Unfolding the direct and indirect effects of social class of origin on faculty income

Abstract

Studies on faculty income have typically focused on disparities associated with gender and race. Surprisingly, much less attention has been paid to the social class background of university faculty and how it might affect their pathways to the professoriate and their opportunities to access high-paying positions. We attempted to address this gap in the literature by looking at a sample of faculty working at Chilean universities. We used a path analysis approach to estimate not only the direct effects of social class of origin on income but also the indirect mechanisms through which social class of origin influences faculty income. We posed two alternative conceptual perspectives with regard to the effects of social class on income—social reproduction and human capital. We found that faculty who come from the upper social class have access to higher-quality undergraduate education and to more prestigious PhD-granting universities and they report higher earnings as compared with those who come from a low social class. These findings resemble a dynamic of cumulative educational advantages that provides grounds to the theory of social reproduction. Although it could be argued that the positive effect of prestige of the PhD-granting university on income is in line with the human capital theory, we claim that such effect cannot be analyzed independently from the direct and indirect relationships that exist between social class of origin and the prestige of the university from which faculty obtained their doctorate degrees.

Appendix A: Latent class analysis (LCA)

Model specification

The purpose of estimating an LCA model was to classify individuals in the sample into social classes using four observed indicators: mother’s level of education, father’s level of education, high school type, and high school location. To specify the LCA model, we followed the steps recommended by Masyn (2013) and Geiser (2013). We first used fit indices and classification diagnostics to decide on the number of classes to be retained. We relied on absolute, relative, and information-based fit indices to assess the degree to which the LCA model reproduced the observed data. We also looked at the relative entropy index and average class assignment probabilities to assess the precision and reliability of the classification. Once the model was fitted, the algorithm sorted individuals into the latent class for which they displayed the highest posterior class probability.

Class enumeration process

To determine the number of k classes, we specified a 1-class model and then fitted additional models incrementing the number of classes by one, until models were no longer identified, as recommended by Masyn (2013) and Geiser (2013). We compiled the fit indices and classification diagnostics of each competing model in Appendix Table 5 to decide on the number of classes to be retained.

Table 5 Fit indices for alternative latent class models

Assessment of the model fit

As for absolute fit, we used Pearson and likelihood ratio chi-square tests to assess the degree to which an LCA model reproduces the observed data. Significant values associated with these tests indicate statistically significant discrepancy between the hypothesized model and the observed data. We also relied on likelihood ratio difference tests (BLRT and VLMR) to assess the relative fit of the models obtained. A significant p value for these difference tests indicates that the model with k classes fits the data better than the model with k − 1 classes. Additionally, we obtained the Akaike’s information criterion (AIC), the Bayesian information criterion (BIC), and adjusted BIC (aBIC) to assess the model in relation to both fit and parsimony. Smaller values of these information indices indicate better balance between fit and parsimony among competing models.

Classification diagnostics

After checking the models’ goodness of fit, we looked at the relative entropy index (E_k) and average latent class assignment probabilities to gauge the precision with which each competing model is assigning individuals to each class; the values for each competing model are displayed in Appendix Table 5. E_k summarizes the overall precision of the classification for the sample across all latent classes (Masyn 2013). Values of E_k close to 1 indicate a high degree of classification accuracy (Geiser 2013), while values close to 0 indicate latent classes are not sufficiently separated. As for class assignment probabilities, values larger than 0.8 in the diagonal indicate high precision or reliability of the classification (Geiser 2013).

Results

The results of this analysis showed that there are four distinct social classes within the sample, based on the fit indices, classification diagnostics, and average class assignment probabilities. The fit indices and classification diagnostics obtained for each competing model are displayed in Appendix Table 5. Bolded values indicate the corresponding value is the best one for each fit criterion. The absolute fit indices show that the 1-class, 2-class, and 3-class models do not fit the data well. For the LR chi-square test of fit, the 4-class model is marginally adequate, while the 5-class model displays a high degree of model-data fit. However, the BLRT and VLMR tests of differences indicate that the 5-class model is not significantly better than the 4-class model. Based on all fit indices, we concluded that the model that best fits the data is the 4-class model.

The classification diagnostics also support the choice of the 4-class model. On one hand, the entropy index E_k (see Appendix Table 5) shows that the more accurate classification is given by the 4-class model. The average class assignment probabilities of the 4-class model indicate high precision or reliability of the classification, as all probabilities in the diagonal are larger than 0.8 (LC1 = 0.87; LC2 = 0.82; LC3 = 0.96; LC4 = 0.92). Finally, the estimated conditional probabilities of class assignment are displayed in Appendix Table 6.

Table 6 Estimated latent class probabilities for each observed indicator