@techreport{ai26-1394, title = {{Why Fadeout is (Probably) Worse Than We Think: Adjusting for Correlated Sampling Error in Meta-Analyses of Behavioral Interventions}}, author = {Joshua B. Gilbert and Zachary Himmelsbach}, institution = {Annenberg Institute at Brown University}, type = {EdWorkingPaper}, number = {26-1394}, year = {2026}, month = {July}, doi = {10.26300/87r9-qm15}, url = {https://edworkingpapers.com/ai26-1394}, abstract = {The extent to which intervention effects persist or fade over time is an important question in the behavioral sciences. In meta-analysis, persistence is often assessed by meta-regressing effect sizes at followup on effect sizes at endline. While common, the standard meta-regression does not adjust for the shared sampling error between effect sizes across time points. We show that in general, estimated slopes from the standard meta-regression are inflated under mild assumptions about correlations between outcomes across time. We show how to adjust for correlated sampling error using a sensitivity analysis approach with meta-analytic data from a series of behavioral interventions. Our results suggest that effect fadeout is likely more severe than current estimates suggest.}, }