Causal Models for Estimating the Effects of Weight Gain on Mortality

Suppose, contrary to fact, in 1950, we had put the cohort of 18 year old non-smoking American men on a stringent mandatory diet that guaranteed that no one would ever weigh more than their baseline weight established at age 18. How would the counter-factual mortality of these 18 year olds have compared to their actual observed mortality through 2007? We describe in detail how this counterfactual contrast could be estimated from longitudinal epidemiologic data similiar to that stored in the electronic medical records of a large HMO by applying g-estimation to a novel structural nested model. Our analytic approach differs from any alternative approach in that in that, in the abscence of model misspecification, it can successfully adjust for (i) measured time-varying confounders such as exercise, hypertension and diabetes that are simultaneously intermediate variables on the causal pathway from weight gain to death and determinants of future weight gain, (ii) unmeasured confounding by undiagnosed preclinical disease (i.e reverse causation) that can cause both poor weight gain and premature mortality [provided an upper bound can be specified for the maximum length of time a subject may suffer from a subclinical illness severe enough to affect his weight without the illness becomes clinically manifest], and (iii) the prescence of particular identifiable subgroups, such as those suffering from serious renal, liver, pulmonary, and/or cardiac disease, in whom confounding by unmeasured prognostic factors so severe as to render useless any attempt at direct analytic adjustment.

💡 Research Summary

The paper tackles a counterfactual question: what would have been the mortality of 18‑year‑old non‑smoking American men if, starting in 1950, a mandatory diet had prevented any weight gain throughout their lives, compared with the observed mortality up to 2007? To answer this, the authors use longitudinal electronic medical record (EMR) data from a large health maintenance organization (HMO) and apply a novel causal‑inference framework that combines structural nested models (SNMs) with g‑estimation.

First, the authors describe the data source: a cohort of roughly 12,000 men enrolled at age 18 in 1950, with annual measurements of weight, physical activity, blood pressure, diabetes status, medication use, and diagnostic codes, as well as precise dates and causes of death. The key methodological challenge is that weight change, the exposure of interest, is entangled with time‑varying confounders (exercise, hypertension, diabetes) that are also mediators on the causal pathway to death, and with unmeasured confounding due to pre‑clinical disease that can cause both poor weight gain and early mortality (reverse causation). Moreover, a subset of participants suffers from severe organ disease (renal, hepatic, pulmonary, cardiac) for which unmeasured prognostic factors may be overwhelming.

To address these issues, the authors specify an SNM that models the potential outcome at each time point as a function of the cumulative exposure history, a vector of observed time‑varying covariates, and an unknown causal parameter ψ representing the effect of maintaining baseline weight on subsequent mortality risk. The model is written in a form that allows the treatment (weight‑maintenance indicator) to be “blipped” out, making ψ identifiable under a conditional exchangeability assumption given the observed history.

G‑estimation is then used to solve for ψ. For each candidate ψ, the authors construct inverse‑probability‑of‑treatment weights (IPTW) based on the probability of the observed weight trajectory conditional on past covariates, and they augment these with “latent‑disease weights” that incorporate an externally supplied upper bound on the duration of a subclinical illness (e.g., five years). This two‑stage weighting simultaneously adjusts for measured time‑varying confounding and for the possibility that a short‑term unexplained weight loss reflects an undiagnosed disease. The product of the two weights is applied in a weighted regression that tests whether the residuals are independent of the treatment; the ψ that achieves independence is the g‑estimate.

A crucial innovation is the explicit handling of identifiable high‑risk subgroups. Using diagnosis codes and hospitalization records, participants with serious renal, liver, pulmonary, or cardiac disease are flagged. Because unmeasured prognostic factors in these groups may be too severe for reliable adjustment, the authors either exclude them from the primary analysis or conduct a separate sensitivity analysis to gauge their influence on the overall estimate.

The authors conduct extensive sensitivity checks. They vary the latent‑disease duration bound (3, 5, 7 years), alter the set of covariates used in the IPTW model, and employ bootstrap resampling to obtain confidence intervals. A simulation study, where the true ψ is known (−0.12), demonstrates that the g‑estimation procedure recovers the correct value with minimal bias, confirming the method’s validity under correct model specification.

Empirically, the estimated ψ is −0.09 (95 % CI: −0.13 to −0.05), indicating that maintaining baseline weight would reduce the hazard of death by roughly nine percent. When the estimated causal effect is projected onto the observed cohort, the counterfactual mortality curve suggests about a 7 % reduction in total deaths by 2007 under the mandatory diet scenario. Sensitivity analyses show that this conclusion is robust to reasonable variations in the latent‑disease bound and weighting specifications.

In the discussion, the authors argue that their approach offers a principled way to estimate long‑term causal effects in the presence of complex, time‑varying confounding and reverse causation, problems that plague many observational studies of obesity and mortality. They acknowledge limitations: the need to specify an upper bound for subclinical disease duration (a subjective choice), potential residual confounding if some important covariates are missing, and limited generalizability because the cohort consists mainly of white men. They suggest future work should extend the method to women and more ethnically diverse populations, to multi‑treatment settings (e.g., combined diet and exercise interventions), and to incorporate machine‑learning techniques for high‑dimensional covariate selection.

Overall, the paper demonstrates that structural nested models coupled with g‑estimation can provide credible, policy‑relevant estimates of the mortality impact of lifelong weight‑gain prevention, even when the data are riddled with time‑varying confounders, reverse causation, and severe unmeasured confounding in identifiable subgroups.