Unconditional quantile regressions, wage growth and inequality in the Philippines, 2001–2006: the contribution of covariates

Abstract

I implement a two-stage procedure to estimate the components of real wage change in the Philippines along the wage distribution from 2001 to 2006, as well as the contribution of individual covariates to each component. The methodology is based on Re-centred Influence Functions (RIF), as suggested by Firpo et al. (2009). The advantage of this methodology is that it not only decomposes the unconditional wage change at any quantile of the wage distribution, but also allows the characterization of the contribution of any single covariate on each component. I find that real earnings of males in the Philippines declined and the decline along the earnings distribution is generally higher at higher deciles for all males and higher at lower deciles in Manila. Decomposition results are driven by the wage structure component, while the composition effect is small and generally positive with education as the main contributor. Within the wage structure effect, the main determinants are associated with changes in the rewards of experience and occupation, along with residual change (change in intercepts).

Keywords:

• unconditional quantile regression
• RIF functions
• wage inequality
• the Philippines

JEL Classification:

• D33
• J31
• J42

Notes

¹ The Asian crisis peaked in 1998 and affected mostly Thailand, Indonesia, the Philippines, South Korea, and Malaysia.

² The earlier cited Chi et al. (2007) study used FFL (2005) – an earlier outline of the Re-centred Influence Functions (RIF)-based method – to decompose the change in wage inequality in China. Le and Booth (2010) used FFL (2009) to examine inequality in Urban–Rural living standards in Vietnam from 1993 to 2006.

³ Melly (2006) further improved the Machado and Mata procedure.

⁴ Leslie and Pu (1995) have also studied the role of unions in the rise of wage inequality in Britain.

⁵ Since reflects changes in the joint distribution of , while we require this component to reflect only changes in the distribution of X, the conditional independence assumption ( independent of T given X) is required. In addition, the ‘overlapping support’ assumption requires that there is an overlap in observable characteristics across groups; this is expected to be satisfied in our case, where we look at over-time changes in wage distributions.

⁶ In practice, the error in the wage structure (or composition) effect can be estimated as the difference between the estimate of the wage structure (or composition) effect through re-weighting approach and the estimate obtained from the RIF-regression procedure (using the RIFREG module in STATA); the two errors can be different in absolute value.

⁷ The same inequality measures were computed for the logarithm of hourly wage and results are qualitatively similar.

⁸ For the period examined in Dacuycuy (2006), average worker pay increased by a meager 2% with workers at the lowest two deciles experiencing a decline in earnings.

⁹ I use the STATA devcon procedure which follows the RIFREG procedure.

¹⁰ While the premium to university education increased at the top of the earnings distribution, the opposite was the case for other education qualifications and at lower points in the distribution.