What the Evidence Can—and Cannot—Promise
The subject here is the evidence for student self-assessment and peer-assessment interventions, not assessment research in the abstract. The specific task is to read their effect estimates by outcome, implementation condition, and study quality, so that a positive average is not mistaken for a universal method or a proven explanation of why learning changes. The reader should leave able to distinguish a relatively consistent pattern from a promising signal and a practical recommendation from an inference the studies do not yet support.
That discipline is necessary because “self-assessment works” is an incomplete statement. Works for what—metacognitive strategy, motivation, self-efficacy, emotion, or an exam score? Compared with what? In a single activity or a repeated process? With criteria, feedback, and revision, or with a grade-only prompt? The recent evidence is valuable precisely because it begins to separate these questions.
Averages reveal different outcome profiles
The broadest recent synthesis examined 75 independent studies and 353 effect sizes. It grouped outcomes into self-regulated learning strategies and affective outcomes rather than treating every desirable change as the same kind of learning. The distinction matters. A student who becomes better at monitoring a strategy has changed something different from a student who reports greater motivation or self-efficacy, even if both changes make future learning more possible.
The review reports positive average effects for both self-assessment and peer-assessment when each is compared with a control condition. For self-regulated learning outcomes, the pooled estimate was g = 0.293 for self-assessment and g = 0.468 for peer-assessment. For affective outcomes, the corresponding estimates were g = 0.393 and g = 0.325. Here, g is a standardized difference between groups or conditions, not a number of extra marks. It allows results from different measures to be placed on a common scale, but it does not make the underlying tasks or outcomes identical.
The average pattern suggests that peer assessment may have been more strongly associated with the available self-regulated learning outcomes, while self- and peer-assessment looked more similar across broad affective outcomes. A mixed intervention that combined both practices showed g = 0.469 for self-regulated learning, but g = 0.130 for affective outcomes based on very sparse evidence. That last estimate should not be used to judge combined assessment in general. It illustrates why the number of studies behind an estimate matters as much as the direction of the estimate.
The detailed results make the profile more informative. For (meta)cognitive strategies, self-assessment produced g = 0.463 and peer assessment g = 0.478 after the review’s outlier handling. Those estimates are close, and they fit the mechanism already established: both practices ask learners to engage with standards, inspect performance, and reason about improvement. The convergence is one of the more consistent patterns in the synthesis.
Motivation looks different. The self-assessment estimate was g = 0.601, while the peer-assessment estimate was g = 0.229, with the reported confidence interval for peer assessment crossing zero. Self-efficacy was more modest: g = 0.298 for self-assessment and g = 0.224 for peer assessment, with the peer estimate again having a confidence interval that crossed zero.