Improving Causal Inferences in Meta‐analyses of Longitudinal Studies: Spanking as an Illustration

Abstract

To evaluate and improve the validity of causal inferences from meta-analyses of longitudinal studies, two adjustments for Time-1 outcome scores and a temporally backwards test are demonstrated. Causal inferences would be supported by robust results across both adjustment methods, distinct from results run backwards. A systematic strategy for evaluating potential confounds is also introduced. The methods are illustrated by assessing the impact of spanking on subsequent externalizing problems (child age: 18 months to 11 years). Significant results indicated a small risk or a small benefit of spanking, depending on the adjustment method. These meta-analytic methods are applicable for research on alternatives to spanking and other developmental science topics. The underlying principles can also improve causal inferences in individual studies.

Does community violence exposure moderate the associations between maternal spanking and early child behavior problems?

“…The magnitude of these effect sizes is modest, but comparable to that reported in extant literature (Gershoff & Grogan‐Kaylor, 2016a). A recent meta‐analytic review of the literature on the effects of spanking has contended that most prior findings have failed to control for behavioral issues at baseline and alternative factors that predict spanking and child outcomes (Larzelere et al, 2018). The effect sizes reported herein from fixed effects regression eliminate these possible methodological flaws by controlling for pre‐existing behavior problems of children and all time‐invariant confounds.…”

“…Interwoven with questions regarding the potential influence of contextual factors on parenting practices are concerns about the degree to which the observed effects of parental spanking and of a disadvantaged social context on child behavior are truly causal in nature, or are instead associations that could be attributed to the presence of other unobserved, heritable family characteristics that may have confounding relations with parenting behavior and child outcomes (Larzelere et al, 2018). Behavioral genetic studies have demonstrated the correlation of aggressive tendencies in the child and the parent (Arseneault et al, 2003).…”

Meta‐analysis and Individual Participant Data Synthesis in Child Development: Introduction to the Special Section

“…One clear message of the articles presented here is that the quantitative synthesis of developmental data is a domain‐general enterprise . Indeed, the articles featured in this special section cover most of the full range of modern developmental science, including developmental cognitive neuroscience (Yaple & Arsalidou, ), language development (Bergmann et al., ; Quinn & Wagner, ), numerical cognition (Hornburg, Wang, & McNeil, ), social and emotional development (Larzelere et al., ; Verhage et al., ), intervention and prevention science (Gardner et al., ; Leijten et al., ), and even generational‐ and age‐related change in children's stereotypes and behavior (Miller et al., ). In short, it would be a mistake to claim that meta‐analysis is only appropriate for select domains of developmental science.…”

“…Likewise, Larzelere et al. () introduced methods for improving (albeit incrementally) on causally ambiguous correlations with nonexperimental data and, in doing so, questioned the magnitude and causal significance of the associations documented in the literature on parental spanking and peer‐directed aggressive behavior.…”

“…One primary emphasis was on methods of meta‐analysis as applied to developmental questions and we thus welcomed articles illustrating these methods with substantive developmental examples (alongside the publication of extensive electronic supplements to maximize reproducibility and future extension to other domains). In the developmental field specific questions arise, for example, about how to conduct meta‐analyses on cross‐sectional and longitudinal studies (e.g., Larzelere, Gunnoe, & Ferguson, ; Leijten et al., ; Miller et al., ), maximize the quality of causal inferences (Gardner et al., ; Larzelere et al., ; Leijten et al., ), and go beyond bivariate associations to testing more complete and realistic developmental models (Quinn & Wagner, ). Articles in this special section touch on these and related topics.…”