You are here: Vanderbilt Biostatistics Wiki>Main Web>Clinics>ClinicalHealthResearch>ThursdayClinicNotes (22 Feb 2024, JacksonResser)EditAttach

- External validation \x{2013} separate study of 300 patients. Same variables were collected. Is the development sample big enough to stand on its own, without validation patients?
- One issue here \x{2013} inclusion / exclusion criteria could vary patient population. You could explore if there is a more narrow range of important predictor variables.
- Drop off from internally-validated to externally-validated was .04 \x{2013} R squared measures may also be used. Frank likes 90
^{th}percentile of absolute differences, as well. - Half patients had outcome \x{2013} the way you analyzed, we don\x{2019}t take into account whether a patient was very close to being cognitively impaired but wasn\x{2019}t. An overall scale (Frank will provide FDA talk; https://www.fharrell.com/talk/cos/) could be a path forward.
- Maximum resolution outcome variable is best. Averaging ranks of two scales, using one scale to predict another to that one scale may be calibrated to another\x{2026} some additional discussion could be done to help here.

- Support study, end of life decision making; if personalized, reliable survival curve is provided, provisions didn\x{2019}t make as big of a difference as one may think. Misinterpretation of results was common, and patients didn\x{2019}t latch on to risk scale.
- Median life expectancy would maybe have been more effective.
- The proposed model will give a risk-based estimate, but volume, amount of cognitive challenge, other unit of measurement. may be effective for patient interpretation.
- Wes\x{2019}s summary: if you take away from this paper that we enable a clinician or team to tell patient that they have X% likelihood of new brain disfunction, and they way we would handle that is support / classes / etc, that would be a win. Calculator may be less important; a distillable statement may be preferable.
- Mark\x{2019}s concern: Many people will fall into a category that, rehab may be needed \x{2013} maybe a very small low-risk and small high-risk population.
- Also worth noting that this is a survival-only model. Ordinal longitudinal analysis may be communicated as % of being included in a cognitive level or worse; if you\x{2019}re excluding the people that die after counseling, as in the current analysis, that is likely misleading.
- That would align with being prospectively defined. Chance of being in good cognitive function + alive, poor function but alive, dead.
- Having one scale would make the biostatistical problems easier to solve. Can estimate median scale for a person.
- If the model was to be changed to be just cognitive impairment, but death was included, with an ordinal scale \x{2013} patients would have a median. Challenge would then be in communication to patients.
- Frank\x{2019}s comments based on similar study; density function with most likely level of disability a year after surgery. Median could be target summary, but 10
^{th}and 90^{th}percentile could be provided as well.- Hui Nian may be able to help.
- 4-8 levels, a stacked bar chart may be effective for discrete / ordinal outcome.
- Adjusting of independent variable; can one or two variables be subbed in and out while remaining methodologically rigorous?
- Just don\x{2019}t try a lot of variables that you then discard, and then in the subsequent validation fail to repeat those \x{201c}tries\x{201d}.
- But yes, adding a few variables at this point (length of stay, for example) is not an issue.
- If many experts were assembled, and various levels of the two scales was given, and you asked each which one is worse; if you can order those combinations such that they agree, that may be effective. 20 or so experts would be needed, though.
- Bare minimum would be five ordinal levels (not including death); ten would be ideal.
- Takeaways: Ordinal scale is best, combination of outcomes or only one, adjustment with independent variables is possible, and communication + interpretation of results will need to be thought about for patients.
- Multiple observations over time as a longitudinal analysis may be another good option.
- Ex: If you died after the first time point, that\x{2019}s an absorbing state.

- Frank may be an author if desired, or perhap acknowledged.

Table 1: Patient and Tumor characteristics comparison between young onset and average onset patients

Table 2: Survival and Recurrence rates between the 2 populations

Table 3: odds ratios, kaplan meir curves vs logistic regression etc Two ways of getting into system: 1) oncologist started database for cancer patients (all cancer patients at the hospital) 2) At every follow-up visit, patient added to database Some women could have failed to enter the study population because the cancer became severe quickly Want to guard against ill-defined denominator - Problem: patients who die before entering population (not a random sample) - Example: cats falling off buildings; cats that died the moment after the fall were excluded Paucity of data for breast cancer in this area of Africa Time-oriented outcome like age of onset prone to bias Could be difference in types of breast cancer in population Value in determining pieces in the data that don't matter and then confirming that they don't matter - Negative controls give you more confidence in positive controls Data exploration: make a model to predict a missing lab value First: dig into data, build demographic tables Multivariable analysis of the differences (logistic regression model) to predict age cohort - Looking for unique differences Pre-cursor analyses: degree of missingness could limit types of analyses you could run - Cluster analysis: understand degree of missingness on the same individual Regression analysis: using R - https://hbiostat.org/rmsc/software Also resources available to help get data from REDCap into R Kaplan-Meier vs logistic regression - LR better when time is not important In some cases, not confident whether participant died from breast cancer or another cause

- Troponin stays elevated for 2 weeks \x{2013} peaks 12 -48 hours after MI
- Dynamic range is very large
- Possibility of secondary analysis \x{2013} log ratio to upper limit of normal. Relationship between log ratio and outcome, as well as same relationship for number of components drawn, to assess if here\x{2019}s bias that makes interpretation difficult.
- Help determine if peak troponin should be adjusted for number of draws

- Andrew: If Frank were reviewing, would he want to see an analysis where size of MI is held constant? Or would he ask for a second stratification (within Type 1, then by size; within Type 2, then by size)?
- The current display is not that helpful due to heterogeneity between Type 1 and Type 2
- Graph that shows log ratio vs. outcome; if adjusted for log ratio vs. outcome, does type add anything to predicting the outcome? (In Andrew\x{2019}s words, MI by size vs. outcome, and see if that is impacted by size of MI)
- May be more useful to see if log ratio interacts with treatment
- May be fit with spline function

- Perpetuating clinical trials to give minimal information to the reader.
- MI\x{2019}s are being treated as \x{201c}equally big\x{201d}. An analysis that looks at relationship between liberal and restricted, and how big of an MI someone got as a second MI, would be encouraged.
- Andrew notes that this is set for a second paper; Frank thinks it may be best used here.

Research in animal studies, retrospective cohorts, and secondary clinical trials analyses suggests that plasma may improve outcomes in traumatic brain injury (TBI). We examined the association between plasma administration and mortality in moderate-severe TBI, hypothesizing plasma is associated with decreased mortality after accounting for confounding, including by indication. Methods:

Patients greater than 18 years with moderate-severe TBI from the 2017-2020 Trauma Quality Improvement Program (TQIP) dataset were included. Patients with anticoagulant/antiplatelet use, specific comorbidities (bleeding disorders, cirrhosis, chronic renal failure, congestive heart failure, chronic obstructive pulmonary disease), outside hospital transfer, and missing hospital mortality were excluded. Multivariable logistic regression examined the association between plasma volume and hospital mortality, adjusting for sociodemographics, severity of injury/illness, neurologic status, and volume of other blood products, including interaction terms of plasma with shock and need for hemorrhage control procedure, respectively (see Table 1 for details). Sensitivity analysis excluded patients with shock and hemorrhage control. Results:

Of 4,273,914 patients in TQIP, 63,918 met inclusion. Hospital mortality was 37.0%. 82.8% received no plasma. Other cohort characteristics were mean age: 44.9; mean Injury Severity Score: 28.5; percent female: 24.4%; percent severe TBI: 69.4%; percent in shock: 7.4%, percent needing hemorrhage control procedure: 12.2%. Unadjusted, each categorical increase in plasma volume (from 0 to 0-2 to 2-6 to 6-12 to greater than 12 units) is significantly associated with greater odds of mortality. Confounder adjustment attenuates this effect (Table 1): the odds ratio (95% confidence interval) increasing from 0 to 0-2 units is 1.23 (1.09-1.38); from 0-2 to 2-6 units is 0.96 (0.83-1.11); from 2-6 to 6-12 units is 1.19 (0.96-1.47); and 6-12 to greater than 12 units is 1.68 (1.20-2.34). Similar results are seen in sensitivity analysis. Shock and need for hemorrhage control procedure significantly (p less than 0.001) modify the relationship between plasma and mortality. Conclusions:

Plasma\x{2019}s effect on mortality in TBI remains unclear. Likely due to residual confounding despite adjustment, plasma is associated with increased mortality in moderate-severe TBI in this retrospective cohort. Interaction term analysis suggests this is confounding by indication, specifically because plasma is usually administered for hemorrhage (which in turn, increases mortality). A prospective randomized study of plasma for nonbleeding patients with TBI would better answer this important clinical question. Discussion notes: Association of plasma and mortality in severe TBI Question: other methods ot account for unmeasured confounding? Instrument variable analysis (generally limited to randomized study), e-value sensitivity analysis Inflection point found between 6-10 units of plasma (categorized exposure at clinically relevant threshold) - Frank: categorization approach counts all values in a category as the same. Categorizing at inflection point does NOT respect the form of the data Survival bias (patients that die early don't receive as much plasma) - Higher resolution data needed to address We want to adjust for confounding of bleeding for the relationship between plasma and severe TBI - Difficult to disentangle bleeding from plasma -- so intimately intertwined (hard to do without randomized design) - Question that can be answered: investigate relationship between bleeding and plasma, characterize by other variables - Could analyze quality of clinical practice, variation in how much plasma was given - Could inform later analysis when you bring in mortality Frank R package (rms) to perform instrument analysis High proportion of participants who did not receive plasma at all - Need to choose knots in spline function. Placing knots difficult when lots of zeroes - Manual override places knots using non-zeroes Retrospective data -- feedback loop

- Implementation of complex, quantitative risk information
- And, how should results of predictive models and other tools be applied?
- SPECTACULAR \x{2013} rapidly & empirically look at design elements
- Primary test: four different ways (across 10 nurses, 12 timepoints) to display content
- Measured preference, what action would be taken from that (Contact RRT, Contact MD, Contact Charge, Contact Peer, Increase monitoring, Continue Same)
- Many other things could be modified
- Can we take a factorial design and merge it with a Bayesian adaptive trial to start eliminating design elements that are not preferred / do not lead to the outcome we want?
- Can this be done on a voucher? Is this project too broad for 90 hours of work?

- 90 hours of work for pre-award work; so, this may not be a good fit for a voucher
- Factorial design may be performed where some factors get dropped
- Ensures balance; gets best average power
- Marianna: From modelling perspective, how should this be done iteratively?
- Possible messy part could be a factorial design where some factors interact (if one thing is red, another doesn\x{2019}t work entirely)
- If these factors can be thought of as independent, that is best for sample size calculation
- Alvin: Similar paper w/ 72 possibilities. Bayesian (non-adaptive), so comparable framework has been done, just not in the field of interest
- Statistical simulation study may be another possibility
- Response-surface design may be comparable. Breakfast cereal industry is notable here.
- Polynomial regression estimating optimum combination of factors (optimize response surface)
- Outcome is average taste test rating
- By solving for optimum, that\x{2019}s how they decide what to market. Could be comparable.
- Per Alvin: Different \x{201c}types\x{201d} of users (field, facility type, etc) may have different preferences, which would be great to parse out
- Thermometer plot may be recommended. \x{201c}People\x{201d} plot (X out of 100) may be preferred by patients, but not preferred by Frank / Alvin / previous sample of nurses.
- Would there be biostatisticians able to work on this?
- May depend on timing. Frank may discuss with other department members.
- If design can be nailed down, that may help determine what statistical support is required.

- Fractional Factorial Design: By not having balance in every possible cell, occasionally of use

Edit | Attach | Print version | History: r1053 < r1052 < r1051 < r1050 | Backlinks | View wiki text | Edit wiki text | More topic actions

Topic revision: r1053 - 22 Feb 2024, JacksonResser

Copyright © 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

Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback