Pass-Through Rates for Alcohol Beverage Excise Taxes: Fixed-Effect versus Random-Effects Meta-Analysis and Meta-Regressions

Jon P Nelson

doi:10.13189/aeb.2021.090201

Abstract

This paper extends an earlier study to compare two methods for meta-analysis of economic data: fixed-effect models and random-effects models. The models differ fundamentally in the ability to generalize beyond the sample in question. Both models are applied to estimates of pass-through rates for excise taxes on alcohol beverages. Using best-set data from 30 primary studies, weighted means are first reported and compared against a fully-passed tax or rate of unity. Dispersion and heterogeneity statistics are used to assess the performance of each model. Second, means and dispersion statistics are reported by subgroups for country source, beverage (beer, wine-spirits), and published status. Third, tests are conducted for publication selection bias using funnel plots and regression asymmetry tests. Fourth, three procedures are undertaken to reduce selection bias: trim-and-fill; cumulative meta-analysis; and meta-regressions. Three conclusions are reached in the paper. First, average pass-through rates are approximately unity regardless of beverage. Primary researchers should compare estimated rates against this value. Second, a random-effects model is more appropriate for these data, reflecting highly diverse estimates of pass-through rates. Third, greater attention needs to be given to the choice of model for meta-regressions in economics and related disciplines.

Highlights

During the past thirty years, numerous meta-analyses have been published in economics and related areas.1 Typically these analyses obtain a sample of empirical studies from the literature on a particular subject and summarize or synthesize the distribution of estimates using precision weighted-means and meta-regressions
It is assumed that estimates within a subgroup are random, but subgroups are fixed, i.e., the pooled model is based on mixed-effects size (MES) that permits generalizations to comparable populations [1: 196]
Correcting for publication bias, meta-analysis using both models and a variety of tests yields the following: (1) funnel plots and asymmetry tests indicate that publication bias affects the sample of estimates; (2) trim-and-fill estimates for bias-adjusted means are close to unity, regardless of beverage type or model; (3) cumulative meta-analysis yields adjusted-means close to unity, except for beer; (4) meta-regressions yield a random-effects size (RES) predicted mean for alcohol of 0.89-0.91; beer, 0.87-0.96; and wine-spirits, 0.90-0.91; and (5) RES predicted mean confidence intervals include unity for alcohol and both beverages

Summary

Introduction

During the past thirty years, numerous meta-analyses have been published in economics and related areas. Typically these analyses obtain a sample of empirical studies from the literature on a particular subject and summarize or synthesize the distribution of estimates using precision weighted-means and meta-regressions. During the past thirty years, numerous meta-analyses have been published in economics and related areas.. During the past thirty years, numerous meta-analyses have been published in economics and related areas.1 These analyses obtain a sample of empirical studies from the literature on a particular subject and summarize or synthesize the distribution of estimates using precision weighted-means and meta-regressions. Despite the high degree of heterogeneity that exists in economic data and studies, an overwhelming majority of meta-analyses in economics employ a conditional or fixed-effect size (FES) model to summarize and analyze study results. FES models assume that sampled estimates represent the only population of interest, which means analytical results should not be generalized beyond these data [1]. In the limiting case of a Keywords Excise Taxes, Tax Pass-Through, Meta-Analysis, Publication Bias

Objectives

Methods

Results

Discussion

Conclusion