OBJECTIVES : Big data provide an opportunity to uncover complex networks of causal relationships that can be used to inform further advances in clinical care. Structural equation modelling (SEM) is a potential solution to quantitatively test models of relationships within these large datasets. Piecewise SEM offers more flexibility than traditional SEM by accommodating a wider range of response types in variable networks. The objective of this study is to provide a comparison and validation of piecewise SEM compared to traditional SEM.

METHODS : Simulated data were generated for a causal network accommodating 8 variables arranged in a causal chain, simulating a series of interventions and outcomes over time. In one scenario all outcome data were normally distributed, while for another a single outcome with a Poisson distribution was introduced. Datasets of 50 cases were simulated 1,000 times and both traditional and piecewise SEMs fit. The methods were compared in terms of their bias (difference between the average estimated parameter values and the underlying true population value).

RESULTS : The bias for both traditional and piecewise SEMs were nearly equal for the first scenario, with an average percentage bias of 2% for the traditional SEM and 1% for the piecewise SEM, indicating that both methods accurately predicted the data. In the second scenario, both traditional and piecewise SEMs were nearly equal for the normally distributed data, with average percentage biases of 7% and 6% overall, but the bias was larger for the Poisson distribution data (28% and 2%, respectively).

CONCLUSIONS : The results show that the piecewise SEM provides a nearly equivalent solution for compared to traditional SEM when using linear regressions; however, the piecewise SEM outperforms the traditional when counts are introduced as outcomes. The piecewise SEM appears to be a reliable alternative to a traditional SEM when modelling a non-linear relationship is necessary.