Methods/Statistics
Marginal Structural Bayesian Relevant Life-course Model: A Novel Approach to Evaluate Life-course Hypotheses in the Presence of Time-varying Confounding Yinxian Chen* Yinxian Chen Emory University
Current life-course models, such as the structural life-course modeling approach (SLCMA) and the Bayesian Relevant Life-course Model (BRLM), are developed to assess life-course hypotheses that best explain the data, thereby informing optimal intervention strategies over the lifespan. Yet, they rely on baseline adjustment and cannot address time-varying confounding, risking biased identification of the best life-course hypothesis. The marginal structural model (MSM) can address time-varying confounding but cannot assess the compatibility of the data with given life-course hypotheses. Thus, we developed a marginal structural BRLM (MSBRLM). Like BRLM, MSBRLM estimates the effect of a life-course exposure and the importance of each period, while controlling for confounding via inverse probability weighting and computing valid 95% credible intervals (CrIs) via weighted finite population Bayesian bootstrapping. We generated simulated data with a time-varying exposure and covariate at three time points under the accumulation hypothesis (i.e., equal importance across periods). Under all scenarios, MSBRLM yielded slightly higher bias in life-course effect estimates than MSM, but substantially lower than BRLM. Under strong time-varying confounding, MSBRLM achieved 98% accuracy in identifying the correct hypothesis, whereas BRLM and SLCMA achieved 63% and 61%, respectively. Using data from the Future of Families and Child Well-being study, we estimated the life-course effect of neglect from ages 3 to 9 on the externalizing behavior (age 15). By accounting for time-varying confounding by prior externalizing behavior, the estimated effect from MSBRLM (θ=2.56, 95% CrI: 0.80, 4.32) was similar to MSM (θ=2.60, 95% CI: 1.22, 3.98). MSBRLM identified accumulation as the best hypothesis, while BRLM and SLCMA suggested recency (i.e., recent exposure matters more). MSBRLM could be a promising tool for assessing life-course hypotheses when time-varying confounding is present.
