Figure 8A: Comparing Results with Alternative Corpora/Methods, Women.

Note: This figure reports the estimated marginal effects and 95% confidence intervals as we vary the corpora or method used to identify stereotyped language. The string matching reports the coefficient of the dummy variable indicating that the stereotype appeared in a job ad. Positive predictions indicate we expect to see more discrimination against older applicants, negative predictions indicate we expect to see less discrimination against older applicants. For age-occupation-gender cells where the word does not appear on any ad or is perfectly correlated with discrimination, no coefficient is reported. This occurs most often for memory, careful, and creative. The spider algorithm model and the complete Wikipedia models report the coefficients from Equation (10), with the complete Wikipedia model being identical to the baseline results. The string matching uses equation (11). Standard errors have been clustered at the job ad level.