pooling treatment groups at baseline

[wlandau]

For MMRMs that analyze raw response (instead of change from baseline), there is sometimes a desire to do what my colleagues call a "constrained longitudinal data analysis" (cLDA) where all treatment groups are pooled at baseline. The rationale is that since baseline is measured prior to dosing, we should really be observing differences among study arms yet.

For models with continuous time, this is easy to parameterize in the formula. For example, the model y ~ time + time:group automatically sets all groups equal at time = 0. But for discrete time models like MMRMs, the formula cannot help us in the same way.

The only implementation strategy that comes to mind is to manually set all groups to placebo at baseline. For the FEV data, this would look like:

suppressPackageStartupMessages(library(dplyr))
data(fev_data, package = "mmrm")
fev_data <- fev_data |>
  mutate(ARMCD = as.character(ARMCD), AVISIT = as.character(AVISIT))

distinct(fev_data, ARMCD, AVISIT)
#>   ARMCD AVISIT
#> 1   TRT   VIS1
#> 2   TRT   VIS2
#> 3   TRT   VIS3
#> 4   TRT   VIS4
#> 5   PBO   VIS1
#> 6   PBO   VIS2
#> 7   PBO   VIS3
#> 8   PBO   VIS4

clda_data <- fev_data |>
  mutate(ARMCD = ifelse(AVISIT == "VIS1", "PBO", ARMCD))

distinct(clda_data, ARMCD, AVISIT)
#>   ARMCD AVISIT
#> 1   PBO   VIS1
#> 2   TRT   VIS2
#> 3   TRT   VIS3
#> 4   TRT   VIS4
#> 5   PBO   VIS2
#> 6   PBO   VIS3
#> 7   PBO   VIS4

^{Created on 2024-06-07 with reprex v2.1.0}

Then the analysis would proceed as usual on the transformed data:

lm(FEV1 ~ ARMCD * AVISIT, data = clda_data)

Is there anything wrong with this approach? I feel like it may be frowned upon because it changes the data and not the model, but it's extremely convenient in all sorts of ways.

[wlandau]

Is there anything wrong with this approach? I feel like it may be frowned upon because it changes the data and not the model, but it's extremely convenient in all sorts of ways.

From insightsengineering/rbmi#384 and openpharma/mmrm#448, it looks like this sort of thing may already be happening in practice.

It would be better to code the model formula to handle this, but R adds one too many group:time interaction terms, c.f. openpharma/mmrm#183.

[wlandau]

I just posted a question about cLDAs for MMRMs: https://stackoverflow.com/questions/78799680/clda-constraint-for-discrete-time-models-in-r-formulas. As it turns out, cLDA is not as simple as recoding the data. Whether or not you recode the data, you would have to compute the model matrix manually, remove some columns, and then fit the model on the hacked model matrix instead of the actual data. Unless there is some clever workaround in base R with formulas or contrasts, we would not be able to use cLDA in regular usage of brms.mmrm.

We could consider it for informative prior archetypes because there brms.mmrm actually does take control of the model matrix, but this may require a lot of work.

[wlandau]

Implemented cLDA for informative prior archetypes in #126.