PH784: Problem Set #5

Author

Koichiro Shiba

Published

November 11, 2025

Read the following article and answer the questions.

Briefly, this study estimated various “effects” of religious service attendance on mortality. They used data from the Nurses’ Health Study, a large prospective cohort study among US women. They examined both all-cause mortality and cause-specific mortality as outcomes, but we will focus on their findings for all-cause mortality.

They estimated the effects of religious service attendance in 1996 (\(A_{1996}\)) and 2000 (\(A_{2000}\)) on all-cause mortality between 2000-2012 (\(Y_{2000-2012}\)). They controlled for various covariates, including sociodemographic characteristics and health conditions in 1992 (\(L_{1992}\)) and 1996 (\(L_{1996}\)), as well as pre-baseline religious service attendance in 1992 (\(A_{1992}\)), using a method called a “marginal structural model.” Suppose their analysis is based on the DAG below.

Table 2 of the paper shows results of this analysis. For example, the estimate for “At least once per week” in 1996 and “More than once a week” in 2000 was HR = 0.50 (95% CI: 0.46-0.54). This corresponds to the following causal estimand.

\(HR=h(t)^{a_{1996}=1,a_{2000}=1} / h(t)^{a_{1996}=0,a_{2000}=0}\)

where \(t\) is time to the onset of \(Y_{2000-2012}\).


Q1. Explain in words what hypothetical interventions this causal estimand is simulating.

Your answer

Q2. Discuss whether any components of \(L_{1996}\) could be affected by \(A_{1996}\). Justify your answer.

Your answer

Q3. Can we fit a conventional outcome regression to estimate this effect? Why or why not?

Your answer


A critique suggested using the conventional Cox regression for the association between the baseline exposure \(A_{1996}\) and time to all-cause mortality through 2012 (\(Y_{1996-2012}\)), adjusting for pre-baseline covariates (\(L_{1992}\) and \(A_{1992}\)). Suppose the DAG above is correctly capturing the underlying causal structure.

Q5. When do you expect this effect and the result in Table 2 above (i.e., HR = 0.50) to diverge?

Your answer


Q6 (Optional) How did the authors motivate the use of time-varying exposure analysis via a marginal structural model? Do you find it convincing? Why? Why not?

Your answer