In One Institution's Records, Dropping Low Scores Nudged Most Grades Up and Changed 29% of Letter Grades

Say you drop each student's lowest test score. Who gains, and by how much?

Robert Erdmann of the University of Minnesota Rochester checked a decade of gradebooks (pp. 1, 4, 15).

What he did

Erdmann studied 3,991 student-enrolments, meaning one student in one course, across 68 courses at one institution (pp. 3–4). Each course documented a fixed rule for dropping low scores (p. 4).

He rebuilt every final grade as a counterfactual, meaning the grade without the rule, changing nothing else (p. 4). The gap between the two grades measured the rule's effect (p. 4).

What the rule changed

By this measure, the rules raised final grades by a mean of 1.16 percentage points, with a median of 0.76 (p. 6). They lifted 92.5% of enrolments, but only 17.7% by more than two points (pp. 6–7).

Even so, 29.0% of enrolments moved into a different letter grade, such as from a B+ to an A- (pp. 5–6).

Gains relative to classmates showed no statistically detectable difference by first-generation status, ethnicity or classification as Black, Indigenous or people of colour (pp. 7–9). A sex difference was detectable but very small, and did not extend to letter-grade changes (pp. 9–10, 13).

Lower prior grade point averages went with larger gains relative to classmates (p. 11). Erdmann calls this consistent with "modest buffering" (pp. 12–13).

Courses differed: some permitted gains above two points, others under half a point (p. 13).

Limits and context

This used past records, with no experiment or random assignment (p. 15). The courses were science and maths at one small, health-sciences-focused institution (pp. 4, 15).

The recalculation assumes unchanged student behaviour without the rule, though students may choose whether to complete droppable work (pp. 6, 15). Some subgroups were small or merged diverse identities, and records held sex, not gender identity (p. 15). Only grades were measured, not learning (p. 15).

Erdmann stresses that similar gains across groups do not show the rule reduces wider inequities (p. 13).

Evidence strength: early signal. One institution's grade records, recalculated without the rule, with no experiment. It shows how the rule moved grades under recorded behaviour, not how it changes learning.