The sleeper effect in persuasion: a meta-analytic review

Psychological Bulletin 130(1): 143–172, doi 10.1037/0033-2909.130.1.143. PMC3100161 (author manuscript). Read by researcher R5b (change with reasons), 5 October 2026, from the publisher's PDF hosted on the senior author's lab site. Provenance: papers/change/kumkale2004_sleeper_effect.provenance.json.

What was read

• Read in full: all 30 pages: theory, method, results, discussion, references, Appendix, and Tables 1–6.
• Conversion note: the read file is pdftotext in reading order, because -layout put the two columns side by side behind a long margin watermark. Table 1 (per-study effect sizes, four pages) is scrambled in reading order and was read as such; its individual rows are not used here. Tables 2 and 5 were checked against the -layout extraction.
• Not inspected as images: Figures 1–8 (the curves, funnel and quantile plots); their content is described in the text.
• Errors in the article noted:
  • Table 5 prints −0.09 for "cue after message, low outcome relevance" with a 95% CI of −0.02 to 0.21, which cannot both be right;
  • the discussion says the 0.25 effect arose when "the discounting cue was first", where the results say last;
  • Figure 8's caption labels panels C and D identically.

What it is

• The sleeper effect: a message paired with a discounting cue (a non-credible source, a disclaimer that it is false, reactance-inducing statements) persuades less at first, but persuasion rises later, as the cue is forgotten or dissociated from the message while its arguments are remembered.
• Two kinds:
  • absolute: the discounting-cue group gains over time while no-message baselines do not;
  • relative: discounting-cue persuasion decays less than, or rises in parallel with, other groups.
• Mechanisms considered:
  • forgetting of the cue;
  • dissociation of cue and message (the cue stays in memory but is less accessible);
  • differential decay (the cue's impact fades faster, especially when it comes last);
  • heuristic–systematic and elaboration-likelihood accounts: arguments processed carefully persist; a cue processed effortlessly is short-lived.
• Data: 24 reports, 72 data sets, 1951–2003, all persuasive messages with arguments measured at least twice after exposure:
  • 88% university students, 89% from the US, 88% journal articles;
  • mean delay to the first follow-up 18 days (SD 16); 85% within-subject;
  • 96% one-sided messages, about 5 arguments and about 510 words;
  • cues: source credibility 71%, disclaimers 15%; cue before the message 53%, after 39%;
  • median discounting-cue group of 27 people, "lack[ing] the statistical power to detect … d = 0.20".
• Analysis: d for change within each condition (later minus earlier, pooled SD; repeated-measures variances using an average test–retest r = .42), and d between conditions at each time; fixed- and random-effects models gave near-identical estimates.

Main results (verified)

Change over time within each condition (Table 4; positive = more persuasion later):

• Overall, a relative sleeper effect only: discounting-cue groups rose by as much as baselines (0.08 each), while acceptance-cue groups decayed (difference B = −0.29, p < .0001). Heterogeneity was significant.
• No boomerang: at the immediate test, discounting-cue groups were already more persuaded than no-message controls (d+ 0.19 [0.10, 0.28], k = 26), and more so later (0.30 [0.20, 0.40], k = 22). Only one data set showed less agreement than control (d = −0.02).

Between conditions at each time (Appendix, fixed effects):

(The article prints the first comparison as −0.31 and −0.02 under a sign convention whose note is ambiguous; the direction, the message alone persuading more at first, is unambiguous.)

• The cue's suppression is gone by the delayed test: the gap between message-only and discounting-cue groups fell from 0.31 to 0.02. The persuasive effect of the arguments themselves persisted: about 0.34/0.56 ≈ 0.6 of the immediate gap over control remained after a mean of 18 days. The authors: "the impact of persuasive communications in these experiments did not disappear within the time period observed."

When the sleeper effect is absolute (Tables 5–6, random effects):

• Cue after the message and high ability or motivation: repetition 0.24 [0.15, 0.33]; high outcome relevance 0.26 [0.16, 0.36]; high prior knowledge 0.25 [0.16, 0.34]. Overall d+ = 0.25 in these conditions, against 0.08 in baselines (each difference B > 0.16, p < .02).
• Cue before the message: −0.06 to 0.02 in every cell; cue after with low ability or motivation 0.05–0.11.
• Interactions of cue timing with repetition and with relevance: B = −0.22 (p < .05), robust to fixed and random effects; with prior knowledge −0.16 (n.s.).
• Initial suppression matters: the closer the initial change was to zero (the cue cancelling strong arguments), the larger the later rise (B = −0.48, SE 0.16, k = 26). Initial change was smaller where the cue came last and ability and motivation were high (0.11 against 0.29); partial mediation (Sobel p < .11).
• Primary studies cited for elaboration: instructions to imagine using the product and think about the arguments 0.59 against −0.26 (Mazursky and Schul 1988); high need for cognition 0.61 against −0.17 for low (Priester et al. 1999).
• Type of cue (exploratory, Bonferroni): disclaimers (an explicit statement that the message is false) produced a sleeper effect (0.22 [0.13, 0.31], k = 15); non-credible sources did not (0.02, k = 45). Disclaimers always came after the message; non-credible sources placed last gave 0.09 against −0.02 placed first.
• Memory:
  • the more people still recalled the cue at the delayed test, the smaller the sleeper effect (B = −0.01 per percentage point, β = −.49, k = 11);
  • recall of the message content itself decayed strongly: d+ −0.63 (fixed) to −0.81 (random) after discounting cues, −0.70 to −0.90 after acceptance cues (k = 16 and 12);
  • in two experiments people recalled more of the arguments than of the source at follow-up.
• No moderation by delay length (the authors infer that arguments and source are forgotten at similar rates over longer times), by attitude against belief measures, or by message discrepancy. No sign of publication bias in the funnel and normal-quantile plots after removing two outliers (d = 1.13 and −1.07).

Limits

• Small, old, student studies, mostly in classrooms; underpowered cells; a few reports dominate (12 of the 72 data sets from Pratkanis et al. 1988).
• Moderators are coded from procedures, not measured (coder κ .68 for knowledge and relevance, .45 for discrepancy); ability and motivation indicators are correlated (.33–.63), so they are not separable.
• Persistence is over weeks (a mean of 18 days), not months.
• Cue recall was measured in only 22% of data sets; the memory mechanism rests on 11 of them.

What it means for Kurisutina (inference)

• Change with reasons persists better than change from cues. Arguments heard once still moved attitudes about 0.6 of their initial gap after about 18 days. A source cue's effect, positive or negative, faded within the same weeks (acceptance cues decayed, discounting cues lost their suppression). A rule of change should treat the argument content and the source cue as two inputs with different decay rates.
• Discredited content can return as the source is forgotten, when people thought hard about strong arguments and the discrediting came after. In a replica this is a source-memory process: the person's access to "who said it" fades faster than the content. That argues for keeping the system's provenance record exact while letting the person's attribution decay, as base specification B already separates them [V base_spec_b, requirement 13].
• For M1: a "discredited argument" family has a checkable time course here only as a direction and a rough size: the cue's suppression vanishes over weeks; with cue-last and high elaboration, belief rises by about d 0.25.
• For the battery: a persistence reference for argument-based persuasion: the message-only effect over control falls from d 0.56 to 0.34 after about 18 days (ratio about 0.6); the test–retest correlation of induced attitudes is about .42.