Decomposing rank- and level-dependent indicators of income-related inequality of health: A comparison of recent regression-based approaches
In recent years several regression-based decomposition methods have been developed in order to identify the main determinants of socioeconomic inequality of health. In my presentation I will analyse and compare three of these approaches.
Wagstaff, Van Doorslaer and Watanabe (2003) have proposed a two-step methodology that is nowadays widely applied in empirical work. First, they identify the factors deemed to be related to health. Second, they link the associations of these factors with socioeconomic status to the observed level of socioeconomic inequality. Because only health-dependent factors are explored and factors related to socioeconomic status are ignored, this may be called a one-dimensional approach. Another two-step methodology has been formulated by Heckley, Gerdtham and Kjellsson (2016). They provide a general framework for decomposing rank-dependent indicators based on the notion of the (recentered) influence function (RIF). Instead of focusing directly on the two variables that determine the indicator, health and socioeconomic rank, they do so indirectly by considering the RIF of the indicator.
First, a RIF value is computed for each individual. Second, the level of socioeconomic inequality is decomposed by regressing the RIF values on a set of explanatory variables using OLS, yielding the marginal effects of these variables on the influence they exert on the indicator. The final approach is the one proposed by Kessels and Erreygers (2018). They suggest an alternative OLS regression approach that does not require the RIF transformation of the indicator. Their method disaggregates the indicator more directly by reformulating it into a product term for each individual that is regressed on the explanatory variables. The product term combines an individual’s outcome in the health and socioeconomic domains. This simple reformulation of the indicator does not require the explanatory variables to be exclusively related to either health or the socioeconomic variable, but allows for a combined relationship.
The three approaches can be applied to both rank- and level-dependent indicators of income-related inequality of health. I will focus on the strengths and weaknesses of the three approaches, and provide an empirical illustration using Australian household survey data.