Question

If you are assessing depression, anxiety, and substance use before and after people disclose their HIV status, what analysis should be run to analyze the data? Before and after measurements are the same people.

Answer 1

Variables are continuous.
Not normally distributed data.

Answer 2

Because the measurements are repeated on the same people (pre and post disclosure), you need a paired/repeated-measures analysis. The exact test/model depends on the type of outcome (continuous, ordinal, binary, count), whether you need to adjust for covariates, and how you want to handle missing data. Short decision guide and recommended analyses:

1. If your outcomes are continuous scores (e.g., depression scale, anxiety scale, substance-use score)
- Simple: paired t-test (if the difference is approximately normally distributed)
- R: t.test(post, pre, paired = TRUE)
- Nonparametric alternative: Wilcoxon signed-rank test (if differences are nonnormal or small sample)
- R: wilcox.test(post, pre, paired = TRUE)
- To adjust for covariates or handle missing data: use a linear mixed‑effects model (random intercept for subject)
- R: library(lme4); lmer(score ~ time + covariates + (1|id), data = df)
- You can also analyze change scores (delta = post − pre) with linear regression: delta ~ covariates.

2. If outcomes are binary (e.g., any substance use: yes/no)
- Simple paired test: McNemar’s test
- R: mcnemar.test(table(pre_bin, post_bin))
- For adjustment and modeling: conditional logistic regression or generalized linear mixed model (GLMM) / GEE with subject ID as cluster
- R (GLMM): glmer(binary ~ time + covariates + (1|id), family = binomial, data = df)
- R (GEE): geeglm(binary ~ time + covariates, id = id, family = binomial, corstr = "exchangeable")

3. If outcomes are ordinal or counts
- Ordinal: Wilcoxon for two time points or ordinal mixed models (clmm in ordinal package)
- Counts: paired Poisson/negative binomial models or GLMM/GEE with appropriate family

4. Multiple outcomes (depression, anxiety, substance use)
- You can run separate paired tests/models for each outcome but correct for multiple comparisons (Bonferroni, Holm, or report false discovery rate).
- Or analyze them jointly with multivariate repeated-measures methods (repeated‑measures MANOVA if assumptions hold) or multivariate/multiple-outcome mixed models.

Other practical points
- Check assumptions (normality of differences, variance, etc.), report effect sizes (paired Cohen’s d = mean(diff)/sd(diff)) and 95% CIs for the mean difference.
- If you have dropouts/missing post data, mixed models or GEE are generally preferable because they handle missing-at-random data better than complete-case paired tests.
- If you want causal inference beyond within-person change, consider potential confounders and use models that adjust for them.

If you tell me the variable types for each outcome (continuous score vs. binary), sample size, whether you need covariate adjustment, and what software you’re using, I can give the exact recommended test and example code.