Building causal connections among job accessibility, employment, income, and auto ownership using structural equation modeling: A case study in Sacramento County

In this study, I proposed two models to integrate job accessibility, employment ratio, income, and auto ownership into a structural equation system, and empirically examined the causal connections between job accessibility, workers per capita, income per capita, and autos per capita at the aggregate level with year 2000 census tract data in Sacramento County, California. Under the specification of the conceptual model, the model implied covariance matrix exhibited a reasonably good fit to the observed covariance matrix. The direct and total effects showed that job accessibility had a negative effect on autos per capita, autos per capita had a positive effect on workers per capita and income per capita, workers per capita had a positive effect on income per capita and autos per capita, and education attainment had a positive effect on workers per capita, income per capita and autos per capita. Job accessibility had a negative total effect on workers per capita, income per capita and autos per capita. These results were largely consistent with theory and/or with empirical observations across a variety of geographic contexts. They suggested that structural equation modeling was a powerful tool for capturing the endogeneity among job accessibility, employment, income and auto ownership, and had other advantages over linear regression in this context.; I also examined the generalizability of the conceptual model built on census tract (CT) data, tested the cutoffs suggested in the literature for reliable estimates and hypothesis testing statistics at the CT and block group (BG) levels, and evaluated the efficacy of deleting observations as an approach to improving multivariate normality in structural equation modeling. The results showed that the conceptual model could be applied to a BG sample. The parameter estimates and standard errors of the parameter estimates at the BG level were less sensitive to changes in sample size and multivariate normality than were those at the CT level. The cutoffs for nonnormality suggested in the literature did not perform well in this study. I argued that pursuing a multivariate normal distribution by deleting observations should be balanced against loss of model power in the interpretation of the results.