One of Us? Identity and Preferences for Controversial Politicians

The experiment is embedded in a survey conducted on 2000 individuals in the Indian state of Bihar. It begins with a priming phase, wherein the participant is randomly assigned to receive one of three conditions: 1) caste identity salience priming, 2) information and individual decision-making priming, or 3) a condition with no priming. Following the priming (if any), respondents are presented with a hypothetical political candidate running as a Member of the Legislative Assembly (MLA) with randomized attributes. The main attributes of interests are caste or religious identity and an attribute corresponding to past controversial behavior. After being presented with the hypothetical candidate, respondents are asked six follow-up questions.

2023-12-03

2024-01-07

The primary outcome is the variable Vote which is a binary variable taking the value 1
if the respondent answered that they were ”Very Likely” or ”Somewhat Likely” to vote
for the candidate and 0 otherwise.

The outcome is based on an ordinal survey question but it will be coded as binary before using it in the analysis.

Secondary outcomes are Advocate, Communal Peace, and Corruption. These are all
binary variables taking the value 1 if the respondents answered that it is ”Very Likely”
or ”Somewhat Likely” that the candidate will advocate for their interests, promote
communal peace, or misuse their position for their personal gain. These questions can
help to further understand individuals’ preferences and how this maps to perceptions of
candidate behavior. Lastly, Socially Acceptable is a binary variable that takes the value
1 if the respondent thinks that it is ”Very socially acceptable” or ”Socially acceptable”
to vote for the candidate and 0 otherwise. This variable will help to understand if
individuals’ preferences are generally aligned with those of their caste community.

The secondary outcomes are based on ordinal survey questions but all will be coded as binary before they are used in the analysis.

The experiment is embedded in a survey conducted on 2000 individuals in the Indian state of Bihar. It begins with a priming phase, wherein the participant is randomly assigned to receive one of three conditions: 1) caste identity salience priming, 2) information and individual decision-making priming, or 3) a condition with no priming. Following the priming (if any), respondents are presented with a hypothetical political candidate running as a Member of the Legislative Assembly (MLA) with randomized attributes. The main attributes of interests are caste or religious identity and an attribute corresponding to past controversial behavior. After being presented with the hypothetical candidate, respondents are asked six follow-up questions. tribute corresponding to past controversial behavior. After being presented with the hypothetical candidate, respondents are asked six follow-up questions.

See detailed description in the attached pre-analysis plan.

In this study, randomization is conducted at the individual level. Each participant is randomly exposed to both priming and a hypothetical candidate, assessing a single profile. The survey experiment was coded using KoboToolbox, and due to software limitations, random assignment of experimental attribute levels involved generating a number between 0 and 1 for each attribute. This number is scaled by the number of levels for the respective attribute, rounded to the nearest integer, and used to determine the assigned attribute level. Each attribute level is randomly assigned with equal probability.

Individual

No

2000 individuals

2000 individuals

This study employs a conjoint survey experiment featuring six experimental attributes, including priming. Each attribute comprises up to three randomly selected levels with equal probability, resulting in over 300 treatment conditions. Consequently, the average number of observations per treatment arm is anticipated to be less than 10. However, the analysis will focus on estimating the average marginal component effect. This effect is defined as the marginal impact of a specific attribute level (in comparison to the benchmark level), averaged across the joint distribution of the other attributes (Hainmueller, Hopkins, & Yamamoto, 2014). Therefore, the primary focus lies in the sum of treatment effects rather than individual comparisons.

To calculate the minimum detectable effect, I utilized the power calculation tool for conjoint experiments developed by Lukac et al. (Lukac & Stefanelli, 2020). With a chosen power level of β = 0.80, a significance level of α = 0.05, and a sample size of 2000 individuals, the minimum detectable effect size for an attribute with 3 levels is approximately 0.04. For interactions, the minimum detectable effect will be approximately 0.07 for a 3x3 interaction and 0.10 for a 3x3x3 interaction.