Abstract / Summary
Abstract Background Digital inclusion is widely promoted as a mental-health prevention strategy for middle-aged and older adults, yet the evidence is inconsistent and prone to selection bias, reverse causation and other methodological problems. We tested whether any protective estimate survives methods that address several of these biases. Methods We reanalysed the China Health and Retirement Longitudinal Study (CHARLS, 2011–2018) among community-dwelling adults aged ≥ 45 years who reported no internet use at baseline and scored below 10 on the 10-item Center for Epidemiologic Studies Depression Scale (CES-D-10). We estimated the association under several designs — target-trial emulation with clone-censor-weighting and inverse probability of censoring weighting, within-person fixed-effects models with lead and placebo-lead tests of temporal direction, and linear mixed models — and probed the estimates with negative-control outcomes and E-values for unmeasured confounding. The primary outcome was annual CES-D-10 change; the secondary, incident elevated depressive symptoms (CES-D-10 ≥ 10). Results Cross-sectionally, internet users appeared less depressed (crude difference − 3.07 points), but the association attenuated with covariate adjustment and was null under within-person fixed effects (+ 0.096; 95% confidence interval [CI] − 0.148 to 0.341; p = 0.440). A linear mixed model found symptoms rising slightly faster among users (+ 0.094 points/year; 95% CI 0.014–0.174; p = 0.021). In target-trial emulation, initiators had a lower hazard of elevated symptoms (per-protocol hazard ratio 0.699; 95% CI 0.559–0.875). Within-person lead analyses pointed in the opposite direction: internet use preceded later symptom increases (+ 0.50; p = 0.01), the placebo-lead was null ( p = 0.68), and the estimate was largest where reverse causation was least plausible (baseline CES-D-10 < 5; +0.83; p < 0.001). The negative-control outcome, arthritis, showed the same protective direction (adjusted odds ratio 0.51–0.75) — a generic selection artefact rather than a depression-specific causal pathway. Conclusions The protective association between internet use and lower depressive symptoms is a selection artefact, mirrored by a negative-control outcome. A modest within-person residual, directionally consistent across designs, is compatible with harm but fragile to unmeasured time-varying confounding (robustness values < 0.005). Internet use should not be recommended to prevent depressive symptoms, and digital-inclusion programmes should not be promoted on selection-confounded evidence — not because harm is proven, but because protection is not.