When is Changing Policy Positions Costly for Politicians? Experimental Evidence

We use the term “character” to generally refer to the politician’s nature or makeup. As such, our use of the term is broader than its use in theoretical models that define a politician’s “character” as being committed to a campaign platform (e.g., Kartik and McAfee 2007).

Sigelman and Sigelman (1986) also note, however, that “out-of-character” acts can raise respondents’ assessments of a (fictional) president’s “flexibility and open-mindedness.”

However, this effect is conditional: people are no more forgiving of financial scandals when they occurred long ago.

All data and replication code for this article are available via the Political Behavior Dataverse: https://​dataverse.​harvard.​edu/​dataverse/​polbehavior.

A multinomial logit model predicting treatment assignment (to one of the five treatment conditions described below) with an array of pre-treatment demographic measures (age, gender, race, education, income, income missing, political ideology, party identification, and political interest) suggested some covariate imbalance (p = .042 for test of joint significance of the model). This imbalance was driven primarily by imbalance in respondent race across conditions. In the analysis we report below we include controls for this vector of demographic measures. Identical models excluding these controls yield substantively similar results.

For example, -1 indicates that the respondent either strongly supported or strongly opposed the policy and was presented with a representative who adopted the opposite position; 1 indicates that the respondent had a strong preference and agreed with the representative’s position. Respondents indicating that they either “somewhat” supported or opposed the policy described were scored as −.33 or .33, depending on whether they (somewhat) agreed with the representative’s position.

In additional analysis, presented in Supplementary Table A3, we assess whether Policy Agreement moderated the effects of the Change of Position treatments. However, theoretically we might expect this interaction to be either positive or negative. If a change of position fosters doubts about the sincerity of a candidate’s true current position we would expect a change of position to attenuate the positive relationship between policy agreement and evaluations of the candidate (negative interaction term). Alternately, current substantive policy agreement may attenuate the negative effect of having changed position (positive interaction term). We find no evidence of a statistically significant interaction between agreement on Abortion and either Change of Position treatment and including these terms does not jointly improve the fit of the model (p = .627; Supplementary Table A3, column [1]). Nor do we find evidence that agreement with the representative’s position on Social Security or nuclear power moderate the effects of a position change on Abortion (Supplementary Table A3, column [2]). Unfortunately, our data do not provide a way to determine whether these null findings are the product of off-setting positive and negative moderating effects or if no such moderating effects exist.

Estimates drawn from a model regressing summary candidate evaluations on indicators for the Change of Position treatments, the Agreement: Abortion indicator, and interactions between Agreement: Abortion and each Change of Position measure indicate that this difference is statistically significant (p = .036; two-tailed test).

Question wording: “Regardless of what you think should or should not be done in each of the following areas, how important (relative to all other issues—not just those listed here) are the following issues to you?”.

Question wording: “How confident are you that you have the knowledge and expertise to evaluate policy in each of the following areas?”.

In additional analysis we find that, as with the analysis presented in the first experiment, the treatments affected each of the five measures used to construct the Evaluation of Candidate scale similarly (see Supplementary Table A4). As in the first experiment we find no evidence of a statistically meaningful interaction between the Change of Position and Policy Agreement measures. See Supplementary Table A3, column (3) and footnote 7 above.

For each issue we estimated models regressing summary candidate evaluations on a Change of Position indicator, a policy Agreement indicator (excluding cases where respondents did not report a directional policy preference), and the interaction between the two. A target who had changed to the respondent’s preferred position was evaluated more favorably than one who was not described as having changed positions, but with whom the respondent disagreed (Social Security: p = 0.277; ISIS: p < .01; Nuclear Plants: p = 0.054; Abortion: p < .01; two-tailed tests).

Models entering only the Importance interactions or only the Confidence interactions yield substantively similar conclusions.

Note that respondents assigned to a “no position change” condition were, nonetheless, randomly assigned to an issue area. We use measures of agreement, confidence, and importance on this randomly assigned issue for these respondents. After creating these measures we rescaled the Confidence and Importance measures to have a mean of 0 and standard deviation of 1.

There are several potential explanations for the fact that the Confidence × Change of Position interaction is substantial and statistically significant in the collapsed analysis presented in Supplementary Table A5, but the interactions between separate measures of Confidence and the Change of Position treatments in Table 2 are not (with the exception of Abortion). Perhaps the most likely explanation is simply that the experimental design does not provide sufficient statistical power to identify these effects once they are disaggregated or to identify real differences in how confidence moderates responses to different types of issues.

These items asked respondents to rate the likelihood of a change on seven additional issues: nuclear power, taxes, Social Security, Medicare, foreign policy, environmental policy, and immigration. Factor analysis indicated that these measures loaded onto a single underlying factor. Rotated factor scores were used to generate a scale of likelihood of change across these issues with a mean of 0 and standard deviation of 1.

Because the policy position a politician adopts matters little to individuals who do not have a preference on the issue at hand, we should expect them to respond to a change in position unfavorably (or be indifferent to the change at best). Thus, all else equal, the larger the share of voters who are indifferent on a given policy matter, the more damaging a change of position is likely to be.

If a politician faces a situation where a prospective new position is strongly supported by a large swath of the public, and only somewhat opposed by the rest, this calculus changes. For example, the estimates in Panel D of Fig. 5 suggest that in a situation where the public is evenly divided between supporting and opposing a prospective new position on abortion, changing positions is expected to yield a net .229 unit penalty. However, if we posit that 40 percent of the public strongly supports the new position and 40 percent only somewhat oppose (with the remaining 20 percent being indifferent), this penalty is reduced to essentially zero (estimate = .013; p = .862).

For example, President Obama’s much publicized change of position on same-sex marriage may have stemmed in part from an expectation that public opposition to same-sex marriage will continue to wane.

Specifically, in columns (1) and (3) of Supplementary Table A7 we estimate models analogous to those presented in column (1) of Tables 1 and 2, interacting the Change of Position indicators with an indicator for whether the target’s current pattern of policy positions indicated ideological consistency (the ideological valence of each issue was determined based on the relationship between respondents’ self-reported ideology and their issue attitudes). Statistically significant coefficients on these interaction terms would indicate that people respond differently to a change of position when that position makes the politician more ideologically consistent. However, the coefficients fall well short of statistical significance. In columns (2) and (4) of Supplementary Table A7 we instead add an indicator for respondents presented with a politician who had been ideologically consistent, but changed positions in a way that appeared to defy his ideology. We find no evidence that people reward (or punish) a representative for adopting a less ideologically rigid policy platform.