.
2015 July 29
Nick Carter, Surgery resident
 I am working on a study trying to show noninferiority of postoperative care provided by community health workers compared to standard postop care with the operating surgeon. We are just getting started in planning the study, and I hoped to discuss study design and power with a statistician."
 Study will take place in Haiti. Prevalence of wound infection is very low (~5%), so getting enough patients will be difficult. Discussed a validity study, showing sensitivity/specificity/PPV/NPV with confidence intervals, using surgeon inperson assessment as gold standard and both a) surgeon via cell phone picture and b) community health worker assessment as new diagnostic tools. Due to low prevalence, will need a lot of patients to get a "reasonable" (according to clinical judgment) confidence interval for each quantity.
 Discussed adding "dirty" infection sites as well to raise prevalence to an estimated 1020%  something to think about.
 Link to PS sample size software (Windows)
 Link to VICTR studio page
2015 July 22
Tracy Marien, Endourology and laparoscopic surgery
 "I am from the Urologic Surgery department. Basically, my primary question is as follows: I was performing a multiregression analysis in Stata to assess which factors are associated with passage of ureteral stones. However, it appears that while two factors are significant they also cancel each other out because they are so strongly associated. Is there a way to control for this?"
 She has ~ 100 patients with kidney stones, some of whom pass the stone without intervention and some that need surgical intervention. Standard of care is to measure the size of the stone in the axial direction with CT scan. Coronal measurements are also made but are not referenced typically. Her hypothesis was to investigate whether both measurements would help predict who requires surgical intervention.
 We discussed looking at the scatter plot of the axial and coronal measurements and found there to be a strong linear relationship. We advised against fitting both variables in the model of interest. We also advised to prespecify the model based on literature and clinical knowledge rather than univariate tests. We also discussed the 10:1 or 20:1 ratio for determining the complexity of the model based on the minimum of events/nonevents.
 Because of the strong correlation observed in the scatter plot, we discussed including an aspect ratio variable in the model with either axial or coronal measurement to assess whether this would be a question of interest.
2015 July 15
Jonathan Siktberg and Mayur Patel, TBI
 Project on diffuse axonal injury; sent PPT slides
 Team has a good list of aims, models and covariates: POLR model for GOSE score, linear regression for quality of life score  these seem appropriate provided assumptions are met with data. Possibly a Cox proportional hazards model for mortality, if this outcome is of clinical interest.
 Discussed multiple imputation, since many patients could not be reached for followup (approximately 35%). Recommended doing both multiply imputed and complete case analyses to compare results; multiply imputed analyses may be less biased. Clinical data other than model covariates can be incorporated into imputation.
 Currently patients are classified as DAI negative or positive, with positive having three possible grades. We recommended doing models with a fourlevel variable for DAI (0/1/2/3); to address the more typical clinical question of any vs. no shear (DAI positive vs. negative), could redo the model dichotomizing into positive vs. negative.
 The group plans to apply for a VICTR voucher. Given the possible complexities of multiple models and multiple imputation, we estimate 90 hours.
Jamie Kuck, Division of Allergy, Pulmonary, and Critical Care Medicine
 "Sepsis patients have high levels of cellfree hemoglobin in their plasma, and these levels are associated with increased risk of mortality. While exploring the possible mechanism of cellfree hemoglobin, I measured levels of oxidized LDL in sepsis patient plasma and found that those patients with high cellfree hemoglobin have low levels of oxLDL, which was a surprise. An endocrinologist suggested that we then look at LDL levels since sepsis patients usually have low amounts, which could explain the low amounts of oxLDL."
 We suggested a linear regression model like this: oxidized LDL = hemoglobin + total LDL. Prism doesn't seem to be capable of this, so get SPSS and look at examples on UCLA stats web site for instructions.
2015 July 8
Nick Kramer, M3 Meharry Medical College
 He would like continued input on his project.
 We discussed how clinical judgment should guide the selection of manuscripts to include in their systematic review. We also discussed how to organize their analysis from the big picture of outcomes by type of fracture (simple versus compound) or type of repair (screws versus plates).
Christopher Brown, Dept Internal Medicine
 The project is fairly simple, we measure labs once a day or twice a day to follow potassium. I would like the primary end point to be regarding this lab, meaning if you measure the potassium twice a day, does the potassium stay in the normal range for more time during the patients hospitalization than if you measure it only once a day. However because I am measuring the value more often in one group than the other I am wondering what the method for accounting for this statistically would be, as there appears to be a sampling bias between the groups. It was suggested to me that this could be accounted for with a "generalized least squares approach" however I do not completely understand how that regression adjustment would help me. In any case, can you tell me if it is possible to compare the time (persondays or personhours) a variable spends between two values (IE potassium between 3.5 and 5.0) when the two groups involved are sampling the variable with different frequencies? (IE I can detect twice as much low or high values in theory twice a day than once a day, so how do I compare these groups)
 We discussed how to approach analysis of the prospective data  create an indicator for whether the subsequent days lab was within normal range or not treating the 1x/day or 2x/day as a treatment group variable (standard of care versus intervention) and how this analysis would require a repeated measures analysis such as GEE. As secondary descriptives, we discussed calculating the proportion of measurements outside of normal range in each of the groups.
 We estimate that this analysis will require 40 hours of a VICTR statisitican's time.
 We also discussed the retrospective data abstracted from the medical record and methods of analyzing it. This also will require some type of repeated measures analysis as well as decisions on how to handle repeat hospitalizations per subject. We estimate that the retrospective analysis would require 60 hours of a VICTR statistician's time.
2015 July 1
Kendra Parekh, Assistant Professor Department of Emergency Medicine
 "I would like to attend a biostat clinic on July 1 to discuss data analysis for a survey that evaluated emergency medical technicians’, nurses’, and physicians’ attitudes toward a new Emergency Medical Services system in Georgetown, Guyana."
 She has responses from 17 EMTs for one survey and about half MDs and half RNs who filled out the second survey. There were about 10 questions between the surveys that were the same. We recommended using the chisquare test with continuity correction to assess whether there was an association between provider type and responses to the question.
 We recommended she use graphs to display the data as well as tables with proportions rather than simply relying on pvalues from tests of association.
Alexander Gelbard, Otolaryngology

I am looking at disease of unexplained scarring in the trachea. We looked at the biology of the fibrotic tissue response, and then investigated the association with defined respiratory bacteria. Finally we investigated activation of immunologic pathways in the tracheal tissue samples.
Expt 1. qPCR results (10 experimental, vs 3 controls) *the controls are 23 pooled donors preformed in triplicate.
Expt 2. PCR results (binary yes/no presence of detectable bacterial) in 10 experimental vs 10 controls.
Expt 3. In isitu hybridiation (binary assessments of staining positive/negative) with 10 experimental vs 10 controls, and 10 normals
Expt 4. Comparision of % positive cells in Transmission Electron microscopy immunogold staining.
Expt 5. Elispot comparison. IFNgamma release in response to antigen specific stimulation. 10 experimental, 10 controls.
Expt 6. Comparison of immunohistochemistry quantification. 10 experimental vs 10 controls and 10 normals.
Expt 7. qPCR results (10 experimental, vs 3 controls)  *the controls are 23 pooled donors preformed in triplicate.
 We recommend that he contact the Friday clinic for a question involving the replication of pooled data for the healthy normal group. Otherwise, we recommended nonparametric tests and chisquare with continuity correction, as appropriate.
Daniel Heath Hagaman, Anesthesiology
 We advised that he get a copy of SPSS rather than using Excel.
 We suggest chisquare for difference in proportions of forms filled out in VPEC before and after intervention. Include a few months of wash out period after to get a more stable estimate. To look at factors to predict forms being filled out among nonVPEC providers, use logistic regression and include covariates based on clinical knowledge. Ideally, a random effect for provider should be included.
2015 June 24
Laura Wilson, Hearing & Speech Pathology
 "I am seeking expertise regarding the data analysis plan for a study on school outcomes after sportsrelated concussion."
 Survey study of ~120 patients (age 1317) who were seen for concussions during the last school year, following up on academic performance, special accommodations (extra test time, sitting out gym), and satisfaction with return to school. Hypothesis is that school absences and accommodations are a) correlated and b) both affect academic performance and satisfaction, which are measured on both parents and children.
 For future studies, mediation or PATH analyses might be appropriate, but these models will need to be simple due to data and sample sizes. Suggest proportional odds or logistic regression depending on distribution of final outcomes (have multiple levels, but if some levels are not well represented, may make sense to combine). Base effective degrees of freedom on final outcome levels Ns  for logistic, the minimum of the two outcome categories. Do not use univariate analyses to prioritize covariates; rather, develop a priority ranking based on clinical knowledge and importance in hypotheses.
2015 June 17
Trisha Pasricha, 4th year medical student
 "I am a 4th year medical student who has completed a study at the Vanderbilt Center for Surgical Weight Loss that looks at the correlation between depressive symptoms and BMI/medical comorbidities in patients who have sought surgical and medical weight loss at the center. I would like some help determining if we have used the correct analysis of our data."
 Info about the study: Complete data on 38 patients, but only 8 medically treated patients  lots of loss to followup in this group. Therefore, any inferences about treatment will need to be approached from the standpoint of very preliminary research. Due to limited sample size and complex research questions, we can't really put all exposures of interest in the same model.
 Possible models:
 posttreatment BDI = (% change BMI) * comorbidities + pretreatment BDI; this answers the question of whether, among patients equally depressed at baseline, there is any association between % change in BMI and/or comorbidities and posttreatment depression, and whether the association for comorbidities changes based on % change BMI, and/or vice versa.
 posttreatment BDI = (% change BMI) * medical vs. surgical treatment + pretreatment BDI; this answers the question of whether, among patients equally depressed at baseline, there is any association between % change in BMI and/or treatment type and posttreatment depression, and whether the association for treatment changes based on % change BMI, and/or vice versa.
 posttreatment BDI = treatment * comorbidities + pretreatment BDI; this answers the question of whether, among patients equally depressed at baseline, there is any association between treatment type and/or comorbidities and posttreatment depression, and whether the association for comorbidities changes based on treatment type and/or vice versa.
 BDI (Beck Depression Inventory) is typically very skewed. Suggested checking histograms but probably using an ordinal logistic regression model (also called proportional odds logistic regression). Trisha is currently using Prism; if Prism can't handle POLR, check into SPSS (helpful link for SPSS: http://www.ats.ucla.edu/stat/spss/dae/ologit.htm)
2015 June 10
Kristy Kummerow, General surgery resident
 "I am a surgical resident and would like to attend a Biostats clinic to request help doing multiple imputation in Stata. I would prefer tomorrow (Wednesday) if there is still space, or Thursday. Please let me know whether either of these are options."
 We discussed what should go in the imputation model (outcome to be included) and found links to help with the syntax for fitting the MI model and the final model with the MI results.
Michael Kenes, PGY2 Critical Care Pharmacy Resident
 "My study is looking at clinical outcomes of Clostridium difficile infections in neutropenic patients. I have collected all of the data and performed univariate analysis. I am in need of assistance in discussing how to handle patients who died within the study timeframe as well as a regression analysis."
 Since time to diarrhea resolution was captured in the data, we recommended they use a competing risk analysis to handle the deaths.
 Due to the low number of events, we discussed prespecifying the model as opposed to using univariate analyses to drive model selection as well as propensity scores for data reduction.
2015 June 3
Jennifer Hale, Pediatric Pharmacy
 "Evaluation of a computerized prescriber order entry protocol for pain management and sedation in a pediatric cardiac intensive care unit"
 We discussed alternate outcomes besides total med dosage, ventilatorfree days (necessary to have a common denominator to alleviate bias in differing stays in the ICU).
 We recommended the Wilcoxon Rank Sum test rather than the ttest for the unadjusted tests with continuous outcomes. We also discussed fitting models to adjust for potential confounding. If normality assumptions are met, then the linear model would be most appropriate; if they are not, then other types of models need to be considered such as the logistic regression or proportional odds model.
 We also suggested investigating average dose/day as an outcome.
Justin Bachmann, Cardiovascular Medicine
 I’d like to attend the health services research biostatistics clinic today in D2221 MCN. I’m conducting an analysis of the association between selfefficacy and physical activity in a cohort of 2000 patients in the Vanderbilt Coronary Heart Disease Study. Physical activity is characterized as both a continuous (METminutes/week) and an ordinal (low, moderate, high) variable. The independent variables include continuous, categorical and ordinal variables. I’m using negative binomial regression as well as ordinal logistic regression and would like to get the statisticians’ thoughts on these models.
 We suggested investigating whether SAS can run a zeroinflated negative binomial model or not and whether this is necessary given his data. We also discussed the robustness of the proportional odds model when the assumptions are borderline, especially when using the Score test pvalue to make the determination.
2015 May 20
Tim Shaver, Biochemistry
 Per the email: "I am a Biochemistry graduate student working to develop a retrospective study of the correlation of novel gene fusion events with biochemical recurrence following radical prostatectomy. I previously attended the Thursday clinic on May 7 and received some guidance regarding sample size justification for an upcoming VICTR proposal. However, I have encountered some difficulties implementing your advice due to incomplete reporting in our test data set. I would like to bring some of the specific numbers and receive feedback on a new plan for our sample size and power calculations."
 This study is a casecontrol study. The main issue is that the pilot data may be inaccurately or incompletely coded, so pilot estimates might be incorrect. We recommended 1) removing patients with no information from the test data set (a sizable number) in order to avoid artificially inflating the denominator; 2) using PS's casecontrol functionality to calculate the difference in proportions they can detect with the 300 patients they plan to sequence at various levels of recurrence rate among the general population.
2015 May 13
Dikshya Bastakoty, Department of Pathology, Microbiology, and Immunology
 Per the email: "I am applying for a VICTR grant (part of which involves analysis of human samples for gene expression), and was looking for help with sample size determination based on recommendation by the biostatician who reviewed my grant."
 We recommended that she check what detectable alternatives she could detect for each of her 3 primary proteins of interest given the maximum number of subjects she could afford.
 We also recommended that she review the literature to see if other studies existed that supported the difference she was using for her current power calculation.
 If the current number she reports is all that can be feasibly recruited and afforded in the time and budget constraints, we recommended she highlight that in the application.
2015 April 29
Jim Jackson and Jo Ellen Wilson, VA Quality Scholar Program
 Per her email: "I am needing a quote for the amount of time it would require to develop an analysis plan and perform analysis for a proposed project of mine. I will be using this information to submit a request for funding from VICTR."
 Recommended keeping it to aims 14, since we're not sure we have adequate data to answer aim 5. Estimate 90 hours for two manuscripts (data set is very clean, and VICTR statistician is familiar with it). We edited Jo Ellen's aims to reflect discussion of modeling specifics in clinic  Jo Ellen has this document.
2015 April 15
Eric Wise, Department of Surgery
 Spearman correlation is univariate; multivariable regression adjusts for multiple potential confounders
 Do NOT use univariate selection to decide which variables to put in models; instead use clinical judgment, literature search, hypothesis to decide which variables to put in  less subject to noise and confounding
 Model selection (forward, backward, stepwise) is fraught with problems; using above approach > no model selection algorithm, less likelihood of "data fishing"
 KaplanMeier is univariate; Cox proportional hazards model good for time to event analyses with adjustment
2015 March 18
Alexandra Fish, Center for Human Genetics
 I have a question regarding approaches to handling sampling zeros. I had previously conducted an analysis in which I used a likelihood ratio test to determine if including interaction terms substantially improved model fit. I am now trying to reproduce that analysis in a new data set, which contains sampling zeros. When I run the analysis, I am getting a pvalue for the LRT, but the program is unable to estimate the betas for individual terms. So, I guess my question is  is the LRT appropriate in this situation? Can I trust the pvalue? I am uncertain which of the clinic themes is most appropriate for this question.
 She is investigating whether the interaction of two SNPs are associated with gene expression. She has fit a model with additive and dominant terms for both SNPs and their interaction; however, out of the 5000 subjects, no one is recessive for both SNPs. We were unsure of how best to approach this; however, we recommended against some suggestions she found while searching for an answer such as simply adding a constant count to all frequencies to avoid cell counts of
 We recommended she contact Yaomin Xu.
2015 March 11
 "I’d love suggestions about data manipulation, in preparation for planned analyses. I have received data pulled from the EDW that includes multiple BP and other measures per subject and over time; this data needs to be combined with prescription data (also over
time), as well as with a 3^rd dataset that includes measures of adherence to the blood pressure medications."
Henry Ooi, Cardiovascular Medicine Heart Failure & Transplant
 "We have a large dataset in Stata format which we are planning to analyze. There are duplicate entrys and we would like advice on how to handle this this in Stata without losing data."
For both of today's clients, a couple of links on aggregating/collapsing data in Stata that may be helpful:
UCLA Stata web page with examples Short example from Indiana University
2015 March 4
Dupree Hatch
 "I am a NeonatalPerinatal Fellow with an interest in Patient Safety in the NICU. I am planning to look at unplanned extubations in the context of the NICU with a future study. I have a data set which contains ~80 unplanned extubations in ~60 patients as part of a larger cohort of all infants that have received mechanical ventilation in our unit for the past year. I am hoping to describe the risk factors for infants to have these events. I have a data set which contains timeto event data for all of the infants as well as various demographic and clinical data. My needs for Biostats clinic are: Help with designing the survival analysis with repeated measures for the patients who have had multiple events. Modeling risk of unplanned extubation with postnatal age as the independent variable Quote for how much time and resources would be needed to have formal biostatistics analysis and help with preparation of the manuscript."
 Ideal scenario would be survival model with competing risks and repeated measures; we are not sure that this exists. Possible solutions: look at time to first unplanned extubation, with death as a competing risk, censoring babies who did not have an unplanned extubation; also look at calculation of ventilatorfree days per ARDSNet definition (babies who die get zero VFDs; otherwise, calculated as [time of interest, eg 28 days]  [time on vent or time after unsuccessful extubation, usually defined as extubation followed by death or reintubation within 48 hours]). Possibly look into data reducation techniques like propensity scores, since low number of events (~60) means only 46 degrees of freedom included in model, and association between many covariates and time on vent is very likely nonlinear.
 Current VICTR policies: http://biostat.mc.vanderbilt.edu/wiki/Main/VICTRBiostatPolicies
Lisa Rae
 "I am in the process of writing a VICTR grant application looking at changes in Plasminogen and the coagulation cascade in burn patients, with a primary outcomes of: development of heterotopic ossification (incidence 13% of burns) and serum levels of coagulation factors after injury and during subsequent wound healing. My data will include serum lab values, xrays and photos of the healing wounds (time to wound closure). "
 Recommend primarily descriptive study (looking at relationship between total burn percentage as continuous variable vs inflammatory markers, eg ddimer); event rate for HO is so low (13% in total population) that it's unlikely we could practically enroll enough patients to get a proportion/CI within a reasonable margin of error.
 VICTR policies: http://biostat.mc.vanderbilt.edu/wiki/Main/VICTRBiostatPolicies
2015 February 4
Lara Harvey
 "...a study of BMI and its influence on FSH levels. I have three large deindentified datasets I have obtained from synthetic derivative and would like some help cleaning and compiling the data and entering it into Stata to use."
 To read in file from SD: "insheet using "filepath"", or Import, ASCII Data Created by Spreadsheet, select file and choose Tabdelimited, OK (using Stata 10)
 If more complicated data management is needed to find closest BMI/FSH combination, ask synthetic derivative folks where to start (bioinformatics core?)  this would fit in a 35hour VICTR voucher, but not sure biostats vouchers can be used for data management and graphics.
George DeKornfeld, VCH Pediatric Heart Institute
 "We are looking at low birth weight infants under 2000kg with complex congenital heart disease. We are attempting to statistically compare the surgical outcomes, in terms of complications experienced, of a group which was treated at initial presentation and a group which was first allowed to mature and grow. Our hypothesis is that it is better to treat earlier. The groups have been divided and the complications have been noted for each patient."
 Two main outcomes: worst complication (scored per patient) and days on ventilation
 Days on ventilation is complex due to mortality (average ~32% in population), so will be artificially truncated for patients who die. Need to account for this  look into ARDSNet definition of ventfree days for one approach.
 Think of potential confounders; birth weight is a major one (include as continuous variable in model). Can adjust for 12 confounders in addition to treatment.
 In SPSS, use ordinal logistic regression (also called proportional odds logistic regression). We caution against Excel for statistics.
 A VICTR voucher might be helpful if more complicated techniques (propensity score) or a manuscript are desired; see policies here.
2015 January 28
Mitch Odom, 4th year medical student
 He has a database of a few thousand; baseline testing only (no repeated measures), looking into neurocognitive and symptom scores for young athletes assessed by a computerized testing battery.
 There are 740 subjects in his data. Each have taken a cognitive assessment (continuous measure ranging from 0  100).
 The primary question of interest is to examine the association between cognitive score, cognitive status (ADHD; LD; ADHD/LD) and hours of sleep the night prior to the exam (categorized as < 7; 79; >9.
 We discussed any confounders that should be included in the linear regression and whether there should be an interaction between cognitive status and hours of sleep.
 He might find helpful code hints for SPSS at UCLA's web site: http://www.ats.ucla.edu/stat/
2015 January 21
Sarah Greenberg, Research Coordinator and Health Policy Fellow
 Orthopaedic trauma  looking for help with a linear regression for determining complications in long bone fractures
2015 January 14
Donald H Arnold, Pediatrics and Emergency Medicine
 The analysis is as follows:
 Pulse oximeter plethysmograph estimate of pulsus paradoxus (PEP) is an electronic measure we have developed to measure the severity of acute asthma attacks.
 Predictor variable: PEP
 Primary outcome variable: FEV1, continuous variable
 Secondary outcome variables: i. Acute Asthma Intensity Research Score (AAIRS), ordinal scored 0 to 16; ii. Airway resistance, continuous
 Will fit 3 baseline models and 3 change models. Need to finish analysis (and possible manuscript) before April 2015. Apply $2500 VICTR voucher (~40 hours).
Cesar Molina, Orthopedic Trauma
 Plans to attend for help with a sample size calculation where the prevalence is low.
 Submitted manuscript and was criticized about small sample size.
 Identify risk factors for deep infection and nonunion in pts with open distal radius fractures . N=62 (only 1 infection); N=54 were followed to be able to get outcome of nonunion (4 nonunion)
 Download PS (Power and Sample Size) software for sample size calculation. Choose dichotomous outcome, prospective, two proportions. For example, two groups will be diabetic and nondiabetic, compare prevalence of deep infection between the two groups. Need 140 diabetic and 140 nondiabetic pts to detect a difference between 15% and 5% with 80% power.
Carolina Pinzon, Surgery
 Feasibility study on prevalence of BMP7 protein in two groups of women. Have done experiments in animals, but no data in human
 If the proteins can be measured in human, want to compare the protein levels between groups. Need standard deviation and a clinical meaningful difference to calculate sample size