Generalized Odds for Ordinal Outcomes: A Causal Inference Framework Adjusting for Covariates
Background: Generalized Odds (Gen Odds) provide summaries for ordinal outcomes without requiring the proportional odds assumption. However, conventional Gen Odds compares independently drawn treatment and control outcomes and targets a population-level contrast rather than an individual-level contrast. We developed an individual-level causal formulation of Gen Odds and examined its identification. Methods: We defined the Causal Gen Odds using the joint distribution of potential outcomes and identified it under conditional independence of the potential outcomes given baseline covariates. We developed a model-based estimator using nonproportional cumulative logistic regression and introduced an odds-ratio sensitivity parameter to assess departures from conditional independence. We evaluated the estimators through simulations and illustrated them using data from the ACCORD Study. Results: In simulations, the conditional-independence estimator was approximately unbiased with 95.5% confidence interval coverage under conditional independence, but bias and undercoverage increased with residual dependence. When the sensitivity-analysis estimator incorporated the data-generating dependence parameter, bias remained below 0.01 and coverage ranged from 93.7% to 96.1% across examined settings. In the ACCORD Study, the Causal Gen Odds for right-foot pre-tibial edema was 0.909 (95% CI: 0.880 to 0.940) under conditional independence and remained stable across the examined sensitivity range. Conclusions: The Causal Gen Odds provides an individual-level estimand that complements the conventional population-level Gen Odds. This framework enables estimation of this contrast while explicitly assessing its robustness to assumptions about dependence between potential outcomes.