Transfer: Why What You Learned Doesn’t Always Apply
A friend who teaches introductory statistics at a small liberal arts college in Ohio once ran an unofficial experiment on her own students. In November, she taught them the chi-square test of independence in the context of survey data about breakfast preferences. They aced the unit quiz. In February, she gave them the same test, structurally identical, but framed as an analysis of whether the color of a car predicts its likelihood of being ticketed. A third of the class did not recognize it as a chi-square problem. Another third recognized it but could not remember how to compute one. They had learned the procedure, at least by the standards of a grade. What they had not done was make it available across contexts.
This is the transfer problem in miniature, and it is one of the least solved puzzles in education. The puzzle is easy to state. A student who learns a technique, principle, or skill in one setting often fails to use it in another setting, even when the second setting is objectively very similar. You would think this would be a narrow anomaly. It is closer to the default.
Psychologists divide the phenomenon into near transfer, when the new situation is close to the training situation, and far transfer, when the new situation differs substantially in surface features or domain. Near transfer is spotty. Far transfer is rare enough that a long tradition in psychology has questioned whether it meaningfully exists at all.
The pessimistic case has its classical expression in a 1901 paper by Edward Thorndike and Robert Woodworth at Columbia. They dismantled the then-fashionable idea of mental discipline, the belief that studying Latin or geometry trained a general faculty of reasoning that would then apply everywhere. Their theory of identical elements argued that transfer depends on specific overlapping features between the original training and the new task. If the overlap is thin, the transfer is thin. Latin, it turned out, trained you mostly to know Latin.
Ninety years later, Douglas Detterman at Case Western Reserve wrote what remains one of the more deflating reviews of the transfer literature. His 1993 chapter, bluntly titled The case for the prosecution: Transfer as an epiphenomenon, surveyed decades of studies and argued that the evidence for general transfer was essentially absent. Training in one task reliably produced improvement in that task. Claims of broad cognitive enhancement, by contrast, dissolved under careful measurement. He went so far as to suggest that when transfer appeared, it usually did so because the researcher had tipped off the participants about the connection.
A more nuanced picture arrived in 2002 with Susan Barnett and Stephen Ceci’s taxonomy, published in Psychological Bulletin. They argued that the question does transfer occur was too blunt, and proposed instead a multidimensional map across content, context, and social setting. Transfer across small gaps in any of these dimensions happens. Transfer across large gaps is unusual. The map helped explain why the literature looked so contradictory: different studies were measuring transfer across different kinds of distance, and their results were not really in disagreement so much as measuring different quantities.
This research has direct implications for how students study, though the implications are uncomfortable. A technique learned in a single context becomes tightly bound to that context. Your brain, sensibly enough, encodes not just the procedure but the smell of the classroom, the font of the textbook, the shape of the problems, the vocabulary of the examples. When you encounter the same underlying structure dressed in different surface features, the cues that would normally trigger the procedure are absent, and the procedure stays asleep.
What helps?
The most reliable answer from the last forty years of learning science is variability. Practice the same skill in many different contexts, with different surface features, different kinds of problems, different framings. This is what Robert Bjork has called interleaving and it is one of the central pillars of his desirable difficulties framework. A math student who does twenty chi-square problems in a row will feel fluent. A math student who does chi-square problems mixed with t-tests, correlation problems, and regression problems will feel confused and will score lower on practice. They will also, weeks later, be the one who recognizes a chi-square problem in the wild.
Kornell and Bjork demonstrated this neatly in a 2008 study on learning painting styles. Students shown paintings by different artists in blocks, all the Monets together, then all the Cezannes, felt they were learning well. Students shown the same paintings interleaved, a Monet then a Cezanne then a Pissarro, felt they were learning poorly. On the final test, the interleaved group was dramatically better at identifying artists from new paintings. The feeling of fluency had misled the blocked group; the messy, surface-varied exposure had given the interleaved group something closer to the structural understanding transfer requires.
A second ingredient is explicit abstraction. Mary Gick and Keith Holyoak’s famous Duncker radiation problem studies in the 1980s showed that students exposed to one analog of the problem rarely transferred the solution to a new analog. Students exposed to two analogs, and prompted to compare them, often did. The mere experience of multiple instances was not enough. Noticing what the instances had in common was. Transfer, in other words, requires a learner to extract structure, and the extraction usually has to be done deliberately.
A third ingredient is retrieval practice under varied conditions. Pulling a concept out of memory in many different disguises builds something closer to a decontextualized representation than reading about it in many disguises ever will. This is also why low-stakes quizzing with varied problem framings tends to produce better real-world performance than uniform homework sets.
None of this makes far transfer easy. The literature is fairly clear that you will not teach students to reason better in general by teaching them logic, or to think mathematically about the world by teaching them algebra. What you can plausibly do is teach them a technique in enough varied contexts that the technique, for them, becomes somewhat context-free. This is a modest claim. It is also, as the cognitive load tradition has documented, nearly everything educational research can honestly promise on transfer. The rest is marketing.
The Ohio statistics teacher now includes one mismatched problem on every quiz, borrowed from a different domain, sometimes biology, sometimes sports, sometimes economics. Her students complain. Her students also, by April, recognize chi-square problems in almost any disguise they meet. The complaint and the recognition, it turns out, are the same phenomenon.
Photo via Unsplash.