Is iron stronger than copper? When data and theories tell different stories about forgetting

We forget. Whether we trade the name of a fish for the name of a student, or whether we eventually forget the name of the U.S. President, the loss of information from memory is nearly always unavoidable.

The reasons that underlie forgetting are, however, still debated. One putative mechanism that has attracted considerable attention is the idea of inhibition. According to this idea, information in memory may be inhibited by a cognitive control process that actively suppresses competing and currently incorrect information in order to facilitate retrieval of the target information.

This inhibitory phenomenon is known as retrieval-induced forgetting (RIF) and is typically revealed with the so-called retrieval practice paradigm.  Experiments in this paradigm involve three phases:

  • Study phase: Participants are presented with pairs consisting of a category member and the label of the category, such as “METAL-copper” and “METAL-iron” and so on.
  • Retrieval-practice phase: During this phase a subset of the studied items is presented for recall. For example, people might have to fill in the missing letters in response to “METAL-ir__”.
  • Test phase: During this final and crucial phase, participants are asked to recall all studied items, usually in a similar form as in the practice phase.

The result of interest involves the difference in recall performance at test between studied items whose categories were never practiced (called NRP, for not-retrieval-practiced), and studied items whose categories—but not the items themselves—were practiced. For example, if people practiced “METAL-ir__” but not “METAL-co___”, then “copper” would be a so-called RP- item (because the category but not the item were practiced).

Across many studies, performance on the RP- items is found to be lower than for the NRP items, even though both types of items were studied under identical conditions and even though both types were equally unpracticed during the second phase. On the inhibition account, this result arises because during retrieval practice, the category cue (“METAL”) activates many category members, not all of which are relevant at the time—for example, only iron is relevant in response to “METAL-ir__” and hence other metals are actively suppressed to keep them from competing with iron.  This suppression, in turn, lowers recall on the final test even though now the item is a valid recall candidate.

There is, however, another account for the same phenomenon: On the competition account, there is no inhibition of RP- items, but the enhanced strength of the practiced items (called RP+) overshadows the RP-items from the same category. Thus, the RP- items fail to be produced not because they are inhibited but because their RP+ counterparts are recalled so successfully that people never get around to reporting the RP-items.

recent article published in the Psychonomic Bulletin & Review provided some insights into the differentiation between these alternative accounts. Researcher Jeroen Raaijmakers focused on one crucial difference between the opposing theories. According to the inhibition account, the RP- deficit should be independent of the degree to which RP+ items had been strengthened during retrieval practice. After all, the presumed inhibition involves the RP- items only, so the strength of the RP+ items should be irrelevant. By contrast, according to the competition account, the extent of strengthening is what gives rise to the observed retrieval-induced forgetting, and hence the more RP+ items are strengthened, the greater the recall deficit should be for their RP- counterparts. In other words, there should be a negative correlation between RP- and RP+ performance.

Raaijmakers began with a review of the existing experimental findings and found that, overall, the data failed to show the correlation between RP- and RP+ that is expected by the competition account.

Case closed? Inhibition but not competition?

Raaijmakers argued otherwise by reporting a large-scale computer simulation. At first glance, one might wonder how a simulation can re-open or resolve an empirical issue. After all, if the data tell a story, what else is there to say?  Intriguingly, this is not always the case: Data always tell a story but the story is not always what it appears to be.

In this instance, Raaijmakers showed that a model based on response competition did not actually predict the expected negative correlation between RP- and  RP+ performance. The model he used, known as SAM (for “Search of Associative Memory”) assumes that recall involves a series of retrieval cycles, each of which in turn comprises a sampling and a recovery process. In the sampling process, an item in memory is selected with a probability proportional to the relative strength of the retrieval cues. The recovery process, by contrast, involves the reconstruction of the name of the sampled item which is a function of the absolute strength of the association to the retrieval cues. In this model, the increase in strength of the RP+ items puts the RP- items at a disadvantage during sampling because even though they are not inhibited, relative to the practiced items their retrieval cues are less effective.

Raaijmakers showed that the model produced the standard retrieval-induced forgetting effect (the RP- items were recalled 54% of the time, compared to 64% for the NRP items). Of greater interest was the finding that across the 10,000 simulated experiments, the correlation in performance between RP- and RP+ was close to zero.

That is, even though the SAM model was indisputably based on competition at retrieval and not inhibition, it does not in fact predict a correlation between the two types of items.

This result does not change the available data, but it does change the meaning ascribed to those data: whereas the absence of a correlation across the existing behavioral experiments had been interpreted as a challenge to the competition view, it now turns out that the competition view actually predicts that very absence.

How can this be the case? If RP+ items overshadow RP- items at test, how can there be no correlation between those two item classes at test?

Raaijmakers conducted further simulations to identify the reasons for this counter-intuitive outcome of the simulation. He discovered that the correlation failed to materialize because the only measure that is available for a given item in an experiment is whether or not it has been recalled. That is, the measure of performance at the item level is binary—the only possible outcomes are 1 (recalled) or 0 (not recalled). It turns out that this binary measure is so impoverished that it is nearly impossible to detect a correlation in an experiment.

Raaijmakers’s work permits two conclusions: First, the debate between inhibition and competition accounts of retrieval-induced forgetting is far from settled. This may be bad news for anyone interested in closure of this issue, but it is exciting news for anyone who can think of better ways to tease the two accounts apart. Second, whenever theories predict associations between measures that involve binary observations for each item, those predictions may be very difficult to test experimentally.

Article focused on in this post:

Raaijmakers, J. G. (2016). On testing the strength independence assumption in retrieval-induced forgetting.Psychonomic Bulletin & Review. DOI: 10.3758/s13423-015-0991-4.

You may also like