02.06.2014 Views

Evaluation of the Australian Wage Subsidy Special Youth ...

Evaluation of the Australian Wage Subsidy Special Youth ...

Evaluation of the Australian Wage Subsidy Special Youth ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

219<br />

6.3 Discussion<br />

The comparison <strong>of</strong> <strong>the</strong> employment effects <strong>of</strong> Heckman versus PSM is shown in<br />

Table 6.9. The former unweighted results are in <strong>the</strong> first columns, to facilitate comparison.<br />

The weighted results are shown in <strong>the</strong> last two columns, shaded.<br />

Comparison between <strong>the</strong> weighted and unweighted results provides an additional<br />

indicator for possible attrition bias. D’Agostino and Rubin (2000) consider <strong>the</strong> issue <strong>of</strong><br />

missing data in item non-response, and note that most propensity score methods are based<br />

on complete data. They suggest that missingness needs to be controlled in PSM. The fall<br />

in <strong>the</strong> employment gain estimated under PSM once attrition is accounted for, is in line<br />

with <strong>the</strong> need to account for <strong>the</strong> missing data that attrition introduces. The significant<br />

differences between <strong>the</strong> results <strong>of</strong> <strong>the</strong> weighted and unweighted models are interpreted as<br />

indicative <strong>of</strong> <strong>the</strong> presence <strong>of</strong> attrition bias. In a general comparison with <strong>the</strong> earlier<br />

unweighted results, it is noticeable that <strong>the</strong> weighted results are much smaller in size,<br />

although still positive, and are also <strong>of</strong> much lower statistical significance. Whereas before<br />

<strong>the</strong> weighting was applied, <strong>the</strong> Heckman and PSM gave very different sizes for <strong>the</strong><br />

employment effect, once weighted to account for attrition <strong>the</strong> effects gained are very<br />

similar in size. This would indicate that under <strong>the</strong> different modelling assumptions <strong>of</strong><br />

PSM and <strong>the</strong> Heckman bivariate probit, similar results for <strong>the</strong> employment effect are<br />

gained once selection due to attrition is accounted for.<br />

The PSM result is significant at <strong>the</strong> 22 per cent level <strong>of</strong> significance, which is outside<br />

normal test bounds. However, <strong>the</strong> variance estimates in <strong>the</strong> weighted estimates are also<br />

far more conservative than for <strong>the</strong> unweighted. 152 If <strong>the</strong> unweighted Heckman probit<br />

results had used <strong>the</strong> more conservative estimate for variance, <strong>the</strong> t statistic for <strong>the</strong><br />

employment effect falls to 1.49. This is <strong>the</strong>n similar in size to <strong>the</strong> weighted PSM, and<br />

also gives about a 22 per cent level <strong>of</strong> significance. Under <strong>the</strong> more conservative variance<br />

estimates, nei<strong>the</strong>r Heckman probit gives statistically significant results at normal test<br />

sizes.<br />

152 See chapter 7 sensitivity analysis for a fuller discussion <strong>of</strong> <strong>the</strong> variance estimates used.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!