REECAP 7th Annual Meeting

Bayesian Causal Inference with experimental survey data: an application to the Nutriscore labelling system

Not scheduled

20m

German Centre for Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig

Applications of experimental methods to agricultural and agri-environmental themes Session Block

Speaker

Alessandro Varacca (Università Cattolica del Sacro Cuore)

Description

This work introduces a foundation modeling strategy that integrates Item Response Theory (IRT) and the potential outcome (PO) framework to estimate the heterogeneous and average treatment effect (HTE and ATE, respectively) of randomly assigned interventions in experimental settings. This approach allows to automatically correct for measurement errors when eliciting individual latent characteristics through multiple Likert scale inquiries, making sure that the resulting uncertainty propagates to the causal estimands of interest. Placing Normal priors on these unobservable features, we show that the distribution of the POs is also Normal, enabling the explicit derivation of posterior predictive distributions for all the counterfactual quantities, which are then used to construct imputation estimators for the HTE and ATE. The latter are validated through simulations, which highlight how the proposed approach effectively recovers the underlying coefficients at different sample sizes under weakly informative priors. We implement the proposed methodology in a simple case study in which we evaluate the impact of a negative Nutriscore (NS) label on the Perceive Healthiness (PH) of a hard cheese product. Using a representative sample of Italian and Dutch consumers, we estimate a negative overall effect of low NS ranking on respondents’ PH. Moreover, we use HTEs to address treatment effect heterogeneity at different levels of predetermined (exogenous) covariates representing the main demographic characteristics of our sample. Using a simple regression tree fir to the posterior mean of the individual-level treatment effects, our results suggest that middle-aged individuals with no child tend to exhibit, on average, the sharpest drop in PH when confronted with a negative NS.