Running head: CAUSAL INFERENCES FROM LONGITUDINAL DATA

Analyses of passive longitudinal data can yield causally relevant evidence only to the extent that plausible alternative explanations are ruled out. This study compared the ability of five types of longitudinal analyses to correct for selection biases confounded with corrective interventions, using a cohort of 1464 4and 5-year-olds from Canadian NLSCY data. Three lines of evidence indicated that apparent child outcomes of corrective interventions were due to selection biases and accompanying regression artifacts. First, all significant effects of corrective interventions indicated apparently detrimental effects when predicting residualized change scores, but apparently beneficial effects when predicting simple change scores. This was true whether analyzing measured or latent variables. Second, results from temporally reversed analyses were consistent with selection and regression artifacts, not with unidirectional causal effects. Third, the findings were similar for empirically supported interventions (e.g., Ritalin, psychotherapy) and for parental interventions considered controversial (spanking, yelling). Typical longitudinal data may suppress the ability to detect causal effects due to lengthy inter-wave intervals and temporal overlap of covariates and causes. When applicable, longitudinal analyses should check for similar artifacts by implementing temporally reversed analyses and by determining whether the substantive results replicate without artifacts biased in their favor.