Reducing withdrawal and failure rates in introductory programming with subgoal labeled worked examples

265 students across multiple sections of one introductory programming course

Undergraduate CS1 at one US university.

Students self-select into lab sections, which then received subgoal-labelled or standard instruction — not randomised at the student level, so section-level selection is uncontrolled. Outcomes are the course's own quizzes and exams, partially aligned with the treatment. The interrater-reliability procedure for open responses is well described. The contribution over prior subgoal work is duration: a full semester rather than a lab session.

Formative quizzes, taken within a week of each new procedure, moved d = 0.44, t(264) = 12.03. Summative exams did not clearly move: total exam d = 0.26, p = .04, with component scores of d = 0.20-0.22 that were not significant. The durable finding is about the tail rather than the mean: about half as many subgoal students as control students had a failing exam average or missed exams (control 7% took one exam and 13% took two; subgoal 5% and 5%), and the subgoal group's exam VARIANCE was significantly lower — fewer students failing badly, not more students excelling.

Medium. Section self-selection is uncontrolled, and students who choose particular lab sections differ.

The subgoal-label effect has a documented failure inside its own research programme — Morrison, Margulieux & Guzdial (2015) report that student performance gains from being GIVEN subgoal labels did not replicate as expected in the introductory CS task, despite replicating in mathematics and science.