Why they quit: a focused look at teachers who leave for other occupations

Abstract

To better understand the factors related to teachers' decisions to leave for jobs outside of education, the study employs a structural equation modeling approach to analyze data from two large national datasets from the National Center for Education Statistics. The focus on occupation switchers is unique, with most studies of teacher attrition failing to differentiate between teachers who leave by their reasons for doing so. Overall, our findings suggest that district- and school-level leaders concerned about keeping good teachers in the classroom can take steps to improve teachers' job satisfaction by enhancing salaries and the conditions in which teachers work. Forced to choose between these levers, administrators may be more successful in boosting satisfaction and reducing the rate by which teachers quit to take a job outside of education by focusing their efforts on improving working conditions.

Keywords:

• teacher attrition
• school improvement
• educational policy
• structural equation modeling

Notes

1. Urban schools also have difficulty filling vacancies, primarily in mathematics (34.7%), foreign languages (30.3%), and special education (31%) (Strizek et al., 2006).

2. The survey uses a complex sampling framework that includes stratification, clustering, and oversampling of teachers with certain characteristics (e.g., new teachers, bilingual teachers) to ensure that samples of these teachers are large enough to produce reliable estimates. In surveys with complex sample designs, direct estimates of the sampling errors based on the assumption of simple random sampling will typically underestimate the variability in summary statistics and distort tests of statistical significance (Hahs-Vaughn & Lomax, 2006; Thomas & Heck, 2001). In order to compensate for this bias, weights assigned by NCES were used in our analyses. Weights are inversely proportional to the probability of selection. Given the relatively small size of our analytic sample, sampling weights can improve our ability to generalize results to the population of the K-12 public school teachers.

3. From a sample of 263 teachers who reported leaving teaching for another job, we excluded 41 part-time or itinerant teachers and teachers who left their teaching position for another job in education (e.g., administrators, counselors, curriculum coordinators, instructors).

4. A latent trait that cannot be measured directly is more reliable and valid when it is measured with two or more indicators (Kline, 2005).

5. We could not use SASS items related to induction/mentoring as they were only asked of first-year teachers.

6. Variation across teacher subgroups (e.g., race, age, gender) and school characteristics in the effect of each exogenous variable on satisfaction and teacher decisions to switch occupations was explored by comparing separate SEM models for each group.

7. The model fit is evaluated using χ ² (chi-square), the root mean square error of approximation (RMSEA) for a measure of absolute fit, and the normed chi-square (χ ²/df ) for a parsimonious fit measure. As rules of thumb, RMSEA values of .05 or less indicate a good fit (Cudeck & Browne, 1993; Hoyle, 1995) as do normed chi-square values of less than 5 (Kline, 2005). Confirmatory factor analysis showed that the fit of the measurement model was satisfactory (χ ²/df = 4.61; RMSEA = .048) with one exception. The significant chi-square statistic [308.94 (df = 67), p < .01] indicates an unsatisfactory model fit. However, the chi-square fit index is highly sensitive to sample size. A model is likely to be rejected when the sample size is large, even though the discrepancy between the sample correlation/covariance matrix and model-predicted correlation/covariance matrix may be small or trivial (Fan, Thompson, & Wang, 1999; Fan & Wang, 1998).

8. Construct validity and reliability are also evaluated (data not shown). Construct validity is evaluated by examining the standardized factor loadings within the constructs (Hair et al., 1998). The standardized factor loadings on all latent constructs are statistically significant at the .05 level. Construct reliability, a statistic that measures the amount of scale score variance that is accounted for by all underlying factors, ranges from .575 and .737 (Hair et al., 1998). The construct validity and reliability for working conditions is just below the acceptable threshold (.6). This may be due to correlated error variances between “principal leadership/administrative support” and “teacher influence over school policy” and between “student conduct” and “teacher control in the classroom”, which we allowed to remain in the model based on evidence from other empirical studies. For example, Blase and Kirby (1992) and others (Clift, Veal, Holland, Johnson, & McCarthy, 1995; Conley, 1991) found that facilitative principal leadership and support that provide teachers opportunities to participate in decisions about policies and practices are positively associated with teachers' job satisfaction and commitment. In addition, teacher authority over instruction and discipline have been found to be related to student behavior wherein fewer student behavioral problems exist in schools where teachers perceive having more control (Evertson & Weinstein, 2006; Ingersoll, 2003).

9. For the multigroup comparisons, we build a final structural model with four significant constructs (excluding professional development experiences) from the relative weight and mediating models. The equivalence of the measurement model is established, and the structural models for subgroups compared (Byrne, 1998; Jöreskog & Sörbom, 1993; Kelloway, 1998). The structural paths of interest among the latent variables are compared by examining chi-square and other fit indices (e.g., χ ² /df and RMSEA) between the fully and partially unconstrained models.