Data and Analysis for Clinical and Health Research Clinic Notes (2021)

2021 December 09

Marianna LaNoue, VUSN

  • I have been wrestling with the issue of statistical modeling of ‘adverse childhood experiences’ (ACE’s) as predictors of health outcomes for many years. In short, since the work of the Felitti et al ACE’s group in the early 1990’s, many many hundreds of papers have been published using a ‘count of events’ as the primary predictor of some health outcome, (usually) all reported at the same time (i.e. self-report of current health status and recall of some set of childhood events occurs simultaneously). I find this approach somewhat puzzling and unscientific in that 1) there is no ‘true zero’ to use as a reference in any regression model (as subject could have adversity not captured in the list), 2) the set of candidate events captured by these surveys is arbitrary and different surveys capture different numbers of and specific sets of events and 3) there are certainly measurement issues related to the self-report nature of the data (i.e. bias from the potential confounding of current health status and recall of childhood events, among others). I need help thinking through alternative models for this type of data specifically…. (for example would a competing risk model work? Other strategies I may be missing?) (I published a short paper on this issue in 2020 ( VICTR Biostatistics voucher.
  • Concern about the bias of self-reported count of ACE’s. People from one region are more likely to experience certain events, therefore, a count of zero is not truly zero as reference. Suggest treat the count of events as restricted domain covariate conditional on region, population, etc and model it as quadratic effect. May use one independent variable to scale another independent variable.

  • If a rich data set is available (for exaple, 1000 events and 9000 non-events), it is possible to have a better score system giving different weight to each type of event. Test for equal weight using chunk test (include both individual event and sum of all events).Combine a few events to find co-occur events and present using tree diagram. Among patients with repeated measures, rank events by finding disagreement within patient.

Jennifer Richmond (Melinda Aldrich), Medicine, Division of Genetic Medicine

  • I am working on a career development award application (K99/R00) and am in the study design phase of the project. I am working to develop and pilot test an intervention aimed at equitably improving lung cancer screening uptake. Currently, the project has the following aims: 1) identify multilevel barriers and facilitators to the equitable implementation of lung cancer screening (K99 phase), 2) engage with community advisors and key stakeholders to identify multilevel implementation strategies (K99 phase), and 3) pilot test two tailored implementation strategies to equitably improve lung cancer screening rates (R00 phase). I am seeking input on Aim 2 and Aim 3. In Aim 2, I am proposing to use a conjoint analysis to help identify the value that key stakeholders (e.g., patients, providers, and health administrators) place on attributes of a potential lung cancer screening intervention. I would like input on the sample size and statistical approach for the conjoint analysis. In Aim 3, I am in the early phases of designing a small study to pilot test the intervention. For this aim, I am seeking input on the study design and power analysis to help me finalize the number of participants needed to recruit and budget for. Mentor confirmed.
  • Currently in the study design phase. There are too few people get screened for lung cancer, only ~6% eligible for it. Aim to develop strategy for patient and for provider as well.

  • First aim is feasibility and plan to use pre-post study design in one clinic to get pilot data to design future cluster randomized study. Pre-post design tends to contaminant the study outcomes and is only good for learning about variability (within patient, within physician). It is important to collect most key variables in pre-post design study.

  • In writing the study protocol, think about the burden of assessment of effectiveness, burden on patients, burden on providers comparing to cluster randomized design.

  • From stats point, this is not hypothesis generating but estimation, use margin of error (precision) for sample size estimation. ~70 patients to estimate SD for continuous variable. See Questions about things in BBR may be posted to under an appropriate topic name that begins with BBR (there is one topic for each chapter)

2021 December 02

Alex Cheng, Biomedical Informatics

  • Review of analysis plan and power calculation for two studies involving ResearchMatch participants. One study is to formulate a model for financial incentives based on a vignette of participant burden. Another is to determine whether allow patients to pull their data from Epic results in more complete ResearchMatch profiles.
  • ResearchMatch profile: Is it worth randomizing the complexity (size) of the data that must be extracted? Outcome is a count variable. Another option is to use the ranked difference test (manuscript ~1970).
  • Model for financial incentives: Proposed sample size is 150. Ordinal regression may also be an option. May try to do more than 150, and stop when desired CIs are hit.

2021 November 18

Meridith Balbach (Bapsi Chakravarthy), Vanderbilt University School of Medicine

  • As a subprotocol of BRE03103, we would like to explore the use of ctDNA in breast cancer. Specifically, we would like to investigate 1.) feasibility of measurement in serum samples and 2.) whether ctDNA detection post-chemotherapy levels may predict patients benefitting from radiation. Dr. Chakravarthy was approached by Natera to utilize their ctDNA detection method. We believe that this study would fill an important knowledge gap. We would specifically like help understanding power calculation & and suggested analyses. VICTR Biostatistics voucher. Mentor confirmed.
  • Industry funded study, in planning stage.
  • Q: can this marker predict who will benefit from radiation.
  • This is a pilot study. Two blood draws before and after treatment. Seeking power analysis. Would like to build on this later. Would like to use presence of cells as an indicator of the need for additional therapy.
  • Limitation. Disease return is a low information outcome. This is proof of concept. Can use dose response, evidence of non-flat curve. Change in number of cells, vs positive/negative. Change measure should be self-contained.
  • VICTR voucher is possible, and this would probably fit in that framework.
  • Collaborator contact is Tatsuki.

Ashruta Narapareddy (David Isaacs), Neurology

  • Tourette Syndrome is a multi-faceted neurodevelopmental disorder. Tics are associated with antecedent sensory phenomena called premonitory urge (PU). Interoception, implicated in PU, is the process of how the brain perceives and integrates bodily signals. The primary objective of our study is to collect data to identify a relationship between interoception and premonitory urge in TS adults. The main question that will be addressed is in regards to our method of statistical analysis: we have n=50 sample size and plan to construct a regression model that would end up having multiple independent variables that impact power. This is because the MAIA-2 scale we will use to measure interoception has 8 sub-scales; however, it has no composite power. Therefore, we were wondering what would be an efficient model to construct in which power would not be lost and we could construct a regression model with the 8 sub-scales in addition to 8 other covariates we plan to include in the model. Biostatistics voucher. Mentor confirmed.
  • Question: how to best create a model with n=~100 (50 per arm) and 16 variables. Advise against using multiple model. Select the correct model, and use that model. Report partial r squared. Data reduction/variable clustering may be a good option. Potential for a lot of over;lap. Could collapse into a score (e.g. principle components). Variable clustering program can be found in Hmisc R package. Remember the data may be bad at telling you the "best" predictor-can be a pitfall. Importance will require ranked CIs.
  • No voucher at this, may return to clinic if voucher is needed.

2021 November 11

Jessica Miller (Laveil Allen), Radiology

  • To understand the rate at which positive findings are identified in imaging studies ordered in the ED, and if specific imaging studies have a higher likelihood of being positive. Further, we will examine how the time to a final read on an imaging study varies with positive findings. Mentor confirmed. VICTR Biostatistics voucher.
  • IRB status is currently exempt, will need to be amended. Multiple data sources will be needed and there will be considerable data management needs.
  • Investigator has done some programming in MatLab to read the text radiology impression.
  • Initial plan is to calculate PPV, sensitivity, etc. But there are limitations to this approach. Another option would be to create a scale with the gradation of possible outcomes.
  • Could also examine tendency to order these tests (time of day, etc). Need to consider people who have tests ordered but not yet completed (death, Leave AMA). Mike Ward could be a good partner.
  • Overall goal is to examine use of resources, ordering habits, are they following national guidelines.
  • Could also consider using redcap data mart or clinic data pull tools.

2021 October 28

Sarah Osmundson, OB/Gyn

  • Therapeutic inertia (TI), or the failure to initiate or intensify therapy when appropriate to do so, has been associated with adverse outcomes in patients with type 2 diabetes. We hypothesize that TI is also associated with adverse outcomes for women with gestational diabetes (GDM). Our project has two aims: 1) To evaluate clinician and patient factors associated with therapeutic inertia during encounters for diabetes management in pregnancy and 2) To examine the association between therapeutic inertia and adverse perinatal outcomes among women with diabetes in pregnancy. We would like advice/support for the analytic plan we have drafted.
  • Must consider all the correlations in the data, and there will be plenty. Challenge will collecting all data with appropiate dates and times, and defining outcomes very carefully.

2021 October 21

Marianna LaNoue, School of Nursing

  • I have been wrestling with the issue of statistical modeling of ‘adverse childhood experiences’ (ACE’s) as predictors of health outcomes for many years. In short, since the work of the Felitti et al ACE’s group in the early 1990’s, many many hundreds of papers have been published using a ‘count of events’ as the primary predictor of some health outcome, (usually) all reported at the same time (i.e. self-report of current health status and recall of some set of childhood events occurs simultaneously). I find this approach somewhat puzzling and unscientific in that 1) there is no ‘true zero’ to use as a reference in any regression model (as subject could have adversity not captured in the list), 2) the set of candidate events captured by these surveys is arbitrary and different surveys capture different numbers of and specific sets of events and 3) there are certainly measurement issues related to the self-report nature of the data (i.e. bias from the potential confounding of current health status and recall of childhood events, among others). I need help thinking through alternative models for this type of data specifically…. (for example would a competing risk model work? Other strategies I may be missing?) (I published a short paper on this issue in 2020 ( VICTR Biostatistics voucher.
  • Inference we will be concerned about unmeasured variables, prediction, we don't care so much. These are explanatory models. Problems with this measure are multiple. Non random error may not matter too much if you don't care about the betas.

2021 October 7

Mary Cella (Jennifer Beavers, Susan Hamblin), Pharmacy

  • Our project is titled “Safety of NSAID Use in Traumatic Brain Injury.” It is a retrospective study that will have a matched cohort. Our study question is "Does NSAID use within 14 days of injury in patients with TBIs lead to significant increase in intracranial bleed compared to patients who did not receive NSAIDs? Data collection will be accomplished via chart review, Health IT data pull, TRACs, and Neurotrauma Registry (IT data pull points submitted to be collected). VICTR Biostatistics voucher. Mentor confirmed.
  • Difficulty looking at drug choice retrospectively, an option could be examining who actually gets NSAIDS. "Confouding by indication" is a concern here as well. Best to understand all variables that are needed, not just what happens to be available. Timing is also important for medication variables. Could examine time to start of medication. Suggest survey of providers for conditions that lead to admin of NSAIDS. VICTR voucher is desired, and probably suitable for any of the discussed pathways. Suggested a return visit to discuss data collection.

RESCHEDULED: Drew Johnson (Robert Tunney), Pharmacy

  • This is an epidemiological, observational study to establish the prevalence of iron-deficiency in hospitalized advanced heart failure patients with elevated ferritin as assessed by reticulocyte hemoglobin equivalent. Hospitalized advanced heart failure patients who are being evaluated for or planned to undergo cardiac surgery will be followed via chart review to identify individuals with elevated ferritin (>300 ng/mL). Patients will be followed via chart review to identify the administration of intravenous iron and/or completion of cardiac surgery per the primary team. Final reticulocyte hemoglobin equivalent, ferritin, and transferrin saturation will be subsequently evaluated prospectively as a study procedure. We would like assistance in developing a statistical analysis plan as we hope to conduct a multivariate regression analysis. Mentor confirmed.

Kayvon Sharif (Michael Topf), School of Medicine

  • The primary objective of this study is to demonstrate the feasibility of incorporating 3D scanning/graphics technology into the surgical pathology intraoperative workflow, and communicating 3D frozen section results to the surgeon in real-time. The secondary objective is to solicit clinical team member attitudes regarding the relative efficiency of this system via electronic surveys. We are doing a pilot study introducing a new 3D imaging modality into the intraoperative workflow in frozen section pathologic analysis during head and neck cancer surgery. For outcomes we are collecting a qualitative survey [attached] from members of the surgical and pathology teams. We are looking for biostatistical help to determine feasibility, satisfaction, and sustainability measures from these surveys. Additionally we will be timing all aspects of the workflow and would like guidance on analyzing and presenting this data. This will be a pilot study of 20 cases, and does not have a designated control/comparator group. Mentor confirmed.
  • This is a feasibility study and framed as a hypothesis testing (3D improves Communication). However outcome is subjective/opinion. This may not be sufficient to demonstrate true improvement. Perhaps consider not trying to measure "improvement". Frame as "to what extent" does this improve communication. In terms of sample size roughly ~70 - 96 respondents needed. Link to course notes that mat be useful: Suggest slider scale in redcap to capture a finer point. Pre/Post requires complete data.

2021 September 30

Jaclyn Tamaroff (Ashley Shoemaker), Endocrinology

  • The aim of this study is to evaluate the relationship between cardiac disease and glucose abnormalities in Friedreich’s Ataxia. I have obtained grant funding and IRB approval but wanted to review my statistical plan as I am starting to recruit. VICTR voucher request. Mentor confirmed.
  • For glucose measures how local should variability be? Such as every hour, every day, etc. Main outcome is CoV for glucose and septal wall thickness (Z score). Concern with Z score, should ensure that it is measuring/adjusting for the correct things. Sample size is fixed at ~40. Correlation is acceptable as an approach—but consider moving away from hypothesis testing. Would be suited to a VICTR voucher.
  • Also see:, Search for plotCorr to see Figure 8.5

Selena McCoy Carpenter (Alecia Fair), Wilkins Health Equity Engagement Lab

  • We are looking at data from a survey of 20,000 individuals on the value of the return of research results. I need some guidance in inferential statistics. Mentor confirmed.
  • All of us study. 100,000 surveys sent, ~20,000 returned. Response bias is a concern, should compare respondents to non-respondents.
  • Can do clustering, stratify by a category of interest. Categorical regression could be an option.

2021 September 23

Jack Walker (Daniel Moore), Pediatric Endocrinology

  • We have noted that persons with T1D have reduced pancreas volume which may mean they have reduced pancreas enzyme function. We are planning a study (blinded, cross-over) assessing affect of pancreas enzyme replacement on glucose response to OGTT/c-peptide response (primary outcome) and response of other hormones that respond to food intake and glucose. VICTR voucher request. Mentor confirmed.
  • Double-blind crossover pilot trial. Outcome is glucose AUC. Will also measure insulin intake.Treatment is PERT vs. Placebo. 5-6 subjects are planned, given the expense. The goal appears to understand the distribution, and 5-6 is probably reasonable. Cannot determine effect size with this number.

Patrick Doyle (Nicholas Kavoussi), Urology

  • We are examining the use of machine learning to predict kidney stone recurrence using 24-hour urine study data and EHR-derived data. Our data set is fully extracted, however we would like to discuss next steps for developing our machine learning predictive model with statistical evaluation of performance and ranking of predictive factors. VICTR Biostatistics voucher. Mentor confirmed.

2021 September 16

Drew Johnson (Bob Tunney), Pharmacy

  • This is an epidemiological, observational study to establish the prevalence of iron-deficiency in hospitalized advanced heart failure patients with elevated ferritin as assessed by reticulocyte hemoglobin equivalent. Hospitalized advanced heart failure patients who are being evaluated for or planned to undergo cardiac surgery will be followed via chart review to identify individuals with elevated ferritin (>300 ng/mL). Patients will be followed via chart review to identify the administration of intravenous iron and/or completion of cardiac surgery per the primary team. Final reticulocyte hemoglobin equivalent, ferritin, and transferrin saturation will be subsequently evaluated prospectively as a study procedure. Specific questions to be addressed during the meeting include the following. How do we evaluate the relationship between initial reticulocyte hemoglobin equivalent and ferritin? How do we compare initial vs. final reticulocyte hemoglobin and ferritin in individuals who do not receive iron? How do we compare initial vs. final reticulocyte hemoglobin equivalent, ferritin, hemoglobin, and hematocrit in individuals who receive iron? What would a collaboration entail and what is the expected cost? Mentor confirmed.
  • Patients will serve as their own controls. Advise to collect all labs, date oriented. Could consider using ferritin level as a continuous level. Issue at hand in confounding and sample size. Advise one more clinic visit before starting VICTR voucher request.

Jen Cihlar (Milner Staub), Infectious Diseases

  • Longitudinal behavioral survey of employees undergoing routine asymptomatic COVID-19 testing assessing their risk behaviors and prevention infection adherence. We would like to know the best statistical method to compare changes in answers over the 4 survey periods. Also should we exclude all employees who filled out less than 4 surveys or impute survey data for missing surveys (ie if employees only filled out 2 or 3 of the total 4 surveys)? The data has a non-normal distribution, so would you use an ANOVA? Finally, we also wanted to look at changes in paired surveys between each individual time period (example changes between first and second survey, compared to second and third survey, etc) using McNemar ’s–is that acceptable? Or is there a better method? Mentor confirmed.
  • 65 completed all four. 440 total surveys done. About 175 can be paired. The number missing the fourth survey is a barrier to success, non-response bias. Alternate analysis. First and fourth responders, see how we can predict response to second/third surveys. Longitudibal ordinal analysis (GEE) for those who completed all four.

2021 September 09

Madison Cook (Mayur Patel), Surgery

  • The purpose of the project, Family Views on Quality of Life for Traumatic Brain Injury (TBI) is to survey surrogates of actual TBI survivors to determine quality of life for certain conditions of health after a TBI based on the preferences and values as a surrogate. This study incorporates an anonymous survey method that will identify and describe the most important themes of these experiences, with the goal of developing a better tool to measure outcomes for brain injury patients. We aim to survey surrogates of actual TBI survivors to understand the realities of post-TBI disability using the Extended Glasgow Outcome Scale (GOSE) metric and our standard gamble approach for alternate GOSE health states. We’d like to address/review best statistical analysis for collected data. VICTR Biostatistics voucher. Mentor confirmed.
  • Family view on QoL after TBI

    Goal is QoL metric following TBI. Interview survivors and loved one of survivors.

    308 people included, about 10% of possible ~3,000. Research Match used to recruit. Prior analysis had order effects, could do a similar analysis with this project. Comparing those who completed and those who did not complete will be important, given the recruitment methods. Sample is not representative, will need to address that as well..

    VICTR voucher request. Team has had a prior project with VICTR (Frank and Li). This will fit in the framework of a voucher.

Leah Dunkel, VICTR

  • We propose to conduct a national survey to evaluate a new, lower-literacy description of randomization for use in clinical trial consenting.

    Community informed description of informed consent. There is concern that trial participants may not fully understand the concept of randomization.

    Participants would be randomized to one of three groups.

    Responses will be presented in random order, we will want to determine how order affects choice.

    Question is about sample size may be needed. Funding is there to

    A correlation will require ~400 subjects, and a difference in proportions will take about the same. Having 386 in the smallest group of interest (for example, low literacy group) is a good way to think about sample size—the other groups can be more.

    This should fit in a VICTR framework, should a voucher be desired.

2021 September 02

Alex Foy (Neal Maynord), Pediatric Cardiology

  • Establish a risk stratification tool to identify low vs high-risk patients undergoing Glenn palliation using the past 10 years of clinical data. I would like to address if our data collection method is sufficient for statistical analysis/power. Mentor confirmed.
  • 10 years of Glenn surgeries, want to stratify low/high risk, outcome will be composite (e.g. Death, ECMO, Transplant, Readmit). This is for a QC/QA project, goal to prevent complication via ICU intervention. Vanderbilt patients only, n=220. A true risk prediction will require a sample in the thousands, so this won't possible with the current number of patients. This is staged palliation: Norwood, Glenn, Fontan. Time 0 is post Glenn procedure. Rather than risk model, could put structure on degrees of “badness”, a separate project. Or also could develop complication score, using non-fatal outcomes.

Katelyn Backhaus (Alexandra Shingina), Gastroenterology/Hepatology

  • We are looking at the effectiveness of norepinephrine infusion protocol for the resolution of hepatorenal syndrome in patients NOT treated in the ICU or step down units. This will be a retrospective study. VICTR Biostatistics voucher. Mentor confirmed. Data to come from the RD/SD.

2021 August 26

Margaret Free (Isaac Thomsen), Pediatric Infectious Diseases

  • The primary aims are to identify and characterize current dominant strains of S. aureus causing invasive disease in children by utilizing 2 primary comparisons. First, current strains will be compared to previously circulating strains to assess which factors remain consistent over time. Second, contemporary, colonization-only strains will be obtained from asymptomatic carriers to assess which factors are represented in invasive disease versus colonization alone. We would like to know about statistical tests to run. Advised to meet with a statistician early on in the project to ensure the data we are collecting is amenable to meaningful results. Mentor confirmed.
  • How do colonized strains differ from disease causing strains? How do the strains change over time, disease causing and colonization causing? Could do graphically, but question is how to do statistically.

  • Logistic regression, model time flexibly. Smoothed relative prevalence over time with confidence bands. For sample size: Sample size is set, could just acknowledge this and accept the limitations.

CANCELLED: Madison Cook (Mayur Patel), Surgery

  • Title: Family Views on Quality of Life after Traumatic Brain Injury (TBI) Study. We aim to survey surrogates of actual TBI survivors to understand the realities of post-TBI disability using the Extended Glasgow Outcome Scale (GOSE) metric and our standard gamble approach for alternate GOSE health states. We hypothesize that health utility valuations will be inversely related to functional outcomes as described by the GOSE, but with unique distribution, as compared to our previous health utility after a theoretical TBI. VICTR Biostatistics voucher. Mentor confirmed.

2021 August 19

David Armstrong (David Merryman), Cardiology

  • VICTR application - requesting help addressing statistical justification for N numbers. VICTR Biostatistics voucher. Mentor confirmed.
  • Clinic Notes: Currently have 50 samples and 10 controls. The primary analysis is to compare the samples and controls – trying to figure out how to deal with this sample size. The secondary analysis combines the cases and controls to look at correlations.
  • Recommendations:
    • For the primary analysis, try to figure out the standard deviation in to order calculate the expected margin of error. If you can get more controls, that would help increase power.
    • For the secondary analysis, look at (page 258) for some guidance and examples

Giovanni Davogustto (Quinn S. Wells), Division of Cardiovascular Medicine

  • Using the Get with the Guidelines - heart failure dataset from AHA I would like to study the structural/social determinants of patients being prescribed neprilysin inhibitors at discharge. Questions: Discuss the appropriateness of my biostatistical plan, Appropriateness of inclusion/exclusion criteria, Discussion of management of missing data. VICTR Biostatistics voucher. Mentor confirmed.
  • Clinic Notes:
    • Main focus: Among those with indication for ARNI therapy, is insurance status associated with prescription of ARNIs?
    • Sample size: About 90,000 with ARNI indication that are eligible for ARNI; 20,000 are prescribed at discharge, 70,000 are not prescribed at discharge
  • Recommendations:
    • Show proportion of EF available at the current visit and the proportion of EF available at an earlier date. Then stratify prescription by if the EF is at the current visit or not
    • Look at histograms comparing the absolute difference for all the variables that are measured at admission and discharge (or look at scatterplots) before modeling
    • In the model, might consider interacting age with insurance. Could consider other interactions if there are clinical reasons to. Also, add some sort of calendar time variable to the model.
    • Missing data will be the biggest challenge. Can start with exploratory analyses (ex: what predicts missingness of insurance status). From here, might have a better idea on if missing insurance should be imputed or if those subjects should be removed. Including calendar time in these analyses can help determine if the missingness is associated with calendar time -> could restrict the timeline to later.

2021 August 05

Madeleine Alder (Lynne Stevenson), Cardiology

  • Assistance with analysis of survey data.

  • Meeting Notes: The data were collected using questionnaires given out during office visits. The questionnaire is about quality of life and preferences for advanced heart failure patients. The sample size is about 1300. This is a sub-study of a larger multi-center study.

  • Recommendations: 1. Could describe how many patients declined to fill out the questionnaire. 2. For each question, could report proportions and means. 3. For two questions, could use rank correlation to describe their relationship. There is no need to group numbers/categories. 4. Ideally for heart failure study, would have longitudinal data where patients fill in for each day/week what’s their worse outcome. 5. For further help could reach out for cardiovascular medicine biostatistics team (Meng Xu) or apply a VICTR voucher.

2021 July 29

Steven Allon, Medicine/General Internal Medicine and Public Health

  • Our project is a multicenter exploratory study of a novel resident journal club curriculum involving gamification. We hypothesized that this curriculum would result in improved subjective experience by residents and objective understanding of critical appraisal. Residents were surveyed pre/post intervention with a survey of items with Likert-style responses (subjective component) and a validated measure of critical appraisal skills (objective component). The question I’d like to address is the appropriate statistical test to best analyze our subjective and objective data recognizing there may be data limitations, including missing data.
  • Meeting Notes: The data come from 3 sites (2 of which are paired data, and 1 that is not paired). All sites have pre and post surveys (just not all are paired). Surveys include questions with Likert scale answers, categorical answers, and qualitative answers. The longitudinal data includes about 30 surveys answered per year (about 120 overall). The other 2 sites have a total of N = 61 pre surveys and N = 70 post surveys from the 1 year these have been collected so far.
  • Recommendations:
    • Hypothesis testing isn’t generally the best approach with survey data (so p-values will not be calculated)
    • Think about this more descriptively: report a proportion and confidence interval for each question
    • Instead of doing a proportion for each possible answer, make the Likert answers 1-5 and calculate the mean and 95% confidence interval
      • For the 95% CI, can use a non-parametric bootstrap interval (there are other bootstrap options that could be considered)
    • Can report the mean and CI for individual questions, and for a difference in means for pre and post survey questions. The difference in means can only be done on the paired data; not on the unpaired, longitudinal data
    • Look at the longitudinal plot of means over time for each site, to compare sites

Olawunmi Winful (Jada Benn Torres), Anthropology

Previous clinic session June 17, 2021
  • I wanted to go over some suggestions given to me at my last appointment concerning the phenotype of the individuals of my study. Mentor confirmed. VICTR biostatistics voucher.
  • Meeting Notes: Returning from a previous clinic (06/17/2021) where the recommendation was to do a retrospective cohort study with a CRP value at baseline and a follow-up time. Main question for the clinic is if this is the best approach and how to determine the times for the CRP values. Main research question is how ADI measures affect health, measured through CRP levels.
  • Recommendations:
    • Overall, there are a lot of approaches to take. Deciding on the main question of interest could help narrow this down.
    • An option is to use all the data and look at it longitudinally
      • Can look at trajectories and averages for one variable at a time
      • Can also do a cross-correlation analysis with each patient for specific variables of interest
      • A good option to avoid having to pick one time point as time zero
    • To do a cross-sectional analysis, would need to pick one time point per patient
      • It is important to consider how to pick this time point and if it is reasonable
    • Think about the exclusion for CRP levels greater than 10. This type of exclusion (during the study time period) will make interpretations hard. For the exclusion for subjects who are taking anti-inflammatory drugs, with the longitudinal data, can just exclude the CRP at the time point when they are taking those drugs; don’t have to exclude the patient over all

2021 July 22

Katherine Wiley and Kathryn Morgan (Brian Christens), Community Research and Action (Peabody - Human and Organizational Development)

  • Determining imputation strategy for a longitudinal analysis (likely a multi-level growth curve analysis) with nine unevenly-spaced waves. Mentor confirmed.
  • Meeting Notes: This is a longitudinal study with 9 waves (unevenly spaced) from age 4 to age 25. The dependent variable is sociopolitical control (measured only once at 25). The independent variables are parents’ sense of community involvement, children’s relationship with parents, etc. The study started with about 1000 participants but for the last 4 waves we wanted to focus on, there are about 600 participants.
  • Recommendations: 1. To increase sample size, could assign a read ID (unique ID) for every participant. 2. Prefer to analyze general trajectories as opposed to change score. 3. Could do an ordinal regression analysis with participants who have the outcome. 4. Could perform landmark analysis. Can include the independent variable at each time point as covariates and include an age-specific variable. Need to deal with missing data in this case and use imputation for longitudinal data.

Audrey Bowden, Engineering

  • What I need help with is design of the study for Aim 3. I just spoke with someone yesterday and learned about something called a “superiority study” and I think that’s what I need. I’ve tried to come up with a design and I want to get feedback on whether this design makes sense (and how to change it), if my hypotheses are testable, and, if possible, how to calculate sample sizes for this.
  • Meeting Notes: Want to prove the proposed device have better sensitivity and specificity than standard of care. The proposed device does the test in office instead of in the operation room.
  • Recommendations: 1. Could check out the diagnosis chapter in 2. Need to speak with experts to determine whether this study is a clinical trial. 3. If a single patient could get both the standard of care and the new device, then there would be paired data. When the two tests do not agree, could explore if one test tends to be more correct than the other. If a single patient only gets one test, then could do a comparison. 4. Could be challenging if there are 4 or 5 readouts. The required sample size would decrease if limit the sample to people who have been diagnosed with cancer. 5. Some possible next steps are: 1) have another clinic with the surgeon; 2) attend a VICTR studio where the surgeon joins formally; 3) use cancer center shared biostatistics resource (contact Tatsuki Koyama). 6. For the sample size, estimating a single sensitivity requires at least 96 patients with positive biopsy.

2021 July 15

Giovanni Davogustto (Quinn S. Wells), Cardiovascular Medicine

  • Using the Get with the Guidelines - heart failure dataset from AHA I would like to study the structural/social determinants of patients being prescribed neprilysin inhibitors at discharge. Questions: 1) Discuss the appropriateness of my biostatistical plan, 2) Appropriateness of inclusion/exclusion criteria, 3) Discussion of management of missing data. Mentor confirmed. VICTR biostatistics voucher.

2021 June 24

Sara Duffus (Vivian Weiss), Pediatric Endocrinology

  • Iodine-123 diagnostic whole-body scan (I-123 WBS) is utilized in combination with TSH-stimulated thyroglobulin (Tg) levels for postoperative staging in children with differentiated thyroid carcinoma (DTC) who are thyroglobulin antibody (TgAb) negative to assess for evidence of persistent locoregional disease. Given the high predictive value for absence of disease with low Tg values and debate regarding the added utility of I-123 WBS over neck ultrasound in identifying structural disease, the usefulness of I-123 WBS has been called into question. We aim to quantify the association between findings on I-123 WBS and disease recurrence when controlling for other pertinent clinical variables within a retrospective cohort of 56 pediatric patients with DTC treated at Vanderbilt. Given the relatively small size of the cohort and considering that disease recurrence is a relatively uncommon event in pediatric thyroid cancer in general and in this cohort, we are seeking assistance with forming a statistical plan.
  • Protocol with no expected funding support, Abstract
  • Meeting Notes: There are only 6 events of recurrence out of 56 subjects. Main question is if there is a way to analyze this data with the small cohort size (and small number of events).
  • Recommendations: With both approaches (both logistic regression), the biggest issue is the small sample size. Try data visualization -> then can maybe get a multi-center study going. This will still be difficult due to the small sample size. Can add confidence limits to the plots to help capture the small sample size issue

Tanya Marvi (Todd Rice), Internal medicine

  • We built a competing risk regression to assess the association between thrombolastography and the development of VTE in critically-ill patients with COVID-19 with death as the competing risk. During a previous office hours meeting, we were able to have the initial code for the competing risk regression checked. We have since build a graph of the predicted relative subhazard using Dr. Harrell’s book for stata, and we would like to have that code checked.
  • Meeting Notes: Main goal is to make sure the STATA code for the relative hazards plots is correct.
  • Recommendations: 1. STATA code for figures looks good. The way the reference was coded in appropriate. 2. With only 27 events, it is a good idea to keep the variables linear. 3. A potential future analysis could use a state transition model (or multi state model), so death can be modeled as an absorbing state (nothing else can happen after death). Discussion for this:

2021 June 17

Olawunmi Winful (Jada Benn Torres), Anthropology

  • The purpose of this study aims to examine how biological inflammatory response mechanisms are shaped by demographic, psychosocial, and other SES-related variables in African Americans. The main question i would like to address is how to calculate power. Ive looked at online calculators but im not sure im grasping what im really doing. this is a case control study. Mentor confirmed.
  • Meeting Notes: The outcome of this study is CRP (c-reactive protein), a continuous variable. The exposures of interest is income, represented by ADI (Area Deprivation Index). ADI is based on census information collected 10 years ago. The study includes African Americans aged 45-65 who are not taking any anti-inflammatory medicine, both acute and chronic patients. Timing of exposure and outcome could be an issue.
  • Recommendations: 1. DAGs could be helpful to organize exposure and outcome and understand the association between ADI and CRP labs. This study is probably better suited as a retrospective cohort study. 2. Need to decide about CRP baseline, a window for baseline variables, and a window for follow-up time. 3. One challenge of the study is incomplete follow-up or censoring by death. Severity of the disease could be correlated to death. Death is going to be an important variable if you do the retrospective cohort design. 4. If stick with matched cohort design, need to come up with definitions for case and control. 5. Better off if think about the association as a matter of degree rather than Yes/No. Model the relationship between the exposure and outcome variables. Avoid dichotomizing makes a more statistically powerful study. 6. Some websites that can be helpful for sample size calculation:, If using R, then don’t have to simplify problems into difference in groups. Could use STATA for graphs.

David Isaacs, Neurology / Movement Disorders

  • Environmental stress is also postulated to impact the course of TS,(20) given its known impact on abnormal brain development(21–23) and adulthood psychopathology.(24,25) However, only a single study has explored the role of environmental stress in TS, finding that it predicted two-year tic and psychiatric symptom severity in a small pediatric cohort (n=37 TS patients).(26) No similar investigations have been undertaken in adults with TS. Identifying modifiable risk factors for tics and psychiatric symptoms is critical for developing interventions to reduce the substantial mental and physical health burden on TS adults. We propose a two-year, longitudinal, observational pilot study of adults (>18) with TS to determine the association of lifetime environmental stress exposure with tic severity (Aim 1), psychiatric comorbidity severity (Aim 2), and health-related quality of life (Aim 3). I would like to discuss the statistical analytic plan and sample size analysis for longitudinal / panel data. I have a draft statistical plan (by aim) already and would be happy to share this with the clinic attendees in advance of the meeting. VICTR voucher request.
  • Meeting Notes: Tic severity will be assessed from baseline and two-year follow-up. This study includes adults from different age groups. All the participants will be patients, no controls. Aim 1 explores correlation at baseline. Aim 2 explores baseline association between TS and chronic stressors. Aim 3 explores baseline environmental factors associated to quality of life two years later.
  • Recommendations: 1. Could be helpful to draw out the DAGs. Identify which variables needed to be controlled for. 2. Wouldn’t worry about collinearity at this point for regressions. 3. At this point, sample size is driven by number of variables, not based on variation of the outcome. Need to talk to the individual who wrote the comments to clarify comments on sample size. 4. For sample size calculation, it is recommended to look at the notes on multiplicative margin of error for an odds ratio. For simplicity in sample size calculation, could use just 0 and 1 for responses and think about the analysis as a t-test. But in the end needs to know how many observations there are for each level of the outcome.

2021 May 27

Kaitlyn Bunn (Heather Pua), Pathology, Microbiology, and Immunology

  • Our dataset consists of repeated measures of a single patient with multiple patients and approximately 30 linear continuous variable parameters. We are interested in measuring correlation over time in this data set, which includes dependent variables. Mentor confirmed.
Meeting Notes:
  • Collected data on 8 patients, with 5 timepoints for each subject and 30 surface markers at each timepoint. The biggest challenge is the dimensionality. Regardless of how this is done, it will be arbitrary. Biggest question is how to use this data.
  • Display descriptive statistics. Can do this with spaghetti plots with all 8 subjects on 1 plot; 30 plots for each parameter. Can also look at cross correlation over time between pairs of parameters (though this will be a lot, with 30 parameters). Something to keep in mind for decreasing the dimensionality, is how valid would this be in the entire population?
  • Some other graphics to consider: could color code a line plot for one parameter by the change in value of another parameter (though this could be difficult with the small sample size

  • Normalization would be difficult because the margin of error would be very high – do not recommend doing this

  • Can look at one parameter (a summary over the 8 subjects – likely the slope) and look at the correlation with a clinical parameter. Would want to report just the confidence interval for this, not the point estimate because that is difficult to estimate

Amany Alshibli (Bantayehu Sileshi), Anesthesiology

Notes from previous clinic session, 05/17/2021:
  • Looking at impact of pandemic on surgical care and outcomes in Ethiopia. Retrospective analysis. 3 exposure groups based on time 0) pre-covid 1) "lockdown", no elective procedures 2) after "lockdown", policy of no elective procedures lifted. Want to understand case volume, referral patterns, outcomes (28-day cumulative mortality) in the 3 different exposure groups. Primary outcome is 28-day mortality. Secondary outcomes are change in surgical case volume in phase 1, 2 compared to phase 0; change in referral pattern.
  • Reviewer feedback: Adjust all outcomes and associations for confounders, Sample size/power justification, Type 1 error and adjustment for multiple comparisons, Before-after design need to use segmented regression, plot over time.
  • Recommendations
    • Plots - For case volume, suggest to create a plot (profile/longitudinal) with cases on the Y, time on the X. Estimate what case volume is in phase 1 with confidence interval. For categorical variables, plot is the same but now Y is proportion of the category over time. Possibly stacked bar chart.
    • Could use risk scores if they will apply in a low resource setting. Present analysis as we are interested in descriptives and degree of change/association.
    • Type I error rate - Could suggest to reviewer that we are not in a setting that would be concerned with family wise error rate and multiple comparisons. Could convert to estimation with 95% CI, rather than formal hypothesis testings.
    • Segmented regression - Helpful in situations where you have administrative thresholds, policy changes. Allowing a jump in outcome at a specific time point. If you do segmented, would report plot.
    • Power/sample size - Recommend to not do this. They are basically asking for post-hoc power. Explain, sample size was determined by all available data. Information about effect sizes and precision is contained within the confidence intervals provided, not a power analysis. Large research suggests this should not be done (provide references).
Notes for 05/27/2021:

Meeting Notes:

  • Questions for the clinic: How to report the CIs for mortality rate? Should pairwise differences be displayed between the phases? Should a power calculation be done and how?


  • For the binomial proportions, you can use Wilson interval IF you have balanced data. Need to look into the cases that were deleted for incomplete follow up. It would be useful to include a table of how many people were lost at each timepoint. If most of the subjects were lost at day 1, then the Wilson interval is best. Will need to reconsider if many subjects are lost somewhere between 7-28 days, because there is some info known about these subjects that should not be deleted
  • In table 1, included the p-values is fine. However, do not use these to determine which parameters to include in the model. It is not recommended to include the standardized differences table.
  • It is not recommended to do power calculation after the analysis. Could instead report the study yield. One way to do this is by reporting the margin of error (half of the width of the CI) of the main quantity of interest. Or could report the multiplicative margin of error of the odds ratio (ratio of upper CI to point estimate) for the main quantity of interest

2021 May 13

Amanda Peltier, Neurology/neuromuscular

  • We collected prospective emg and ncs data on neuropathy patients as to usefulness of emg in addition to ncs. would like guidance on appropriate analysis. VICTR biostatistics voucher.
  • Meeting Notes: Wanted to assess if doing EMG adds diagnostic sensitivity or specificity for suspected neuropathy patients. Patients were excluded if they had known diagnoses for neuropathy. Have 89 patients’ data and each patient’s data contain both NCS and EMG. Would like to complete analysis by October.

  • Recommendations: 1. Using NCS only, patients would be diagnosed as “every nerve is normal”, “one nerve is abnormal”, or “two or more nerves are abnormal”. The first approach is to see for patients in each “bucket”, does EMG provide additional information? Identify the combination of findings that would be clinically meaningful. 2. Prespecify the information. For each “bucket”, what proportion of patients that find EMG useful is clinically meaningful? 3. Second approach is to present the 89 patients in two different ways: one with only NCS and one with both NCS and EMG (now we have 178 cases). Mix the cases up and give them to specialists to make diagnosis. Also ask the specialists their certainty for each diagnosis (could use slider scale or Likert scale). Could see what proportion of cases are discordant. 4. For the second approach, most of the analyses would be descriptive statistics with proportions and confidence intervals. Could look at patient characteristics and see if any characteristic is associated with discordance. Could run a logistic model with discordance as the outcome with patient characteristics as predictors. 5. Recommended to submit a VICTR voucher request. We will help to write the analysis section for the VICTR resource request. May want us to review data collection forms before collecting data.

2021 April 29

Jason Springer, Rheumatology and Immunology

  • We are working on an NIH grant. We are looking at a precision approach to dosing rituximab in ANCA associated vasculitis. We will need a power analysis for 2 of the aims.
  • Meeting Notes: The overall goal is to develop a dosing calculator. The discovery cohort included 28 subjects, and gender, BMI, and dose frequency were found to be associated with dosing levels. Some of the subjects have multiple samples (longitudinal data). First question is how to do a power analysis to validate this? Another aim was to look at b-cell depletion in groups with 2 different treatments. The binary outcome is based both on the literature and clinical decision, verses using a continuous outcome.

  • Recommendations: For the first aim, use a precision calculation instead of a power analysis. The precision calculation will validate the absolute difference within a pre-specified level. For the original regression, it is important to consider over-fitting (no more than N/15 parameters in the model). Do an internal validation using a resampling procedure. With a small sample size, a large cross validation will be necessary.

Margaret Compton, Pathology, Microbiology, and Immunology

  • I am trying to assess the impact that COVID-19 had on cervical cancer screening. I would like to discuss how to analyze data with multiple variables (e.g. age, race, normal vs. abnormal pap). VICTR biostatistics voucher.
  • Meeting Notes: It was hypothesized that since COVID, numbers for cervical cancer screening would decrease, but they actually have increased. There was a sharp decline in March – May 2020, but it quickly jumped back up, and higher than the past years. The goal is to determine what could be leading to these increases and if there are specific demographic groups that did or did not increase with the rest.

  • Recommendations: Need to consider reasons why the numbers could be increasing (population increase, new clinics, new providers, changes in the health system, etc.). Defining some kind of denominator will help with making group comparisons. Some ideas for a denominator: get enrollment numbers or number of well women visits. If a denominator can be defined, you can look at a subset of your data that has had multiple screenings over the years and look at gaps between screenings (as a function of age) to see if there is a change. If a denominator cannot be determined, an epidemiologist or someone from the health policy department could be good contacts for discussing a different approach. It may also be possible to look at different events relative to each other, especially if a denominator cannot be defined. If this is a project you would like to make generalizable (and not just for an internal decision), it would be useful to apply for a VICTR voucher.

2021 April 22

Dakota Vaughn (Sean Donahue), Ophthalmology

  • Previous clinic session: June 25, 2020 (Thursday); October 19, 2020 (Monday); February 15, 2021 (Monday), March 25, 2021 (Thursday)
  • Developing a predictive model for amblyopia given results of preschool vision screening. Wanted to touch base to make sure that the validated model looks good after feedback at our last meeting. Mentor confirmed.
  • Recommendations:
    • Odds ratios from the report are based on interquartile range by default. Interpretation of odds ratio is “N times the odds of having ***”.
    • Report the current slopes for worst sphere. Prime means it’s not a linear relationship, and the slope for the prime is the sum of the two slopes. Standard error corresponds to each variable. Report two digits for coefficients and SEs from model.
    • Recommends doing partial effects plot for all the variables. ggplot(Predict(model)).
    • Recommends doing a model that assumes age is not linear. Use restricted cubic spline and add 3 knots at clinically important ages.
    • Could do variable clustering to detect the relationship between variables. varclus/redun. Could also do means of continuous variables by race-ethnic groups.
    • On the clinical utility from this model, could 1) do partial effects plot. Could make these under risk scale. 2) do a nomogram showing people how to get predictive risks. n <- nomogram(model, fun = list(P = plogis)) plot(n) or fun = list(plogis), funlabel = ‘P’ 3) do a histogram of predicted risk. p<- predict(model, type = “fitted”) histSpike or hist(p, nclass = 50)
    • ROC curve is not recommended in this case.

2021 April 15

Jacqueline Antoun (Ashley Shoemaker), Pediatric Endocrinology

  • Pseudohypoparathyroidism (PHP) is a rare, genetic disorder usually diagnosed in childhood. The most common forms of hormone resistance in PHP are resistance to PTH and TSH . Recent international treatment guidelines now recommend efforts to normalize PTH levels in patients with PHP, along with maintaining normal calcium levels. Despite these guidelines, patients continue to present to our clinic and research center with PTH levels >3x the upper limit of normal. There are currently no published dosing guidelines for levothyroxine or calcitriol in patients with PHP and our experience suggests that high medication doses may be required to overcome the intrinsic hormone resistance. In order to better understand the natural history of hormone resistance in PHP and the range of medication dose requirements, we undertook a retrospective chart review of patients followed at our institution for clinical care or enrolled in one of our PHP research studies. Using data from a retrospective chart review, do weight-normalized, prescribed doses of levothyroxine and calcitriol for patients enrolled in PHP research studies at Vanderbilt University Medical Center between 2012-2020 who meet the clinical criteria for a diagnosis of pseudohypoparathyroidism differ from current dosing guidelines of levothyroxine and calcitriol for hypothyroidism and hypoparathyroidism, respectively? Mentor confirmed.
  • Meeting Notes: Data include 30 subjects with PHP. Of the 30 subjects, longitudinal data is available on 7 of the subjects. The data include age, sex, weight, clinical data, and dosing over time for the 7 subjects. Dosing can be in different forms (pill or liquid) and can be in set quantities or continuous.
  • Recommendations:
    • Start by making a spaghetti plot looking at the raw dosing data over time
      • Can make the x-axis age or time since start of medicine (whatever makes the most sense)
      • Can color code by sex, or any other variable that would be of interest
    • Plot hormone level by dose, and color code by age group
    • Can do analyses showing what the average patient looks like
    • To look at what the average patient looks like over time, random intercepts and/or random slopes can be added
      • In order to do this, a larger sample size is needed
    • Can potentially use the cross-sectional data to estimate the average dose at a certain age
      • Can calculate age-specific means and confidence intervals
      • Would be possible if the longitudinal data is comparable to the cross-sectional data

2021 April 01

Isik Turker (Matt Alexander), Cardiology

  • We would like to analyse before and after treatment BP data djusted for multiple confounders. Mentor confirmed.
  • In the current dataset there are 193 patients with 2 measurements of blood pressure that are 2 years apart. The primary question is whether or not BP (or TIS*SBP) differs in the individuals before and after ICI therapy. Would like to incorporate the paired nature of the data into the analysis. There are 4 types of ICI therapy and everyone received one of the four types. Potential confounders are type of cancer, age, sex, weight change, race, history of HTN, time on therapy, etc.
  • Recommendations: 1. Post-time variables could be hard to interpret. Adjust for baseline weight or take out change in weight in the model. 2. Ideally randomize some patients to ICI, others to no treatment. 3. Graph BP of all individuals over the two years and compare it to the change of BP in the general population. 4. Could also get all the BP data from all individuals in the two years and the medication they were on at each BP point. The current data have a problem of survival bias by excluding people who dropped out before 2 years. Define the cohort at baseline. 5. Do not recommend doing TIS*SBP in the model. It’s hard to interpret and confusing. 6. Could treat TIS as the outcome or create an ordinal scale. 7. Could do a mixed effects model. There would be two rows of data for each patient and patient would be the random effect (cluster). Create a variable for pre/post, adjust for baseline variables, and include independent variables that vary over time. 8. Could also do a logistic regression with the outcome hypertension Yes/No for the individuals who did not have hypertension to start with. Find the factors that predict the new hypertension. 9. Recommended coming back to the clinic after running more analyses or applying for VICTR voucher.

2021 March 25

Dakota Vaughn (Sean Donahue), Ophthalmology

  • Previous clinic session: June 25, 2020 (Thursday); October 19, 2020 (Monday); February 15, 2021 (Monday)
  • Predictive Model for Amblyopia Risk Factors given results of vision screen, as discussed at previous clinics. Would like to address model progress and validation strategies. Mentor confirmed.
  • Meeting Notes: Following the previous clinic, wanted to review if the approach was appropriate, what the next steps are, and how to validate the training set. Originally generated a predictive model, placing a linear spline on sphere, and then used ‘fastbw’ to find the important variables to keep in the model.
  • Recommendations:
    • Since it is not expected for cylinder to be linear, add a restricted cubic spline to cylinder: rcs(cylinder,3)
    • Instead of doing a split sample validation, do a resample validation: refit model ~300 times with a different sample w/ replacement, each sample has same number as full dataset: validate(model, B=300)
    • Don’t need to do backwards step down. Instead, just include all the variables that were originally considered before this.
    • Do not need to run univariate analyses.
    • Make race/ethnicity all into one variable, instead of having separate binary indicator variables for each race and ethnicity.
    • For the odds ratios, use the ‘summary’ function. You can change the ‘Low’ and ‘High’ values by stating those in the summary function: summary(f, worst_sphere=c(-2,2))
    • As a sub-analysis, do left and right eye separately. Compare this model to original model, using the likelihood ratio chi-square to compare.

2021 March 18

Kevin Patel (Trent Rosenbloom), Neonatology

  • We intend to develop and validate a Neonatal Discharge Model (NDM) that uses clinical data from the electronic health record to identify patients approaching discharge with 3-5 days from the Neonatal Intensive Care Unit (NICU). We want to review the statistical considerations in performing an ROC curve and the format of data that would be required. VICTR Biostatistics voucher. Mentor confirmed.
  • Meeting Notes: Main question: Can we develop a criteria model that can predict/forecast patient discharge within 3-5 days from the neonatal intensive care unit? The aim is to find a collection of patient variables that once they are true could identify patients as approaching discharge. Based on clinicians’ practice and current literature, there are 10-12 variables that are important to include in the criteria model. For continuous variables (temperature, weight, etc.), the investigators wanted to set threshold values. And for binary variables, they will be converted to true/false.
  • Recommendations: 1. Treat all continuous variables as continuous probably will get better predictions. A hard boundary usually causes problems. 2. Consider which values need to have their trajectories taken into account. For example, for temperature, could use the last temperature and the average among all the previous temperatures. 3. Before going into data collection/data analysis, consider how do you want to state the result (probability that a patient will be discharged in 3 days/expected time for stay/median time for discharge given the length of stay). When stating the result, use estimates and prediction intervals. 4. Think about how to handle death. Use time as successful discharge (count death as unsuccessful discharge). Can sensor on the death, count as incomplete treatment. 5. In terms of sample size, probably need a few thousands considering the complexity of the model. 6. It is recommended to think more about research question and come back to clinic.

CANCELLED: Soha Patel, OB/GYN, Maternal-Fetal Medicine

  • Fetal outcomes in COVID+ patients. Feasibility question. VICTR Biostatistics voucher.

2021 March 11

Raymond Zhou, Vanderbilt Eye Institute

  • Previously clinic session February 20, 2020
  • Vision screening in young children helps identify important eye diseases. Our goal was compare the positive predictive value of a failed vision screening in children <3 years old vs. children 3-5 years old. We completed a data pull on all patients forwarded to VEI after failed vision screening. Our primary analysis was a multivariate regression analysis, to assess for the effects of various demographic/clinical variables (age, race, ethnicity) on the prevalence or absence of Amblyopia Risk Factors (a defined set of eye diseases). Was also completed multiple imputation for instances of unreported race and ethnicity. We submitted the manuscript for publication and were rejected. I had a couple questions about how to best present our findings, using the comments from reviewers.
    • We presented differences in the PPVs of children <3 vs. 3-5 years old through Odds Ratios. Odds Ratios were presented before and after regression analysis +/- multiple imputation. Is this the best way to present our findings?
    • We presented baseline characteristics in children <3 vs. children 3-5 years old. Would it be valuable to split these numbers into even smaller categories?
    • What is the best way to refer to the RMS package in a manuscript?
  • Meeting Notes: The PPV is the probability of final positive diagnosis given that the test was positive. The odds ratios listed in the table are for the association between age group and final positive diagnosis.

  • Recommendations: List the PPV for each age group, with their respective CI’s. Also list the difference in the two PPVs, with the respective CI. Keep the current odds ratio table. To cite the RMS package, use the link:

2021 March 04

Chelsea Gorsline (Gowri Satyanarayana), Infectious Disease

  • Previous clinic session July 16, 2020; July 30, 2020; October 8, 2020; January 21, 2021
  • Antimicrobial de-escalation in patients with high-risk febrile neutropenia: attitudes and practices of adult hospital care providers. This is follow-up to recent biostats clinic where proportional odds ratio analysis was recommended. Would like to review results and interpretation with the group. Mentor confirmed.
  • Meeting Notes: Most of the data are from hematology/oncology department and that was used as the reference group.

  • Recommendations: 1. For the reference group, either include it in the table or in the footnote. 2. The interpretation of a proportional odds model is “the odds of feeling more comfortable is…”. 3. For tables, present odds ratio with CI is enough. P-values and model fit statistics are not necessary. 4. Recommend PS software for calculating power ( ).

Lynsa Nguyen (Soha Patel), ObGyn /Maternal Fetal Medicine

  • We are interested in looking at fetal and neonatal outcomes in COVID-19 positive pregnancies. We would like to review our study design with statisticians to ensure study design is sound. We plan to review outcomes of all pregnancies within the last year and compare COVID-19 positive pregnancies to COVID-19 negative pregnancies stratified by gestational age at the time of delivery. Mentor confirmed.
  • Meeting Notes: A retrospective study looking at COVID negative and COVID positive pregnancies at VUMC last year. Variables collected include patient demographics, gestational age of delivery, maternal age at delivery, and other comorbidities.

  • Recommendations: 1. The next step after getting the data would be schedule another clinic to determine the feasibility of the project. 2. Propensity score matching could be used. 3. Recommend applying for a VICTR voucher if the project is feasible.

2021 February 25

Dylan Williamson (Ashley Shoemaker), Pediatric Endocrinology

  • We are performing a retrospective chart review of patients who have had a genetic test for obesity. We’ve collected longitudinal height and weight data on two groups of patients and want to explore if there are any different trends between the two. We also have comorbidity and prescription drug data between these groups, and would like to compare these as well. We are wondering if we have sufficient data to begin analysis, and if so what the best methods of analysis will be. We plan to apply for a VICTR biostatistics voucher for assistance with this project. Mentor confirmed. VICTR voucher request.
  • Meeting Notes: The goal is to compare pediatric patients with a positive genetic test for obesity to those with a negative test. The data is about 50-50 for positive and negative tests, with about 90 subjects total. Height and weight are both collected over time for all subjects, but there is variability in the number of data points and the overall time for each subject. The hypothesis is that a patient with a positive test for this gene will have more rapid weight gain and an earlier onset of obesity.

  • Recommendations: Start with plotting the data overtime by the 2 groups to explore the data. This project is appropriate for a VICTR voucher – recommended to apply.

2021 February 18

Aniket Rali (Sandip Zalawadiya), Cardiovascular Medicine

  • We are working on analyzing the ELSO registry to look at predictors of ECMO wean success. We need help with our statistical analysis. Mentor confirmed. VICTR biostatistics voucher.
  • Meeting Notes: There are two aims of this study: 1. Find early (within the first 24 hours) predictors for success (survive the entire hospitalization). 2) Define a modified SAVE score with additional parameters. Since ELSO is an international registry, the data are not always complete. Some of the baseline characteristics of the ELSO registry include height, gender, race, hours on heart and lung machine, primary diagnosis. There are more than 16,000 subjects in the ELSO registry.

  • Recommendations: 1. Check what proportion of each variable is missing. If more than 30% of a variable is missing, might consider drop that variable from the model. For the rest of variables use multiple imputation for the prediction model. 2. Could do a logistic regression model that includes clinically important variables with binary outcome success/failure. Could divide data into training/test sets or do bootstrapping. 3. It would be ideal if we have time to death and time to discharge. In that case we could do survival model instead of logistic model. 4. Apply for VICTR voucher or work through collaboration.

David Brooks, Neonatology

  • My study is evaluating the accuracy of a point of care glucose meter compared to lab based glucose testing in preterm and ill neonates. I submitted my application for VICTR funding and received feedback from the pre-reviewer that my analysis plan is under-developed and I would benefit from attending biostats clinic.
  • Meeting Notes: POC glucometers are less reliable during hypoglycemia, which frequently occurs in neonates. One of the aims of the study is to quantify the difference in plasma compare device and lab. The sample size was calculated to be 60.

  • Recommendations: 1. The PS Power and Sample Size Calculation should be able to calculate sample size. Specify power and test used to get sample size in the application. 2. Need 10-20 patients per variable. The multivariate linear regression could include 3-6 variables.

2021 February 11

Joseph Vanterpool (Jamaine Davis), SOM

  • I’m doing a retrospective data analysis on mistrust’s affect on patient behavior and health status. I’ve collected some data with chi-square and glm tests, interpreted them, and written a draft of an abstract. At this point, what other statistical tests, if any, would I run? Mentor confirmed.
  • Meeting Notes: Mistrust is answered on a scale 1-5. It was condensed to a categorical variable (1-3) and (4-5). Goal is to compare mistrust groups to other variables (healthcare choice, risk of high cholesterol, and other clinical variables).
  • Recommendations: For the 2x2 tables, percentages should be based on column totals. The chi-square tests help to answer if the choice of health care (or other variable) is different among mistrust groups. From the 2x2 tables, you can report the column percentages between the mistrust groups, and can report the p-value from the chi-square test. For missing data, we want a high response rate (around 90% or higher).

2021 February 04

Luke Laws, (Ben Palmer, Dan Clark), Internal Medicine

  • We are studying cardiac magnetic resonance imaging (CMR) in heart transplant patients. We are comparing the imaging findings of Hepatitis C (HCV) positive donors against a control population of HCV negative donors. We would appreciate assistance with matching HCV+ patients with appropriate controls to minimize confounders of CMR data. What is the most statistically sound way to match patients between “intervention” (HCV+) and controls (HCV-)? Would propensity scores be the best way to match and if so, how would this affect data analysis? Mentor confirmed.
  • Clinic Notes: Characterize imaging findings using scores (range 800-1200). Have 26 HCV+ cases and 70 HCV- controls. Covariates include age, gender, CMV viremia within 1 year, rejection, and number of times treated for rejection. Can build a propensity score (PS) model using logistic regression (HCV+ vs. HCV-) and adjust for patient characteristics (using rule of thumb 26/10 ~ 3, or up to 5 variables), then match patients based on PS score, and build linear regression model using matched cohort with approximately 9 covariates. If any variables are highly correlated, then only need to include one in the model. This can reduce the number of variables in the model.
  • Recommend applying for VICTR Award for biostatistics support (90 hours). Application website ( and research proposal template (

Jessica Willis, Internal Medicine

  • I am doing a retrospective chart review of approximately 150 patients. I have an excel file of the data but would like some assistance with a linear regression analysis and further interpretation of the data.Assistance with Analysis of retrospective chart review data. Would like help making basic characteristics chart.
  • Clinic Notes: Want to assess whether inflammation increases risk for developing ventricular arrhythmia. Also recorded gender, race, age, QT interval at event, CO2, creatinine, type of inflammation, and treatment. Recommend combining Excel worksheets and adding column for Arrhythmia (Y/N). Then run chi-square test (categorical variable), t-test (continuous variable), or Mann-Whitney U test (non-parametric for continuous variable, if available) in Excel. A logistic regression model for arrhythmia (Y/N) can adjust for confounders; report odds ratios for each covariate in the model.

2021 January 28

Rachel Fortner (Jennifer Lindsey), Ophthalmology

  • I am looking at whether or not the delay from the COVID pandemic lead to a progression of glaucoma. I did a chart review from 2020 and then from 2018 patients to see how their eye disease changed between two appointments. I would like help on how to best analyze the data I have collected. I have talked with a few statisticians and they recommended different things. I was hoping to see if there were any statistical conclusions (logistic regression or anything else) that I could draw instead of my results just being descriptive. Mentor confirmed.
  • Clinic Notes: Main outcome is eye disease progression.
    • Are baseline severity between two groups same? Interval between two appointments depends on patients' severity. Collect baseline severity, time scheduled, whether second appointment was delayed due to pandemic, want to assess whether delay caused worsening. The 2020 group has longer time interval between two appointments due to the delay. Pandemic effect is confounded by delay.
    • Glaucoma: Mild, moderate, severe. How to quantify progression? Any progression yes/no, use Chi-square test to compare between groups. Since this is observational, try logistic model to adjust for confounding factors (baseline severity, length between two visits, etc.)
  • Data in excel. Suggest apply for VICTR voucher to help with the analysis.

Tray Hunley, Pediatrics/Nephrology

  • I am a clinician. I am writing a paper (small case series) with a couple authors at other institutions. I wanted to check to make sure my very basic statistics are valid (paired t test).
  • Clinic Notes: A small case series (N=8) on kidney patients. Out of the 8 patients, 5 of them responded to a new medication and 3 did not within the first month of using the new medication. The outcomes include protein level in urine and serum albumin level. A potential confounding factor is that patients were on immunosuppressants.
  • Recommendations
    • Observe the patients’ outcomes and define responders/non-responders clearly.
    • The effects of the medication could be the result of regression to the mean.
    • In future (larger) studies, compare change of outcomes with pre-medication values.
    • Instead of doing multiple statistical tests (multiple testing issues, small sample size), calculating and reporting confidence intervals would be a better idea.
    • Due to small sample size, the data may not be normally distributed. A Wilcoxon test (a nonparametric version of paired t-test) would be more appropriate in this situation.

2021 January 21

Melanie Whitmore (Jessica Anderson & Julie Pingel), Pharmacy

  • Assessing weight based versus non-weight based vasopressor doses (norepinephrine, epinephrine) on time to achieve MAP goal, length of hospital stay, various other end points. Mentor confirmed.
  • Clinic Notes: There was a change in practice a few years ago to use non-weight based doses. Secondary outcomes include all-cause mortality, ICU length of stay, and number of vasopressors required to achieve goal. Have 106 subjects total. Can plot time to achieve MAP goal with calendar time on the x-axis. Use interrupted time series model to assess difference in time to achieve MAP goal between pre/post time periods. May consider running a sensitivity analysis using only patients who survived. Since there were so few deaths, recommend reporting mortality rates stratified by pre/post time period.

Chelsea Gorsline (Gowri Satyanarayana), Infectious Disease

  • Previous clinic session July 16, 2020; July 30, 2020; October 8, 2020
  • Presenting updates to prior Biostats Clinic discussions in 7/2020 and 9/2020, with particular focus on statistical analysis of knowledge, attitude and practice survey data (KAP). The survey was distributed internally to heme/onc, ID and pharmacy providers regarding antimicrobial use in the management of febrile neutropenia. Would like to discuss next steps for analysis and presentation of survey data. Mentor confirmed.
  • Clinic Notes: Survey included two clinical scenarios to assess attitudes toward de-escalation of antimicrobials. There were a total of 45 respondents. Can use proportional odds model for ordinal responses (outcomes). For most important factors, concerned with empty cells to be able to run a statistical test. It will be best to report descriptive statistics only.

2021 January 14

William McEachern (Eric Austin), Pediatric Cardiology

  • The project is a prospective cardiac MRI study of unaffected BMPR2 mutation carriers (which predisposed to pulmonary hypertension) with comparison to existing repositories of cardiac MRIs for pulmonary hypertension patients as well as healthy controls. I applied for VICTR funding, and statistical questions were raised about the appropriate analysis that we should perform. We had proposed more traditional measures to detect a difference between the groups, but the VICTR pre-review understandably encouraged use of predictive metrics with confidence intervals, c-index/AUROC, and ultimately a multivariable regression model. I would like advice on how best to reframe our statistical proposal, including: is sample size justification necessary if we use the Efron-Gong optimism bootstrap to correct for over-fitting? Mentor confirmed. VICTR voucher request.
  • Meeting Notes: Looking at Cardiac MRI’s in three different groups: Patients who are unaffected carriers (at risk for pulmonary hypertension), patients with pulmonary hypertension, and healthy controls. 25 subjects in each group. Have 10-12 meaningful cardiac markers to look at. Study question: Is there a difference between these groups?
  • Recommendations: Use a conditional logistic regression model to find which markers are associated with pairwise comparisons of the 3 groups, matching on age and gender. Sample size is a limiting factor – could only include 2-3 markers in the model based on the rule of thumb, 10-15 events per marker. Will need to rank the markers and then run multiple regression models. Can use Benjamini-Hochberg’s approach to control the FDR to account for the multiplicity issues. Can report the C-Index.

2021 January 07

Amany Alshibli (Bantayehu Sileshi), Anesthesiology

  • Previous clinic session: 12/17/2020
  • We are conducting a retrospective analysis of perioperative data collected in REDCap as part of the ImPACT Africa program to understand the effect of the COVID-19 pandemic on surgical care and outcomes. We previously attended a Thursday clinic on Dec 17 and want to follow-up on the analysis we have done since then based on recommendations given during that time. We are anticipating submitting a conference abstract by mid-January. Mentor confirmed.
  • Meeting Notes: Goal to compare 28-day mortality rates across three phases. There are 88 deaths out of 3131 patients. Each level of a categorical variable counts as one covariate (degree of freedom) for the rule of thumb to include one covariate in a logistic regression model per 10-15 cases.
  • Since there were fewer elective surgeries during Phase 1, need to look at a table of baseline patient characteristics stratified by the three phases. Recommend including clinically relevant covariates in the model (phase, patient age, etc.) regardless of statistical significance from a univariate analysis. Can report odds ratios and 95% CI for phase to quantify the effect size and report C-index to assess how well the model fits the data.

Reena Jayani (Adetola Kassim & Harvey Murff), Medicine/Hematology/Oncology

  • My proposal is looking at the association of biologic markers of aging and severe toxicity after allogeneic hematopoietic cell transplant. The exploratory aim is looking at change in the biomarkers over time and I am interested in understanding the best statistical analysis to evaluate change over time. This proposal includes 6 biomarkers and will be at 2 time points: at baseline and 100 days after transplant. There is a potential for collecting the data at 30 days and 60 days post-transplant, but this is not finalized. Mentor confirmed. VICTR biostatistics voucher.
  • Meeting Notes: Plan to record baseline patient demographics, comorbidities, and transplant characteristics. Note that patient survival may impact availability of post-transplant biomarker measurements; there will be an interaction between survival and biomarker. Time of transplant is the anchor point. Recommend joint modeling with survival analysis (Cox proportional hazards) and longitudinal model (mixed effects model with a random effect for subject) on the subset of patients who survive to time t . Can create spaghetti plot of biomarker measurements for each patient over time.
  • Recommend applying for VICTR Award for biostatistics support (90 hours). Application website ( and research proposal template ( Feel free to contact Chang Yu with questions.
Topic revision: r2 - 13 Dec 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