Recollection rejection of false narrative statements

Abstract

Our research was focused on a false-memory editing operation that is posited in fuzzy-trace theory—recollection rejection. The main objectives were (a) to extend model-based measurement of this operation to a narrative task that ought to ensure high levels of recollection rejection and (b) to study five manipulations that ought to influence recollection rejection by affecting the accessibility of verbatim traces of narrative statements: recency of narrative presentation, narrative repetition, type of false-memory item, testing delay, and repeated testing. The results showed that the narrative task did indeed yield high levels of recollection, with an estimated 49% of gist-consistent distractors being rejected in this way on initial memory tests. Consistent with current theoretical conceptions of false-memory editing, the results also showed that recollection rejection increased as a function of manipulations that should enhance the accessibility of verbatim traces of narrative statements, with repeated testing delivering especially large increases in verbatim accessibility.

Acknowledgments

This research was supported bya National Science Foundation grant (SBR-9730143) and by a National Institute of Health grant (NIH31620).

Notes

¹Because it was necessary to administer eight distinct types of recognition probes to unconfound surface familiarity from meaning resemblance, it is necessary to use three-letter acronyms for these probes to avoid numerous repetitions of phrases such as “true premises with original wording”. As these acronyms appear, their meanings can be regenerated merely by remembering that they follow a 2×2×2 factorial structure: The first letter indicates whether or not a probe is true, in the sense of being consistent with the gist of a narrative (T = true or F = false); the second letter indicates whether a probe mentions two objects that were connected in one of the premise sentences or whether it is an inference that involves two objects that were not connected in the premises (P = premise or I = inference); and the third letter indicates whether the probe contains only old words from the narrative or contains a new word (O = old or N = new).

²Because the conjoint-recognition model's parameters are estimated by the method of maximum likelihood, between-condition significance tests of those estimates also involve likelihood ratios. The specific procedure for comparing the estimated values of a given parameter, R in this instance, between pairs of conditions involves three steps (see Brainerd et al., 1999, for details). First, the model's likelihood function is used to compute the joint likelihood of the data of the two conditions when all of the model's theoretical parameters are free to vary. Call this value L ₁₈ because a total of 18 parameters, 9 for each condition, are estimated when all parameters are free to vary. Second, the model's likelihood function is used to recompute the joint likelihood of the data, subject to the single constraint that R has the same value in each condition. Call this value L ₁₇ because one less parameter is free to vary. Third, the value of the test statistic −2ln[L ₁₇÷L ₁₈] is computed, which has an asymptotic χ²(1) distribution and a critical value of 3.84 for rejection of the null hypothesis of no between-condition difference in parameter values at the .05 level of confidence.

³Mathematically, these tests are the same as between-condition tests, apart from the fact that they are computed for pairs of parameters using the data of individual conditions, rather than for a single parameter using pairs of conditions (see Brainerd et al., 1999, for details). First, the model's likelihood function is used to compute the likelihood of the data of a target condition when all nine of the model's theoretical parameters are free to vary. Call this value L ₉. Second, the model's likelihood function is used to recompute the likelihood of the same data, subject to the constraint that two of the nine parameters (in this instance, β_V and β_G or β_V and β_G) have the same value. Call this value L ₈ because one less parameter is free to vary. Third, the value of the test statistic −2ln[L ₈÷L ₉] is computed, which has an asymptotic χ²(1) distribution and a critical value of 3.84 for rejection of the null hypothesis of no difference in parameter values at the .05 level of confidence.