Automated scoring models for e-rater were built and evaluated for the GRE argument and issue-writing tasks. Prompt-specific, generic, and generic with prompt-specific intercept scoring models were built and evaluation statistics such as weighted kappas, Pearson correlations, standardized difference in mean scores, and correlations with external measures were examined to evaluate the e-rater model performance against human scores. Performance was also evaluated across different demographic subgroups. Additional analyses were performed to establish appropriate agreement thresholds between human and e-rater scores for unusual essays and the impact of using e-rater on operational scores. The generic e-rater scoring model with operational prompt-specific intercept for the issue-writing task and prompt-specific e-rater scoring model for the argument writing task were recommended for operational use. The two automated scoring models were implemented to produce check scores at a discrepancy threshold of 0.5 with human scores.
