Linear growth deficit and risk of hypertension among the elderly: a generalized propensity score evaluation based on knee-heel length

Linear growth deficit, reflected by reduced knee-heel length, may represent early-life nutritional and developmental constraints with long-term implications for cardiovascular health. This study evaluated the association after adjustment for measured confounders between linear growth deficit and the risk of hypertension among elderly individuals using a generalized propensity score framework. Generalized propensity scores (GPS) were estimated using four approaches: ordinal logistic regression (OLR) model and machine learning models [XGBoost, ordinal random forest (ORF), and ordinal support vector machine (OSVM)]. Covariate balance after weighting was assessed using standardized mean differences and variance ratios to examine the extent to which the methods reduced baseline imbalance across growth categories. The performance of the various models were evaluated in terms of weighting structures, overlap and covariate balance as well as average treatment effect (ATE). Pattern across ordered knee-heel length (KHL) categories were estimated using marginal structural model (MSM) and covariates adjusted with GPS regression model. The Mexican national health surveys data on risk of hypertension was adopted for the study with both outcome models revealing a nonlinear relationship between linear growth deficit and risk of hypertension. The individuals in the optimal linear growth category showed lower predicted risk of hypertension compare to other categories. In terms of weighing structure, ORF and OSVM had more stable weight, however, stabilized-trimmed IPW provided the most controlled and stable weight distribution across all four models whereas XGBoost achieved better overlap, covariate balance and yielded more stable ATE estimates than other models. Overall, the study demonstrates that KHL may serve as a useful life-course marker for assessing hypertension risk in later life. The application of GPS methods also provides a robust framework for evaluating ordered exposure effects while addressing confounding in observational health data.