Stochastic modified treatment policies for evaluating hypothetical disability interventions on dental affordability — The Association Specialists
  1. Schedule
  2. Oral Presentations - Advancing Epidemiological Methods 2
  3. Stochastic modified treatment policies for evaluating hypothetical disability interventions on dental affordability

Stochastic modified treatment policies for evaluating hypothetical disability interventions on dental affordability (23875)

Upul Cooray 1 , Gagandeep Kaur 1 , George Disney 2 , Saman Khalatbari-Soltani 1 , Barbara Janssens 3 , Ruby Cole 1 , Ankur Singh 1
  1. The University of Sydney, Camperdown, NSW, Australia
  2. Melbourne School of Population and Global Health, University of Melbourne, Melbourne
  3. Department of Oral Health Sciences, Ghent University Hospital, Ghent, Belgium

Background: A persistent challenge in causal inference with longitudinal observational data is defining realistic, policy-relevant estimands. Deterministic contrasts (all-exposed vs. all-unexposed) often violate positivity assumptions and poorly reflect achievable interventions. This study demonstrate a application of stochastic Modified Treatment Policies (MTPs) to estimate the causal impact of partial, probabilistic reductions in disability on dental unaffordability among Australian adults.

Methods: HILDA survey data from waves 18-22 ( n=9,635 working-age adults) we used. We implemented Incremental Propensity Score Interventions (IPSI) that multiplicatively scaled each individual's probability of disability by pre-specified factors; representing 25% and 50% odds reductions, rather than imposing fixed deterministic exposure values. These stochastic MTPs were contrasted with deterministic extremes (continuous disability; no disability) and the observed natural course. TMLE with a Super Learner ensemble (GLMs, GAMs, XGBoost, neural networks) with five-fold cross-fitting was used for estimation. A directed acyclic graph informed the adjustment set of baseline and time-varying confounders. Inverse probability of censoring weights were used to account for dropouts. Multivariate imputation by chained equations (five datasets) addressed missing covariate data; results were pooled using Rubin's rules. Sensitivity to unmeasured confounding was quantified using E-values.

Results: Stochastic interventions produced dose-dependent reductions in dental unaffordability: a 50% odds reduction yielded RR 0.66 (95% CI 0.59-0.75); a 25% odds reduction yielded RR 0.79 (95% CI 0.74-0.85), both compared to the natural course. Deterministic elimination of disability reduced risk by 48% (RR 0.52; 95% CI 0.40-0.67), while sustained disability increased risk by 72% (RR 1.72; 95% CI 1.27-2.33). E-values confirmed robustness to moderate unmeasured confounding.

Conclusion: Stochastic MTPs operationalise achievable population-level shifts in exposure, offering better positivity and policy interpretability compared to deterministic contrasts. This approach demonstrates that even moderate, realistic reductions in disability prevalence produce meaningful causal gains in oral health affordability.