Introduction to estimating a marginal estimand with beeca

What is an average treatment effect?

In general, an average treatment effect is the effect of moving a population from untreated (control treatment) to treated (experimental treatment). An average treatment effect can be estimated either with or without covariate adjustment.

Different flavours of average treatment effect exist, depending on the specific population that is moved from untreated to treated. For an overall superpopulation (denoted S), and for a given sample of patients in the trial (who are assumed to be randomly sampled from the superpopulation), we define the following potential estimands.

1. Population Average Treatment Effect PATE^S. The average outcome of everyone in the superpopulation if they were treated with the experimental treatment vs. the average outcome if they were treated with the control treatment.
2. Conditional Average Treatment Effect, CATE^S(x). The average outcome of everyone in the superpopulation with covariate value x if they were treated with the experimental treatment vs. the average outcome if they were treated with the control treatment.
3. Conditional Population Average Treatment Effect CPATE^S. The average of CATE^S(x_i) for the patients i = 1, …, N in the trial sample. Can also be thought of as the expected difference in outcome means (experimental vs. control), conditional on the baseline covariates x₁, …, x_N.
4. Sample Average Treatment Effect SATE. The average outcome of everyone in the trial if they were treated with the experimental treatment vs. the average outcome if they were treated with the control treatment.

In the FDA guidance, and in other recent discussions on covariate adjustment, average treatment effects are often classified as either unconditional or conditional. (Sometimes marginal is used interchangeably with unconditional). In this context, what is usually meant by “conditional” is CATE^S(x). That is, a subgroup specific average treatment effect. Often, an assumption is made that subgroup specific average treatment effects are identical across subgroups. If this assumption is wrong, then interpretation can be challenging. What is usually meant by “unconditional” (or “marginal”) is an average of subgroup-specific average treatment effects. This second averaging step gives corresponding estimation procedures a level of robustness to model misspecification. Hence the following distinction in the FDA guidance:

1. Sponsors can perform covariate-adjusted estimation and inference for an unconditional treatment effect in the primary analysis of data from a randomized trial.
2. Sponsors should discuss with the relevant review divisions specific proposals in a protocol or statistical analysis plan containing nonlinear regression to estimate conditional treatment effects for the primary analysis.

Under the coarse classification of average treatment effects as either “unconditional / marginal” or “conditional”, the PATE^S, CPATE^S and SATE can be viewed as “unconditional / marginal”, while the CATE^S(x) can be viewed as “conditional”.

How to estimate “unconditional / marginal” average treatment effect?

For a binary endpoint, an average treatment effect, whether it be a PATE^S, CPATE^S or SATE, can be estimated from an adjusted logistic regression using the g-computation method. That is, by estimating the average counterfactual outcomes on each arm based on predictions from the working model. Based on these average counterfactual outcomes, different summary measure such as risk difference, relative risk or odds ratio can be evaluated.

Regardless of whether the estimand is PATE^S, CPATE^S or SATE, the same point estimator can be used.

How to estimate the variance of the g-computation estimator of average treatment effect?

To avoid ambiguity regarding how the variance of the g-computation estimator should be calculated, it is necessary to be precise about whether one is estimating the PATE^S, CPATE^S or SATE.

{beeca} implements variance estimation for PATE^S based on the methods of Ye et al. (2023), and CPATE^S based on Ge et al. (2011).

See Magirr et al. (2024) for discussion of the estimands PATE^S and CPATE^S, and the corresponding methodology for variance estimation.

Ge et al (2011)

# pre-process trial01 dataset to convert treatment arm to a factor and handle missing value
data01 <- trial01 |>
  transform(trtp = as.factor(trtp)) |>
  dplyr::filter(!is.na(aval))
fit <- glm(aval ~ trtp + bl_cov, family = "binomial", data = data01)
beeca_ge <- get_marginal_effect(object = fit, trt = "trtp", method = "Ge", 
                                contrast = "diff", reference = "0", type = "model-based")
cat("Point estimate", beeca_ge$marginal_est, "\nStandard error estimate", beeca_ge$marginal_se)

## Point estimate -0.06836399 
## Standard error estimate 0.06071641

Alternative version 1: code provided in Ge et al. (2011) paper. Note the corresponding SAS code is also available in the paper.

ge_var_paper <- function(glmfit, trt) {
  pder <- function(ahat, vc, x) {
    #### ahat: logistic regression parameters
    #### vc: variance-covariance matrix of ahat
    #### x: full model matrix of the logistic regression
    #### return mean of phat on x and its se
    phat <- plogis(x %*% ahat)
    pbar <- mean(phat)
    pderiv <- t(phat * (1 - phat)) %*% x / nrow(x)
    sepbar <- sqrt(pderiv %*% vc %*% t(pderiv))
    return(list(pbar = pbar, sepbar = sepbar, pderiv = pderiv))
  }

  difP <- function(glmfit) {
    #### estimate the proportion difference and its standard error
    df <- glmfit$model

    vc <- vcov(glmfit)
    df[, trt] <- 1
    mat <- model.matrix(glmfit$formula, data = df)
    pderT <- pder(coef(glmfit), vc, mat)
    df[, trt] <- 0
    mat <- model.matrix(glmfit$formula, data = df)
    pderC <- pder(coef(glmfit), vc, mat)

    difb <- pderT$pbar - pderC$pbar
    sedif <- sqrt((pderT$pderiv - pderC$pderiv) %*% vc %*% t(pderT$pderiv - pderC$pderiv))

    return(list(
      pT = pderT$pbar,
      pC = pderC$pbar,
      dif = difb,
      sedif = sedif,
      var = sedif**2
    ))
  }

  return(list(est = difP(glmfit)$dif, se = difP(glmfit)$sedif[[1]]))
}
paper_ge <- ge_var_paper(fit, "trtp")
cat("Point estimate", paper_ge$est, "\nStandard error estimate", paper_ge$se)

## Point estimate -0.06836399 
## Standard error estimate 0.06071641

Version 2: using {margins} package

if (requireNamespace("margins", quietly = T)) {
  margins_ge <- margins::margins(model = fit, variables = "trtp", vcov = vcov(fit))
  cat("Point estimate", summary(margins_ge)$AME, "\nStandard error estimate", summary(margins_ge)$SE)
}

## Point estimate -0.06836399 
## Standard error estimate 0.06071641

Version 3: using {marginaleffects} package

if (requireNamespace("marginaleffects", quietly = T)) {
  marginaleffects_ge <- marginaleffects::avg_comparisons(fit, variables = "trtp")
  cat("Point estimate", marginaleffects_ge$estimate, "\nStandard error estimate", marginaleffects_ge$std.error)
}

## Point estimate -0.06836399 
## Standard error estimate 0.06071641

Version 4: using SAS %Margins macro

%Margins(data      = myWork.trial01,
         class     = trtp,
         classgref = first, /*Set reference to first level*/
         response  = avaln,
         roptions  = event='1', 
         dist      = binomial,  
         model     = trtp bl_cov,
         margins   = trtp, 
         options   = cl diff reverse, 
         link      = logit);
    
** Store output data sets ; 
data myWork.margins_trt_estimates;
  set work._MARGINS;
run;
         
data myWork.margins_trt_diffs;
  set work._DIFFSPM;
run;

cat("Point estimate", margins_trial01$Estimate, "\nStandard error estimate", margins_trial01$StdErr)

## Point estimate -0.06836399 
## Standard error estimate 0.06071641

Ye et al (2023)

beeca_ye <- get_marginal_effect(object = fit, trt = "trtp", method = "Ye", 
                                contrast = "diff", reference = "0")
cat("Point estimate", beeca_ye$marginal_est, "\nStandard error estimate", beeca_ye$marginal_se)

## Point estimate -0.06836399 
## Standard error estimate 0.0608836

Version 1: from RobinCar package

if (requireNamespace("RobinCar", quietly = T)) {
  robincar_ye <- RobinCar::robincar_glm(data.frame(fit$data), response_col = as.character(fit$formula[2]),
      treat_col = "trtp", formula = fit$formula, g_family = fit$family,
      contrast_h = "diff")$contrast$result
  cat("Point estimate", robincar_ye$estimate, "\nStandard error estimate", robincar_ye$se)
}

## Point estimate -0.06836399 
## Standard error estimate 0.0608836

References

• Arel-Bundock V (2023). marginaleffects: Predictions, Comparisons, Slopes, Marginal Means, and Hypothesis Tests. R package version 0.14.0, https://marginaleffects.com/.

• Bannick M, Ye T, Yi Y, Bian F (2024). RobinCar: Robust Estimation and Inference in Covariate-Adaptive Randomization. R package version 0.2.0, https://CRAN.R-project.org/package=RobinCar.

• European Medicines Agency. 2020. “ICH E9 (R1) addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials.” https://www.ema.europa.eu/en/documents/scientific-guideline/ich-e9-r1-addendum-estimands-and-sensitivity-analysis-clinical-trials-guideline-statistical-principles-clinical-trials-step-5_en.pdf

• FDA. 2023. “Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products. Final Guidance for Industry.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/adjusting-covariates-randomized-clinical-trials-drugs-and-biological-products

• Ge, Miaomiao, L Kathryn Durham, R Daniel Meyer, Wangang Xie, and Neal Thomas. 2011. “Covariate-Adjusted Difference in Proportions from Clinical Trials Using Logistic Regression and Weighted Risk Differences.” Drug Information Journal: DIJ/Drug Information Association 45: 481–93.

• Magirr, Dominic, Mark Baillie, Craig Wang, and Alexander Przybylski. 2024. “Estimating the Variance of Covariate-Adjusted Estimators of Average Treatment Effects in Clinical Trials with Binary Endpoints.” OSF. May 16. osf.io/9mp58.

• SAS Institute Inc. “Predictive margins and average marginal effects.” https://support.sas.com/kb/63/038.html (Last Published: 13 Dec 2023)

• Thomas J. Leeper (2021). margins: Marginal Effects for Model Objects. R package version 0.3.26.

• Ye, Ting, Marlena Bannick, Yanyao Yi, and Jun Shao. 2023. Robust Variance Estimation for Covariate-Adjusted Unconditional Treatment Effect in Randomized Clinical Trials with Binary Outcomes.” Statistical Theory and Related Fields 7 (2): 159–63.