RDP 2018-06: The Effect of Minimum Wage Increases on Wages, Hours Worked and Job Loss 8. Robustness Checks

I now consider a set of robustness checks on the baseline DD and DDD models. These tests are described in Sections 8.1–8.4 below, with the results for the full sample presented in Table 5.

8.1 Different Controls for Unobserved Group Effects

The baseline models assume that the effects of group-specific heterogeneity are linear (as captured by the continuous FWC_k control). A more flexible parametisation is to use separate dummy variables for each award wage decile prior to FWC decisions. The baseline models also impose that the effects of group-specific heterogeneity are constant across each of the 11 FWC decisions. I relax both of these assumptions by including a set of wage decile dummies, along with interactions between these dummies and a set of dummy variables for each of the 11 FWC decisions (Table 5). The results are very similar to the baseline model.

8.2 Excluding Jobs in Queensland

My baseline results also assume that award wages are adjusted shortly after the FWC announcement. However, employees on certain state awards experienced a lag in the adjustment of their award wage due to a delay by their state industrial relations commission in ratifying the national decision.^[15] This was particularly true for employees in Queensland, which was slower than the other states to ratify the increases in 2003, 2004 and 2005. My results are robust to excluding all jobs in Queensland from the sample (Table 5).

8.3 Adjusting for Pre-trends

I also consider the robustness of the baseline estimates to re-specifying the dependent variable in Equations (1) and (2) to be in terms of log changes over the previous six months, rather than in log levels. This specification will compare growth in, say, wages in the period immediately before the FWC decision with growth in wages over the six months after the FWC decision, across low-wage and high-wage jobs. This type of ‘triple differencing’ approach is common in UK studies of minimum wages (Dickens et al 2009; Bryan et al 2013). An advantage of this approach is that it controls for the possibility that the labour market outcomes of low-wage and high-wage jobs had different trajectories in the lead-up to each decision. However, I find that re-specifying the dependent variable in this way makes little difference to the findings (Table 5). Results are not presented for the job destruction rate, which is already constructed using transitions over six-month windows.

8.4 Controlling for Firm-specific Shocks

An alternative approach to controlling for violations of the parallel trends assumption is to augment Equation (1) with a full set of firm dummies, and interactions between these dummies and the after period dummy, d_t. This specification only uses variation within firms to identify the DD coefficient of interest, and thus controls for many potential violations of the parallel trends assumption (for example, a shock in the after period that disproportionately affects low-wage industries). The estimated wage elasticity (0.73) is slightly lower than in the baseline model (Table 5). For hours worked and the job destruction rate, the standard errors on the coefficients of interest are several times larger than in the baseline model, which limits the conclusions that can be drawn.

Footnote

Prior to the WorkChoices legislation, employees were covered by a patchwork of federal and state awards. For essentially historic reasons, some jobs were covered by federal awards, while others were subject to state awards (Stewart 2015, p 7). In practice, however, prior to 2006 the state industrial relations commissions had almost always announced the same increases as that announced by the federal commission, albeit with a lag of several months. [15]