Does Merit Aid Push Students Out of STEM?

Programs like Georgia's HOPE scholarship make college affordable, but keeping the money requires keeping a minimum GPA - and STEM courses grade harder than most alternatives. That creates a quiet incentive: choose the major you can keep the scholarship in. Sjoquist & Winters found exactly that pattern nationally. Our job was to rebuild their headline analysis from raw microdata and see if it survives.

The sample comes from ACS microdata via IPUMS (2009-2011): ages 24-39 with at least a bachelor's degree, roughly 570,000 person-year observations. Majors were classified STEM or non-STEM by merging the paper's Appendix A1 crosswalk on the ACS degfieldd code. The hard part was treatment assignment: no dataset records when each state enacted its merit program, so enactment years were researched state by state and cross-checked against the paper's cited sources.

Exposure is then a cohort rule: you were treated if you turned 18 after your birth state's program started - encoded in Stata as

with separate strong and weak program classifications taken from the paper, plus state dummies, birth-year dummies, and state-by-birth-year time trends.

A linear probability model of STEM completion on merit exposure with sex, race, Hispanic-origin, age, and education controls; state fixed effects; birth-cohort fixed effects; and state-of-birth × year-of-birth time trends. Standard errors are clustered by state of birth in the specification. The fixed-effects structure makes this a difference-in-differences design: identification comes from cohorts inside the same state who straddle the program's start date. LPM over probit/logit because the coefficients read directly as percentage-point effects - the same choice the authors made.

Each specification varies who counts as treated and control - strong states only, weak states folded into control, state-year trends added, a within-strong pre/post design, and any-merit as treatment. The effect on STEM completion:

1 · strong vs. non-merit

-1.32 pp

2 · weak states in control

-0.73 pp

3 · + state-year trends

-1.23 pp

4 · strong states, pre/post

-1.29 pp

5 · any merit as treatment

-0.38 pp

Sjoquist & Winters' published estimates for the same five columns are -1.3, -1.3, -1.2, -1.4, and -0.5 - our replication lands within 0.15 points on four of five and tells the identical story on all of them, every model significant at p ≈ 0.000.

Students on merit aid still graduate - they just graduate in different fields. The original authors show the displaced majors surface in business, social sciences, and liberal arts, consistent with students protecting their GPA to keep the scholarship. For policy, the implication is uncomfortable: a program built to expand opportunity may quietly tax the STEM pipeline it claims to feed, and program design (GPA thresholds, major-specific provisions) is where that trade-off gets decided.