Give me enough time to rehearse: presentation rate modulates the production effect

Abstract

This paper uses the production effect to test one of the important predictions of a view of memory that is embodied in the Revised Feature Model (RFM). When to-be-recalled lists contain items both read aloud and silently, words read aloud are less well recalled at the beginning of the list and better recalled at the end. According to the RFM, producing the items by reading them aloud adds distinctive features which supports recall, but production also interferes with rehearsal – a process that operates more significantly at the start of a list. This critical role assigned to rehearsal has never been systematically tested. We do this here through a systematic literature review and an experiment that manipulates presentation rate. With a faster presentation rate, rehearsal is less likely; the implication is that the advantage observed for silently read items in the primacy positions should vanish, while the recency advantage for produced items should remain. The systematic review collected an initial sample of 422 unique articles on the production effect in immediate serial recall and revealed the predicted pattern. In addition, in our experiment, the presentation rate was manipulated within an immediate serial recall task (500, 1,000, and 2,000 ms/word). As predicted, the recency advantage for produced items was observed for all presentation speeds. Critically, the production disadvantage for early serial positions was only present for the two slowest rates, but not at the fastest speed. Results were successfully modeled by calling upon the RFM.

Appendix A: Modelling details

The RFM is too complex for an analytic expression for the likelihood to be derived, so as with all previous attempts to fit the model to data we used a version of Approximate Bayesian Computation (ABC) (see Turner & Van Zandt, 2012, or Marin et al., 2012, for a review). Following Poirier et al. (2019), Saint-Aubin et al. (2021), and Cyr et al (2022) we used ABC Partial Rejection Control (ABC-PRC) (Sisson et al., 2007, 2009). ABC-PRC works by repeatedly sampling from a prior over the parameter space until it finds a set of parameters which generate a set of summary statistics (in our serial position curves) sufficiently close to the data, as determined by the discrepancy function. When this happens, the algorithm stores these parameter values, and moves on to the next particle in the generation. Once all particles in a generation have been associated with parameter sets, the algorithm gives each particle a weight depending on the prior, and then begins a new generation, sampling from the previous generation with probabilities given by the weights, and repeatedly perturbing around the previous parameter values until a set is found producing summary statistics even closer to the data. Once the required number of generations have elapsed posterior estimates for the parameters can be obtained as the fraction of particles in the final generation with that parameter value. Posterior predicted distributions of the summary statistics are also easily obtained. For full details see Sisson et al. (2007) (note also the errata, Sisson et al., 2009).

As explained in the main text, we are fitting six data sets, one for each condition, but our hypothesis is that only the rehearsal rate varies between the different presentation rate conditions. We therefore split the parameters into two groups, we fit lambda and tau but demanded these be the same for each condition, and we fit a number of rehearsal rates which we allowed to vary between conditions. In addition, as is standard, we assumed the aloud conditions had more modality dependent features. Full details are given in Table 1.

The important parameters for ABC-PRC are the number of particles (set to 1000 for all fits reported here), the details of the prior, the proposal distributions, and the minimum tolerances for each fit. The proposal and tolerances can be found in the code on the OSF. Priors, and resulting posterior distributions are summarized in Table 1.

Table 1 Priors and descriptives for the posteriors for the fitted parameters in the model fitting