Retrieval-mediated directed forgetting in the item-method paradigm: the effect of semantic cues

Abstract

Item-method directed forgetting is widely considered a storage phenomenon. However, by applying a multinomial model, which separates storage and retrieval effect components, Rummel et al. (J Exp Psychol Learn Mem Cogn 42(10):1526–1543, 2016) recently provided evidence that item-method directed forgetting effects are reflected by both storage and retrieval changes. The current investigation demonstrates that supposedly intentionally forgotten information can still be retrieved to some extent when semantic cuing facilitates retrieval of this information. Participants studied word pairs, with some pairs being followed by a “forget” and others by a “remember” instruction. A subset of items shared the same superordinate semantic category. In Experiment 1, a sub-portion of to-be-forgotten items was semantically related and less forgetting occurred selectively for these items when the category was reinstated during test. This finding was replicated and extended to reinstatement effects for to-be-remembered items in Experiment 2. The application of the storage–retrieval model confirmed that providing a category cue facilitates retrieval of to-be-forgotten as well as to-be-remembered information. The results are discussed in light of existing theories of directed forgetting.

Appendices

Appendices

Appendix 1: Description of the storage–retrieval model and its application

The storage–retrieval model used in the present investigation is a multinomial processing tree (MPT) model first proposed by Riefer and Rouder (1992) that can be applied to memory paradigms employing a free-then-cued recall memory test. The model has been used to investigate storage and retrieval processes in many different memory domains, including the bizarreness effect (Riefer & LaMay, 1998; Riefer & Rouder, 1992), lag effects (Küpper-Tetzel & Erdfelder, 2012), age-related memory differences (Riefer & Batchelder, 1991a), clinically related memory deficits (Riefer et al., 2002), and recently directed-forgetting effects (Marevic et al., 2017; Rummel et al., 2016).

To apply the model to the current data, the free-then-cued recall test results of Experiments 1 and 2 were scored as either correct or incorrect paired free recall, singleton free recall, or cued recall. The combination of the resulting recall possibilities yields six possible recall events (E₁–E₆) for each studied item-pair: E₁, successful free recall of the complete pair and successful cued recall; E₂, successful free recall of the complete pair but failed cued recall; E₃, successful free recall of a single item from a pair (singleton) and successful cued recall; E₄, successful free recall of a singleton but failed cued recall; E₅, failed free recall but successful cued recall; E₆, failed free recall and failed cued recall. From these outcome frequencies the model parameters (see Fig. 4) were estimated:

Associative storage

Storing and maintaining an item-pair association until the free recall memory test. These processes occur with probability a, 0 ≤ a ≤ 1.

Associative retrieval

Retrieval of both items of a pair, given that the pair was stored in the first place. The pair does not necessarily need to be retrieved associatively, as singleton-linked retrieval is also possible. The model does not differentiate between these two types of associative retrieval. These retrieval processes occur with probability r, 0 ≤ r ≤ 1.

Stored singleton retrieval

Retrieval of only one item of a previously stored pair. These processes occur with probability s, 0 ≤ s ≤ 1.

Memory loss of stored association

Even though a successfully free recalled association is assumed to be associatively stored, memory loss from free to cued recall can occur. The probability l, 0 ≤ l ≤ 1 accounts for such memory loss.

Non-stored singleton retrieval

If an item-pair is not stored associatively, singletons from the pair can still be stored and retrieved independently. These processes occur with probability u, 0 ≤ u ≤ 1.

The model has five parameters (a, r, s, u, and l), resulting in 6 × 5 = 30 parameters in total. There were six event categories per condition and therefore 5 degrees of freedom per condition, totaling 6 × 5 = 30 degrees of freedom. If the number of parameters and the number of degrees of freedom are equal, model parameters can be estimated but model fit cannot be assessed (Riefer & Batchelder, 1991b). As in Experiments 1 and 2 the cued recall test immediately followed the free recall test, l parameter estimates that reflect memory loss from free recall to cued recall should be similar between conditions and should generally be close to zero. Therefore, in the current studies, the l parameter was set to be equal across all item types and groups. In Experiment 2, the s parameter had additionally to be set to be equal across groups for each item type in order to deal with a cell frequency of zero.

In fitting our data to the storage–retrieval model by using the software multiTree (Moshagen, 2010), the following steps were performed (see Erdfelder et al., 2009, for a detailed description of hypothesis testing with MPT models): First, parametric order constraints were imposed when necessary (i.e., in Experiment 2) that allow to evaluation of proportional changes from one item type to the other [e.g., a(TBF_UR items) < a(TBR items) for the a*(R-F_UR difference) for each experimental group]. Second, model parameters were estimated via minimization of the log-likelihood ratio statistic G²(df) using the expectation–maximization (EM) algorithm (Hu & Batchelder, 1994). Third, the model’s fit to the data was assessed by comparing the previously minimized G² statistic against the χ²(df) statistic. When the fit statistic felt below the (1 − α) percentage of the distribution, the model was retained. In our case the model fitted the data well, both in Experiment 1, G²(5) = 4.53, p = 0.475, and in Experiment 2, G²(20) = 27.76, p = 0.114. The last step involved hypothesis testing through parameter comparison. To this end, we imposed restrictions on the parameters of interest and compared the resulting restricted version of the model with the superordinate model by assessing the ∆G² difference statistic (Batchelder & Riefer, 1999).

Appendix 2: Recall frequencies of all recall events by experiment

Condition	E 1	E 2	E 3	E 4	E 5	E 6
Experiment 1
Category-Cue group
TBFUR items	27	3	1	60	13	276
TBFSR items	51	12	2	58	22	235
TBR items	151	21	4	71	16	117
No-Category-Cue group
TBFUR items	31	5	2	73	8	261
TBFSR items	37	4	1	65	18	255
TBR items	162	16	5	64	21	112
Experiment 2
Category-Cue Forget group
TBFUR items	21	2	1	46	13	183
TBFSR items	29	3	2	35	28	183
TBRUR items	106	3	11	49	23	88
No-Category-Cue Forget group
TBFUR items	35	2	3	63	6	191
TBFSR items	31	2	1	70	7	189
TBRUR items	135	8	3	57	22	75
Category-Cue Remember group
TBFUR items	32	1	4	37	15	211
TBRSR items	98	11	1	37	28	125
TBRUR items	64	8	5	58	15	150
No-Category-Cue Remember group
TBFUR items	32	1	1	36	9	211
TBRSR items	87	9	0	64	30	100
TBRUR items	74	11	4	55	18	128

1. E₁ = both items freely recalled, correct cued recall; E₂ = both items freely recalled, incorrect cued recall; E₃ = one item freely recalled, correct cued recall; E₄ = neither item freely recalled, correct cued recall; E₅ = one item freely recalled, incorrect cued recall; E₆ = neither item freely recalled, incorrect cued recall. TBF_UR items were items that did not share a superordinate category cue and were post-cued as TBF. TBF_SR items were semantically related items in that they shared a superordinate category cue and were post-cued as TBF. SR items were semantically related items in that they shared a superordinate category cue; in the Category-Cue Forget and No-Category-Cue Forget group these items were post-cued as TBF and in the Category-Cue Remember and No-Category-Cue Remember group they were post-cued as TBR. TBR items were unrelated items that were post-cued as TBR

Appendix 3: Output order analysis for semantically related items

Output order analyses were conducted to test whether semantically related items were recalled earlier in the recall test by the Category-Cue groups than by the No-Category-Cue groups. For these analyses, we calculated the proportion correct recalled for each group and item type at each output position aggregated over participants. The proportion correct recalled as a function of group, item type, and output position is displayed in Fig. 5 for Experiment 1 and Fig. 6 for Experiment 2. As output order of semantically related items was of main interest, in the following output order analyses we compared whether the proportion recalled for the semantically related items in the first quarter (items 1–3), second quarter (items 4–6), third quarter (7–9), and last quarter (> 10) differed between the Category-Cue and No-Category-Cue groups.

In Experiment 1, the proportion of semantically related items recalled in the first quarter did not vary between the Category-Cue and No-Category-Cue groups, χ²(1) = 1.98, p = 0.159. Thus, participants in the Category-Cue group were generally not more likely to recall more TBF_SR items at early output positions during recall. For the second quarter, χ²(1) = 10.26, p = 0.001, and third quarter, χ²(1) = 8.71, p = 0.003, however, the proportion recalled was higher in the Category-Cue compared to the No-Category-Cue group (see Fig. 5). For the last quarter the pattern was reversed, with higher recall rates in the No-Category-Cue compared to the Category-Cue group, χ²(1) = 8.10, p = 0.004. From inspection of Fig. 5, it is evident that participants are most likely to recall TBR items in the first quarter, followed by TBF_SR and TBF_UR items. Thus, participants seem to initiate recall with items highest in memory strength, followed by semantically related TBF items and lastly unrelated TBF items.

In Experiment 2, we first compared the proportion of semantically related items recalled in each quarter between the Category-Cue Forget and No-Category-Cue Forget group followed by comparisons between the Category-Cue Remember and No-Category-Cue Remember group. For the Forget groups, the proportion of semantically related items recalled in the first quarter, χ²(1) < 0.01, p > 0.999, and second quarter, χ²(1) < 0.01, p > 0.999, did not differ between the Category-Cue Forget and No-Category-Cue Forget groups. Thus, participants in the Category-Cue and No-Category-Forget groups were similarly likely to recall TBF_SR items at early output positions during recall. For the third quarter, however, semantically related item recall was higher in the Category-Cue Forget than in the No-Category-Cue Forget group, χ²(1) = 16.89, p < 0.001. For the last quarter, recall proportions did not differ between the two groups, χ²(1) = 0.58, p = 0.445. Again, these results indicate that, on average, participants in the Category-Cue Forget and No-Category-Cue Forget groups recalled similar amounts of TBF_SR at the beginning of the recall episode, and from inspection of Fig. 6 it is evident that participants were most likely to start with recalling TBR items. For the Remember groups, the proportion of semantically related items recalled in the first quarter was higher in the Category-Cue Remember group compared to the No-Category-Cue Remember group, χ²(1) = 5.05, p = 0.024. For the second quarter, χ²(1) = 0.06, p = 0.797, and third quarter, χ²(1) = 2.75, p = 0.097, the groups did not differ, and for the last quarter, χ²(1) = 7.36, p = 0.006, the pattern was reversed. That is, the category cue increased the likelihood to recall TBR_SR items at early output positions. This is in line with the notion that participants start recalling items that are high in memory strength at early output positions (Rundus, 1971), and is consistent with the pattern seen for forget groups.

Fig. 4

Multinomial processing tree model (MPT) for a free-then-cued-recall paradigm, to separate storage and retrieval processes (based on Rouder & Batchelder, 1998; adapted from Rummel et al., 2016). The processing tree represents the different latent cognitive processes that lead to six observable recall events (E₁–E₆). Rounded rectangles represent latent cognitive states with transition probabilities being described by the model parameters: a = probability of associative storage; r = probability of associative retrieval; s = probability of singleton retrieval given association was stored; u = probability of singleton retrieval given association not stored; l = memory loss due to time delay between free and cued recall. Parameter l was restricted to be equal across all conditions to render the model testable

Fig. 5

Recall probabilities from Experiment 1 as a function of group (Category-Cue Group, No-Category-Cue Group), item type (TBF_SR, TBF, TBR) and output position (1–16)

Fig. 6

Recall probabilities from Experiment 2 as a function of group (Category-Cue Forget Group, No-Category-Cue Forget Group, Category-Cue Remember Group, No-Category-Cue Remember Group), item type (SR, TBF, TBR) and output position (1–16)

Appendix 4: Data repository (Zenodo)

To access/download the data and analyses scripts please visit the data repository http://zenodo.org/record/1217615.