Intent to leave, a dichotomous dependent variable (yes/no) was modeled using a logistic regression approach. Two different analytical populations were studied: the whole analytical sample of 119,500 aides, and a subset of 63,070 aides who were not extremely satisfied. While only 53 percent of the sample, this population of aides nonetheless accounts for roughly 85 percent of workers who intend to leave their jobs. Subpopx in SAS-callable SUDAAN was used to ensure that weights were applied correctly. Due to the limited number of intended leavers among the aides who were “extremely satisfied”, multivariate analysis for that population could not be reliably estimated.
Two alternative modeling approaches of intent to leave were included for a number of reasons. First, we had an a priori expectation that the factors that drive intent to leave are likely to differ for those workers who are satisfied with their jobs and those workers who are dissatisfied with their jobs. Second, the key independent variable in the intent-to-leave model--job satisfaction--was expected to be influenced by workers’ personality traits (e.g., attitudes and background) that are not directly observable. Such a model would be affected by omitted variable bias. However, the whole population approach was included for its policy relevance: agencies that wish to retain aides would find it difficult to identify workers with higher and lower satisfaction and to apply different approaches to them.
A two-stage model was not included due to the difficulty of constructing an instrumental variable that is highly correlated with job satisfaction. As we discovered in conducting multivariate analyses of job satisfaction, independent variables in these models were generally found to have poor predictive power.
1 Job satisfaction was initially modeled using ordered multinomial logit regression. This model assumes a common slope with separate intercepts for each level of the dependent variable. The ordered model failed to meet the proportional odds assumptions, and job satisfaction was subsequently modeled using a nominal model.