Hi,
I am conducting a conjoint experiment, and my theory predicts that treatment effects vary by respondents’ socioeconomic status—specifically, whether they are poor and employed. I have approximately 1,000 respondents completing a 5-task forced-choice conjoint, yielding about 10,000 profile-level observations.
I am interested in identifying the best approach to estimating heterogeneous treatment effects in this setting. Many applications examine heterogeneity by interacting conjoint attributes with a single pre-treatment covariate (e.g., poor vs. non-poor) or by estimating AMCEs separately by subgroup. However, because my theory involves two conditioning variables (poverty status and employment status), I am unsure about the preferred strategy.
Specifically, should I:
1. Estimate a model with triple interactions between conjoint attributes, poverty status, and employment status; or
2. Conduct a subgroup analysis, estimating effects separately for poor–employed, poor–unemployed, non-poor–employed, and non-poor–unemployed respondents?
I would greatly appreciate guidance on best practices, including tradeoffs related to interpretability, statistical power, and inference in conjoint designs.
Thank you very much.
I am conducting a conjoint experiment, and my theory predicts that treatment effects vary by respondents’ socioeconomic status—specifically, whether they are poor and employed. I have approximately 1,000 respondents completing a 5-task forced-choice conjoint, yielding about 10,000 profile-level observations.
I am interested in identifying the best approach to estimating heterogeneous treatment effects in this setting. Many applications examine heterogeneity by interacting conjoint attributes with a single pre-treatment covariate (e.g., poor vs. non-poor) or by estimating AMCEs separately by subgroup. However, because my theory involves two conditioning variables (poverty status and employment status), I am unsure about the preferred strategy.
Specifically, should I:
1. Estimate a model with triple interactions between conjoint attributes, poverty status, and employment status; or
2. Conduct a subgroup analysis, estimating effects separately for poor–employed, poor–unemployed, non-poor–employed, and non-poor–unemployed respondents?
I would greatly appreciate guidance on best practices, including tradeoffs related to interpretability, statistical power, and inference in conjoint designs.
Thank you very much.
Comment