What Drives ROI Differences?
Using a two-way ANOVA, we decomposed program-level ROI variance into four components: field of study, institution, the interaction between field and institution, and everything else (credential level, cohort effects, measurement noise). The results are clear.
Two-way ANOVA: field (CIP 2-digit, 41 groups) × institution (3,424 groups). Sum of squares decomposition on 20-year ROI. Additive share: 57.3% (field 19.2% + institution 31.4% + overlap 6.7%). Interaction: 9.2%. Residual: 33.5%.
Top 10 Fields by Median ROI
Field choice still matters — it just matters less than most people think. Engineering and computer science programs dominate the top, but note how wide the ranges are. The best liberal arts program beats the worst engineering program by a mile.
ROI by Institution Type
The institution effect isn't just about prestige — it's structural. Technical institutes produce the highest median ROI, followed by public universities. Community colleges, despite low tuition, have the highest negative-ROI rate because of lower post-graduation earnings.
What This Means for Regulation
Institution-level regulation would catch more of the problem than field-level regulation — but you'd miss the cases where a decent school has one terrible program. The variance decomposition shows that institution quality is the single largest explainable factor in ROI. Regulating at the institution level (e.g., gainful employment rules applied to entire schools) would address the largest source of poor outcomes. But the 9.2% interaction term means some programs at otherwise-good schools are still destroying value. Program-level accountability catches what institution-level misses.
How We Decomposed Variance
Two-way ANOVA (Type III sum of squares). We modeled 20-year ROI as a function of field of study (CIP 2-digit code, 41 levels) and institution (3,424 levels), plus their interaction. The residual captures within-cell variation: credential level differences, cohort effects, and measurement noise from the Scorecard's earnings data. The additive model (field + institution, no interaction) explains 57.3% of variance. Adding the interaction term brings total explained variance to 66.5%.
Source: College Scorecard program-level earnings, IPEDS institutional data. N = 63,368 programs with complete ROI data. Analysis excludes programs with fewer than 10 completers (Scorecard suppression threshold).