Magnitude and sources of proactive interference in visual memory

ABSTRACT

Proactive interference – the disruptive effect of old memories on new learning – is a long-established forgetting mechanism, yet there are doubts about its impact on visual working memory and uncertainty about the kinds of information that cause proactive interference. The present study aimed to assess these issues in three experiments using a modified recent probes task. Participants encoded four target images on each trial and determined whether a probe matched one of those targets. In Experiment 1, probes matching targets from trial N-1 or N-3 damaged responding in relation to a novel probe. Proactive interference was also produced by probes differing in state to a previously experienced target. This was further assessed in Experiments 2 and 3. Here, probes differing in colour to a previous target, or matching the general target category only, produced little proactive interference. Conversely, probes directly matching a prior target, or differing in state information, hindered task performance. This study found robust proactive interference in visual working memory that could endure over multiple trials, but it was also produced by stimuli closely resembling an old target. This challenges the notion that proactive interference is produced by an exact representation of a previously encoded image.

KEYWORDS:

• Proactive interference
• visual working memory
• forgetting
• recent probes task

Acknowledgements

We are grateful to Professor John Krantz for hosting Experiment 1 on Psychological Research on the Net (https://psych.hanover.edu/research/exponnet.html) and Dr Talia Konkle for permission to use the stimuli employed in both experiments.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are freely available in the Open Science Framework project “Magnitude and Sources of Proactive Interference in Visual Memory” at https://doi.org/10.17605/OSF.IO/F7SG4

Notes

1 There is some confusion in the literature concerning the distinction between short-term memory and working memory, but the former is often used to describe passive maintenance of information and the latter refers a more active system that both maintains and manipulates information (Aben et al., 2012). For consistency with similar previous studies, we use the term “working memory” in this manuscript, but we are referring to a limited capacity system that maintains and uses visual information over relatively short delays.

2 The Bayesian ANOVA in JASP is based on default priors, which are designed to be computationally convenient and applicable in a variety of situations. However, it is also possible to measure the impact of the chosen priors by running an identical analysis that swaps default priors (r = 0.5) for wider (r = 1) or narrower (r = 0.2) priors. Crucially, the conclusions did not change at all for the analysis on accuracy when either narrow or wide priors were used. The same held for the response time analysis, except under narrow priors the interaction effect became insensitive (BF_Inclusion = 0.58), though there was still very little support for retaining it within the model.

3 The issue with the negative skew for accuracy scores could not be corrected using typical transformations in the NRN condition, due to high performance. Yet exploring the data with Friedman's test found significant PI when comparing the three probe types, χ²(2) = 27.09, p < .001. Holm-Šidàk corrected Wilcoxon tests showed higher performance on NRN trials compared to RN (Z = −4.67, p < .001) and State RN (Z = −3.83, p < .001) conditions, but not between RN and State RN (Z = −1.90, p = .058). A further Wilcoxon test confirmed lower performance for N-1 compared to N-3 trials (Z = −5.02, p < .001). Thus, the effects from a non-parametric analysis supported those emerging from the ANOVA.

4 The positive skew was corrected through a log transformation of the data. Analysing the log-transformed response times in a 2 × 3 ANOVA revealed the same effects as the untransformed data: there were significant effects of the trial on which the probe was last encountered, F(1, 36) = 15.70, MSE = 0.001, p < .001, ${\eta }_p^2$ = 0.30, and probe type, F(2, 72) = 38.49, MSE = 0.002, p < .001, ${\eta }_p^2$ = 0.52, but no interaction, F(2, 72) = 1.79, MSE = 0.001, p = .175, ${\eta }_p^2$ = 0.05. Post-hoc analyses also revealed the same pattern of results as the ANOVA based on untransformed data. The latter is reported in the main analysis as it aids interpretation of the effects.

5 Friedman's test found a significant PI effect when comparing the five probe types, χ²(4) = 11.46, p = .022. Holm-Šidàk corrected Wilcoxon tests showed significantly worse performance in the RN (p = .033) and State RN (p = .012) conditions, in comparison to the NRN control, whereas Colour RN (p = .429) and Exemplar RN (p = .238) did not differ from NRN.

6 Following Experiment 1, the robustness of the Bayesian ANOVA was tested by re-running the analysis with narrow and wide priors. For both the accuracy and response time analyses, there remained convincing support for the alternative hypothesis regardless of the priors.

7 The positive skew was corrected through a log transformation of the data. Analysing the log-transformed response times in a one-way ANOVA revealed the same effects as the untransformed data, with a significant probe type effect, F(4, 100) = 5.98, MSE = .001, p < .001, ${\eta }_p^2$ = 0.19. Holm-Šidàk corrected paired-samples t-tests revealed responses on NRN trials were faster than both RN (p = .005) and State RN (p = .032) trials. Conversely, Colour RN (p = .587) and Exemplar RN (p = .849) response times did not differ from NRN. Responses on RN trials were also slower than Colour RN (p = .035) and Exemplar RN (p = .020) trials. All other comparisons were non-significant.

8 Friedman's test found a significant PI effect when comparing the five probe types, χ²(4) = 51.86, p < .001. Holm-Šidàk corrected Wilcoxon tests showed significantly worse performance in the RN (p < .001) and State RN (p < .001) conditions, in comparison to the NRN control, whereas Colour RN and Exemplar RN did not differ from NRN. Additionally, RN and State RN conditions led to significantly lower performance than Colour RN and Exemplar RN conditions (all p values <.005).

9 Re-running the Bayesian ANOVA using narrow and wide priors did not change the outcome of the analysis. Extreme support for the alternative hypothesis remained.

10 The positive skew was corrected through a log transformation of the data, and analysing the transformed data with a one-way ANOVA revealed a significant probe type effect, F(4, 240) = 8.08, MSE = 0.002, p < .001, ${\eta }_p^2$ = 0.12. Holm-Šidàk corrected paired-samples t-tests showed that responses on NRN trials were faster than both RN (p = .003) and State RN (p < .001) trials. Conversely, Colour RN and Exemplar RN response times did not differ from NRN. Responses on Colour RN trials were also faster than RN (p = .007) and State RN (p = .003) trials. All other comparisons were non-significant.

11 Support for the alternative hypothesis remained when the priors were either narrow or wide. The Bayes factor was smaller under wide priors, BF¹⁰ = 7.53, but decent evidence for the alternative hypothesis remained, in comparison to the null hypothesis.

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.