Retrieval Practice Usually Beats Another Read

The result reversed with time

Rereading won. Then, two days later, it lost.

In a 2006 experiment, undergraduates studied short prose passages under three conditions: repeated study, one study followed by one test, or one study followed by three tests. Total time was equated across conditions.

Repeated study produced the best performance on an immediate test. But after two days and again after one week, repeated testing produced better recall than repeated study. The practice tests provided no feedback, so another explanation was needed beyond simply seeing the answers again (Roediger and Karpicke, 2006).

That reversal is the testing effect. A test can be a learning event, not merely a measurement taken after learning.

How large the advantage is

The dramatic laboratory result is not the average result everywhere.

Rowland’s 2014 meta-analysis compared testing with restudy across experiments measuring performance on later tests. Its overall picture was a small-to-moderate advantage for retrieval practice, with substantial variation among materials, delays and test formats (Rowland, 2014).

Recall-based practice generally produced larger benefits than recognition-based practice. Asking someone to reconstruct an answer tended to beat asking them to recognize it. The advantage was also more pronounced after longer delays than at very short intervals.

Small-to-moderate is a standardized difference, not a promise of a particular number of extra exam points. The same average can contain large gains in one design and little benefit in another.

The clearest number is about transfer

Remembering the practiced material is one outcome. Using it on a different task is harder.

Pan and Rickard synthesized 192 transfer effect sizes from 122 experiments. Retrieval practice produced an average transfer effect of d = 0.40, with a 95% confidence interval from 0.31 to 0.50, compared with nontesting re-exposure (Pan and Rickard, 2018).

That moderate average came with a warning. Transfer was stronger when the practice and final responses were congruent, when retrieval included explanations or self-generated mediators, and when learners could retrieve enough during practice.

After publication-bias corrections, estimated baseline transfer was substantially smaller. When alignment, elaboration and sufficient initial performance were absent, some estimates indicated no positive transfer.

Where the evidence stops

The evidence supports retrieval practice, not testing in the abstract. A prompt that requires reconstruction is not equivalent to an easy recognition question. An answer successfully retrieved is not equivalent to an impossible prompt. A test aligned with the eventual task is not equivalent to one that rehearses a different response.

Here is the practical inference: replace some rereading with an attempt to answer before looking, then check the answer. That follows the studied contrast between retrieval and re-exposure, but it is not a universal recipe with a fixed payoff.

The honest estimate remains conditional. Being asked usually beats being told, often by a small-to-moderate amount. How much depends on what the question demands, whether retrieval succeeds, how long memory must last and what the final test asks the learner to do.