Ambiguity and valence of descriptive norms: Causal evidence from climate-policy messaging

Each respondent is randomly presented with descriptive-norm information about earlier public support for a climate policy. The information independently varies valence--high endorsement (74% or 75% of a past sample) versus low endorsement (26% or 25%)-- and ambiguity--a precise percentage versus a vague phrase ("more than half" or "less than half"). This 2x2 factorial manipulation precedes belief-elicitation and attitude questions for both a ban on combustion-engine cars and a debt-financed green-infrastructure programme, with one information condition assigned between subjects for each policy block.

At the start of each policy block, we introduce respondents to a piece of descriptive-norm information that reports selected responses from an earlier study. For the car-ban block, the message states that, out of 100 earlier respondents, either 74 or 26 people supported the ban, or, in vaguer language, that "more than half" or "less than half" did so. The infrastructure block mirrors this structure, substituting 75, 25, "more than half", and "less than half" for the corresponding numbers or phrases. The treatment arms jointly manipulate two features of the social norm: valence (a favorable vs. an unfavorable signal) and ambiguity (numeric precise vs. linguistic imprecise information). Respondents are randomly assigned to the four treatment conditions, guaranteeing that each respondent encounters only one of the four formulations per policy and that assignment is orthogonal across blocks. The information appears on its own screen immediately after the policy description, and functions as the sole experimental treatment whose causal effects the study aims to measure.

2025-09-18

2025-12-31

Our primary outcome is the point estimate ("best guess"; introspection) of current peer support. Respondents face an all-or-nothing accuracy bonus, i.e. they receive an incentive of approx. EUR 10 (paid out as exchangeable points of the panel provider) if they guess the number correctly.

The primary outcome variable is directly observable. No additional measures have to be constructed.

We elicit the subjective distribution over support levels by asking respondents to distribute 20 tokens into five bins. In addition, we also ask for respondents' confidence on their point estimate (on a 5-point Likert scale), their confidence in the full distribution (5-point Likert scale) and their own policy stance (5-point Likert scale). Responses on the secondary outcome variables are not incentivized.

Regarding the belief distribution, we operate with two statistics, the weighted mean (WM) and the normalized entropy (NEI). The WM is computed as the sum of the proportion of tokens allocated to each interval times the interval's midpoint. The NEI is a normalized measure on belief dispersion with 0 indicating that all tokens were allocated to a single interval, and 1 indicating a uniform distribution of tokens across all the five intervals. We may also compute binary support indicators, taking 1 if the respondent chooses "Somewhat support" or "Strongly support" and 0 otherwise.

This study is an online survey experiment with 3,200 quota-balanced French adults recruited through a panel provider in summer 2025. Each respondent answers two policy blocks--first, either a ban on combustion-engine cars or a debt-financed green-infrastructure programme, with block order randomized. At the start of each block, the respondent receives descriptive-norm information about earlier public support; this information varies orthogonally in valence (high vs. low support) and ambiguity (exact percentage vs. vague verbal phrase), creating four between-subject conditions per policy. Immediately afterward, the survey elicits a point estimate of current support, a full subjective distribution, confidence rating, and the respondent's own stance. One market point-estimate question is randomly pre-selected for payment; respondents earn approx. EUR 10 bonus if their answer exactly matches the realized proportion in the sample. The project is cleared by the IÉSEG ethics board; all data are collected and stored in accordance with data privacy standards.

Each policy block opens with a description of its climate measure and then displays descriptive-norm information that purports to summarize a national survey conducted three years earlier. For the combustion-engine-car-ban block, the information states that, out of 100 earlier respondents, either 74 or 26 supported the ban or--using vague wording--that "more than half" or "less than half" did so. The green-infrastructure block mirrors this structure, substituting 75, 25, "more than half", or "less than half" for the corresponding values. A screener appears after the first policy block. If answered correctly, the experiment proceeds to the second policy block.

We randomly assign each respondent to one of these four formulations independently for each policy, using single randomization that fully crosses social-norm valence (high vs. low support) with ambiguity (exact figure vs. vague phrase) at the respondent level. The information appears on its own screen, and serves as the sole experimental treatment.

Immediately afterward, the questionnaire collects five belief-and-attitude measures in a fixed order, 1. a 0-100 slider estimate of current peer support, 2. confidence in that estimate, 3. a five-bin subjective probability distribution over possible support levels, 4. confidence in that distribution, and 5. the respondent's own stance on the policy. Timers prevent rapid progression and enforce minimum viewing times.

One of the two slider questions is randomly pre-selected ex-ante for an all-or-nothing approx. EUR 10 accuracy bonus. Respondents are reminded of this incentive directly on the respective page.

Beyond the core treatment and outcomes, the questionnaire also measures risk aversion (through list elicitation of the certainty equivalent of a bet on a color in a deck of cards with half winning cards and half loosing cards), ambiguity aversion (through the difference of a list elicitation of the certainty equivalent of a bet on an deck of cards with unknown composition and the known 50-50 deck) , three-item risk literacy, political ideology and first-round presidential vote, an extensive climate-attitudes battery (importance, perceived consequence, macro-economic expectations, personal behavioral intentions, and perceptions of norms and policy instruments), two attention checks, and basic demographics (gender, age, region, urbanity, occupation, education).

The panel sample will be roughly representative of the French adult population with regard to gender, age, education and region.

Random draws via the Qualtrics randomization function.

Respondents are randomly allocated to one of four treatment arms.

No

800 respondents per treatment condition, i.e. a total of 4x800 = 3,200 respondents

800 respondents per treatment condition, i.e. a total of 4x800 = 3,200 respondents

800 respondents per treatment arm

The minimum detectable effect size in terms of Cohen's d is 0.2 for a statistical power of 0.8 and p-value of 0.01 when corrected for multiple hypothesis testing (i.e., 0.01/6). The minimum detectable Pearson correlation coefficient (within each treatment with 800 respondents) is 0.12 for a statistical power of 0.8 and p-value of 0.01.