Date of Degree
Doctor of Public Health (DPH)
Epidemiology and Biostatistics
Epidemiology | Public Health
Causal Inference, Marginal Structural Models, Inverse Probability Weighting, Randomized Controlled Trials, Intention to Treat
Background: Unlike traditional regression used in the Intention to Treat (ITT) approach, Marginal Structural Models (MSM) can account for joint effects of baseline and subsequent treatments as well as the presence of time-dependent confounding influenced by prior treatment and selection bias due to censoring. In addition, MSMs have been theorized to be able to assist investigators in determining the overall benefit of a drug in the total population as they are able to provide a summary effect size across all strata of an effect modifier which cannot be done via tradition regression techniques. The overall goal of this dissertation is to demonstrate the advantages and disadvantages of using MSM to 1) control for time-dependent confounding and 2) detect effect modification in a randomized controlled trial (RCT) with a time-varying exposure, non-adherence and missing data.
Methods: The ITT analysis consisted of a logistic regression model linking the annual rate of acute asthma exacerbations (outcome) to assigned asthma treatment. Weights for the MSM analysis were derived from a pooled logistic regression assessing the probability of staying on assigned treatment (adherence) and, in Aim 1 and 3, of remaining uncensored for subjects at each visit by treatment arm. Poisson regression models using PROC GENMOD were fitted for the annual rate of acute asthma exacerbations (outcome) as a function of the assigned treatment using the weighted sample and a generalized estimating equation (GEE) with an independent correlation matrix in Aims 1 and 3. The final outcome model in Aim 2 also included a treatment covariate interaction term. In all aims, the final models were fit to uncensored cases with complete data.
Results: Despite the theoretical advantages of MSMs, my research found that the approach failed to invalidate previous ITT analyses, regardless of adherence level. In Aim 1, the ITT analysis found a 22% increased risk of EPACs for theophylline compared with montelukast (RR=1.22, 95% CI: 0.82-1.86, p=0.35), no increased risk between theophylline and placebo (RR=0.99, 95% CI: 0.67-1.50, p=1.00), and an 18% decreased risk of EPACs between montelukast and placebo (RR=0.82; 95% CI: 0.55-1.21, p=0.31) for the ITT approach. This was in comparison to a 24% increased risk of EPACs for theophylline compared with montelukast (RR=1.24, 95% CI: 0.83-1.84, p=0.28), no increased risk between theophylline and placebo (RR=1.01, 95% CI: 0.70-1.48, p=0.95), and a 17% decreased risk of EPACs between montelukast and placebo (RR=0.83; 95% CI: 0.57-1.19, p=0.27). In Aim 3, despite finding a statistically significant difference in adherence rates between the self-reported group and the blood assay group over time (p=0.001), adjusted rate ratios and corresponding 95% confidence intervals obtained were nearly identical and in both cases non-significant. In the self-report group, those on theophylline were 28% more likely to have an asthma exacerbation than those in the montelukast group (95% CI: 0.85-1.94, p=0.24) compared with 24% in the blood assay group (95% CI: 0.84-1.84, p=0.28). In Aim 2, the MSM analysis was able to detect effect modification by race in one of the treatment groups (montelukast). In the unadjusted analysis, non-whites were twice as likely to have an EPAC on montelukast as their white counterparts (5.75 vs. 2.66 episodes per person year, p=0.0034). Similar findings were seen for increased medication use and health care visits. Results of the MSM also indicated the presence of effect modification for overall EPACs, medication use and unscheduled health care when treated with montelukast instead of placebo. Compared with whites, non-whites were more than twice as likely to suffer from an EPAC on montelukast as on placebo (RR=2.13, 95% CI: 1.08-4.46, p= 0.04) and almost 3 times as likely to increase medication use (RR=2.86, 95% CI: 1.10-7.42, p=0.03). Non-whites were over 5 times more likely to have unscheduled health care visits than whites while on montelukast compared with placebo (RR=5.01, 95% CI: 1.36-18.97, p=0.02).
Conclusions: In theory MSMs hold much potential for further analyses of RCTs as it allows adjustment for time-varying exposures, time-dependent confounding and selection bias, issues more traditional regression based methods cannot account for. However, it remains unclear as to whether this is the case in practice. At the very least, MSMs should be conducted as a sensitivity analysis to the ITT approach in RCTs where there is preliminary evidence suggesting the presence of time-varying exposures, time-dependent confounding and/or selection bias and when MSM’s limitations can be reasonably ignored. Conducting a MSM as a sensitivity analysis of the ITT can only bolster one’s confidence of the estimated effects of treatment on an outcome. In terms of effect modification, more research is needed to determine the most appropriate way to calculate inverse probability treatment weights propensity scores as there is no consensus in the literature on how best to calculate the propensity scores required for weighting and achieve accurate subgroup results.
Lancet, Elizabeth A., "USING MARGINAL STRUCTURAL MODELS TO CONTROL FOR TIME-DEPENDENT CONFOUNDING AND DETECT EFFECT MODIFICATION IN A RANDOMIZED CONTROL TRIAL WITH A TIME-VARYING EXPOSURE, NON-ADHERENCE AND MISSING DATA" (2019). CUNY Academic Works.