However as noted above very few studies included in

This preview shows page 28 - 30 out of 83 pages.

estimates with the lowest level of estimated variance. However, as noted above, very few studies included in our sample report variance in price and income elasticity estimates. We thus use primary studies ’ sample sizes as proxies to these variances, which is a common procedure used in MRA studies and notably in environmental economics (Nelson and Kennedy, 2009). Other issues can arise in the presence of correlations of effect size estimates within and between primary studies. Indeed, our data contain several elasticity estimates collected from each primary study. However, if most characteristics distinguishing estimates from the same study (product category, household income level, demand functional form, econometric estimation method, etc.) are introduced as explanatory variables and thus are controlled for in our MRAs, some unobservable characteristics may give rise to correlated error terms across elasticities collected from the same primary study. In the same way, primary studies conducted by the same author may share unobservable characteristics and may lead to between studies correlations. To overcome this issue, we follow the same approach that was used by Disdier and Head (2008) and Cipollina and Salvatici (2010) and introduce random study/author effects into the MRA models. This results in the generation of mixed-effect models, which can be defined as multilevel regression models (Bateman and Jones, 2003) and are formally expressed as: 𝜃 ?? = 𝛼 0 + ∑ 𝛼 ? 𝑋 ? 𝐾 ?=1 + 𝑢 ? + 𝜀 ?? (1) where 𝜃 ?? is the dependent variable and denotes the j-th (price or income) elasticity estimate collected from the i -th primary study (or i -th author), 𝛼 0 is a fixed intercept and 𝛼 ? ( 𝑘 ∈ {1, … , 𝐾} ) is the fixed effect coefficient associated with 𝐾 explanatory variable 𝑋 ? ( 𝑘 ∈ {1, … , 𝐾} ). 𝜀 ?? is normally distributed with constant variance and can be interpreted as a sampling error term. 𝑢 ? is a random study (or author) effect that is normally distributed with constant variance independent of 𝜀 ?? and is assumed to be uncorrelated with the explanatory variables. Adding this random effect to the MRA model allows one to account for correlations