A systematic search across PubMed, PsycINFO, Web of Science, Scopus, ProQuest Dissertations & Theses, and Google Scholar (accessed January 2025) used five complementary search string families: (1) “retrieval-enhanced suggestibility,” “reversed testing effect,” and “RES effect”; (2) “testing effect” or “retrieval practice” combined with “misinformation” or “false memory”; (3) “initial interview” or “repeated retrieval” combined with “eyewitness” and “misleading information”; (4) “test-induced priming” or “retrieval-induced facilitation” combined with “forensic”; and (5) targeted author searches for researchers active in this domain (Chan, LaPaglia, Gordon, Thomas, Bulevich, Wilford, Saunders, Pansky, Karanian, Wulff, Torrance). Supplementary searches included reference lists of identified articles and prior reviews (Chan & LaPaglia, 2013), forward citation searches for seminal RES articles (Chan et al., 2009; Gordon et al., 2015), conference programs from SARMAC, Psychonomic Society, EAPL, and AP-LS (2013–2024), and direct correspondence with researchers requesting unpublished data.

Studies were included if they (1) employed an experimental design with at least one retrieval-practice condition versus a no-practice control; (2) used forensic-relevant stimuli (crimes, accidents, or violent events); (3) included a misinformation phase following the retrieval manipulation; (4) assessed misinformation acceptance on a quantifiable final memory test; and (5) reported sufficient statistics to calculate effect sizes. Both peer-reviewed articles and dissertations/conference proceedings with full methodological detail were eligible. Studies were excluded if they used exclusively non-forensic stimuli (word lists, prose, neutral objects), lacked a misinformation phase or no-practice control, used clinical samples, or provided insufficient statistics despite author contact.

The systematic search identified 312 potentially relevant records. After removing 89 duplicates, 223 records underwent title and abstract screening by two independent reviewers (κ = 0.91). Following abstract screening, 63 full-text articles were assessed; 37 were excluded (14 non-forensic stimuli, 9 lacked control conditions, 8 had no misinformation phase, 4 examined special populations, 2 insufficient statistics). This resulted in 26 eligible studies. Figure 1 presents the PRISMA flow diagram.

Figure 1.

Note. Initial database searches (PubMed, PsycINFO, Web of Science, Scopus, ProQuest, Google Scholar) identified 312 records. After removing 89 duplicates, 223 records underwent title and abstract screening. Following screening, 63 full-text articles were assessed for eligibility; 37 were excluded (14 used exclusively non-forensic stimuli, 9 lacked control conditions, 8 had no misinformation phase, 4 examined special populations, 2 had insufficient statistics). The final meta-analysis included 26 studies contributing k = 72 effect size comparisons from N = 4,218 participants.

Two independent coders extracted data using a standardized coding protocol. Coded variables included study characteristics (publication year, type, sample size), stimulus event characteristics (event type, modality), initial test characteristics (format, timing, number of tests), misinformation characteristics (source format, detail type, number of items), temporal intervals (test-to-misinformation delay, misinformation-to-final-test delay), warning manipulation, and final test characteristics. Inter-rater reliability was high: κ = 0.92 for categorical variables (range: 0.86–0.97), ICC (2,1) = 0.998 for continuous variables.

The primary effect size was Cohen’s d, calculated as the standardized mean difference in misinformation acceptance between retrieval-practice and no-practice conditions. Positive d values indicate RES effects (greater misinformation acceptance following retrieval practice); negative values indicate protective testing effects. Effect sizes were calculated from means and SDs when available, otherwise from t- or F-values, or conservatively estimated from p-values and sample sizes (Lipsey & Wilson, 2001). All effect sizes were corrected for small-sample bias using Hedges’ (1981). correction; we report Cohen’s d notation using bias-corrected values throughout. Studies reporting multiple relevant comparisons were coded as separate effect sizes with dependency accounted for in analyses.

All meta-analyses were conducted in R (version 4.3.2) using the metafor package (Viechtbauer, 2010) with random-effects models and restricted maximum likelihood (REML) estimation. For categorical moderators (test format, misinformation source, warning), we conducted subgroup analyses using mixed-effects models and tested between-group heterogeneity with Q_between. For the continuous moderator (temporal intervals), we used meta-regression: d_i = β₀ + β₁._i + ε_i. Dependent effect sizes were handled via robust variance estimation (RVE) with cluster-robust standard errors treating each study as a cluster (Hedges et al., 2010); sensitivity analyses compared RVE results to single-effect-per-study and within-study aggregation approaches, yielding consistent conclusions. Publication bias was assessed using funnel plot inspection, Egger’s regression test, trim-and-fill analysis, p-curve analysis (Simonsohn et al., 2014), and three-parameter selection models (Vevea & Woods, 2005). Sensitivity analyses examined influential studies via Cook’s distance and leave-one-out analysis, and tested robustness of results to restrictions by publication type and design type.

The final meta-analysis included 26 studies published between 2002 and 2025 (median year: 2014), yielding k = 72 independent effect size estimates from N = 4,218 participants. Sample sizes ranged from 60 to 498 per study (median = 162). Twenty-three studies were peer-reviewed journal articles; three were dissertations or conference proceedings with full methodological detail. The majority (85.7 %) were from three primary research groups: Chan and colleagues (Iowa State University), Thomas and colleagues (Tufts University), and LaPaglia and colleagues. Most studies (85.7 %) used undergraduate samples; 62 % female on average. All studies used forensic-relevant stimuli, most frequently a terrorism/crime scenario from the television program 24 (57.1 % of studies), with additional event types including bank robbery scenarios (14.3 %), theft or perpetrator identification scenarios (9.5 %), and other criminal/violent events (19.1 %). Studies predominantly used between-subjects designs (71.4 %). Table 1 presents detailed characteristics of all included studies.

Table 1 Table 1 Characteristics of Included Studies (k = 26 studies, 32 experiment entries) Note.

N = sample size; d = Cohen’s d effect size (positive values indicate retrieval-enhanced suggestibility; negative values indicate protective testing effects). NR = not reported in accessible source materials; effect size direction confirmed from statistical tests reported in text. Exp. = Experiment. Studies with multiple experiments are listed separately when different moderator conditions were examined. Several included studies contributed multiple effect size comparisons (k = 72 total) based on separate moderator conditions within study. Warning column refers to the presence of an explicit warning about misinformation possibility; for warning studies, d values reflect the warning effect (tested+warning vs. control), not the baseline RES effect. Full study details available from authors upon request.

The three units of analysis are nested and should not be confused. The 26 studies are the independent published reports that met inclusion criteria. Because several studies reported more than one experiment, these yielded 32 independent experiment entries (i.e., separate samples). Within these experiments, the meta-analytic model was estimated over 72 individual effect-size comparisons, as a single experiment frequently contributed multiple comparisons (e.g., different retrieval formats, delays, or misinformation conditions measured on the same or on independent subsamples). Thus, 26 studies → 32 experiment entries → 72 effect-size comparisons. Dependencies among comparisons drawn from the same sample were handled through the robust variance estimation / multilevel approach described in the Method section. The full list of the 72 comparisons, together with the corresponding standardized mean differences (Cohen’s d) and their 95 % confidence intervals for each comparison, is provided as online Supplementary Material.

Random-effects meta-analysis across all 72 comparisons revealed a significant overall RES effect: d = 0.34, 95 % CI [0.23, 0.45], p < 0.001 (Figure 2). Retrieval-practice participants accepted an average of 45.4 % of misinformation items compared to 36.8 % in control conditions –an absolute increase of approximately 8.6 percentage points. The 95 % prediction interval [−0.04, 0.72] indicates that across populations the true RES effect ranges from near-zero to medium-large, justifying moderator analyses.

Figure 2.

Note. Random-effects meta-analysis across all 72 comparisons yielded d = 0.34, 95 % CI [0.23, 0.45], p < 0.001. Each horizontal line represents one effect size comparison, with length proportional to the 95 % confidence interval and square marker area proportional to inverse variance weight. The diamond at the bottom represents the weighted mean effect size and its 95 % confidence interval. Positive values (right of zero) indicate greater misinformation acceptance in retrieval-practice conditions relative to no-practice controls (RES); negative values indicate protective testing effects. The vertical dashed line at zero represents no effect. Heterogeneity indices: Q (71) = 163.4, p < 0.001, I^² = 56.4 %, τ^² = 0.058.

Test format was a significant moderator, Q_between (2) = 22.3, p < 0.001 (Table 2, Figure 3).

Table 2 Table 2 Moderator Analysis: Initial Test Format Note.

k = number of effect sizes; d = weighted mean effect size; CI = confidence interval; I ^² = percentage of variance due to heterogeneity. Q_between (2) = 22.3, p < 0.001. Recognition includes multiple-choice, yes/no, and forced-choice questions; Cued Recall includes short-answer and specific-detail questions; Free Recall/CI includes open-ended narrative recall and Cognitive Interview protocols.

Figure 3

Note. Subgroup analyses showing separate pooled effects for recognition-format initial tests (k = 14, d = 0.66), cued recall (k = 42, d = 0.38), and free recall/Cognitive Interview (k = 16, d = 0.17). The between-group test is significant: Q_between (2) = 22.3, p < 0.001. Effect sizes are ordered from largest (recognition) to smallest (free recall). The figure illustrates the clear gradient from maximal RES (recognition) to minimal/protective effects (free recall). Subgroup diamonds show weighted mean effects for each format; the dashed vertical line at zero marks the null effect boundary.

Misinformation source significantly moderated RES, Q_between (2) = 18.7, p < 0.001 (Table 3).

Table 3 Table 3 Moderator Analysis: Misinformation Format Note.

k = number of effect sizes; d = weighted mean effect size; CI = confidence interval; I ^² = percentage of variance due to heterogeneity. Q_between (2) = 18.7, p < 0.001. Negative d values indicate protective testing effects. The distinction between questions with and without context follows LaPaglia & Chan (2019): questions with context (surrounding narrative information reinstating event structure) produce RES, while decontextualized questions produce a protective effect.

Narrative misinformation (k = 48): d = 0.48, 95 % CI [0.35, 0.61], p < 0.001. Misinformation acceptance increased from 37.4 % to 50.8 % (13.4 percentage point increase).

Questions with context (k = 8): d= 0.55, 95 % CI [0.34, 0.76], p < 0.001. Questions embedded within event-reinstating contextual information produced RES comparable to narratives.

Questions without context (k= 16): d = −0.19, 95 % CI [−0.34, −0.04], p = 0.013. Decontextualized questions produced a significant protective testing effect. LaPaglia and Chan (2019) demonstrated this narrative versus question distinction is mediated by contextual reinstatement: it is not the surface form but the reinstatement of original event context that drives misinformation encoding. Source monitoring analyses (k = 7 studies) showed 65 % of accepted misinformation was attributed to the original event in retrieval-practice conditions versus 49 % in controls, supporting enhanced source confusion following retrieval practice.

H3 was strongly supported.

Detail type significantly moderated RES, Q_between (1) = 14.6, p < 0.001 (Table 4).

Table 4 Table 4 Summary of Moderator Analyses: Detail Type and Warning Effects Note.

k = number of effect sizes; d = weighted mean effect size; CI = confidence interval; I ^² = percentage of variance due to heterogeneity. For Detail Type: Q_between (1) = 14.6, p < 0.001. Central details include perpetrator characteristics, weapons, and primary crime actions; Peripheral details include background objects, bystander characteristics, and environmental features. For Warning Effects: Q_between (1) = 29.4, p < 0.001. Warning conditions include both pre-warning (given before misinformation exposure) and post-warning (given shortly after misinformation, before final test); both produce near-zero or negative effects (Chan et al., 2022; Karanian et al., 2020; Thomas et al., 2010; Torrance et al., 2025). A negative d in the warning condition indicates that warned retrieval-practice participants accepted slightly less misinformation than warned controls, though this difference is not statistically significant (p = 0.41).

Peripheral details (k = 38): d = 0.50, 95% CI [0.36, 0.64], p < .001. Misinformation acceptance increased from ~39.4% to 52.4%.

Central details (k = 34): d = 0.16, 95% CI [0.03, 0.29], p =.014. Only a 3.9 percentage point increase. Studies coding retrieval success (k = 8) revealed that peripheral details were retrieved less often than central details (~44% vs. 72%), and RES was substantially greater for non-retrieved details (d ≈ 0.58) than successfully retrieved details (d ≈ 0.06, n.s.), consistent with the selective attention account.

H4 was supported.

The temporal interval between retrieval and misinformation significantly moderated RES; meta-regression revealed a significant negative relationship, β = −0.021, SE = 0.007, p = 0.004.

Immediate (k = 31, < 1 hour): d = 0.46, 95 % CI [0.31, 0.61], p < 0.001.

Short delay (k = 27, 1–24 hours): d = 0.28, 95 % CI [0.13, 0.43], p < 0.001.

Long delay (k = 14, > 24 hours): d = 0.22, 95 % CI [0.04, 0.40], p = 0.017.

Immediate versus long-delay conditions differed significantly, Q (1) = 7.8, p = 0.005. Studies examining misinformation within the proposed reconsolidation window (~6 hours) showed d = 0.49 versus d = 0.26 beyond 6 hours, Q (1) = 6.2, p = 0.013. Critically, Thomas et al. (2017) demonstrated that a 48-hour delay before the final test reversed RES into a significant protective effect (d = 1.44 on accuracy), indicating that the final-test interval is also a key determinant. Chan and LaPaglia’s (2013) six-experiment PNAS study provided the most controlled reconsolidation evidence: disruption occurred only when misinformation followed retrieval within approximately 20 minutes.

H5 was supported.

Explicit warnings dramatically altered RES, Q_between (1) = 29.4, p < 0.001 (Table 4, Figure 4).

Figure 4.

Note. Subgroup analyses showing separate pooled effects for no-warning conditions (k = 55, d = 0.41) and explicit warning conditions (k = 17, d = −0.07). Q_between (1) = 29.4, p < 0.001. The figure illustrates the dramatic impact of warnings: standard RES in the absence of warnings (positive d significantly different from zero) is completely eliminated when warnings are provided (negative d not significantly different from zero, p = 0.41). The overlap between the individual effect sizes in the warning subgroup and the zero line illustrates the absence of RES in warned conditions.

Multiple complementary methods converged on minimal publication bias. Funnel plot inspection revealed slight asymmetry with a modest gap in the lower-left region, but Egger’s regression was non-significant (b = 1.11, SE = 0.79, p = 0.17). Trim-and-fill analysis imputed only 5 potentially missing studies, yielding a minimally adjusted effect of d = 0.31 versus observed d = 0.34. p-curve analysis showed significant right-skew (Z = −3.84, p < 0.001), indicating genuine evidential value. Three-parameter selection models yielded an adjusted d = 0.32, 95 % CI [0.19, 0.45], nearly identical to the unadjusted estimate. Overall, the RES effect and moderator patterns are robust to potential publication selection.

Exploratory analyses examined event type (terrorism, robbery, theft), event duration, participant age, sample type (student vs. community), and number of initial tests. Event type did not significantly moderate RES, Q (2) = 2.1, p = 0.35, nor did event duration (β = 0.002, p = 0.84), suggesting generalizability across forensic scenario types. Student (d = 0.35) and community (d = 0.30) samples produced comparable effects, Q (1) = 0.4, p = 0.53. Number of initial tests showed a positive dose-response relationship with RES magnitude (β = 0.16, SE = 0.07, p = 0.026 on log-transformed tests), consistent with Chan and LaPaglia’s (2011) finding of monotonically increasing suggestibility from 0 to 5 tests –a finding with direct forensic implications for repeated re-interviewing.

Our findings support multi-mechanism accounts of RES (Johnson et al., 1993; Schacter et al., 2011): no single framework fully explains the observed moderator pattern; instead, different mechanisms operate under different conditions.

The test format gradient (recognition > cued recall > free recall) is best explained by source monitoring failures (Johnson et al., 1993). Recognition questions maximally encourage guessing, generating internal representations that subsequently serve as competing memory sources alongside actual memories and external misinformation; when misinformation confirms prior guesses, it is misattributed to the original event (Lindsay, 2008). Free recall minimizes guessing –witnesses omit uncertain details rather than committing to potentially wrong answers –reducing source confusion. The memory reconsolidation account best explains the temporal interval effects: RES is strongest within approximately 6 hours of retrieval, consistent with a labile reconsolidation window (Chan & LaPaglia, 2013; Nader et al., 2000). However, RES persisting beyond the reconsolidation window (Chan & Langley, 2011) indicates that source confusion and TPL continue to operate at longer delays.

The narrative versus question distinction and the peripheral versus central detail gradient are best explained by test-potentiated learning and context processing (Gordon & Thomas, 2013; LaPaglia & Chan, 2019). Initial testing directs attention toward tested event aspects; when a subsequent narrative reinstates event context, tested details serve as retrieval cues enhancing processing of related misinformation content (Gordon et al., 2015, 2020). For peripheral details, failed retrieval creates gaps later filled by misinformation; for central details, successful retrieval strengthens original traces, reducing susceptibility. Narratives reinstate temporal and spatial event context, fostering gist-level integration of misinformation; decontextualized questions prompt verbatim retrieval that discriminates original from post-event information.

The complete elimination of RES by explicit warnings demonstrates that RES reflects controllable metacognitive processes rather than automatic, irreversible memory mechanisms. Neural evidence from Karanian et al. (2020) shows that warnings redirect memory retrieval toward original encoding episodes, explaining how warnings counteract RES without necessarily preventing misinformation encoding.

The following recommendations are offered as tentative, evidence-informed suggestions rather than as prescriptive protocols. They are derived from laboratory analog studies conducted largely with university samples and should be interpreted as hypotheses to be validated in field settings before being adopted in operational practice. Read with this caveat, the present findings suggest the following directions for investigative interviewing:

Free recall produces minimal RES (d = 0.17) and protects central details (d = −0.21), while recognition formats produce strong RES (d = 0.66). Initial interviews should prioritize open-ended questions eliciting free narrative recall (“Tell me everything you remember”) rather than recognition-format questions. Avoid yes/no, multiple-choice, or forced-choice questions about uncertain details, as these encourage guessing and create internal representations that later interfere with accurate memory.

RES effects are strongest when misinformation follows retrieval within one hour (d = 0.46) and decrease substantially after 24 hours (d = 0.22). Witnesses should be cautioned against discussing the event with others, consuming media coverage, or reviewing social media immediately after giving statements. When multiple interviews are necessary, waiting at least 24 hours is preferable; Thomas et al. (2017, Experiment 2) demonstrated that a 48-hour delay can reverse RES into a protective effect.

Given that RES increases misinformation acceptance by ~8.6 percentage points overall, determining which details originate from original memory versus post-event sources strongly supports the use of verbatim documentation (audio/video recording preferred, detailed written documentation minimum) before any follow-up or potential contamination. This establishes a baseline, enables assessment of contamination timing, and protects investigative integrity. RES findings provide additional scientific support for jurisdictional policies mandating recording of investigative interviews (Kassin & Gudjonsson, 2004).

Explicit warnings completely eliminate RES (d = −0.07, p = 0.41), validated across laboratory (Chan et al., 2022; Karanian et al., 2020; Thomas et al., 2010) and online (Torrance et al., 2025) samples. Before follow-up interviews, provide warnings such as: “During this interview, I may ask about details you haven’t thought about before. It’s important that you rely only on your own memory of what you actually saw. If you’re not sure about something, say you don’t know rather than guess.” Warnings are most effective when given shortly before potential misinformation exposure; warnings given 48 hours after misinformation lose effectiveness (Chan et al., 2022).

RES effects are substantially stronger for peripheral (d = 0.50) than central details (d = 0.16). Initial interviews should focus primarily on central details via free recall. Peripheral details (background objects, bystanders, environmental features), which are more vulnerable to RES and more likely to go unretrieved initially, should be documented early and treated as higher contamination risk. These recommendations align well with the Cognitive Interview (Fisher & Geiselman, 2010), the NICHD Protocol for child witnesses (Lamb et al., 2007), and the PEACE Model (Milne & Bull, 1999), extending their scientifically grounded safeguards against iatrogenic memory contamination.