Biostatistics applications in surgery, anesthesiology, and emergency and critical care medicine Clinic Notes (2017)

2017 December 20

Rachel Labianca, Pharmacy resident

  • "I previously attended biostatistics clinic in August in preparation for VICTR grant application for my residency research project on time to antibiotics and open fracture trauma patients. I am in the process of collecting data during December, and would like to attend clinic again to ensure that I am formatting the data most appropriately to be ready for statistical analysis. I also would like to determine more specific statistical endpoints to guide my data collection now that I will a definitive number of patients and event rates."
  • Primary exposure is amount of time between admission and administration of antibiotics; recommend analyzing this as a continuous exposure vs dichotomizing (<60 vs >=60 minutes), because you lose information and power when you dichotomize (but guidelines state 60 minutes)
  • Estimated total sample size 230-240 with ~12% infection rate; some patients are missing exposure, and these are more likely to come from outside hospitals
  • Some patients die prior to full followup/opportunity to have infection; could consider doing a Cox model with competing risks (outcome = time to infection, censored = survived and never had infection, competing risk = death); collect dates of death and infection to allow for this
  • Resources on using spreadsheets for sharing data: Broman & Woo ( ); "spreadsheets from heaven/hell" (on Dan Byrne's web page: )

Gary Owen, Pharmacy resident

  • "Follow up from 11/29 re: assessing pain, agitation, and delirium practices in an international cohort. Would like to solidify a plan for VICTR application."
  • Important limitation/confounder to note with delirium outcome: If delirium screening went up between 2010 and 2016, it's possible that delirium rates could increase due to more screening, rather than a clinically meaningful increase. Look at rates of missingness for this outcome as well as rates of delirium itself.
  • Can look at delirium as a secondary outcome (duration, -free days), but make sure to look at coma and mortality as well to get the whole picture
  • If sticking to descriptive statistics and multinomial logistic regression, this project should fit within 90 hours for a VICTR voucher (adding multistate model would add time to this, if VICTR statistician has not done this before)

Jackson Cabo, medical student

  • "I am doing a project with Dr. Bailey (Surgical Oncology) regarding healthcare disparities in colorectal cancer and how they may vary depending on setting of care (i.e. type of treating hospital). A Cox Proportional Hazard Regression will be used to estimate the effects of race, income status, and insurance status in the context of hospital facility (type) on overall survival. We are looking for assistance in generation of this regression model."
  • Five types of treatment facility; data from 2004-2014, with one record per patient
  • Primary question: are social determinants still independent predictors of survival after taking type of treatment facility (academic, community, etc) into account?
  • Could take fixed or random effect approach (do not do separate models for each facility type); which to use depends on whether you want to make inferences about facility type or merely control for variability/allow different baseline hazards, as well as sample size (need more events [deaths] to reliably estimate fixed parameters for facility types; not yet sure what mortality rate is) (statisticians lean toward fixed effect in this case)
  • Data is already restricted to patients with specific type of colorectal cancer and have adequate followup information; next step is to come up with a list of potential confounders/risk factors, in order of importance
  • Funding is available via medical school research program; suggest talking to Dr Shyr about how to arrange short-term collaboration (maybe with Sam Nwosu?), if possible; if that doesn't work out, could apply for a VICTR voucher, and clinic statisticians believe this would fall within the 90-hour category

2017 December 13

Timothy Olszewski, Research dietitian

  • "Dr. Heidi J Silver is the PI on an IRB-approved retrospective study on sex differences in sarcopenia surgical outcomes. We would like to discuss data that we have analyzed."
  • The cohort of interest includes subjects with surgery for benign tumors or malignant tumors. The analysis will separate out these two groups. If someone had a benign surgery and also a malignant surgery, a covariate indicating prior surgery may be in order. Or only the first surgery should be included. The number of subjects who fall into this category need to be determined. For subjects who had >= 2 surgeries in one cohort or the other (ie., benign or malignant), only the first surgery should be considered.
  • Several outcomes are of interest -- dichotomous ones such as ED revisit, rehospitalization, death will be modeled using logistic regression; time to event outcomes will be modeled using Cox proportional hazard models, provided the proportional hazard model assumptions are met.
  • For all models, clinical knowledge should guide the choice of potential confounders to include in the model.
  • This project should fit in a standard 90-hour VICTR voucher. approved.

Carsen Petersen, Registered dietitian, MS student

  • "I have completed data collection and now need to start data analysis. Dr. Heidi Silver is my mentor on the IRB-approved retrospective study. We are examining the agreement between common prediction equations for estimation of resting energy expenditure (REE) and measured REE using the KORR ReeVue indirect calorimeter in free-living obese adults. Statistical analyses will be completed to understand if age, sex, race, or obesity grade influences the agreement."
  • The first question of interest is the agreement between the measured REE and several already established predictions of calories expended in a day. We discussed the use of Bland-Altman plots to illustrate agreement as well as using different colored points to illustrated the relationship for different sub-groups in their cohort (e.g., men/women, race).
  • In order to assess the relationship between the different factors of interest (e.g., BMI) and measured REE, a linear model with mREE as the outcome and BMI fit with restricted cubic splines could be fit (adjusting for any potential confounders). Plots of mREE by BMI would best illustrate this relationship.
  • This project should fit in a standard 90-hour VICTR voucher. approved.

2017 December 6

Kaitlyn Works, Emergency Medicine

  • "QI for the waiting room: gauging patient expectations before and after implementation of an informative video."
  • Make answer options as clear as possible (eg, if patient would say they expect to wait 2 hours, make options clear about where 2 hours would go)
  • Timing of survey: watch video and take survey after triage; need to think about how to get survey to patients who might go immediately from triage to treatment room, or come in via ambulance
  • Discussed specific changes to survey made during clinic
  • Possibility: X months of baseline data collection; X months of "passive" intervention (play video at intervals in waiting room, patient treatment rooms); X months of "active" intervention (sending patients to a specific room to watch video after triage, eg)

2017 November 29

Gary Owen, Pharmacy/Critical Care Resident

  • "We are using a large, international database of critically ill patients (collected from the international study of mechanical ventilation) to study how clinical practice compares to pain, agitation, delirium guideline recommendations. Our objectives, planned analyses, and questions are below:
  • (1) Objective: Describe current practice and changes between 2010 and 2016 (before and after 2013 guidelines) Key variables (measured daily): RASS scores, amount of sedation/analgesia, daily spontaneous awakenings, occurrence of delirium Analysis: Descriptive stats for years 2010 and 2016, pertaining to amount of analgesia/sedation, depth of sedation, spontaneous awakening, along with outcome of delirium. May compare groups before/after guideline implementation in 2013, as well." Question: options for describing adherence (e.g. how to quantify daily spontaneous awakenings for the group - average, frequency, per pt, etc.). dealing with missing data considering any delirium during admission or day-to-day delirium
  • (2) Objective: Determining which aspects of clinical practice are associated with delirium Key Variables: daily measures as above; facility, region/nation; pt specific factors Analysis: Multivariate logistic regression of delirium vs available covariates. Question: how to consider presence of delirium - any delirium during admission vs daily risk
  • (3) Objective: Identifying factors associated with deviations from guidelines. Key variables: year, location, patient specific Analysis: Multivariate analysis. Question: How to quantify adherence (all or none during admission, day to day, each component of guidelines)."
  • Prospective data collection for a month at each site, first intubation only for all intubated patients
  • First goal: describe changes in clinical practice. Can do this well for sedation/analgesia practices; RASS was only collected in 2016 so can't really look at change for this. SATs may be mandated as part of the study protocol, so much harder to determine whether usual care changed between the two time points (since these practices are prescribed by the study). For medications, most reasonable to describe daily doses (not totals over intubation).
  • Delirium is collected as daily yes/no (no ICDSC/CAM-ICU available). 2016 data has about 25,000 patients (can have >1 day each); if logistic regression is used, limiting sample size = minimum of [delirium, no delirium]; divide this by 10-20 to get number of covariates you can include reliably.
  • Fixed vs random effects: Use a fixed effect if you want to make inferences about a covariate (eg, relationship between age and delirium); use random effects if you want to account for variability (eg, between study sites) without making inferences.
  • Multinomial regression may be a better choice than logistic regression - this looks like separating normal vs delirium vs coma, instead of delirium vs [anything else]
  • Data are clustered by both patient and site; it'll be important to adjust for this (could use sandwich estimation, bootstrapping, and/or random effects)
  • Key questions for next clinic: 1) What descriptives are both most interesting and most available/reliable in dataset? 2) Do you really want to look at adherence risk factors? - this will take lots of time due to need to define "adherence," could be confounded, etc; 3) Logistic (delirium vs anything else) or multinomial (n/d/c, but more complicated) for delirium outcome

2017 November 15

Bradley Kook, Obstetric Anesthesia Fellow

  • "We aim to perform a retrospective chart review assessing variability in timing of nurse administered PRN opioids in post-cesarean patients."
  • Lots of value in a descriptive analysis here, once you have ability to classify patients together
  • Could look at intraclass correlation between nurses and patients...?
  • Can easily get lots of data from EMR
  • Would be interesting to look at years of nursing experience, possibly
  • Defining outcomes is hard
  • Planning to get a VICTR voucher; may come back to help refine outcomes and analysis plan once more data is collected

2017 November 8

Andrew Perez, Medical Student

  • "We are looking at rates of complication-related port removal when patients are neutropenic vs normopenic at the time of port placement."
  • Data already collected on cases with clinically defined neutropenia, including some demographics, whether port was removed due to infection, whether 60-day followup obtained, and reason for incomplete followup (including death)
  • Main question today: how many controls to collect data on (SD is not tenable for this situation - no ready-made variable in SD that says "port removed due to infection")
  • No data available currently on controls other than year of port placement (ie, can't stratify on much); make sure that cases and controls represent roughly the same time frames
  • Other, preferred option if feasible: Collect data on all controls, so you don't have to choose a random sample, and use neutrophil as a continuous exposure as opposed to a case/control dichotomous variable
  • To help with feasibility, could restrict data (exposure, outcome, confounders) to, say, last five years; risk losing power because you lose events, but you're still gaining some by including neutrophils as continuous; probably still net gain
  • Clinical messaging: think about a figure showing neutrophil value on the X axis and probability of port removal on the Y axis - sends a concise message, whatever the relationship ends up being (recommend modeling with restricted cubic splines)

2017 November 1

Rachel Coleman, Endocrinology Fellow

  • "Review of patients behavior regarding their insulin pumps, specifically looking at how their bolus behaviors (for example: how manual bolus events and bolus wizard events affect their a1c). Have data for 224 pump downloads (two weeks of data for each of 224 patients) in a de-identified excel file as well as R file. I would like to discuss which statistical analysis is best and how to run the statistical analysis."
  • Key question is how the manual vs wizard users are different - manual users are probably very different, have been using the pump and doing manual calculations for a long time
  • If more a1c measurements were available, probably want to do some kind of longitudinal analysis - a1c (measured every 3 or 6 months) over time vs manual vs wizard, or a continually varying measure of override vs wizard, since patients can do both
  • If stick to single download per patient that corresponds to one measured a1c, maybe look at % of overrides over previous two weeks vs a1c at that visit; if you stick with the two groups, try various levels of cutoffs vs sticking with only 90%
  • Important to take potential confounders into account: there may be factors that affect both whether the patient decides to override the wizard and the a1c (eg, patients with disease for longer may be better at managing their a1c and more likely to manually override at times). Examples - time since diagnosis, comorbidities, age, etc

Shayan Rakhit, Medical student

  • "The Sequential Organ Failure Assessment (SOFA) score, is commonly used to dynamically evaluate a patientís severity of illness over the course of their ICU stay and contains six components measuring six organ systems. The neurologic SOFA component is the Glasgow Coma Scale (GCS), but the GCS has high inter-rater variability and is not routinely collected in ICUs. For this reason, Vasilevskis et al published (at Vanderbilt in 2016) a validation of a SOFA score utilizing the Richmond Agitation Sedation Scale (RASS), a much more reliable and reported measure of consciousness. Specifically, construct validity was determined with correlation with regular SOFA score and predictive validity was determined with regards to mortality (compared to regular SOFA score).
  • We have access to Respira/4th ISMV, a multi-center cohort across 42 countries. Our aim is to 1) validate (as done previously) the modified SOFA score utilizing RASS against the original SOFA utilizing GCS in this larger, more diverse population; and 2) evaluate if specific patient characteristics, such as direct neurologic injury, and practices, such as light sedation, affect the validity of a modified RASS utilizing RASS."
  • High rates of missingness, which are probably informative - patients without RASS/GCS are likely to be different
  • Recommend adjusting for site (or country, if site is untenable), in main or sensitivity analysis
  • Data is pretty clean
  • This project should fit in a standard 90-hour VICTR voucher. approved.

2017 October 18

Paula Smith, Surgery

*" I am using a large multi-institutional retrospective database to look at the relationship between insurance status and oncologic outcomes in Gastrointestinal Neuroendocrine Tumors. I have dome some preliminary work with this data but would like assistance from a biostatistician gaining more sophisticated understanding of my data."
  • Outcomes of interest include time to mortality and time to disease relapse.
  • Given that multiple sites are involved, fitting Cox models stratified by site would be most appropriate.
  • For the disease relapse outcome, we discussed treating death as a competing risk and calculating cumulative incidence.
  • She would like to apply for a VICTR voucher; the scope of work, along with any manuscript revisions, will fit easily into the 90 hour time frame.

Tanya Marvi, Medical student

*"We wrote a paper looking at factors associated with increased blood loss in pediatric scoliosis surgery. We are looking to address some fo the reviews from the journal regarding our statistical analysis as we prepare to resubmit the manuscript."
  • The primary concerns were lack of details regarding how the model was determined as well as whether any transformation of the data were necessary to meet linear model assumptions. Step-wise methods were originally used; we discussed how this is not the best approach and how to go back and redo the analysis using clinical knowledge/literature to determine which covariates to include. We also discussed how to check if model assumptions are met and if not how to address.

2017 October 11

Konrad Sarosiek, Plastic Surgery -- Cancelled

  • "We have a large data set of ~60,000 patients who underwent different surgical procedure combinations under the guise of Ďmommy makeoverí and we are looking to see if there is added risk when combining procedures. We are looking to find relative risk & to isolate frequent complications & identify risk factors."

Andrew Perez, Medical Student

  • "We are comparing port removal rates in patients who were neutropenic vs normopenic at the time of port placement. We are looking for a statistician to help us analyze data and I was told this clinic was the place to start getting help with that."
  • 3500 patients with ports placed; patients excluded if there was no neutrophil measurement within two weeks of port placement
  • 184 neutropenic patients out of ~3500; hypothesis is that these patients have higher rates of port removal than normopenic patients
  • Ports could get removed for different reasons
  • Recommend talking to synthetic derivative team to see if they can quickly extract data on all these patients, to save Andy from having to build dataset himself; drawback = deidentified, so no going back and getting extra info later
  • First step: come up with a detailed list of fields you'd want from the EMR/SD - potential confounders, reason for port removal, list of infections, anything you can think of
  • Take that to SD team and see how feasible it would be to get that data
  • When it's time for data collection + analysis, recommend applying for VICTR biostatistics voucher; this should fit within the 90-hour project time frame
  • Recommend data collection in REDCap, pending discussions with IRB

2017 October 4

Breanna Thomas, Meharry

  • Looking at relationship between subjects being of multiple minorities (LGB + racial minority) and anxiety, depression and substance abuse
  • Parent study is longitudinal, plan is to look at a cross section; need to make sure you know how the data (from UM) was subsetted and sent
  • Data is from parent study of 18-59-year-olds; detailed codebook and data info are available online
  • Anxiety, depression, and substance abuse are currently coded in data as yes/no variables with 45-55% prevalence (based on quick glance); make sure you know exactly how these are determined (self-report, questionnaire...?)
  • Potential confounders: age (leave as continuous if possible! - looks like it is categorized in data; ask if raw data is available); employment status; possibly others
  • Come up with complete list of confounders and prioritize them in order of importance

Justin Shinn, Otolaryngology

  • "Data is now collected regarding neck cancer metastasis in those with smaller tongue cancers. Want to compare retrospective group in those who recurred in the neck to those who did not based on pathology results."
  • 15 years of data with about 75 patients with T1/T2 tongue cancers with some followup of neck observation
  • Have some followup out to 60 months; main time period of interest is two years
  • Main outcome of interest is recurrence (particularly in neck); 10 patients died within 5 years, but most recurred prior to death (two died without recurrence)
  • Idea is that doing more neck dissections in certain patients could help prevent recurrence; no one in database had a neck dissection
  • Recommend time to event analyses with outcome = time to recurrence; patients who never have recurrence will be censored at end of followup period (two years?) or last contact
  • With above approach, patients who died without recurrence would be considered under competing risks, but with small sample size and very few of those patients, probably not worth worrying about
  • Multivariable approach = Cox proportional hazards model
  • 36 recurred out of 74 patients
  • Main exposure = tumor depth, measured in mm
  • Possible confounders: margin (positive - did they get it all? - and/or close); T1 vs T2 (use tumor size instead of categorization? - but only 7 patients in this cohort had T2)
  • Write analysis plan a priori and stick to that (can include secondary analyses, but don't push your data too hard - with 36 events, it will be hard to tell much in detail)
  • More data would be available looking at patients with neck dissections, but would be hard/long to get; getting that would allow you to say "what are the chances of recurrence if I do vs don't have the dissection, assuming all other factors [tumor size, etc] are the same?"
  • Software: SPSS is most user-friendly - if you use it, make sure you turn on a log so you can reproduce results; if you go with Stata, UCLA has good examples/docs ( )

2017 September 27

Jeffrey Weiner, Pediatric Cardiology

  • "I am evaluating a database with clinical risk factors for post-operative thrombosis in congenital heart surgery. I am evaluating known risk factors (age, weight, severity of disease/surgery, cardiopulmonary bypass time) with genetic data (genotypes for known prothrombotic SNPís) to see if I can create a novel risk prediction model. I am having trouble (mostly software related as I am new at this), and would love a biostatisticianís insight."
  • Data on ~1000 patients, with 11% prevalence of thrombosis
  • Among other predictors, genotype is of particular interest - 7 SNPs, each with three subtypes; different SNPs have variable prevalence rates
  • Current approach: lump all SNPs into a single yes/no variable; covariates are age, weight, surgery, genotype
  • Planning to get VICTR voucher; data is in REDCap
  • Suggest looking at relationships between the genotypes: if someone with gene X always has gene Y, including both in model can be problematic; also, if very few patients have
  • Complexity of the model will be limited by minimum of (events, non-events) - in this case, roughly ~110 patients have thrombosis, so can fit 10-11 degrees of freedom max ("df" is kind of like a variable, but not exactly)
  • Clotting rarely happens in the first few days after surgery; for this reason, could consider time to event model (Cox) where event is time to clot, but would ideally account for a) time-varying covariates (severity of illness, etc) and b) competing risks (if patient dies before having a clot)
  • Abstract deadline in two weeks; for that, recommend current logistic regression model among all patients and only among hospital survivors. Hopefully those results are similar, but if not, emphasize results among survivors, because mortality could be a big source of bias and confounding in this study.
  • We believe this project would fit into the 90-hour VICTR voucher category.

2017 September 20

Pooja Santapuram, Hearing & Speech Sciences

  • "The purpose of this study is to examine the relation between language development and eye gaze patterns to audiovisual speech specifically in infants at risk for autism spectrum disorders (ASD). ASD is a developmental disorder characterized by social and communication deficits in addition to repetitive and restricted behaviors. It is known that infants at 1 year of age who later go on to be diagnosed with autism look at individualís faces less frequently (Osterling et al., 2002) and that toddlers (18-24 months) later diagnosed with ASD use fewer vocalizations with speech sounds and greater ďatypical vocalizationsĒ when compared to typically developing (TD) toddlers (Plumb & Wetherby, 2013). Yet, ASD is typically not reliably diagnosed until 2-3 years of age. Therefore, characterization of eye gaze patterns to audiovisual speech and vocalizations in high-risk infants may facilitate earlier identification of ASD and may also allow for future studies on potential treatments in this clinical population. Questions Iíd like to address or basically how best to approach an analytic plan for this study."
  • For first project, recommend calculating sample size for correlation statistic using precision (confidence interval width). Use Spearman correlation (nonparametric; does not assume that variables are normally distributed).
  • Linear regression with skewed variables: 1) Can run model and check assumptions - not interested so much in individual variables as in whether overall model fits well and meets assumptions. Try RP plots, QQ plots. 2) If assumptions are not met, can try transforming individual variables to improve overall model fit; also recommend using spline terms (or polynomials, if restricted to SPSS). This allows associations to not be straight lines, which is usually more accurate.

Joshua Arenth, Pediatric Critical Care

  • "Follow up session regarding best approach to log data into redcap for analysis as discussed a previous clinic."
  • Previous clinic notes
  • Planning to reapply for expired VICTR voucher to analyze pilot data of provider communication intervention. Discussed a longitudinal REDCap database with one demographic form (filled out at session 1 only) and a questionnaire form, filled out at sessions 1, 2, and possibly 3 (for control group only).
  • Each questionnaire's score will be an integer, 0-11. Recommend a Wilcoxon test (nonparametric version of paired t-test).
  • Planning to pitch multicenter trial for this intervention; sample size will be determined once this data is analyzed.

2017 September 13

Brian Adkins, Pathology

  • "Allo-antibodies against red cell antigens in pregnant women lead to poor fetal outcomes. As such OB/GYNs follow serial antibody titers.Traditional tube titration in slow and subjective. Automated gel titration is available but testing requires further understanding before clinical implantation. We are trying to figure out sample size and number of tests we should be running to determine clinical cut offs for antibody levels."
  • Suggest weighted kappa to address agreement between level of titration that each sample method detected antibodies at.
  • Need to get data into a software-readable format; look at the "spreadsheet from heaven" example here.
  • Variables are all categorical (1/2/4/16/etc for levels, 0/1 for differences) so nonparametric tests will likely not help.
  • If VICTR voucher is requested, this will fit under the 90-hour limit.

Paula Smith, Surgery

  • "I have a data set I am trying to run some stats on using the Stata program and I have questions about the best tests to run and how to make my data set compatible with Stata."
  • General recommendations: Use a do file in Stata to save analysis approach; write analysis plan a priori to define cohort; keep continuous information as much as possible rather than categorizing (eg, if raw BMI data is available, use that rather than categorizing)
  • Planning to submit abstract in October for conference in April; may submit abstract based on analysis already done, and work with VICTR on multivariable regression/competing risks
  • Main research question: Do adrenaocortical carcinoma patients with more resection have better/worse survival and risk of recurrence after surgery?
  • Currently analysis does not adjust for confounders or account for competing risk of death in recurrence outcome
  • Currently: used descriptive stats, KM curves; can get logrank p-values for KM curves, but need to look at proportional hazards assumption (do the curves cross? - but look at this in context of how many patients are "left" when they do cross)
  • Possible future recommendations: multivariable Cox proportional hazards model adjusting for potential confounders (age, BMI, surgery type, etc); for recurrence model, may need to use competing risks
  • This would fit into a 90-hour VICTR voucher.

2017 September 6

Shriji Patel, Ophthalmology

  • "I am conducting an analysis of Medicare Part B Claims Data and would like assistance regarding which statistical methods would be helpful in identifying trends in claims data."
  • Medicare Part B only has five years of data, only summary data available. Might be able to look into time series, but not certain that those methods will be helpful. Recommend good descriptives/visuals.

Nick Dantzker, Orthopedics

  • "Study to establish which radiographic parameters correlate with functional outcome and patient satisfaction in operative distal radius fractures. Need assistance with model for intra/interobserver reliability of radiographic measurements and overall statistical model for project"
  • ~55 wrists (final total could be up to 165, but more likely to be ~60) with injury x-rays pre-op, post-op and long-term; have 7 radiographic parameters on each at three time points (VAS pain scores, radial inclination, etc), as well as three injury ratings (one per patient); some patients have both wrists included
  • Main questions: 1) how good is interrater agreement on measures, often in degrees, and 2) are these measures (at one or both time points) actually predictive of outcomes?
  • Because a few (~4) patients have both wrists included, recommend randomly selecting one wrist from each - otherwise, confounders and outcomes from those patients will be more correlated than outcomes from different patients
  • Interrater agreement: suggest Bland-Altman plots (kappas are for categorical measures) - you don't want to see a pattern (eg, differences in agreement based on true value) ( additional link)
  • Intrarater agreement: suggest repeating measurements on ~15 wrists
  • Could do multivariable regression: outcome = [confounders - age, injury type, etc] + any x-ray info that will always be present + [one measurement, eg radial inclination]; run separate model for each measurement of interest
  • Regression is limited due to sample size - if you fit too many things in your model, it will not be generalizable to any other study (won't replicate), and we anticipate about 60 patients
  • Type of regression model will depend on exact outcome you're looking at
  • Three time points - could include an interaction term, but would likely be underpowered. Could also look at each time point with a separate model.
  • How many models are we talking? 10 measurements; __ outcomes; three time points - lots of models
  • Cohort is limited to a select subset of patients who have all three followup time points, isolated injury, respond to followup question - eg, need to be careful about how you generalize this to a general ortho population (patients with complete followup will be different from patients who are lost to followup)
  • Look at demographics of patients who responded to survey vs those who didn't - these will likely be different, which could bias results
  • Suggest applying for VICTR voucher; this will be <90 hours (typical manuscript project)
  • Feel free to come back for input on REDCap database, further discussion

2017 August 30

Jennifer Watchmaker, Radiology

  • "I would like to perform I believe an ordinal logistic regression. I have outcome data and also a continuous variable. I would like to know if the continuous variable (obtained pre-procedure) predicts outcome. I would like to also gain a sense of what additional analysis I can do with my dataset. I have 300 procedures worth of data on redcap."
  • 300 procedures (multiple procedures per patient - about 175 unique patients) are being reviewed by radiologists, currently only one read each; recommend having at minimum a random sample of these reviewed by multiple readers to gauge interrater reliability
  • Main outcome is ordinal, ranging from no response (0) to full response (3)
  • Main exposure is NLR (neutrophil:lymphocyte ratio) - hypothesis is that patients with a higher NLR are less likely to have a treatment response
  • Collect info at time of procedure and two months post-procedure; so far have excluded patients who are lost to followup for any reason
  • Recommend a proportional odds logistic regression model due to ordinality of outcome; will need to adjust for the fact that there are multiple procedures per patient
  • Do *not* do univariate testing to determine what covariates to include in the model; rather, decide based on clinical knowledge/literature review what are important potential confounders. Rough estimate is that you could include 10-15 parameters in this model.
  • Examples on how to do ordinal logistic regression in R ( and Stata (; look into clustered sandwich estimation of variance to account for within-patient correlation
  • If not enough time to figure out clustered sandwich estimation before abstract submission, recommend using just first procedure per patient and applying for VICTR voucher
  • Additional reason to get VICTR voucher: likely that NLR has a nonlinear relationship with the outcome, which makes an accurate model more complicated to fit and interpret
  • If VICTR voucher is required, this project will require 90 hours or less

2017 August 23

Rachel LaBianca, Critical Care Pharmacy

  • "Studying open fracture orthopedic trauma patients to determine whether there is a difference in infection rates for patients receiving antibiotics within 60 minutes of presentation versus those receiving antibiotics >60 minutes from presentation. Would also like to conduct analysis to identify other factors impacting infection rate. Would like assistance with statistical design and help in determining whether VICTR application will be needed for this project."
  • They anticipate approximately 100-200 subjects in their analysis with about 20% infection rate. Their primary outcome is infection rate; therefore, they are somewhat limited on the complexity of the model they can fit without using data reduction techniques.
  • They have identified potential confounders and will put them in order of importance.
  • We discussed the use of splines for the time to antibiotic if that is their primary exposure of interest. We also discussed whether they could fit a model with type of antibiotic, categorized into 3 main categories as their primary exposure of interest.
  • They will return to clinic to discuss setting up the data to ensure a smooth transition to the analysis phase.
  • The analysis is fairly straight-forward although could be a bit more involved if some kind of data reduction technique is used, such as propensity scores. The analysis should easily be completed within 90 hours giving enough additional time for manuscript preparation/revision.

2017 July 26

Joshua Chew, Pediatric Cardiology

*"We are completing a retrospective study evaluating a new echocardiographic measure (pulmonary pulse transit time; pPTT) in pediatric pulmonary arterial hypertension (PAH). Our cohort includes roughly 20 PAH patients with 2:1 age/sex matched controls. Our initial analysis demonstrated a difference in pPTT between PAH patients and controls. We also saw an association between pPTT and a crude measure of right ventricular function. We are now performing follow-up measurements to obtain an objective measure of RV function (myocardial performance index; MPI). We would also like to explore the relationship of pPTT with hemodynamic data and clinical outcomes in PAH patients over time. The questions we would like assistance with are as follows: 1. What is the most appropriate approach to evaluate the relationship between pPTT and MPI, taking into account that we suspect it may not be linear? 2. Would appreciate recommendations on analysis plan for the longitudinal analysis of pPTT in PAH patients. What sorts of hemodynamic data/outcome measures are most appropriate? How do we deal different follow-up times and different times between echocardiograms? How do we account for patients being on different therapies during the follow-up time?"
  • We discussed the use of splines to relax the linearity assumption in any models that may be fit. In order to avoid over fitting with linear regression, we follow the rule of thumb of estimating 1 parameter for every 10-20 subjects.
  • We discussed the use of spaghetti plots to illustrate the trajectories of the pPTT over time in the 20 PAH patients, including using colors to indicate those with different therapies or who may have died.
  • As a potential secondary analysis in the 20 PAH patients, a mixed effects model with a random intercept could be used with pPTT as the outcome and time and type of therapy as the covariate, including an interaction. The number of classes of therapies will need to be discussed given the small sample size and the potential for over-fitting.

Rachel Sosland, Urology

  • "Urinary tract infections may affect as many as one third of patients undergoing intradetrusor onabotulinumtoxinA (BTX-A) injection for medication-refractory overactive bladder (OAB). We have retrospectively collected data on 70 patients undergoing intravesical botox injection in 2016 and seek to identify potentially modifiable risk factors for post-operative UTI in patients with non-neurogenic OAB. Several of these patients have undergone multiple injections. We would like to assess their risk for UTI over time and over multiple different injections. We are seeking statistical support to assist in determining the best way to analyze this data in the same patient over time with multiple injections."
  • We discussed several options to address their question of interest. One potential option is to use Poisson or negative binomial regression with the number of UTIs as the outcome for a given person adjusting for covariates of interest such as where the injection was received (OR or clinic), class of antibiotic received (number of classes will need to be discussed to avoid over fitting), and including the varying follow-up times as an offset.
  • We also discussed how to address the question of risk factors for multiple UTIs. Subsetting on those who had at least 1 UTI, a logistic regression with the outcome indicating whether the subject had > 1 UTI and adjusting for pre-determined covariates would potentially address this.

2017 July 19

Cyrus Adams, Surgery/Urology

  • "We are currently investigating patient factors (demographic and clinical) in a group of adult patients with congenital genitourinary disorders. We currently have a redcap database of ~150 patients who recently presented to the adult clinic meeting this criteria. We are interested analyzing patient demographic and clinical factors that may be associated with renal dysfunction at the time of presentation to the adult clinic (measured by decreased GFR and/or the presence of hydronephrosis or renal scarring)."
  • Second question: Renal dysfunction measured by GFR - typical cutoff is <60, but we recommend also analyzing with GFR as a continuous variable (allows you to keep all information and not make false dichotomies). Only 9 patients had renal dysfunction when categorized, so definitely recommend keeping that outcome continuous. This outcome is fairly normally distributed, which is helpful.
  • First question: predictors of being followed as pediatric patients. Outcome is determined by MDs via chart review. Have 58 Nos ("non-events") and 93 Yesses ("events"); can reliably fit about six degrees of freedom (roughly equal to six variables) in a logistic regression model.
  • Recommend doing graphs of descriptive statistics, overall and by pediatric followup status.
  • Prioritize potential covariates in order of importance/relevance; consider missingness when doing this (if a variable is clinically important but only measured in the hospital, eg education and health literacy, it is less helpful here).
  • Plan is to apply for VICTR voucher.

2017 May 17

Sara Nelson, Anesthesiology

  • "I visited about a month ago to talk about our drug cost project. I've fit a model, but have some concerns that the assumptions haven't been met. I would like to meet and talk about the current model and possible alternative options such as random/mixed regression. If possible, it would be great to pull up the analysis in R (I can bring my laptop)."
  • Collinearity: try Hmisc::redun() for a redundancy analysis, or varclus() from same package.
  • Might try a negative binomial model due to distribution of cost outcome. To incorporate random effects, might try lme4::glmer.nb()
  • Available variables: procedure code; attending anesthesiologist (248 unique); in-room provider (CNA/resident; 572 unique); surgeon; surgery start time; duration; ASA class; age; gender; case type (surgical specialty; 9 unique); institution (VUMC/MGH); cost; base relative value units (?); provider team (3500 unique)
  • No data available on specific medications used
  • Is duration a strong surrogate for complexity? Strong enough to leave out case type?
  • Possibility: cost ~ ASA class + age + gender + duration + institution + age*institution + ASA*institution, random effect = attending; adjusting for other variables (case type, etc) is likely an artifact of things like duration and will result in collinearity
  • Adjusting for attending would get at what is likely driving at least part of the cost, which is sedation choice, but no way to tease that out

2017 May 17

Tanya Marvi, SOM, Medical student

  • "My project is looking at platelet count in patients with musculoskeletal infection. Patients were categorized into local, disseminated, and complicated infection, and I used an ordinal logistic regression to see if we could predict how they would be categorized based on their day 1 platelet count. Additionally, I used rloess to look at the trend of the platelet counts among the different groups. I want to make sure I am interpreting the results correctly and see what other analysis I should consider."
  • The outcome of interest is a 3-level outcome describing severity of infection in patients who present to the ED and are subsequently admitted. One question of interest is whether there is an association of platelet count and type of infection diagnosed. A proportional odds model was fit with platelet count as the single covariate. We recommended fitting splines to the platelet count (3 knots should be fine, given the sample size), and clinically determining what potential confounders are of importance to include in the model. Proportional odds assumptions should also be checked.
  • Because subjects had differing lengths of stay in the hospital, a Cox model could also be fit with time to discharge as the outcome and adjusting for pre-selected confounders.
  • A second question of interest is whether time-varying platelet count is associated with type of infection. A proportional odds model can be fit with robust standard errors. There should be an option in Stata to request the robust estimates.

2017 May 10

Sara Nelson, Anesthesiology

  •  =We performed a study assessing the variation in anesthetic drug costs. We did so by creating multivariate linear regression models in R. We've received reviewer comments with various suggestions we would like to implement, such as combining two of the models. I believe this will involve creating a nested variable; however, I am struggling with getting this to work in R.=
  • Current models have a fixed effect for in-room provider (561 df). Suggest replacing this with a random effect, or possibly a nested random effect (attending -> in-room provider). R package nlme or lme4 (newer) might be good for this.
  • Might also look at interaction between case type and duration.
  • Make sure that costs are the same per patient regardless of insurance.
  • Make sure to check model assumptions, even after outcome is transformed.
  • For all model covariates, show a point estimate + CI for the effect on cost. For age especially (modeled with restricted cubic splines), would be good to also show a visual of age vs cost.
  • Potentially reframe paper as "potential predictors of cost" vs "how much variance in cost can we explain."

Bryan Hill, OB/GYN surgical fellow - CANCELLED

  •  =Reporting complications after surgery are important for quality improvement. Two methods of finding complications are: 1) administrative data from diagnosis codes and 2) key-word search from a manual chart review. We suspect the administrative reporting method, under-reports complications. Primary aim: determine sensitivity, specificity of the administrative method compared to the manual reporting method Secondary aim: Determine which risk factors are associated with having a complication. #1: Question for statisticians: would the best way to look at our secondary aim be to create a regression model with the outcome "complication" and variables age, body-mass index, setting (outpatient or inpatient), sling type, attending, anesthesia time, operation time, smoking history, diabetes? #2 Sling type is heavily dependent on attending (they like to chose a particular brand or type). How do we adjust our model for that?= 

2017 May 3

David Leverenz, Internal Medicine

  • "We have developed an educational podcast for our internal medicine residency program. We are studying the effects of this project through pre and post-intervention surveys. I would like assistance in the statistical comparison of pre and post-intervention survey results."
  • Emphasize descriptive statistics (pre vs post) over p-values. If p-values are needed, chi square tests for categorical variables and Wilcoxon/Mann-Whitney tests for percentages are useful. Also think about boxplots to show variability in data instead of a single summary statistic.
  • Include measures of variability (interquartile ranges) as well as summary statistics (median).
  • Make sure and address differences in response rate in pre vs post, and describe any differences in patient populations.

2017 April 26

Susan Smith, Critical Care Pharmacy

  • "Purpose of project is to examine the efficacy of a short versus long duration of antibiotics for the treatment of intraabdominal infections. We would like help with our binary logistic regression model."
  • This is a really complex analysis due to immortal time bias, confounding, etc. Option A would be a Cox model for time to treatment failure with primary exposure = daily antibiotic use. This is a complex model to fit for a non-statistician; suggest contacting VICTR to see about stats support for this.
  • Option B... maybe a Kaplan-Meier curve removing patients from N at risk as they go off antibiotics?

Jamie Robinson, Surgery

  • "I would like assistance with a regression analysis looking at factors that may affect survival after portoenterostomy for biliary atresia."
  • Data set has 48 patients with this rare condition; about 20 had the outcome of interest (transplant or death). This limits what can reasonably and reliably be put in a survival model.
  • Consider lag time between being placed on the transplant list and actual transplantation - would be good to describe this.
  • When fitting model, use a Cox proportional hazards model (coxph in R's survival package; outcome will be created with the Surv() function; look at vignettes and/or look for UCLA tutorials). Descriptives and qualitative info will be helpful with a small population.
  • Choose a common followup time - maybe five years, two years? Look at minimum/maximum followup time to determine.

2017 April 12

Nishant Ganesh Kumar, Plastic Surgery/Medical School

  • "Would like to conduct a multi-regression analysis of opiate use and hospital length of stay against other variables being studied in an Enhanced Recovery after Surgery protocol for microsurgical reconstruction."
  • Primary outcomes are hospital length of stay and total opioid use. Hospital LOS has a very skewed distribution; original analysis used linear regression. We recommend checking RP plots and, if assumptions are not met, using either ordinal logistic regression or a Cox proportional hazards model with time to hospital discharge as the outcome.

2017 April 5

Ashley McCallister, Pharmacy

  • "My research project is in the NICU on Vitamin A use. I need help identifying what types of statistical tests should be run on the data."
  • Primary outcome is BPD (yes/no); secondary outcomes are discharge on oxygenation and days on the ventilator
  • Currently patients who died in the NICU are excluded; this will present severe limitations due to confounding, but without statistical support it is complicated to account for death when including all patients
  • For dichotomous variables, can use chi square test (vitamin A exposure vs BPD, eg). For days on the ventilator, use a Wilcoxon rank sum test (like a t-test, but does not assume normality)

Ida Aka, Clinical Pharmacology

  • "I need help with my sample size calculation for my PPI and SSRI projects. Both projects are looking at CYP2C19 *2 and *17 variants."
  • Extended discussion on how sample size/power can vary depending on genotype proportions in the sample; will need to investigate distributions of both genotypes and outcomes to decide how many patients are feasible to genotype and what kind of tests to use in eventual analysis

2017 March 23

Kristy Broman, Surgery - No Show

  • "The question I am trying to answer is whether there is a way to compare two incidence ratio. I am using the SEER database and SEER Stat which has built in modules for calculating age standardized incidence ratio for specific events. The output I get is the total N, the total event number, and the standardized incidence ratio. This is essentially the ratio of observed to expected, but I cannot know how the expected is determined (this is a "black box" within the module. So I want to know if there is a way to essentially compare the already calculated standardized incidence ratios."

2017 March 15

Leslie Fowler, Anesthesiology

  • "Prior to developing a Residents as Teachers curriculum within our department, Dr. Robertson and I sought to gain insight into the teaching perspectives of our residents by administering the Teaching Perspective Inventory (TPI). We administered a follow-up survey to gather information regarding dominant and recessive teaching perspectives."
  • "Our manuscript was accepted for publication with revisions. One reviewer notes indicated we should consult a statistician to see if raw data can be used for other statistical analysis as well as a T test. Another reviewer comments stated to consider researching a theoretical framework to base the research design. Should we conduct a T Test with data we collected? Is that the most appropriate? Can the raw data be used for other analysis?"
  • An already validated survey was used to evaluate teaching mode preference in 2nd, 3rd, and 4th year residents. This validated survey converts the raw scores to weighted scores in each of five different teaching modality preferences. Frequencies of primary teaching modalities were computed according to whether the resident planned an academic or private practice career. In order to assess whether there was a difference in the distributions of the frequencies of primary teaching modalities across the academic/private practice groups, we recommended a chi-square test. In addition, we recommended doing a Wilcoxon Rank Sum test using the raw scores across the two groups. Finally, in order to better visualize the distribution of the raw scores, we recommended creating box plots of the raw scores for each group.

2017 March 8

Amol Utrankar, Anesthesiology

  • "I am a medical student working with several members of Department of Anesthesiology on a project examining factors associated with in-encounter mortality among patients who are escalated to the intensive care unit following multiple rapid response team activation events, using a sample of 80 patients from 2016 VU rapid response data. I have several continuous and categorical variables of interest (Sepsis Related Organ Failure Assessment, organ dysfunction by system, age, gender, referring rapid response team, and hours elapsed between rapid response events. I would like to double-check my statistical methods with someone who has more experience in statistical analysis and Stata; I've been using chi-squared tests, Fisher's exact tests, and logistic regressions to assess associations, but want to make sure that I'm applying these tests correctly and setting up my variables properly."
  • There are about 87 subjects in their data with about 20 deaths. They are interested in exploring the association of different risk factors with mortality. We discussed the rule of thumb governing how complex of a logistic regression model could be fit (include roughly one covariate for every 10-20 deaths).
  • They have more covariates of interest to include in the model than degrees of freedom allowed without over fitting. Therefore, we discussed avoiding using univariate analyses to drive model selection. Rather, clinical knowledge and literature reviews should help govern what selecting the models of interest.
  • There are several complications that are of potential interest as covariates. One way of including all of them in the model is simply to create an indicator for whether any complication occurred or not or to sum them up and include the total number of complications.
  • We also discussed ways of displaying the data to help tell the story. One suggestion was to create boxplots and strip charts of the number of hours between the first call to the ICU team to when the patient was elevated to the ICU floor, stratifying by survival status. Points on the graph could be coded by shape and/or color to indicate sex or age or any other categorical variable of interest.
  • A potentially more complicated analysis that would account for variability in the different ICU teams' threshold for elevating a patient to ICU would be to fit the logistic regression model with a random effect. This may not converge due to the small number of ICU teams (~4-5).

2017 March 1

Susan Smith, Critical Care Pharmacy Resident

  • "This is a retrospective study examining the effects of neuomuscular blockers on time to abdominal closure in trauma patients undergoing damage control laparotomy managed with an open abdomen. I would like help determining what type of regression is most appropriate to answer two different questions regarding my data set: 1. Does neuromuscular blockade affect the time to abdominal closure following damage control laparotomy? 2. Does neuromuscular blockade affect the time to goal RASS? For the first question, at least one of the covariates is time-dependent. I also have a few specific questions regarding how to interpret the results form these analyses."
  • Recommend a Cox proportional hazards model for all outcomes (time to...). No need for time dependent covariates (all covariates are baseline).
  • Also recommend including patients who died before primary outcome (abdomen closure) - this will make results more generalizable to all patients who receive this procedure, vs. those who survive (which we can't know when a patient is admitted).
  • For help with SPSS, look for UCLA tutorials on Cox regression. Interpreting output

Brian Adkins, Pathology

  • "I am comparing rates of atopy in patients with allergic transfusion reactions. I need help calculating significance."
  • Strongly recommend collecting data for a control group (patients who received a transfusion and did not have an allergic reaction). Might be possible to do this using BioVU. With current data, you can describe the prevalence of allergies among patients who had a transfusion reaction, but can't draw any statistical conclusions about a difference in allergy rates between them and other transfusion patients.

2017 February 22

Sara Nelson, Anesthesiology

  • "We are looking to determine the effect of the pain consult service on mortality and morbidity in rib fracture patients. The protocol for the consult service was implemented in 2013. Our data is from 2010-2015, so I think there needs to be a before and after analysis utilizing matching. Mortality is the primary analysis, there are numerous secondary analyses--pneumonia, respiratory failure, 30-day vent free days, 30-day ICU free days, length of stay and tracheotomy."
  • 1152 patients seen by consult service after implementation; total data set has ~5000 patients, but not all data is available before protocol implementation
  • Raw mortality rates are 7% prior to implementation and 3% after; about 400 deaths in the data set
  • Recommend excluding patients seen after chest service began, but before official protocol implemented - too many unknowns and variables in this group to allow for clean conclusions
  • Main research question: are outcomes different among patients who met screening criteria, or would have met screening criteria, before and after implementation of the screening protocol? -> Need to exclude patients who never would have met screening criteria
  • Recommend Cox model for mortality, proportional odds logistic regression for other continuous outcomes, logistic for pneumonia, etc; limiting sample size is number of events (Cox model) or minimum of events/non-events (logistic)
  • Matching is probably the cleanest way to do this - match on age, number of rib fractures, ISS? Or match on propensity score: create model for propensity of being screened (among patients in post-implementation period) using data available, then use that model to calculate propensity score for all patients in pre- and post-implementation periods and match on that

2017 February 15

Elena Nedelcu, Pathology

  • I need assistance with choosing the right test to interpret correlation between variables and outcome and perform them
  • Data was collected on 324 liver transplant patients over three phases: baseline, practice changes, and post-implementation; main exposure is blood utilization, outcomes include LOS, mortality, discharge disposition
  • Potential for mixed effects model - want to account for surgeon
  • With multiple outcomes and end goal of manuscript, statisticians recommend 90-hour VICTR voucher ( )

2017 January 25

Laurie Tucker, Department of Pediatrics - Postponed

  • Follow-up to previous clinic visit.
  • Data looks great. There are a few additional pieces on the list that Laurie is trying to obtain from StarPanel. Notes from previous sessions are below.
  • Clinic statisticians estimate 90 hours for VICTR application.

Johnny Wei, Medical Student/Anesthesiology

  • "I am a 3rd-year medical student who is working with the Department of Anesthesiology on a project investigating demographic and clinical factors associated with post-operative opioid use. In short, we are looking at what factors (i.e. age, sex, type of surgery, etc) are associated with having an opioid or benzodiazepine prescription at various time points in the 12 months after a procedure. I have a rough idea of what types of figures I would like to create, and have already created the initial iterations of them. However, because I am a relative novice regarding biostatistics and using my analysis software (Stata), Iíd like to discuss my methodology and my analysis process and see if Iím doing anything inappropriate with my data management or analysis. Most of the tests I have been running are chi-square/Fisherís and logistic regressions, and I would appreciate advice on the appropriateness or the optimization of these tests. In short, Iím more interested in someone looking over the work and code I have done so far, and seeing if there are any major red flags in my methodology rather than coming up with the analysis plan itself (although advice on the latter would be very much appreciated)."
  • Recommend removing p-values from Table 1, un-collapsing outcome to regain information lost in categorizing, and using Kruskal-Wallis test instead of one-way ANOVA.
  • One resource for sample size determination:

2017 January 18

Niels Johnsen, Urologic Surgery

  • "We are working on a project that attempts to determine predictors of bladder rupture in patients following blunt-trauma pelvic fractures. A prior study was performed at an outside institution with similar (or intended to be similar) methods using a smaller cohort of patients. We chose all bladder rupture patients plus control pelvic fracture patients without rupture (4:1) and have the data on these patients. The hope is to identify clinically significant predictors of bladder rupture based on fracture configurations and then to devise a clinical prediction model to risk-stratify patients who present with pelvic fracture for having bladder ruptures. I have attached the previously published similar study that I'm referring to as a reference and will bring the deidentified database with me on Wednesday."
  • Motivating paper uses univariate variable selection and stepwise backwards selection to create the final model. We do not recommend either of these.
  • We do recommend choosing a pool of potential predictors based on clinical knowledge and available data, prioritizing based on potential clinical importance.
  • 140 bladder rupture cases (minimum event size)
  • Reference - Frank Harrell's Regression Modeling Strategies (chapters on predictive modeling and data reduction)
  • Planning to submit VICTR voucher when mechanisms are available again (check with Lesa Black); we estimate 90 hours to develop and validate the prediction model and prepare manuscript

Joel Musee, Department of Anesthesiology

  • "We have put together a study to examine whether a commonly used perioperative device (Lifebox), can be used to alert clinicians of hypoperfusion. The lifebox is pulse oximeter and measures oxygen saturation on extremities. The monitor has a graphical read out made up of 15 bars with more bars associated with a better signal, a proven surrogate for perfusion. We hypothesized that mean arterial pressures of 55 or less would not lead to a perfusion signal of 5/15 bars. The biggest questions is how to best analyze the data to test our hypothesis and also what kind of power we would need for a study like this."
  • Recommend not dichotomizing unless absolutely necessary - above scenario, for example, treats Lifebox measurement of 6 and 15 as exactly the same, which is likely not true
  • Data will be manually collected by staff looking at blood pressure and Lifebox at the same time
  • Likely repeated measurements on each patient (during preop, while administering anesthesia...)
  • Recommend collecting information on other variables - age, amount of sedation at a given measurement, etc; make sure some kind of patient identifier is present
  • Use longitudinal database in REDCap for data collection
  • Recommend fitting a spline (search "restricted cubic splines") - device might be more closely associated with MAP at certain points than others
  • Could be completely separate patient populations - one could be women undergoing C-sections; might be advisable in this case to do subgroup analyses?
  • Will probably apply for VICTR voucher - estimate 90 hours
  • Think about how many patients (in each group?) are feasible to enroll and how many time points could be measured
  • Next step would be to repeat the study in Kenya - possible that we'd see different results if MAP tends to be different, for example

2017 January 11

Chelsea Isom, General Surgery

  • Updates to project.
  • Chelsea has cleaned the data and done some preliminary analysis - project will now probably take ~60 hours to complete. See notes below for details.

Maie El-Sourady, Internal Medicine

  • "I am an attending physician on the Palliative Care consult service and I have been collecting data from the learners that rotate on our service for the past 4 years. They compete a pre-test and a post-test evaluating their comfort level with basic palliative care topics. I would like help doing the statistical analysis with this data. I have several cohorts (medical students, internal medicine residents, visiting fellows, etc) but the test is the same for all of them."
  • Recommended starting with boxplots of raw data to look at differences in overall distribution for pre and post scores on each individual question. Excel doesn't have a template for this - look into SPSS.
  • Possible further investigation could involve collapsing into categories using Cronbach's alpha or similar and looking at differences in pre/post by type of learner (resident/student/other).
  • Recommend asking if there is access to a biostats collaboration plan; if not, and if further analysis is needed, can apply for a VICTR voucher (probably 40 hours).

2017 January 4

Laurie Tucker, Department of Pediatrics - Postponed

  • Follow-up to previous clinic visit.
Topic revision: r1 - 18 Jan 2021, DalePlummer

This site is powered by FoswikiCopyright © 2013-2022 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback