The elusive backfire effect: mass attitudes' steadfast factual adherence

Political Behavior 41: 135–163 (online 2018), doi 10.1007/s11109-018-9443-y. Replication syntax: Harvard Dataverse doi 10.7910/DVN/AGRX5U (not downloaded). Read by researcher R5b (change with reasons), 5 October 2026. Provenance: papers/change/wood2019_backfire.provenance.json.

What was read

• Read in full: the publisher's version of record (29 pages), from a public third-party copy (gwern.net), because the SSRN preprint page and the Springer page both returned bot checks, which were not circumvented. Converted with pdftotext -layout: abstract, introduction, theory section, Studies 1–5, the counterargument test, robustness checks, discussion, references, Table 1 (issues and corrections), Table 2 (Study 5 samples), and every figure caption.
• Figure 1 (the summary of all correction effects) was rendered and inspected as an image, because it holds the effect sizes; its values below are read off the plot (about ±0.1). Figures 2–10 were not inspected as images.
• Not read: the electronic supplementary appendix (statement texts, items and the regression tables cited as Tables 3–14), which Springer serves only behind a client challenge; the Dataverse files.

What it is

• Backfire, as defined here (following Nyhan and Reifler 2010): the average respondent becomes less accurate on a factual question after seeing a false claim plus its correction than after seeing the false claim alone. Excluded from this definition: changes only in policy preferences, in behavioural intentions, factual polarization among the corrected, fluency effects without a randomized correction, and framing effects.
• Design. Real misstatements by politicians of both parties (Clinton, Obama, Sanders, Trump, Cruz, Ryan, Bush and others), each followed or not, at random and issue by issue, by a correction citing neutral government data. All respondents then rated agreement with the misstatement on a 5-point scale. Model: agreement ~ ideology (7-point) × correction, OLS per issue, as in Nyhan and Reifler. Issue order randomized (an adapted Latin square); party of each speaker displayed.
• Five studies, 52 issue tests, more than 10,100 respondents:

• Plus 261 MTurk raters who scored each statement–correction pair for "accordance" (how closely related, 0–100), to test counterargument theories.

Main results (verified)

• No backfire in any of the 52 tests, in any ideological group. The share of issues with a significant correction effect toward the facts: 85% among liberals, 96% among moderates, 83% among conservatives; "for about nine issues in ten, factual information significantly improves the average respondent's accuracy."
• Size (read off Figure 1): corrections lowered agreement with the misstatement by about 0.1 to 1.7 points on the 5-point scale, mostly 0.3 to 1.4. The absolute gap between the ideological extremes, averaged over corrected and uncorrected respondents, was 0 to about 2.3 points. 37% of correction effects were larger in absolute terms than the average ideological effect.
• Partisanship still shapes the size, not the sign. Respondents updated less when a co-ideologue was corrected and more when an opponent was. In Study 3 (corrections embedded in long articles, smaller effects), the most conservative respondents responded least to corrections of conservatives and most to corrections of liberals, and conservatives were less responsive overall.
• The original WMD backfire depended on the item. With Nyhan and Reifler's long item (an active programme, the ability to produce weapons, stockpiles, "but Saddam Hussein was able to hide or destroy these weapons"), all respondents ignored the correction: no update, but no backfire either. With the simple item ("US forces did not find weapons of mass destruction"), liberals adopted the correction.
• Item complexity (Study 4): none of 18 combinations backfired; conservatives were far less factually adherent when asked complex items with preambles that explain away the fact. Liberals showed no comparable variation. A correction citing the very agency Trump accused (the Bureau of Labor Statistics) still improved accuracy on his "real unemployment over 30%" claim.
• Sample (Study 5): MTurk respondents were slightly more responsive on average; the difference was significant only among moderates, not at the ideological ends.
• Counterargument does not explain the pattern: correction size was unrelated to how closely the correction contradicted the statement (accordance on 0–100; examples: 81.7 for a proximate pair, 47.2 for a distant one; Nyhan and Reifler's three pairs sat mid-range). Perceived accordance was unrelated to the rater's ideology. The authors' reading: people treat facts as "banal" and use ideology as a group heuristic rather than counterarguing; adopting an unwelcome fact costs little because it need not change their preferences.
• Robustness: same pattern using party identification instead of ideology; randomization balanced the number and runs of corrections seen; correction effects were as large on the first issue as on the last (no demand build-up).

Limits

• Immediate measurement only. The authors: acquiescing "does not mean that they have retained this information; … the facts we provided may quickly become inaccessible."
• Mostly MTurk; one quota sample. No preregistration is mentioned.
• Agreement with a misstatement, not belief in a fact, on a 5-point scale; the effect is a between-group mean difference. Significance counts across issues are not a meta-analysis, and no correction for multiple tests is mentioned.
• Updating facts is not updating attitudes. The authors argue factual corrections are cheap because they need not move preferences; this paper does not measure whether preferences moved.
• The supplementary tables, including exact coefficients, were not read here.

What it means for Kurisutina (inference)

• The base must not backfire on facts. A human-like empty slot moves toward a clear, sourced factual correction, even one that hurts its side. Partisanship should shrink that move (corrections of allies move people less than corrections of opponents), never reverse it.
• Ambiguity is where "not updating" lives. People ignored the WMD correction only when the question let them explain the fact away. For M1, a correction that leaves an alternative explanation open should move belief less than one that closes it.
• For the battery: a correction probe with a matched human reference exists in this format: a sourced correction of a politician's misstatement, agreement on a 5-point scale, effects of about 0.3–1.4 points in session, no reversal in 52 tests, and an asymmetry by ally against opponent. Persistence is not covered.