Target trial emulation  ·  NZAVS 2018–2023  ·  Lab report

Religious Attendance: Selective, Not Sweeping

Target trial emulation shows that supported causal effects of religious attendance on well-being are selective

Joseph A. Bulbulia · Don E. Davis · Crystal Park · Kenneth G. Rice · Geoffrey Troughton · Daryl R. Van Tongeren · Chris G. SibleyEvolutionary Human Sciences · 2026

Religious service attendance is associated with better well-being, but observational associations do not establish causation. Across six annual waves of a national panel, we estimate the causal effects of monthly attendance on 24 well-being indicators — and the supported effects concentrate in meaning, forgiveness, and intimate life, with little change in somatic health, psychological distress, or social support.

46,377participants at baseline, NZAVS
6annual waves, 2018–2023
24well-being indicators
2–3%of non-attenders initiate attendance per year

The emulated trial

What is our question?

For New Zealand adults, what are the causal effects of monthly religious service attendance on 24 well-being indicators, measured across somatic, psychological, existential, social, and relational domains?

What is our population?

46,377 NZAVS participants at baseline, followed across six annual waves, 2018–2023. The primary contrast concerns the 38,477 who were not attending at baseline, with censoring adjustment carrying inference to that cohort.

What interventions do we compare?

Fails positivity

Make everyone attend monthly, versus none. Only 2–3% of non-attenders initiate attendance in a given year; the data cannot identify a policy this far from what occurs.

Adopted · δ = 5

Multiply each non-attender's chance of taking up monthly attendance fivefold, versus the natural course — the censoring-adjusted course actually observed.

The central result

Small, supported gains appear in meaning and purpose, forgiveness, and sexual satisfaction; distress and somatic health barely change.

Estimates come from the supported stochastic contrast above, fitted with a sequentially doubly robust estimator and cross-validated machine learning among baseline non-attenders. Cross-sectional associations between attendance and well-being are broadly favourable across most outcomes; the causal contrast narrows this profile, and the narrowing reflects both confounding adjustment and the change in causal question.

Supported · where attendance acts

Meaning, forgiveness, intimacy

Meaning and purpose rises by 0.085 standard deviations (E-value 1.25), sexual satisfaction by 0.086 (1.25), and forgiveness by 0.075 (1.21). Smaller supported gains appear for meaning and sense (0.071), short-form health (0.061), and body satisfaction (0.058).

Null · where it does not

Distress and the body barely change

Body mass index, sleep, and alcohol use, and the distress indicators — anxiety, depression, fatigue, rumination — show point estimates near zero with confidence intervals spanning the null. Social belonging and perceived social support do not meet the E-value reliability threshold.

The estimates

meets the E-value reliability threshold interval excludes zero; E-value modest -0.05 0 +0.05 +0.10 sexual satisfaction: 0.086 [0.046, 0.126], E-value bound 1.25 SEXUAL SATISFACTION +0.086 · E 1.25 meaning and purpose: 0.085 [0.044, 0.126], E-value bound 1.25 MEANING AND PURPOSE +0.085 · E 1.25 forgiveness: 0.075 [0.034, 0.116], E-value bound 1.21 FORGIVENESS +0.075 · E 1.21 meaning and sense: 0.071 [0.025, 0.117], E-value bound 1.18 MEANING AND SENSE +0.071 · E 1.18 short-form health: 0.061 [0.022, 0.100], E-value bound 1.16 SHORT-FORM HEALTH +0.061 · E 1.16 body satisfaction: 0.058 [0.021, 0.095], E-value bound 1.16 BODY SATISFACTION +0.058 · E 1.16 gratitude: 0.046 [0.002, 0.090], E-value bound 1.05 GRATITUDE +0.046 · E 1.05 body mass index: -0.031 [-0.060, -0.002], E-value bound 1.05 BODY MASS INDEX -0.031 · E 1.05

Standardised effects of the δ = 5 supported stochastic intervention among baseline non-attenders, with 95% confidence intervals, as reported in the paper. Green marks effects meeting the E-value reliability threshold; gold marks gratitude and body mass index, whose intervals exclude zero but whose E-value bounds fall short of it. The remaining outcomes cluster near zero.

Inference & robustness

01

Estimation

A sequentially doubly robust estimator with cross-validated machine learning adjusts for 62 baseline covariates and 32 time-varying confounders, with inverse-probability-of-censoring weights for dropout.

02

Sensitivity

Each supported effect clears the study's E-value reliability threshold (E-values 1.16–1.25); an unmeasured confounder would need associations of that strength with both attendance and outcome to explain the effect away.

03

Measurement

Attendance is self-reported, and measurement error in self-reports would pull estimates towards zero — actual effects may be larger than we report.

04

Scope

We estimate total effects of specified changes in attendance rather than mediated pathways, and population averages rather than subgroup effects. Effects probably vary by tradition, belief strength, upbringing, and relationship status; childhood religious socialisation lies outside what adult panel data can identify.

Citation

Bulbulia, J. A., Davis, D. E., Park, C., Rice, K. G., Troughton, G., Van Tongeren, D. R., & Sibley, C. G. (2026). Target trial emulation shows that supported causal effects of religious attendance on well-being are selective. Evolutionary Human Sciences, 8, e16. doi:10.1017/ehs.2026.10043

Data: the New Zealand Attitudes and Values Study. Estimation: supported stochastic interventions with a sequentially doubly robust estimator.