Biostatistics applications in surgery, anesthesiology, and emergency and critical care medicine Clinic Notes (2016)

---

2016 December 28 - canceled due to holiday

2016 December 21

Chelsea Isom, General Surgery -- Postponed

• Updates to project.

2016 December 14

Laurie Tucker, Department of Pediatrics -- Canceled

• Follow-up to previous clinic visit.

Andrew Smith, Pediatrics/Anesthesiology - No show

• I am currently embarking on a multicenter look at variations in value delivery to critically ill children across congenital cardiac surgical centers across the US, using data merged from two data streams, a clinical registry (Pediatric Cardiac Critical Care Consortium or PC4) and an administrative data set (Pediatric Health Information Service) to try and pull together the numerator and denominator of the value equation… I was wondering who would be able to help me think about how best to look at cost (and value) comparisons from a statistical standpoint, with respect to outcomes including mortality. Specifically, given that some children die relatively soon after surgery, they may not incur substantial cost though one would also argue that they didn’t get the “ alue" they wanted from their episode of care… I’m thinking about this from a censoring and survival curve/Kaplan-Meier standpoint, but I’m sure it is more complex than that… which is where I think some healthcare economic statistical prowess would come in handy.

Kazeem Oshikoya, Clinical Pharmacology

• Requesting help with interpretation of a data analysis.
• Looking at risk factors for composite adverse event (change in BMI, increase in blood sugar, others) among pediatric patients prescribed risperidone for at least four weeks. Observation period is 16 weeks (>=4 weeks of risperidone + additional weeks up to 16). Eventually will look at genetic variants but focusing on this for now.
• Currently has data on 210 patients; among these, has 45 events. Number of parameters that can be included in a logistic regression model is the minimum of (events, non-events) / 10-20. So, with 45 events, can include 4 (maybe 5) parameters; any more, and the model will be overfit, meaning it will be perfectly fit to this data set but will have radically different results if applied to a different cohort.
• Recommend not doing testing on univariate descriptives: for example, might be OK to describe age among patients prescribed risperidone on vs off-label, but don't test this difference. Can be misleading due to presence of confounders. Use clinical judgment/literature to prioritize which covariates should be included in the model.

2016 November 30

Laurie Tucker, Department of Pediatrics

• Follow-up to previous clinic visit.
• Went over spreadsheets of data collected since last visit and made suggestions: data dictionary to indicate what each variable level means; get ICD9 codes in addition to CPT codes, and think about how to cluster ICD9 codes; think about clinical outcome variables to represent general questions of interest (eg, we can model language category vs level of ED triage).
• Plans to straighten out data issues and come back to clinic 12/14.

Alexander Hawkins, Department of Surgery - had to cancel

• "Working with patient satisfaction scores and looking at association between disease processes. Would like help with how to interpret scores and adjust for providers, pain scores, etc."

2016 November 23 - canceled due to holiday

2016 November 9

Chelsea Isom

• " Approximately 20% of patients with colorectal cancer (CRC) present with metastatic disease-most commonly to the liver or lungs. Successful resection of these metastatic foci leads to significant long-term survival. Less commonly, patients present with isolated metastasis to non-regional lymph nodes (NRLN) and little is known regarding the role of resection in these patients. The primary aim of this study is to evaluate the outcomes of patients with CRC who undergo resection of NRLN metastasis. A retrospective cohort study of patients diagnosed with CRC and NRLN metastasis was performed using the Surveillance, Epidemiology, and End Results database (2004-2012). Demographic and clinical factors will be compared for patients who underwent resection of NRLN metastasis and those who had not. Kaplan-Meier and log-rank analysis will be used for survival analysis. Logistic regression analysis will be used to assess factors associated with resection of NRLN metastasis."
• Data set is one record per patient, 829 patients of interest. Limited data available on potential predictors (registry data). Suggested things to look into: propensity for getting surgery vs not, competing risks, multiple mortality models looking at patients who have had opportunity for at least 1 year of followup and the subset who has had at least 5 years of followup.
• Estimate about 90 hours for data management, analysis, manuscript writing and revisions.

Mark Clay & Ashley Newell, Pediatrics (Cardiology/Critical Care)

• "Restrospective project looking at increased BMI as a risk factor for increased resource utilization in patient after Bidirectional Glenn procedure. The data was previously analyzed use a Loess Regression using R software. The data has been edited and we are seeking help with repeat analysis and graph generation."
• Need to look at model assumptions for prior analyses - for example, residual vs predicted plot. Based on distribution of LOS and ventilator hours, we have concern that model assumptions are violated and therefore the model results would not be reliable.
• If that's the case, look into perhaps a negative binomial model in R (function glm.nb() in the MASS library.
• Continue keeping Z score for weight as a continuous variable in the model. Consider adding patient location (followup at VCH vs clinics in other areas) to model, but may need to prioritize covariates: With 109 patients in a linear regression model, can only have ten degrees of freedom (roughly corresponds to covariates) and still trust model results.

Leah Hauser, Otolaryngology

• "Studying olfactory (smell) dysfunction in CRS. There is some prior evidence that tissue eosinophilia contributes, but this role is controversial. Our 3 major questions are: 1. Does objective olfactory function measured by age/sex adjusted UPSIT score correlate with tissue eosinophil counts?; 2. Is olfactory function in eosinophilic CRS due to tissue eosinophilia or disease severity?; 3.Is the effect of eosinophilia on olfactory function associated with type of CRS (CRS vs CRSwNP)? We think that preliminary data analysis shows that eosinophils counts (column J) correlate moderately with UPSIT score in CRSwNP but not at all in CRS(without NP), but we suspect this may be due to worse disease rather than the eosinophils themselves. We are not sure how to best analyze our data to determine the etiology of olfactory dysfunction."
• Looked at distribution of outcome (UPSIT scores, raw and adjusted); distribution is bimodal, which makes linear regression problematic. Consider other regression options like proportional odds (aka ordinal) logistic regression for multivariable associations.
• For univariate associations, Pearson correlations are probably invalid for the same reason; use Spearman (rank) correlations instead.
• Clinically investigate reason for bimodal distribution.

2016 October 26

Katie DesPrez, Critical Care?

• "Retrospective clinical project on ARDS. Briefly, I am interested in understanding whether my correlation between the variable I've called OSI and mortality is valid even in patients who have no blood gas (i.e., in this data set, patients who do not have the variable OI). Preliminarily it does not seem to be, but I'm wondering whether this is because the data is underpowered for that particular analysis."
• Using Stata to compare non-nested ROC curves: http://www.ats.ucla.edu/stat/stata/faq/roc.htm
• Could also do a model with both oxygenation variables and see whether blood gas version adds additional predictive value after adjusting for pulseox version.
• Also suggested looking at time to death (vs died/survived), and looking at SSDI for death dates especially for patients discharged to hospice.

Laurie Tucker, Department of Pediatrics

• "A project looking at the acute health care utilization patterns of non-English speaking patients in comparison to English speaking patients. The study is set up as a retrospective cohort study. We have gather data from Star Panel, and I would like a bit of help determining the next steps in analyzing the data."
• First step is to determine who exactly data has already been collected on: patients who were already established as of July 2013, or does it include patients who were born or were established after that date? If the former, everyone should have the same followup time; if the latter, need to deal with different followup times in analysis.
• Also think of potential confounders for relationship between language group and rate of acute care visits - does one group have higher severity of illness, for example.
• Recommend coming back to clinic after discussion with data colleagues. Planning to apply for VICTR voucher.

2016 October 19

Billy Cameron, Surgical ICU/Trauma

• "I am currently working on a project to justify a Nurse Practitioner team in the Trauma division at VUMC. We performed a 12-week pilot, for which we have good data showing decreased length of stay. One of the data points was to compare an acuity scale: Injury Severity Score (ISS) to show that acuity remained pretty even from the previous year before the pilot compared to the pilot period. I am trying to figure out what the statistical significance of the difference is (we are wanting to show that the level of patient acuity according to the ISS was relatively stable). The comparison period prior to the pilot, the ISS was 12.62 (scale of 0-75) for n= 281. For the pilot period, the ISS was 11.99 for n= 332."
• Preparing for presentation to leadership and want to show that difference in LOS is not due to clinical factors like difference in mortality or ISS between retrospective and pilot period. For these purposes, recommend t-test or Wilcoxon test (depending on distribution of data) comparing ISS scores between the two periods, and chi-square test for proportion of deaths during each time period.
• For eventual manuscript, will need more advanced analyses; recommended going through VICTR for statistical analysis support.

Joseph Kuebker, Endourology

• "We are attempting to design a study to see if the effective dose (radiation exposure) for a particular type of xray we do is comparable to the generally accepted historical average. Specific questions are how many patients we should enroll to detect differences of >10% (if possible) and what tools to use given we are comparing against a generally accepted number and not against an actual groups of patients/exams."
• Main "punch" will be plotting the data to describe it. Try boxplots with raw data, perhaps seeing Example 2 here for guidance in Stata: http://www.ats.ucla.edu/stat/stata/code/twboxplot.htm
• If a test is absolutely necessary, a one-sample t-test (or nonparametric version, depending on distribution of the data) would be most appropriate. But main value of project will be describing how VUMC patients are dosed compared to current guidelines (single number). Recommend minimum of 20 patients per clinical category (eg, female overweight, male normal weight...).

2016 October 12

Jessica Grahl, Pharmacy

• Basic question: whether antimicrobials are associated with an increased risk of delirium among critically ill patients, using an established cohort (BRAIN-ICU) plus additional data collected from StarPanel.
• Mental status changes daily and can be normal, delirious or comatose. One idea: multinomial regression looking at status "tomorrow" vs antimicrobial use + covariates "today." How to account for repeated measurements within patients? (Jennifer and Rameela have done cluster bootstrapping; this is complicated and takes awhile)
• Simpler idea: take out all comatose days from outcome, only look at delirium vs normal status. This limits what we're able to say from the study, but does simplify analyses.
• Antimicrobials = antibiotics, antivirals and/or antifungals. Patients are often on >1 of these classes and in the case of antibiotics, could easily be on >1 drug of the same class. Also could be lots of interactions between subclasses and other confounders/modifiers.
• Depending on the type of analysis chosen, this could be a very time-consuming project that might bump it up to the highest level of VICTR projects. If the logistic regression approach (or something similar) is chosen, estimate 90-100 hours for a typical project with data management, manuscript revisions, etc.

Joseph Kuebker, Endourology

• "We are attempting to design a study to see if the effective dose (radiation exposure) for a particular type of xray we do is comparable to the generally accepted historical average. Specific questions are how many patients we should enroll to detect differences of >10% (if possible) and what tools to use given we are comparing against a generally accepted number and not against an actual groups of patients/exams."
• Missed clinic

2016 October 5

Debra Braun-Courville, Pediatrics

• Project is looking at weight gain among adolescents given a specific type of birth control; VICTR application was sent back due to lack of control group. Have access to medical records; suggest matching cases (girls who received a specific type of birth control) to controls (girls who did not receive birth control) as well as possible, using age, race, BMI, and any other available factors.

2016 September 28

Luis Huerta, Pulmonary/Critical Care

• "I am working on a pilot single center cluster-randomized multiple-crossover trial of contact precautions in the Vandy Medical ICU which has not begun yet, and would like assistance with designing the statistical analysis plan, particularly taking into account the cluster-randomization of the planned trial."
• Brought up several concerns that have been thought through: seasonality, proximity between patients/presence of infection on floor, how to recruit sites for larger study, different patient populations in different ICUs, admission diagnoses
• Suggested following up on Frank's email and meeting with him, Dan B, Robert, who have experience with these types of trials

2016 September 21

Ashley Kroeger, Pediatric Critical Care

• >1000 patients discharged from pediatric cardiac ICU; looking at risk factors for readmission
• Two different versions of Pediatric Early Warning Signs score: validated, and VCH-specific (extra components)
• Question 1: is PEWS score, or the difference in PEWS between ICU discharge and floor arrival, predictive of time to ICU readmission?
• Question 2: do the extra components added at VCH add helpful information when it comes to predicting readmission?
• Possible analysis: Cox proportional hazards model with outcome = time to readmission (patients who were never readmitted are censored at hospital discharge), covariates = standard PEWS score + score on VCH-specific component + confounders
• could do two versions of above, one using score at ICU transfer and one using score at floor arrival
• be wary of multiple hospitalizations per patient - may need to deal with this in analysis or just take first hospitalization per patient
• estimate ~90 hours for VICTR voucher

Alex Hawkins, Surgery

• Trying to sort out statistical analysis for three separate cohorts- a group that got radiation pre-op, a group that got it post op and a group that never got radiation from data using the NCDB
• Question 1 - does neoadjuvant radiation improve rate of R0 resection and/or overall survival?
• Question 2 - does preop radiation improve rate of survival?
• Suggest time-varying Cox model to incorporate post-op radiation properly, since patients will start post-op radiation at varying times after surgery (time 0) (including it as a single value would introduce immortal time bias - patients who die faster don't have opportunity to have radiation postop)

2016 September 14

Maya Yiadom, Emergency Medicine

• Questions about how to phrase results of study with a small sample size and trending-but-not-significant results
• Key suggestions: make sure to compare patients with and without missing data; try to compare 54 sites with data vs national characteristics (region, hospital type, anything else you can get); discuss limitations of sample size/power, potential bias and overfitting - don't overstate results

Alice Hensley, Pediatric Critical Care

• "I am working on a project looking at multitasking abilities and comparing scores on an online multitasking test to residency milestone assessments, but would like guidance on how best to analyze the data that I will be collecting. "
• Out of 100 residents, currently only have data on 22, so priority #1 is getting people to take the multitasking test as soon as possible (needs to be far enough before six-month evaluation to truly be baseline measurements).
• Get as much data as possible on demographics and resident characteristics, to be able to compare residents who did and didn't take the test and possibly impute for missingness at the time of analysis
• Probably use proportional odds logistic regression (outcome is a score 1-5)

2016 September 7

Shaun Mansour, med student/global health

• won't be here until 12:45; see preliminary info in email

Joshua Arenth, Pediatric Critical Care

• "I am planning a study evaluating the effectiveness of a curriculum on communication skills in the ICU. I would love to talk with someone about project design and potential statistical analysis requirements. "
• Descriptive statistics will tell most of the story: percentages in each of the three communication styles before and after intervention by intervention group
• To get p-value, suggest logistic regression: [optimal vs suboptimal style, after intervention] = [style before intervention] + [intervention, yes/no]; odds ratio for intervention is what will tell you whether the intervention group is different from the non-intervention group
• Spaghetti plot showing change from before to after intervention, with two groups in different colors
• Estimate about 40 hours for analysis and manuscript through VICTR

2016 August 31

Vance Albaugh & Georgina Sellyn, Surgery

• Powering a longitudinal study examining cognitive function after surgical (two types) or medical weight loss surgery
• Could do one/both of the following for a very simplistic approach: paired t-test or two-sample t-test (combined surgery groups vs medical); this should give the "worst case" scenario
• Feasibility: could easily enroll ~100 patients in each of three groups, cost not a major issue
• Main interest is whether differences in cognition are seen very early (1 or 3 months) as opposed to the known differences at 12 months after surgery; could do a mixed effects model (longitudinal) for final analysis

Shaun Mansour, medical student/Global Health

• moved to next week at 12:45

2016 August 17

Chelsea Isom, General Surgery

• "I need help with an NCBD data base set." - general Stata
• Recommend saving all code for data management and analysis in a do file. See: http://www.stata.com/manuals13/u16.pdf

Uche Anani, Division of Neonatology

• "I am currently refining my IRB protocol for my mixed method study on clinical decision-making during the perinatal period. I am using a validated survey for my patient population but need some help to determine how big my sample size needs to be to have a power of 80% or p < 0.05."
• Prospective; planning to administer a decision-making survey to both patients and clinicians (OB/GYN, genetic counselors, and in some cases, neonatalogists) to measure amount of struggle patient is having with making pregnancy-related decision. Main question - how well do clinicians predict how much the patient is struggling. Survey outcome is a continuous score ranging 0-100.
• For descriptive analyses, not much need for power calculations; get as much data as is feasible, then do (for example) scatterplots and correlation statistics comparing genetic counselors to patients and (separately) OB/GYN to patients.
• Predictors of a closer relationship are also of interest (education, religiosity, health literacy, years of clinician experience, etc). Could do linear regression for this.
• Suggest contacting Frank Harrell to determine if there is an available collaboration plan with pediatrics; if not, apply for VICTR assistance. End goal is a manuscript; clinic staff believes this would fit in the <90-hour VICTR category.

2016 August 10

Erin Powell, Pediatric Critical Care

• Followup visit from 7/13
• Suggest performing Cronbach's alpha to make sure all questions are informative. Also check to see if original instrument has a validated scoring system. If the alpha works out and there is no other scoring system, suggest creating a single score that is the sum of all 15 questions.
• Suggested model: post score = pre score + experimental/control (or score = pre/post*experimental/control, using an interaction term to determine whether there is a difference between groups, but first option uses fewer degrees of freedom, and N = 17)
• Goal is to publish both curriculum and a manuscript about the curriculum's efficacy, probably submitted to an educational journal

Debra Braun-Courville, Pediatrics

• Followup visit from 2/25
• Goals: manuscript and conference poster/presentation in March/April 2017 (deadline in November)
• 282 adolescents (age 12-23) on progesterone-only implanted birth control known to have side effect of weight gain
• Ideally, see about pulling data from entire population "eligible" to receive this type of contraception (maybe talk to synthetic derivative folks?); that would allow comparison of weight gain/BMI/etc between similar patients who did and did not receive this type
• If that isn't possible, some ideas...
  • discriminant analysis/risk factor model for weight gain among patients who did get this type
  • separate subgroup analyses among (eg) 12-17yo and 18+yo, because these populations are so different in terms of the outcome
• For VICTR purposes, actual plan will depend on what data is available, but likely to need 90-100 hours either way for data management, analysis, manuscript writing and revisions

2016 July 27

Susan Dickey, Pharmacy

• She has questions related to a logistic regression for a retrospective critical care research project regarding the duration of antimicrobial therapy for intraabdominal infections in critically ill surgical patients.
• 240 patients total who met inclusion criteria, approximately 70 events (event is a composite outcome defining "treatment failure"); excluded transplant patients and those in SI <24 hours
• Lots of confounding due to ICU/hospital LOS. Suggestions:
• Look into doing a Cox model for time to event data instead of logistic regression; patients who did not experience event will be censored at hospital discharge
• Sensitivity analysis only including patients who stayed >=8 days (long enough to potentially be included in the "long" treatment group); this will reduce sample size, but still leaves enough for a reasonable analysis, and would reduce confounding from patients who only stayed a few days
• Look into whether patients in the long antibiotic group but never got vasopressors were withdrawal of care patients

2016 July 20

Melissa Warren, Critical Care Medicine

• "We have created a new chest x ray scoring system and are seeing how this score can/may be used in patients with critical illness to assess prognostication and outcomes. We have currently scored all of the chest x rays and are analyzing patients in the FACTT database (a former critical care study looking at conservative vs liberal fluid strategies in critical care). I was hoping I could sit down with a biostatistician to discuss which tests would be best to use/how to perform them in SPSS in order to look at the correlation between score/outcomes."
• We discussed how to show agreement with continuous measures (Bland-Altman plots) and recommended she create plots for the overall score and by component to see if any portion of the score is driving any disagreement. If needed, we also recommended looking by quadrant because of potential issues in the lower left quadrant.
• Due to the scope and technical complexity of her questions, we recommended that she check with Tatsuki Koyama (tatsuki.koyama@vanderbilt.edu) to see if this work would be covered under a collaboration plan. Otherwise, we encouraged her to apply for a VICTR voucher.
• Some of the outcomes of interest are time to successful extubation, ventilator-free days, time to death, LOS, etc. We think the scope of the work should take at least 90 hours.

2016 July 13

Erin Powell, Pediatric Critical Care

• They would like to discuss the data for their project evaluating the effectiveness of a curriculum to teach communication skills to pediatric critical care fellows.
• We discussed the different measures they are using to evaluate their chosen outcomes and appropriateness of methods suggested to use for analysis (t-tests).
• We suggested factor analysis might be most appropriate to answer their questions of interest and suggested that if this approach is beyond the scope of their abilities to apply for a VICTR voucher.

M. Frances Wright, Medical student

• They have questions related to their project on blood product utilization during liver transplant surgery.
• They had several questions related to outliers and what to do with them in the analysis. We highly encouraged them not to remove them from any analysis unless they can prove that these measurements were made in error. We also tried to help them understand analyses done for them previously. We also encouraged them to consider applying for a VICTR voucher if the scope of the future analyses seems beyond their abilities. They were going to check whether a collaboration existed between our two departments.

2016 June 29

Justin Gregg, Urology

• He has questions regarding sample size calculations.

Jamie Felton, Pediatric Endocrinology

• "My questions are regarding the best way to statistically analyze a data set from an ELIspot assay."

2016 June 8

Ravi Bamba, Plastic Surgery

• "I have a dataset that I finished collecting from a previous project. I needed help running my stats but I do not have funding."
• Investigating risk factors for recurrence of pressure sores in subjects who had a surgical intervention to fix the problem.
• He has data from 1997 - 2015. We advised that he restrict follow-up to ensure that all have had equal opportunity to have a recurrence observed.
• We recommended that he use non-parametric tests (Wilcoxon Rank Sum test, e.g.) for the univariate analyses.
• His primary question of interest is whether there is a difference in time to recurrence between different pre-specified risk factors. We showed him the UCLA website as a resource as to how to fit Cox models in SPSS. We also advised him to organize the covariates of interest from most to least important based on clinical knowledge and literature.

2016 June 1

Ravi Bamba, Plastic Surgery

• "I have a dataset that I finished collecting from a previous project. I needed help running my stats but I do not have funding."
• No show for clinic

2016 May 25

Deborah Jacobson, General Surgery - canceled due to OR schedule

• "We have data including complications/pt/year for 10 years of data and want to see if there is a significant decline in complication rates over time."

Viraj Mehta, Ophthalmology - moved to Thursday clinic

• "I'm evaluating eye motility outcomes after surgery for orbital floor fractures in children. I have collected all the data, and needed help figuring out the best way to analyze it."

2016 May 18

Vance Albaugh, Department of Surgery

• "I have a question about powering a clinical research study, as well as some specifics about the data analysis."
• Looking at gastric bypass patients' glucose tolerance tests at multiple time points after surgery. Plan: give them a regular glucose tolerance test, measure response (every 15-30 minutes, so we get a curve for each test), then a few days later give the same test supplemented with salt. Hypothesis is that salt will make the glucose response worse (higher) initially, but by a year after surgery, the salt/no salt responses will be roughly equivalent. No-salt response will also change over time as patients become less insulin resistant.
• Of possible interest: DeLong et al, 1988: http://www.ncbi.nlm.nih.gov/pubmed/3203132
• Jeffrey Blume in biostats might be a good resource for AUC curves - this project is especially complex due to longitudinal measurements + AUC measurements
• Number of patients could be pretty high - this is a relatively easy study to do compared to other gastric bypass studies
• Describing and plotting this data will likely be as or even more informative than statistical testing. Something like one row per patient, one panel per time point, with salt vs no salt at each time point in each panel. 5-10 patients' worth of pilot data would be highly informative for future sample size/analysis discussions.
• Maybe a mixed effects model along the lines of: glucose response = salt * time + visit * time + covariates (# covariates restricted by sample size)
• Jackie is looking into latent growth curves

Maya Yiadom, Emergency Medicine

• I am submitting a K23 proposal and could use help identifying:1) Whether I’ve selected the right study design for may aims; 2) The right analysis method should be for Aims 2 and 3; 3) How do I get an appropriate ED (Aim 2) and patient (Aim 3) sample size for Aim 2 and 3?
• Could fit one model using patient-level data to answer both aims 2 & 3, including both ED- and patient-level characteristics. This would allow you to get estimates for, say, academic vs non-academic institutions, or patient age, after adjusting for all other factors.
• Recommend plotting time to diagnosis & time to treatment for any available pilot data to help inform model choice. If outcome is normally distributed and patient-level data is used, a linear mixed effects model could be good (site is random effect).
• Time to treatment gets very tricky because some patients get treated via medication and some via procedure - procedure inherently has longer time to treatment. Look at distribution of times separately and together - will likely need two separate models to answer treatment question.

2016 May 11

Jordan Rupp, Emergency Medicine

• "I have a couple quick statistics questions for a small QA study in which I am participating. We will be assessing the lung ultrasound abilities of the emergency medicine residents at the Nepali hospital after a brief 2 week teaching session given by Bales and I in March. I need some help making sure our sample size calculations, etc. are correct."
• Studying pneumonia; typical gold standard is CT, but not feasible in Nepal
• Original plan: do chest x-ray (standard, but can take up to six hours) and ultrasound on all suspected pneumonia patients, compare to discharge diagnosis
• No data from before the class is available, so can't compare pre- and post-training. If that is the true main question of interest, the study needs to involve equivalent providers or sites who didn't get the training course.
• Possible main question: does doing the ultrasound at presentation add value to the standard chest x-ray, in terms of accurately predicting whether the patient is diagnosed with pneumonia?
• Need to refine research question before continuing with sample size calculations.
• If money is available to do CTs on everyone, sample size will depend expected sensitivity/spec/PPV/NPV and on how wide a margin of error would be clinically acceptable (eg, if we expect something like a point estimate of 80%, would a 95% CI of [70%, 90%] acceptable?)

Amelia Maiga, General surgery resident

• "I have two specific R coding questions for a survival analysis I'm doing on a multi-institutional retrospective cohort of surgically-resected distal cholangiocarcinomas. I've tried stack overflow and perusing the Hmisc source code without success.
• 1. I am using aregImpute to impute missing covariates, but run into errors when I attempt to include any factor variables with 6 or fewer observations per factor level. When I attempt to specify group=d$site (where site is a factor covariate with 10 levels, one of which only has 6 observations), I get a different error message about not all values of d$site represented in observations with non-missing values of another covariate.
• 2. I would like to use fit.mult.impute to fit a Cox proportional hazards model utilizing the data imputed by aregImpute, but despite specifying the xtrans object appropriately, I keep getting an error message "imputed=TRUE was not specified to transcan", suggesting that R thinks I intend to use transcan rather than aregImpute to impute the data to fit the model."
• Suggest an interaction term between log10(followuptime) * death in the aregImpute() (might need to create log variable beforehand)
• Suggest pooling sites by region in a new variable to use in imputation, and/or include site in analyses, possibly by stratifying (strat = 'site' in cph()), which would allow the baseline hazard to differ between sites (still get one HR for each covariate)

2016 May 4

Drew McKown, Pulmonary/Critical Care

• Hoping to discuss test selection/power calculations prior to IRB submission
• "The idea is to perform a physiologic assessment of the patient to determine an ideal ventilator setting and then assess if that setting is different from one prescribed by an algorithm."
• Basic question: There is an ARDSNet algorithm for setting tidal volume based on PEEP. Want to compare the tidal volumes recommended by the algorithm by tidal volumes determined by stress index (measure of how much stress is on the lungs).
• Recommendation - calculate power/sample size using a paired t-test and SD of the difference for the means between ARDSNet/stress index results. However, for actual analysis, recommend nonparametric (Wilcoxon) test, since especially with a small sample size, assumptions for t-test are likely to not be met.

2016 April 27

Andy Brooks, Center for Human Genetics Research

• Wants to discuss power calculations for proposed research project
• No current power methods exist for these genetics methods - gave some suggestions about simulations, etc. Suggest coming to future Tuesday omics clinics for more long-term discussions (except first Tuesday of the month).

2016 April 20

Vance Albaugh, Surgery

• "I am planning a clinical study and would like my sample size to be reviewed by a biostatistician before I submit for VICTR funding. The study is a randomized, double-blind experiment in human volunteers examining the effects of a drug commonly given to liver failure patients on oral glucose tolerance."
• First study: ileostomy patients, want to see if they respond differently to placebo vs. treatment given directly to the small intestine. Patients with and without diabetes will likely respond very differently - need to know how to power this (for VICTR application).
• Most conservative approach - power separaately for diabetic and non-diabetic patients. Complication is that there is no pilot data on diabetics; can guess that variance is twice as much for these patients as non-diabetics. For analysis, suggest doing one model with interaction term to get most efficient/accurate treatment effect (AUC = tx + diabetes + tx*diabetes).
• Second study: bowel length vs. weight loss (and potentially other outcomes) in gastric bypass patients. Suggest longitudinal approach: model with patient ID as a random effect, with weight at each time point as the outcome and baseline weight and bowel length and time as independent variables. Data will be structured with multiple records per patient. Can show results graphically by showing a line over time for patients at (for example) the 25th, 50th, and 75th percentile of bowel length. Allow at least bowel length to have a nonlinear association with the outcome (restricted cubic splines is a popular approach). Could apply for VICTR voucher if this analysis is too complex to do himself.

2016 April 13

Flavio Silva, Orthopedics

• Project: Scapular and cervical neuromuscular deficits in musicians with and without playing related musculoskeletal disorders (case-control study)
• Asked for help with regression and descriptive statistics
• Case-control study - original plan is to match; we suggest using entire cohort and adjusting for confounders (effective sample size of ~70)
• Outcome: chronic pain; covariates: three test scores (two are closely related, one is less closely related)
• One test: six unique values (20, 22, 24, 26, 28, 30; he has dichotomized this based on previous literature); neck flexion: number of seconds (less than a minute; normative means are 24 or 38 depending on gender - might dichotomize this); scalpular dyskinesis is dichotomous yes/no
• Suggest looking at Spearman correlation between two neck flexion tests to see how closely related they are - if very closely, might not make sense to include both (adjusting for one would make the other meaningless)
• Additional analyses use test scores (above) as dependent variables. For linear regression with test with 20, 22... as outcome, need to carefully look at diagnostics to make sure results are reliable. Some guidance for SPSS might be here: http://www.ats.ucla.edu/stat/spss/webbooks/reg/chapter2/spssreg2.htm. If assumptions aren't met, could consider ordinal logistic regression for this outcome.

Tony Qiu, Anesthesiology

• "I'm doing a research with anesthesiology department and currently in process getting IRB approval, I have a few questions regarding data analysis part. My question is what model can I use to assess mortality data across different institutions?"
• 30-day mortality is often used as a quality marker for non-emergent surgeries. Question is whether institutions are "gaming the system" - keeping patients alive long enough to make that marker, then transferring to palliative care, or just not taking more severe cases due to the risk of not making that marker.
• For IRB application, could do Kaplan-Meier plot for general time to death across all institutions. This does not get at the question of whether different institutions are "gaming the system" (keeping patients alive to the 30-day marker and then being less careful).
• But one K-M plot is not going to fully answer the question (might get IRB approval, but won't actually answer the question).
• When it comes time for the analysis, might suggest a Cox proportional hazards model with time to death as the outcome, could adjust for potential relevant confounders (severity of case, etc).

2016 March 30

Flavio Silva, Orthopedics - Canceled

• Project: Scapular and cervical neuromuscular deficits in musicians with and without playing related musculoskeletal disorders (case-control study)
• Asked for help with regression and descriptive statistics

2016 March 23

Ravi Bamba, Plastic Surgery

• Working with burn patients, looking for association between age and a) number of cytokines and b) % change in 2nd vs. 3rd degree burn between initial and final assessments.
• Initial plan was to collect data only on patients <30 and >65 years old, then to look for a difference in the two groups. We recommended collecting data on a spectrum of patients and then looking for an association between patient age and the two outcomes. This will allow the results to be more generalizable and have more power (less loss of information). Hoping to collect data on ~60 patients if time/logistics allow.
• Also consider what confounders to collect data on and adjust for: possibly total burn surface area, comorbidities, burn mechanism, other clinical factors.
• Plan to apply for VICTR voucher for both lab funding and statistical support, with end goals of pilot data for a grant and a manuscript. Suggest around 60-75 hours of statistical support for data management, modeling and diagnostics, manuscript writing/editing and revisions.

2016 March 16

Lyly Nguyen, Critical care

• Comparing burn ICU outcomes from time period before a specific drug was administered for inhalation injury (2002-2008) and after that drug became part of standard care (2008-2014).
• For univariate comparisons, recommend describing variables using median and interquartile ranges (rather than or in addition to mean/SD) and using Wilcoxon nonparametric tests rather than t-tests.
• Outcomes include ICU LOS, probably hospital LOS, vent-free days, and pneumonia (ever/never during ICU stay).
• For continuous outcomes, all of these are highly skewed, so need to transform them before running a linear regression model (see this link for help in SPSS: https://statistics.laerd.com/spss-tutorials/transforming-data-in-spss-statistics.php).
• Number of variables you can put in the model: For continuous outcomes, it's the number of complete cases (no missing data) / 10-20. For pneumonia, it's the minimum of (pneumonia, no pneumonia) / 10-20.
• Calculating vent-free days: Pick a common denominator among all patients (say, 28 days). If a patient dies, they automatically get 0. If they survive, they get (28 - number of days on vent in first 28 days of ICU stay; assume not on vent after ICU discharge).
• Missing data methods in SPSS may not be robust.
• Limitations/things to be aware of:
• - Missingness can strongly bias results and affect number of covariates that can be included in the model.
• - Mortality rate is about 20% and can also bias results - as one example, this may mean that patients with a shorter ICU LOS are actually doing worse (dying earlier) than patients with a longer ICU stay.
• - Temporal confounding can limit interpretation - can't say that lower pneumonia rates cause fewer vent days, for example, since we don't have timing of either event; treatment effect is also confounded by time. Also clinical care may have changed fairly drastically over the 12-year study period.

2016 March 9

Oliver Gunter, General Surgery

• "I have a question regarding a large database study I’m conducting. This is an IRB approved study that I’m trying to finalize for submission for publication. I have some questions regarding possible propensity score matching to eliminate problems I have with differences in patient characteristics."
• Given that this is survey weighted data, and there is plenty of sample size (N = 186,000 with event rates of 10-15% for the two outcomes), there doesn't seem to be a need for propensity score adjustment, and it could add complications due to survey weighting. Just adjust for individual covariates in the main model.

2016 March 2

Scott Boyd, Surgery

• Main research question: whether short (24h) vs. long-term (7-day) antibiotic use is associated with a difference in infection rates after a specific type of oral surgery at two sites (retrospective cohort). Most infections concentrated within 30 days of surgery date; infections observed after this tend to be different and in different types of patients.
• Major limitation: antibiotic use is constant across each site, so antibiotic duration is completely correlated with study center. Suggestion: describe rates of other types of infection at those sites to (hopefully) show that those are similar, so that any association found in this analysis is more likely to be due to antiobiotic duration than just study center effect.
• Suggested logistic regression (outcome = infection, yes/no) and also Cox proportional hazards model, where outcome = time to infection. No patient has >1 infection. In either of these models, effective sample size is ~53 (number of infections), so could adjust for up to five parameters to account for potential confounding.
• Make sure to discuss in limitations section the idea that despite doing everything we can, it is not possible to completely tease out the association of antibiotics vs. the association of study center and unmeasured confounders that go along with that.
• Planning future prospective study which will hopefully better address these issues.
• For VICTR planning purposes, this should fit in a regular 90-hour VICTR project.

2016 February 24

Debra Braun-Courville, Pediatrics

• She is working on a clinical research project looking at contraceptive usage among adolescents from chart review data and needs guidance regarding the analysis.
• Recommend doing KM curve for up to 12 months (or whenever a large proportion of patients have data up to this point - majority of patients had device inserted >x months ago).
• Could do a Cox proportional hazards model with time to removal as outcome (patients with device still in are censored at 12 months, or whatever time point is used), and baseline variables as covariates: age, previous pregnancies, etc.
• For variables such as bleeding, weight gain, etc which are collected during followup, recommend doing descriptive statistics for reason the device was removed - analysis with these variables is going to be biased due to lots of missingness in the clinical record.
• Try looking at UCLA's stats web site for examples in SPSS, or apply for VICTR voucher

Dan Wang, Hematology-Oncology fellow

• "I’m a first year Hem-Onc fellow and am doing an epidemiology project and had a quick question on how to calculate a p-value for comparing two APC (annual percentage change) using data from a SEER-like database (Texas Cancer Registry)."
• Because the data is very aggregated (rates per year by demographic), not much we can do statistically. Descriptives and figures are probably the best bet.

2016 February 17

Meredith Stocks, Medical student

• "I am a medical student assisting Dr Sarah Krantz with a project looking at short interpregnancy interval and counseling at antepartum and postpartum appointments. We already have a population of short IPIs but need help setting up our control group. Dr Krantz has done a bit of work regarding the design of the project and I will email specifics closer to the date of the clinic as we are meeting this week to have everything ready."
• Have 300+ women with short IPIs within five years; need to know how many cases to pull. Ideally, pull all data within same five-year period, but if this isn't feasible due to logistics, try PS software to see how many patients give adequate power. http://biostat.mc.vanderbilt.edu/wiki/Main/PowerSampleSize#PS:_Power_and_Sample_Size_Calculation
• Perform one logistic regression model: short IPI = attendance at antenatal visit + confounders (demographics, provider type, etc). Interpretation: odds ratio for antenatal visit is the odds of short IPI for those who attended antenatal visit vs. those who didn't, adjusted for all other confounders.
• Matching is also a possibility; can be more clinically straightforward, but is more work on the front end.

Clint Leonard, Vanderbilt Burn Center

• "We are a team from the Burn Center currently working on a manuscript entitled "Assessment of Outreach by a Regional Burn Center: Utilization of resources should be part of education for referring providers." We had some questions about analysis and interpretation of our results that we were hoping to discuss with you this Wednesday. "
• Essentially, we studied all interfacility transports to the Burn Center from Jun 2012 - Jul 2014. For the 623 patients that met our inclusion criteria, we recorded:

           Method of transport (helicopter, airplane, or ambulance)
           Burn Size
           Outside hospital estimate of burn size
           Actual burn size (as determined by our burn attending)
           Burn Mechanism
           Fluid resuscitation data including fluid type, rate, and bolus administration. 
           Intubation status
           Length of stay (both ICU and total hospital) 
          The difference between our estimate of burn size and outside hospitals' estimates
          Trends in fluid resuscitation rates
          Trends in air versus ground transport
       

• As this is a retrospective study we are only identifying trends rather than making sophisticated inferences, so the majority of our findings are simple declarative statements such as "Of 143 patients who arrived by air, 18 (13%) and 49 (31%) were discharged from the hospital within 24 and 48 hours, respectively." However, there are a few areas that I would appreciate your input on:
• Is there a good way to represent the relationship between overestimation of TBSA and overresuscitation? What is the best way to see which demographic factors (age, TBSA, mechanism) affect the likelihood of air vs. ground transport? Similarly, what is the best way to see which demographic factors (age, TBSA, mechanism) affect the likelihood intubation? - suggested boxplots of delta(TBSA) for each overresuscitation group
• Related to the above, what information will we glean from a chi square test that we will not get from a logistic regression, and vice versa? Would it be worth it to perform both? - logistic regression allows adjustment for confounders, and gives direction and magnitude of association. Chi square only gives association and does not account for confounding at all.
• I want to doublecheck the validity (and utility) of making certain statements without controlling for other variables, e.g. "18% of patients who were burned while smoking on O2 died, while all other mechanisms combined had a mortality rate of 4%" - definitely need to adjust for confounders in this case, since patients with this burn mechanism will have inherent differences from overall population. Use logistic regression if you have certain death data on everyone, or if you only have (for example) in-hospital death, could use Cox proportional hazards regression and censor at hospital discharge.

2016 February 10

Jin Han, Emergency Medicine

• "I have a cohort of delirious and non-delirious patients (230 patients). I want to preliminarily develop novel subtypes of delirium based upon clinical and biomarker data."
• Have data on several delirium characteristics (severity, arousal level, etiology) on 228 patients. Have functional outcomes at 6m on ~160, cognitive outcomes on ~110, and mortality rate of ~30%. Interested in risk prediction score for outcomes using delirium characteristics (whether or not patient meets criteria for full delirium) as well as patient characteristics.
• Planning to submit R01 to develop the full risk score. Suggested VICTR design studio with clinical + statistical experts to figure out how best to use this pilot data in grant submission.

Justin Godown, Pediatric Cardiology

• 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.
• Main goal would be to develop a risk prediction score for mortality, with VAD vs. medical management as a key component
• Data comes from two databases with a wide variety of cardiac patients; suggest limiting patients included to those who are sick enough where this decision would have to be made.
• Could do a Cox regression model with time to death = baseline factors + VAD vs. medical management; not sure about getting a risk probability from this, though.
• Could also do a logistic regression model with, say, one-year mortality as outcome; more straightforward to get a probability, but lose the time information.
• Propensity scores could be useful here to either match VAD patients with medically managed patients with similar propensity of VAD, or as a data reduction technique if number of events is low. It's possible that neither of these are necessary.

2016 February 3

Jin Han, Emergency Medicine

• "I have a cohort of delirious and non-delirious patients (230 patients). I want to preliminarily develop novel subtypes of delirium based upon clinical and biomarker data."

Kiersten Brown Espaillat, Emergency Medicine

• "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 neuro care unit."
• Wants to assess validity of a VUMC-developed bedside swallow assessment compared to video fluoroscopy among stroke patients who can safely have swallow assessment. Have retrospective data on patients who failed swallow screen (thus required fluoroscopy), but no fluoroscopies currently on patients who passed swallow screen.
• Sample size needed will be determined by confidence interval width that's clinically meaningful - for example, if point estimate is 90% sensitivity but CI goes down to 75%, is that clinically OK, or is that too low? Kiersten will look up validation study for only validated tool and use that as a starting point.
• Swallow tool failure rate is ~15%.
• VICTR could be a good resource.

2016 January 13

Mike LeCompte, Surgery and Critical Care

• Mike wants further input on his application to VICTR regarding a surgical resident education project.