Partisanship shapes how people understand the world. Social identities generally help humans process information effectively (Kunda, Reference Kunda1990), and in politics, partisanship is a key identity guiding this process (Campbell et al., Reference Campbell, Converse, Miller and Stokes1980; Huddy, Reference Huddy2001; Lodge and Taber, Reference Lodge and Taber2013; Mason, Reference Mason2018). Partisanship shapes how people genuinely perceive the economy (Bartels, Reference Bartels2002; Graham, Reference Graham2025), moderates their interpretation of anti-democratic acts (Graham and Svolik, Reference Graham and Svolik2020), influences who they are willing to date (Huber and Malhotra, Reference Huber and Malhotra2017; Turnbull-Dugarte and López Ortega, Reference Turnbull-Dugarte and López Ortega2025), socialize with (Engelhardt and Utych, Reference Engelhardt and Utych2020), hire (Gift and Gift, Reference Gift and Gift2015), and biases the inferences partisans make about others’ politics (Turnbull-Dugarte and Wagner, Reference Turnbull-Dugarte and Wagner2026). Partisan cues, statements, or acts that connect information to partisanship often trigger such reasoning (Zaller, Reference Zaller1992; Druckman et al., Reference Druckman, Leeper, Slothuus, Lavine and Taber2018; Groenendyk and Krupnikov, Reference Groenendyk and Krupnikov2021). By telling citizens how to interpret an issue through a partisan lens, a simple statement by a party or politician can suffice to shape people’s attitudes (Bolsen et al., Reference Bolsen, Druckman and Cook2014; Bisgaard and Slothuus, Reference Bisgaard and Slothuus2018; Barber and Pope, Reference Barber and Pope2019; Slothuus and Bisgaard, Reference Slothuus and Bisgaard2021).

However, no prior study has provided causally identified evidence of partisan cue effects in a naturalistic setting with fine-grained temporal resolution. While survey experiments routinely measure cue effects within minutes of delivery (Bolsen et al., Reference Bolsen, Druckman and Cook2014; Broockman and Butler, Reference Broockman and Butler2017; Barber and Pope, Reference Barber and Pope2019; Graham, Reference Graham2025), the external validity of these designs is not without limitations: stimuli are artificial, information environments are controlled, and exposure to research-administered stimuli is essentially guaranteed. The rare quasi-experimental designs that maximize external validity, like the pioneering work by Bisgaard and Slothuus (Bisgaard and Slothuus, Reference Bisgaard and Slothuus2018; Slothuus and Bisgaard, Reference Slothuus and Bisgaard2021), for example, demonstrate cue effects in externally valid settings, but their panel designs measure shifts in public opinion after weeks or months—time in which in- and out-group rhetorical nudges have proliferated via the media and social networks and have been internalized by group members, making it impossible to isolate the initial cue from its downstream amplification. Crucially, it remains unclear whether citizens apply partisan schemas in the moment they encounter a cue from a co-partisan leader, or whether opinion change requires the mediation of subsequent partisan debate and reinforcement. The competing view, implicit in observational designs with week- or month-long intervals, is that naturalistic cue uptake is not instantaneous but requires the amplification that follows an elite signal: news coverage, social media diffusion, peer discussion, and repeated partisan framing (Zaller, Reference Zaller1992; Bisgaard and Slothuus, Reference Bisgaard and Slothuus2018). On this account, what looks like a partisan response to the cue itself may in fact be a response to the ecosystem of partisan commentary that develops around it. If true, immediate partisan updating would be unlikely outside experimentally controlled conditions. We test whether this mediation is necessary, or whether the partisan screen is applied the moment citizens encounter information about a co-partisan leader’s action.

The lack of externally valid, temporally fine-grained evidence is problematic because the effects of partisan cues should theoretically occur immediately. Partisan cues are powerful because they are affectively charged (Lodge and Taber, Reference Lodge and Taber2013), allowing people to immediately evaluate information as positive or negative depending on whether it is congruent with their salient political group-based identity. While this theoretical problem implies that opinion shifts should ideally be measured right after a cue, longer gaps between cue and opinion measures also create empirical problems: outside experimentally controlled contexts, partisan cues usually prompt responses from opposing parties and media coverage (Bisgaard and Slothuus, Reference Bisgaard and Slothuus2018; Slothuus and Bisgaard, Reference Slothuus and Bisgaard2021) that complicates causally identifying the immediate effects of the initial cue. Moreover, cues are subsequently diffused through social networks, further complicating whether opinion shifts are a product of elite messaging or the social diffusion of it.

In this paper, we therefore ask whether we can observe immediate effects of a partisan cue in an externally valid setting. To answer this question, we employ the unexpected event during survey design (UESD) identification strategy (Muñoz et al., Reference Muñoz, Falcó-Gimeno and Hernández2020),Footnote ¹ analyzing how US President Trump’s unexpected announcement of the US bombing of Iranian nuclear sites shapes public opinion. These data provide a unique opportunity to examine partisan reasoning in real time: our data assess how specific policy support, in this case support for US military intervention in Iran, is shaped en masse, and in a partisan-oriented asymmetric direction, within the minutes and hours immediately after Trump announced US military intervention. This allows us to test the effect’s immediacy that we would theoretically anticipate from a partisan cue, and it isolates the effect from subsequent partisan debates and media coverage.

We find that partisanship instantly and strongly conditions how respondents interpret the bombing. Republicans surveyed after Trump’s post are more than 100% more supportive of the military intervention, whereas support among Democrats falls by 68% following the cue. This divergence in public opinion is robust across a multiverse analysis of 28,672 specifications, cannot be explained by pre-treatment trends or Type I errors, and is unique to the issue discussed in the cue. These results contribute a crucial empirical extension to the literature on partisan reasoning: they demonstrate how partisan cues prompt immediate and strong opinion divergence outside an experimentally controlled setting.

These findings have important substantive implications. First, partisan reasoning is so ingrained in American voters that it happens powerfully and immediately. The cue created stark divisions on an issue that was neither important nor polarized for most Americans. While foreign politics is generally secondary to many Americans (Bennett et al., Reference Bennett, Flickinger, Baker, Rhine and Bennett1996), it was overshadowed by discussions about the economy and democracy at the time of data collection in June 2025. Both Republicans and Democrats were united in their skepticism of military interventions in Iran (23 vs. 15%) and preferred diplomatic solutions (61 vs. 58%) days before the cue.

Yet the cue prompted strong opinion divergence, nonetheless. This suggests that partisan schemas are applied near-automatically at the moment of exposure. Citizens activated an ingrained partisan heuristic before any party elite commentary, news framing, or peer discussion had time to develop. The rapidity of the effect is consistent with affective intelligence accounts of partisan reasoning (Lodge and Taber, Reference Lodge and Taber2013), in which co-partisan cues trigger immediate evaluative responses rather than deliberative updating. We acknowledge, however, that our design cannot distinguish between several theoretically plausible mechanisms: automatic in-group/out-group heuristics, motivated reasoning to protect political identity, dissonance reduction given the prior anti-intervention stance of Republicans, or expressive responding. What our evidence establishes is the speed and scale of partisan divergence in a naturalistic setting, not its underlying mechanism.

Second, our findings speak to the established implication of strong partisan responsiveness as a threat to democratic accountability. While Trump repeatedly suggested his administration would refrain from military interventions should he be granted a second term,Footnote ² his partisans immediately tolerated this violation of a campaign promise. This pattern is consistent with a top-down model of public opinion (Zaller, Reference Zaller1992; Achen and Bartels, Reference Achen and Bartels2016; Bartels, Reference Bartels2023): rather than holding the president accountable for a reversal of a campaign commitment, Republican identifiers updated their own positions to align with his action. In this policy domain at least, the president faced no bottom-up opinion constraint (from supporters). Given the opposing direction of opinion among Democrats, we can rule out the alternative explanation that the increase in support is purely a reflection of well-documented rally-round-the-flag effects (Mueller, Reference Mueller1970, Reference Mueller1973; Lai and Reiter, Reference Lai and Reiter2005) associated with military conflict (but see Seo and Horiuchi (Reference Seo and Horiuchi2024) for an empirical rebuttal).

1. Identification strategy

This study exploits an unexpected event during survey fieldwork design (Muñoz et al., Reference Muñoz, Falcó-Gimeno and Hernández2020) to estimate the causal effect of elite cues on public opinion. The UESD identification strategy has been leveraged to causally identify the effect of conflicts (Balcells and Torrats-Espinosa, Reference Balcells and Torrats-Espinosa2018; Hernández and Ares, Reference Hernández and Ares2023; Balcells et al., Reference Balcells, Tellez and Villamil2024), democratic backsliding (Chan, Reference Chan2025), election results (Chan, Reference Chan2024), political violence (Nai et al., Reference Nai, van Erkel and Bos2025), judicial decisions (Turnbull-Dugarte and Devine, Reference Turnbull-Dugarte and Devine2022), and other political events (Unan and Klüver, Reference Unan and Klüver2025; Schulte and Bischof, Reference Schulte and Bischof2026; Frese, Reference Frese2026a).Footnote ³

During the data collection for an online survey administered by YouGov among a representative sample of Americans, the United States conducted military strikes against Iranian nuclear facilities (June 21, 2025). The event, which received extensive and immediate media coverage globally, represents a real-world information shock: not only did it inform citizens of US military intervention in Iran, but it was also a deviation from the Trump 2.0 administration’s campaign promise to limit and reduce the US military’s involvement in other nations. Because the airstrikes occurred while interviews were actively underway, some respondents completed the survey immediately before the announcement, while others completed it within minutes afterward.

Formally, respondents interviewed before the public announcement serve as the control group, while those interviewed afterward serve as the treated group. The running variable is the precise interview start time, measured in minutes relative to the announcement. Under the assumption that, absent the event, potential outcomes would have evolved smoothly across the threshold, any discontinuous jump in attitudes observed at the cut-off can be interpreted as the causal effect of exposure to the airstrike announcement. The design is underpinned by the as-if random assignment of interview timing around the event (see Figure 1), which shows a continuous distribution of interviews before and after the threshold with no evidence of bunching or manipulation.

Figure 1.

Distribution of survey respondents on either side of the treatment threshold.

1 Long description

Temporal distribution in minutes of interviews by treatment condition. The horizontal axis label is, Minutes relative to treatment, 30 minute bins. The horizontal axis range shows tick labels from minus 1320 to 2280. The vertical axis label is, Number of interviews. The vertical axis shows tick labels at 0, 20 and 40. A vertical dashed line is labeled, Treatment threshold, at 0 minutes. Text labels inside the plot read, Control, less than 0 minutes and, Treatment, greater than or equal to 0 minutes. The histogram shows many bins with nonzero counts both before 0 minutes and at or after 0 minutes. The tallest bar appears in the control group near minus 1320 minutes and rises above the 40 tick. Several other control bins between about minus 780 and minus 420 minutes rise to around the 40 tick. At or after 0 minutes, multiple bins rise above the 20 tick, with several peaks between about 660 and 1560 minutes and another cluster of taller bins between about 2100 and 2280 minutes. Several bins close to 0 minutes are short, with bars near the baseline immediately before and immediately after the threshold.

Figure 1 confirms the validity of the research design. Interviews are evenly distributed around the treatment threshold, with no discernible discontinuity in survey timing. This pattern is consistent with as-if random assignment of respondents into the pre- and post-announcement conditions, satisfying the central identifying assumption of the UESD framework. A balance test modeling treatment assignment as a function of observable covariates is reported in Appendix Table A4. In addition to observing partisan balance among the control and treatment groups, we observe no variation of note or significance between the two groups in terms of gender, race, sexuality, age, education level (and status), urban/rural location, income, or voter registration status. We can be confident that the difference between the two groups is exposure to the announcement of military action.

\begin{equation*}{Y_i} = \alpha +{\delta_1}{\mathrm{Treatment}_i}+{\delta_2}\,\mathrm{Partisan}\, {\text{ID}_i}+{\delta_3}({\mathrm{Treatment}_i}\text{*}\,\mathrm{Partisan}\, {\text{ID}_i})+{\epsilon_i}\end{equation*}

To test whether partisans react differently to the same elite cue, the analysis models the outcome variable—support for US military intervention in IranFootnote ⁴—as a function of treatment exposure, partisanship, and their multiplicative interaction (Equation 1).Footnote ⁵ The estimands of interest are the conditional average treatment effect (CATE) of quasi-random exposure among each group of partisans (Democrats and Republicans), as well as the difference in the CATE between the two. Our modeling approach relies on linear regression and reports robust (HC2) standard errors.Footnote ⁶

2. Results

Figure 2 plots mean levels of support for military intervention in 30-min intervals relative to the announcement, separately for Democrats and Republicans.Footnote ⁷ Among Democrats, support for intervention is, if anything, modestly increasing in the pre-treatment period—consistent with the small positive pre-treatment trend identified in the placebo analysis (Appendix Figure C.1)—but drops significantly, sharply and immediately afterwards. Among Republicans, support is flat before the announcement and spikes significantly, sharply and immediately after the announcement. The fitted local linear regressions on either side of the cut-off reveal a clear discontinuity at time $t = 0$, indicating a partisan divergence triggered by exposure to the airstrike news. In short, the same factual event is interpreted through a partisan lens, consistent with theories of motivated reasoning and elite cue-taking: Republicans instantly and dramatically update their support for US military intervention in Iran, whereas Democrats, reacting in direct opposition to the out-group party actions, instantly become less supportive of the same issue in response to the same information. This diversity in reactions does not increase as time passes (the post-treatment linear trends are flat), which indicates that the results do not emerge or compound as additional media reporting diffuses over time. In other words, the party gap that is observed 30 min after the news broke is statistically symmetrical to that which is observed 10 h or 40 h after it broke.

Figure 2.

Descriptive illustration of discontinuity in time.

Figure 2 Long description

Discontinuity in partisan support at the announcement threshold Question: Do you think the U. S. military should or should not bomb Iranian nuclear facilities? The y-axis is labeled, Support for US military intervention in Iran, with tick labels at 0 percent, 25 percent, 50 percent, 75 percent and 100 percent. The x-axis is labeled, Minutes relative to treatment (30-minute bins). Tick labels shown are minus 1320, minus 1140, minus 960, minus 780, minus 600, minus 420, minus 240, minus 60, 120, 300, 480, 660, 840, 1020, 1200, 1380, 1560, 1740, 1920, 2100 and 2280. A vertical dashed line marks the threshold at 0 minutes. A legend inside the plot labels two series: Republicans and Democrats. Circular markers are plotted across the full x-axis range for both series. Two fitted lines are drawn, one for Republicans and one for Democrats, each with a shaded band labeled as 95 percent confidence intervals. For Republicans, the fitted line to the left of the dashed line slopes downward from about 45 percent near minus 1320 minutes to about 35 percent near minus 60 minutes. To the right of the dashed line, the fitted line starts near about 62 percent at 0 minutes and rises gradually to about 70 percent by about 2280 minutes. For Democrats, the fitted line to the left of the dashed line slopes upward from about 5 percent near minus 1320 minutes to about 25 percent near minus 60 minutes. To the right of the dashed line, the fitted line starts near about 7 percent at 0 minutes and stays near about 10 percent through about 2280 minutes. Text below the plot reads, Outcome coded: ′should′ (1) vs. ′should not′/′unsure′ (0). 95 percent confidence intervals. Overlaid observations indicate mean support in 30-minute bins.

Figure 3 formalizes these differences using a regression model which estimates support for US intervention as a function of treatment allocation conditional on party identification using a global bandwidth (see Table B.1 for regression output and Figure C.2 for bandwidth sensitivity tests). Figure 3 presents predicted probabilities of support for intervention conditional on treatment and partisan identification and, inset, the (difference in) CATE. The predicted outcomes illustrate that support for bombing Iran increases from 0.34 to 0.70 among Republicans following the announcement, while it decreases from 0.16 to 0.05 among Democrats. Note that, among the control group, the mean position among both Democrats and Republicans is opposition to intervention, indicated by predicted values below 0.5. Among the treated, there is a clear disparity between the exceptionally opposed position of Democrats (0.05) and the strong endorsement among Republicans (0.7). That is, Republicans (on average) become strongly supportive of intervention only after the announcement, while Democrats harden their opposition further. In magnitude terms, these are not small shifts in opinion. The estimated Democrat CATE is a decline of 10.9 percentage points, which corresponds to roughly 0.37 standard deviations ( $d = - 0.37$), while the Republican CATE is an increase of 36.0 percentage points, equivalent to $d = 0.77$. By conventional benchmarks, this implies a moderate effect among Democrats and a large effect among Republicans.

Figure 3.

Causal evidence of partisan responsiveness in real-time (see Appendix Table B.1).

Figure 3 Long description

Predicted support and conditional average treatment effects. Main chart with two categories on the horizontal axis labeled, Control and Treatment. The vertical axis has tick labels 0.00, 0.25, 0.50, 0.75 and 1.00. A legend below the plot shows Democrat marked with a circle symbol and Republican marked with a triangle symbol. Democrat series: At Control, a circle marker is labeled 0.16 with a vertical confidence interval bar. A dashed line connects to Treatment, where a circle marker is labeled 0.05 with a vertical confidence interval bar. Republican series: At Control, a triangle marker is labeled 0.34 with a vertical confidence interval bar. A dashed line connects to Treatment, where a triangle marker is labeled 0.7 with a vertical confidence interval bar. Inset chart title text reads, Conditional average treatment effect. The inset horizontal axis shows tick labels from minus 0.5 to 0.5 in steps of 0.1 and a vertical dashed reference line at 0.0. Two point estimates with horizontal confidence intervals are shown: one labeled minus 0.11 and one labeled 0.36. A callout in the inset reads, Difference in CATE: 0.469 (p less than 0.001).

The estimated difference in CATE ( $0.47$) indicates that exposure to the same information cue produces sharply divergent attitudinal shifts across partisan lines. The divergence is not small. In real terms, Republicans respond to news of the Republican president’s action by increasing their support for the issue by more than 100% compared to the control group baseline, and Democrats depress their support for the same issue by 68%. This pattern illustrates real-time partisan updating in response to unexpected international events which, as a consequence, results in an increase in the disparity in public opinion upon polarized party lines.

To probe the stability of these findings, Figure 4 presents a multiverse specification curve (Simonsohn et al., Reference Simonsohn, Simmons and Nelson2020) covering 57,344 identified conditional treatment effects (28,672 per partisan group) and 28,672 tests of the difference between these two CATE estimates. Each specification varies both the inclusion of covariates (i.e., age, gender, race, LGBT (lesbian, gay, bisexual or trans) identity, level of education, income, student status, urban/rural location, religiosity, voter registration status, ideology, and having a social media account on X Footnote ⁸) as well as the temporal bandwidth around the event window. The specification curves demonstrate remarkable robustness. Across virtually all model variants, estimated treatment effects remain negative among Democrats (median ATE $= -0.104$, $p = 0.032$) and positive among Republicans (median ATE $= 0.315$, $p \lt 0.001$). These results confirm that the observed partisan asymmetry is not an artifact of model specification or sample restriction.

Figure 4.

Multiverse analysis across alternative specifications.

Figure 4 Long description

Multiverse analysis testing specification sensitivity. Three specification curve line plots are labeled, Democrats, Republicans and Difference Republicans minus Democrats. Each plot uses an ordered specification index on the horizontal axis with tick labels 0, 10000, 20000 and 30000. Each plot uses a vertical scale with tick labels minus 0.4, minus 0.2, 0.0, 0.2 and 0.4. Democrats: One curve increases gradually across the ordered specifications. It starts near minus 0.35 and ends near minus 0.10. The curve stays below 0.0. Republicans: One curve increases gradually across the ordered specifications. It starts near 0.10 and ends near 0.35. The curve stays above 0.0. Difference Republicans minus Democrats: One curve increases gradually across the ordered specifications. It starts near 0.20 and ends near 0.45. The curve stays above 0.0. Below the line plots are two aligned rows of rectangular grids. The first row is labeled, Controls in specification. The second row is labeled, Bandwidth in specification. Each row contains three grids aligned under the three plot labels. Each grid is made of many small cells with varying light and dark shading.

As stated above, it is again worth emphasizing that the effects are far from trivial. Even under the small proportion of specifications among Democrat respondents where the p-value is above conventional thresholds (4.3% of estimations), an 11 percentage point estimate represents an effect that is statistically improbable under the null of no effect (see Appendix Figure C.3 for more detail). Updating among Republicans is more precisely estimated. That said, there are zero specifications among those considered where the difference between how Republicans and Democrats react to the news is not statistically significant. In other words, we can confidently and precisely reject the null of effect homogeneity between the two partisan camps. In terms of the magnitude of the partisan divergence, the median difference in the CATE from 28,672 comparisons is 0.42 ( $p \lt 0.001$).

In addition to the multiverse analysis reported here, which demonstrates that our causal estimates are not sensitive (in magnitude or significance) to alternative specifications, we also demonstrate that the identified effects are not the result of pre-treatment trends, nor are they of a comparable size to those that might be observed by chance (type I error).

First, as shown in Appendix Figure C.1, we re-estimate the analysis using the median interview start time as a faux treatment threshold, as recommended by Muñoz et al. (Reference Muñoz, Falcó-Gimeno and Hernández2020), to test for the presence of pre-treatment linear trends. The results of this analysis demonstrate that the identified treatment effects are not the result of pre-existing time trends in public opinion.Footnote ⁹ Second, given the potential increased risk of Type I errors in UESD applications (Frese and Riaz, Reference Frese and Riaz2025), we assess whether the observed effects are distinct from those which might be identified by chance. We rely on randomization inference (Gerber and Green, Reference Gerber and Green2012) and identify zero instances in 10,000 where effects comparable to those causally identified would be observed under the null of no treatment effect and the null of effect homogeneity between partisans.Footnote ¹⁰ Third, we test for the conditional effect of treatment on distinct placebo items (evaluations of inflation and immigration) which, theoretically, should not be asymmetrically impacted by treatment. As shown in Table B.2, the overall partisan responsiveness we observe in the case of support for US military intervention in Iran is not a product of broader general partisan division that emerges in response to treatment. There is treatment-induced partisan responsiveness in evaluations of Trump’s handling of the Israel-Iran conflict but no effects on evaluations of inflation or immigration.