A double-hurdle approach to modelling tobacco consumption in Italy

Abstract

This article analyzes the determinants of tobacco expenditures for a sample of Italian households. A Box–Cox double-hurdle model adjusted for heteroscedasticity is estimated to account for separate individual decisions concerning smoking participation and tobacco consumption and to correct for nonnormality in the bivariate distribution of the error terms. Nested univariate and bivariate models are found to be excessively restrictive, supporting the adequacy of a generalized specification. Estimation results show that consumption decisions are significantly affected by income and demographic characteristics. In particular, income positively impacts tobacco expenditure, while participation probability substantially declines as age increases. The existence of significant gender differences in both smoking participation and tobacco consumption patterns is found, while high education and white-collar occupation reduces the likelihood to smoke and tobacco expenditure levels. Single adult households have a lower probability of smoking initiation even if, conditional on smoking, they consume more. Finally, complementarity between tobacco and alcohol beverages suggests the necessity of joint public health strategies.

Acknowledgement

We would like to thank Federico Perali and Luca Piccoli for their useful comments and suggestions. This article is part of the research project ‘Dynamic Analysis of Addiction: Intra-household Resources Allocation, Social Welfare and Public Health’, University of Verona.

Notes

¹ Only few countries give a panel data structure to their household expenditure surveys. Recently, Labeaga (1999) and Jones and Labeaga (2003), using a panel of Spanish households (the Continuous Family Expenditure Survey) have tested rational addiction simultaneously accounting for censoring and unobservable heterogeneity.

² This decomposition follows the approach proposed by McDonald and Moffitt (1980) for the decomposition of the unconditional mean of the dependent variable in the Tobit model.

³ Here, for simplicity, we focus on the independent Box–Cox double-hurdle model. Details on the derivation of the conditional mean for the Box–Cox double-hurdle model with dependent errors can be found in Jones and Yen (2000).

⁴ Analytical details on the derivation of conditional and unconditional marginal effects for the Box–Cox double-hurdle model can be found in Yen (1993) and Jones and Yen (2000).

⁵ In estimating the final model we started with a specification that included all explanatory variables in both hurdles; insignificant variables were gradually dropped, with exclusion restrictions giving identification higher reliability.

⁶ Jones (1989) included the individual's age and its square as explicative variables, while Yen and Jensen (1996) used both household age composition and the age of the household head, showing significant life-cycle patterns for both participation and consumption decisions.

⁷ Details on distributional tests in censored and limited dependent variable models can be found in Bera et al . (1984), Pagan and Vella (1989) and Wells (2003).

⁸ In principle, all explanatory variables can be included in the heteroscedasticity specification; however, doing so would considerably increase the number of parameters to be estimated. So we focused our attention only on the variables that are more likely to cause heteroscedasticity and then we tested alternative specification excluding those variables that are not significantly different from zero.

⁹ The Heckman sample selection model can also be obtained as a restricted specification, assuming that participation decision dominates consumption decision. Vuong specification test for nonnested models supports the inadequacy of the Heckman model. The results are not presented here, but they are available from the authors.

¹⁰ Smith (2003) puts into question the relevance of the dependent double-hurdle model itself, asserting that this model contains too little statistical information to support estimation of dependency, even when dependency is truly present.