In Eq. 1 , y i 6 represents the i th mother's outcome at the 6-month interview, X is a vector of covariates, W is a vector of lagged measures of maternal employment measured at the 3-month interview, ν represents unmeasured family/maternal characteristics that affect y , ε is a random error term, and α and β are parameters to be estimated. As discussed below, lagged employment is used to help reduce the possibility of bias resulting from reverse causality. 6 For continuous outcomes (., depressive symptoms, parenting quality, parenting stress), we use OLS for estimation of Eq. 1 , while we use standard probit models for binary outcomes (., poor/fair health, depressed) and an ordered probit model for self-assessed health rating (., 1 = fair/poor health, 2 = good health, 3 = excellent health). We estimate Huber–White standard errors adjusted for clustering on site.