Method and results Method We use the standard approach to decomposing pay disparities in the literature, as introduced by Oaxaca (1973) and Blinder (1973). This involves initially estimating, separately, the wage models for two ethnic groups. These are labelled in the following equations as group 1 and group 2. ln(w )= β1 X + ε (1) ln(w )= β2 X + ε (2) In the above wage models, the I subscript refers to the ith wage earner, w stands for hourly wages, X is the vector of explanatory variables (as shown in Table 1). The outcome in the wage models is the natural logarithm of usual hourly wages. The ethnic pay gap is calculated in (3) and decomposed in (4): ln (w1) – ln (w2) = β1X1 – β2X2 (3) ln (w1) – ln (w2) = β1X1 – X2) + (β1 – β2) X2 (4) Based on the decomposition shown in (4), the first part of the right-hand side is the component of the ethnic pay gap that can be explained by differences in average characteristics of the two ethnic groups. This is essentially the ‘explained’ component of the pay gap, and as will be shown in the results, this can be further broken down to the contribution of each of the domains in Table 1. The second part of the right-hand side of equation (4) is the component of the ethnic pay gap that is left unexplained. This equates to differences in the returns to characteristics in the labour market. Why are there unexplained differences? There are several possible reasons. These include: (i) unobserved differences in characteristics not captured in the current data; (ii) ethnic differences in the non-pecuniary elements of jobs; (iii) discriminatory behaviour; (iv) unconscious bias, etc. A recognised issue in the literature in implementing decompositions is whether the estimated β coefficients used to weight the explained part of the model should relate to Europeans or to the comparator ethnic group, or be estimated from a pooled regression of all workers (i.e. both ethnic groups). The choice of which weights to use can lead to substantive variations in results. We choose to use the estimated β coefficients from a pooled regression as weights, which requires less strict assumptions over the alternative choices regarding the counterfactual wage structure. Another often acknowledged issue in the literature with the Oaxaca-Blinder approach is that it may suffer from sample selection bias (Heckman, 1979), as wages are only available for employed individuals. Since the decision to enter the labour market is systematically linked to the wages an individual is likely to receive, by omitting non-employed from the analysis, we may bias our results. Therefore, to correct for sample selection bias we apply the Heckman procedure and do this for both ethnic groups. The following variables are used in the Heckman selection model; individual characteristics (age, born in NZ); educational attainment (6 dummy variables); regional council (12 dummy variables); and household characteristics (sole parent, partnered, number of dependent children and the income decile of the household). The variables included in the main model are all those illustrated in Table 1, except for household characteristics. Household characteristics are excluded to allow identification of the Heckman selection model. Empirical analysis of Pacific, Māori and ethnic pay gaps in New Zealand 8