Joint effects in cross-lagged panel research using structural nested mean models

A popular approach among psychological researchers for investigating causal relationships from panel data is cross-lagged panel modeling within the structural equation modeling (SEM) framework. However, SEM models are critiqued in the causal inference literature for relying on unnecessarily many parametric assumptions, increasing the risk of model misspecification and bias. Instead, the use of structural nested mean models (SNMMs) with G-estimation is promoted as an approach that relies on fewer assumptions and therefore, in principle, leads to more valid causal conclusions. However, the uptake of SNMMs and G-estimation in the psychological literature is lacking, hampered by a disconnect between the causal inference literature, and the modeling practices that psychological researchers are familiar with. We bridge this divide by introducing joint effects, linear SNMMs, and G-estimation in the context of cross-lagged panel research, and comparing these to cross-lagged panel modeling approaches from SEM. A substantive example from psychological practice is used throughout.