Ceasar Aim 2: What patient factors modify the effect of tx on function?

Study questions/goals

Main patient factors of interest are race and socioeconomic status: income, education, insurance, employment

  • Sub Aim #1: How the effect of treatment on functional outcomes differs by socioeconomic status (education and race)
  • Sub Aim #2: How the effect of treatment on functional outcomes differs by comorbidity and age
  • Sub Aim #3: How the effect of treatment on functional outcomes differs by baseline function
  • Sub Aim #4: How the effect of treatment on functional outcomes differs by disease risk (D'Amico criteria)

Questions/Discussion items

  • Review how I collapsed some socioeconomic variables.

Resolved items

  • Are we going to have separate models for each individual item? In aim 1, we looked at these nine: problems with sexual function, erection quality, problem with urinary function, problem with urinary leakage, problem with pain or burning, problem with frequency, problem with bowel function, problem with bloody stools, and problem with bowel urgency. Yes
  • For the adjusted models, what is our question of interest? Is it what is the difference between races at 12 months? Assuming that we've fully controlled for all important factors, this cross-sectional difference should implicate differences in the effect of treatment by race. Really want to focus on differences in change from baseline.
  • Review whether to have a series of two-way interactions vs. a series of three-way interactions. Discussed this with Mark. Based on the shapes we see on the unadjusted figures, we would not have a great need for three-way interaction with race, tx, and time. Won't consider 3-way interactions. Will have interaction between race and tx.
  • Consider if/how we include the propensity score in these models. Will drop prop score and directly adjust.
    • See detailed points in separate section below
    • Previous prop score included income, race, and employment, but not education.
    • I think we need to recalculate the prop score without race, employment, or income. In fact, I think we should remove all the variables that we are already controlling for in the model.
    • Need to look at numbers of missing variables and keep in mind.
  • With regard to how we define and model treatment, (how) will we handle hormone therapy? Will control for it. Will probably also to subset analysis.
    • 13% of patients had any hormone therapy.
    • In aim 1, we controlled for hormone therapy, meaning that patients kept the same tx group "membership," (surg, ebrt, or observation), but we included an additive, constant term for hormone tx.
    • Also in Aim 1, we did a sensitivity analysis excluding hormone therapy altogether.
  • We've decided to exclude Asians and other races besides whites, blacks, and Hispanics.
  • We changed categories of insurance based on MT's guidance.

Analysis planning

  • One strategy: first fit model several 2-way interactions: treatment by each of the race and each socEcon. Make some conclusions about that regarding size/rank of each of the interactions. Fit a second main model with only the important interactions.
  • Should we force a knot in the time variable to be right near time zero?
  • How each treatment effect differs between the races. Figures should show each treatment in a separate panel, with separate curves for the races.

Propensity score

  • Proposal:
    • Make a new prop model.
    • For vars that were in both the prim model and the propensity score, take them out of the new prop score.
    • For all SES variables, take them out of the prop model.
    • For any vars that were in the prop score but not in the main mod, they can be the new prop score
    • Need to think about the psa, gleason, and stage separately, because they determine the damico risk score. We controlled for damico in the prim analysis but controlled for psa, gleason, and stage in the prop model.
  • In primary analysis, we adjusted for tx, time since tx, baseline tx (spline), propensity of getting tx (spline), any hormone tx, age (spline), race, tibi, damico, sf36 physical, social support, cesd, pdm, site
  • In propensity model, we controlled for age, race, tibi, sf36 phys, social support, cesd, pdm, site, ALL of the 5 bl function scores, and sf36 mental, pc burden, pc worry, marital status, employment, income, PSA, gleason, stage.

  • Remove from the propensity model, but is already in the main model:
    • age, race, tibi, sf36 physical, social support, cesd, pdm, site,
  • Remove entirely, or consider moving to primary model. My thought was that they would be correlated with ses variables race and education
    • marital status, employment, income.
  • Remove gleason, psa, and stage from propensity model. Control for EITHER damico OR psa, gleason, and stage.
    • We are already kind of overfit.
  • What's left In new propensity model (or could be moved to main model):
    • sf36 mental, ps burden, ps worry

  • We've decided to eliminate the prop score for now.
  • Update: we decided to do a sensitivity analysis to hedge against reviewers asking for a propensity score. I think we should use the imputed data to fit this. However, the imputed data is in the long format. Earlier, we put the propensity scores on the wide data. I think now we can use the imputed data and then merge it back in to SRA long. However, then, the propensity scores would still not be in the aregImpute object.

Bother items

There were three bother items: urinary, sexual, and bowel. We fit identical models using log odds of having moderate to large bother as the outcome for each of these items. In PCOS, blacks had worse epic sexual domain scores than whites but less bother (or was that backward?). We wanted to see if that was the case here. (Are we talking about different interaction effects?? Or just about the race effects?) Have asked Mark.

Are bother score outcomes' results are consistent with the domain scores'.

Meeting notes

2015 November 4

  • In PCOS, they did a stratified analysis, so the bother scores' direction in comparison to the summary scores question cannot be answered the same way. Here we just want to check whether the bother score outcomes' results are consistent with the domain scores'.

2015 October 7

  • Need to add WEIGHTED AVERAGE tx effects to the text.
  • Sensitivity analysis: do results hold in subset of patients who got only the modern version of the tx types: IMRT and robotic, nerve sparing.
  • Importancy plot (just email the 4 / 5 plots.)
  • Talked about formatting the table.
  • Add nerve sparing to table 1. Also add overall treatment.
  • Exclude individual items.
  • We will exclude hormone domain from results except summary.
  • Make table of baseline function by race. domain scores. FOR NOW JUST DO THE UNADJ. Age-adjust. Can we standardize to OUR cohort?
  • Talked about controlling for open robot. One way to handle this is to do a sensitivity analysis. "Does surgical technique impact our results?" Exclude open AND non-nerve sparing. Discussed also excluding hormones? For now we will discuss with the group Monday.
  • Consort diagram
  • Response rate. Table this for now. We need more specs about what they want.

2015 September 28

  • Dan surprised by lack of race differences.
  • Discussed ways to make manuscript abstract easier to understand. There are too many comparisons and too many different types of comparisons.
  • talked about combining tables 2 and 3 to make 3 new tables.

2015 September 14

  • MT wants models for total control (item 15 on bl survey), which is just 1 category, and also erection sufficient for intercourse.
  • This is from epic.uf.control. We have a binary version, but we need a version that only has "total control." Must then add it to the longitudinal variables.
  • We talked about the figures: unadjusted or adjusted and the layout. MT will talk to DB tonight.

2015 September 10 JoAnn's note:

  • Compared primary analysis models with aim 2 models on the expected differences in treatment at 1 year post tx within whites (or averaged over races in the aim 2 model), and the results agreed closely. The differences between the model sets were (1) no propensity score, instead adjusted directly, (2) adjusted for psa, stage, and gleason instead of damico, and (3) new interactions.

2015 September 3 JA, MT, DB

  • DB is thinking about having an online calculator
  • DB liked the graphs with the panels for tx type and curves for race.
  • Truncate the AS curves.
  • DB prefers we report on the difference in treatment function at 12 months rather than the change from baseline.
  • Put back in the curves with treatment-education interactions
  • Add sensitivity analysis with adjusted models that exclude any hormone tx and exclude non-nerve-sparing surgeries.
  • Add epic hormone as one of the function outcomes.

2015 August 31 TK, MT, JA

  • Explained lots of raw model output
  • talked about how to group some variables: change insurance groupings to: none and medicaid, other and va/military, etc
  • Send MT a list of variables from which to choose defaults for adjusted figures and possibly table values, etc. He wants the bl function set to the average.
  • MT is thinking about including the plot of y_t ~ y_0 in the appendix.

2015 August 17

  • Dave wants to include income and retirement status, and insurance status.
  • Make correlation table with all these factors. Made graphs showing relative frequencies. Did redundancy analysis and verdict is to keep all. However, excluding income because 400 missing.
  • We can throw away the Asians and others. Done.
  • Check how marital status is coded This was asked as a 5-category question with married, never married, etc. There wasn't anything about a "commited relationship."
  • Yay! Ok, Mark's objective iss to comment on the (racial differences in) change from baseline rather than racial differences in the function at a particular time. With that in mind, we can use the same model (with y_t only post tx), but show the estimated change from baseline (for different specific baseline values, and also give the difference in differences, which is the interaction term).
  • Mark will send us the draft of the results.

2015 July 27 (JoAnn and Mark Tyson)

  • I had sent plots of predicted function over time by tx and race, by tx and edu, by tx and employment, by tx and income, and by tx and insurance, based on several "univariate" models with tx*timeSpline*3rd variable.
  • We looked at these plots in detail and talked about what these models mean in terms of not being adjusted for other factors.
  • talked about which socioeconomic factors to include in adjusted models. MT says race and education.
  • I explained the difference in modeling race with an additive term, with an interaction with tx, and with a 3-way interaction with tx and time.
  • Discussed some rationales for the latter 2 (based on prior knowledge/expectation of whether the tx by race interaction is a change in the level or a change in the rate of change over time, or based on the shapes we saw on the unadjusted trajectory plots).
  • Mark asked me to take a look at his draft of the intro.
  • talked about the propensity score and possibly changing it.
  • Mark said he will work on sketching the structure of the results section (like subheadings)

2015 July 20

  • Talked about which interactions to include.

2015 July 13

  • Dan wants to meet with Lynne and me about how to format a data dictionary for later investigators
  • To do: send Dan a data set with all the patient ids, site, and primary tx recieved done
  • Talked some about how to define the different variables like socEcon status. Mark will be thinking about this.
  • talked a lot about definition of treatment. How many patients switched tx in the first year.
  • DB says that both the ITT analysis and one that accounts for different adjuvant tx have clinical relevance.

-- JoAnnAlvarez - 13 Jul 2015
Topic revision: r26 - 05 Nov 2015, JoAnnAlvarez
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