To compile the data, a search using EconLit, AgEcon Search, and Google Scholar, as well as a perusal of Kuznets, Reeves and Hayman, Tomek, Raunikar and Huang, Smallwood et al., Alston and Chalfant, Moschini and Moro, and Asche et al., led to an initial set of candidate studies. Subsequent to surveying the reference sections of these studies, 362 studies were identified as estimating the price elasticity of meat in North America, Asia, or Europe. These 362 studies provided 3755 estimates of the price elasticity, with the largest number corresponding to North America, followed by Europe and Asia.
Similar to meta-analyses of Espey, Dalhuisen et al., Gallet and List, Gallet, and Gallet, we collected information on several characteristics of the 362 studies. First, the price elasticity has been estimated for a number of meats, including beef, pork, lamb, poultry, fish, and a composite of multiple meats.5 Second, with respect to demand specification, although many studies estimate linear or double-log functional forms, many others estimate functional forms that are consistent with consumer theory, such as the linear-approximate almost ideal demand system (AIDS-Linear), which uses a price index to linearize Deaton and Muellbauer’s AIDS form. Other less common forms that have been estimated include the following: the traditional nonlinear AIDS form (AIDS-Nonlinear), the quadratic AIDS form (AIDS-Quadratic) of Banks et al., the generalized AIDS form (AIDS-General) of Bollino, the Rotterdam form, the CBS form, the translog form, the S-Branch form, the Box-Cox form, the generalized addilog form, and the quadratic expenditure form.
Third, in addition to functional form, a number of other specification issues have been addressed in the literature. Several studies estimate compensated price elasticities, which provided meat is a normal good are expected to be lower in absolute value compared to uncompensated price elasticities. A number of studies include other meats as substitutes in the specification of demand, or estimate a dynamic specification of demand (i.e., include lag terms in the demand equation), or estimate a two-step specification (i.e., the demand for meat is modeled as (i) the choice of whether or not to consume meat followed by (ii) the choice of how much meat to consume).
Fourth, we note several data and estimation characteristics of the 362 studies. Studies not only differ in terms of whether time-series, cross-sectional, or panel data is used, but also differ in terms of the level of data aggregation, be it at the multiple country, country, region of country, city, firm, or individual consumer levels. We also collected information on the median year of the sample used to estimate meat demand, which if found to impact the price elasticity estimate could signal changes in consumer preferences over time. Various methods have been used to estimate the demand for meat, including ordinary least squares (OLS), two stage least squares (2SLS), three stage least squares (3SLS), full information maximum likelihood (FIML), single-equation maximum likelihood (MLE), seemingly unrelated regression (SUR), generalized method of moments (GMM), generalized least squares (GLS), and sparingly the minimum distance and maximum entropy estimators.
Common to many meta-analyses, we also collected information on several characteristics of the publication outlet. Specifically, we note whether or not the study was published in a premier journal, such as a top 36 economics journal (identified by Scott and Mitias or the top-ranked American Journal of Agricultural Economics (AJAE), as well as whether or not the study was published as a chapter in a book. See Gallet for the frequencies of these study characteristics in the literature.
The estimated price elasticity of meat is used as the dependent variable in a series of meta-regressions, with study characteristics serving as determinants of the price elasticity. Specifically, since it is common for studies to report multiple price elasticity estimates, we follow Rosenberger and Loomis, Gallet and List, Johnston et al., and Gallet by considering an unbalanced panel meta-regression model, given by:
Pij = ai + pXij + Sij,
where Pij is the jth price elasticity estimate from study i, ai is a “random researcher” effect, P is a vector of coefficients, and Xij accounts for study characteristics. In Xij, we include the median year of the sample used to estimate the respective price elasticity, as well as a series of dummy variables accounting for the study characteristics mentioned in the previous subsection (i.e., variable equals 1 if the respective study characteristic holds, 0 if not).6 Finally, sij is an iid error term with zero mean and variance c2E.
There are a number of issues concerning the estimation of equation. First, OLS and random effects versions of equation are estimated separately for North America, Asia, and Europe. This allows us to not only see how sensitive the results are to addressing panel data issues, but to also draw comparisons in the nature of the price elasticity across the three regions. Second, since some study characteristics do not apply in all three regions, the dummy variables corresponding to those characteristics must be dropped from the meta-regressions in which they do not apply.8 Some sets of study characteristics (e.g., type of meat, functional form of demand, nature of data, and estimation method) encompass all observations, and so to avoid perfect multicollinearity we must drop several dummy variables from each meta-regression. Along with these dummy variables, setting all included dummy variables equal to zero defines the baseline to which the estimation results are compared.9 Third, we use White’s procedure to adjust standard errors for heteroskedasticity. Fourth, since the price elasticity is typically negative, a negative (positive) coefficient of a particular study characteristic implies that characteristic makes the price elasticity more (less) elastic.