- see [[230502_143020 decay effects for first 8 days - condition-day interaction|condition-day interaction or decay effects]] # fixed-effects meta-analysis (3 experiments' effect sizes) - frequentist model for each campaign/experiment: `t1 ~ condition * t0`, block and day fixed effects - Bayesian meta priors: t-distribution, `t(degrees_of_freedom, mean = 0, scale = 0.05)`, where `degrees_of_freedom` is 3, 5, 15, or 30 ![[_meta_priors_term-conditionC_outcome-count_rfeffect-fixed 2.png]] ![[_meta_priors_term-conditionC_outcome-count_rfeffect-fixed 3.png]] subtract Bayesian posterior samples: count badness minus count goodness posterior samples/estimates ![[bad-minus-good-count 1.png]] ![[_meta_priors_term-conditionC_outcome-sum_rfeffect-fixed 2.png]] ![[_meta_priors_term-conditionC_outcome-frac_rfeffect-fixed.png]] ## example priors ### prior_df = 3 ![[meta_term-conditionC_prior-03_outcome-count_threshold-80_rfeffect-fixed.png]] ### prior_df = 5 ![[meta_term-conditionC_prior-05_outcome-count_threshold-80_rfeffect-fixed.png]] ### prior_df = 15 ![[meta_term-conditionC_prior-15_outcome-count_threshold-80_rfeffect-fixed.png]] ### prior_df = 30 ![[meta_term-conditionC_prior-30_outcome-count_threshold-80_rfeffect-fixed.png]] # differences in pre-campaign activity across experiments Variable: pre-campaign proportion of days with at least 1 tweet/retweet/quote/reply. Experiment 2 had many users with relatively little activity before the campaign. Distribution looks very different from the other experiments, so should it's excluded from the meta-analysis. ![[s20230410_180353.png]]