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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

136<br />

4.8.2 Mean standardised bias statistic<br />

The distributions for <strong>the</strong> remaining comparisons when compared to <strong>the</strong> SYETP group are<br />

not perfectly balanced in terms <strong>of</strong> <strong>the</strong> separate independent variables after propensity<br />

score matching. This is because <strong>the</strong> propensity score allows some variables to be less<br />

well matched for a case, while o<strong>the</strong>rs variables are well matched, leading to an overall<br />

weighting scale across each <strong>of</strong> <strong>the</strong>se. A measure that summarises <strong>the</strong> overall balance is<br />

<strong>the</strong> standardized mean bias. For each variable included in <strong>the</strong> probit to estimate <strong>the</strong><br />

propensity score, <strong>the</strong> measure is calculated as:<br />

100*[ | µ treated - µ comparison | / √{( var treated + var comparison )/2} ]<br />

where µ is <strong>the</strong> mean and var is <strong>the</strong> variance observed for <strong>the</strong> covariates in <strong>the</strong> sample.<br />

The summary measure <strong>of</strong> <strong>the</strong> distribution <strong>of</strong> bias is found as <strong>the</strong> mean standardised bias<br />

across all <strong>the</strong> variables. 89 This measure can be used to gauge <strong>the</strong> matching performance<br />

(Lechner (2000)). It can be interpreted as bias as a percentage <strong>of</strong> <strong>the</strong> standard deviation.<br />

A smaller average bias is generally considered better. It is clear that a bias <strong>of</strong> zero is ideal,<br />

but <strong>the</strong> tolerance level for any non-zero standardized bias is not defined. Matching<br />

generally strongly improves mean bias. A poor match quality would indicate that <strong>the</strong><br />

matched groups are not genuinely balanced on <strong>the</strong> observed characteristics and that no<br />

outcome comparisons based on <strong>the</strong>se matches would be justified (Frölich et al. (2000):<br />

39). However <strong>the</strong>re is no clear gauge as to what level <strong>of</strong> standardized bias indicates a<br />

match is poor.<br />

Comparison with o<strong>the</strong>r studies, which report <strong>the</strong> standardized bias, gives ano<strong>the</strong>r gauge<br />

as to whe<strong>the</strong>r <strong>the</strong> levels are acceptable. Some recent studies that report balance measures<br />

are Sianesi (2001), Larsson (2000), Frölich et al. (2000), and Gerfin and Lechner (2000).<br />

Sianesi (2001) p47 appendix D found in Swedish administrative data analyses for 31,975<br />

cases a mean standardized bias <strong>of</strong> 0.86, but some variables were poor such as age at 2.86.<br />

This was a large data set, which enhances <strong>the</strong> ability to match. Larsson (2000) p27 and<br />

89 The use <strong>of</strong> <strong>the</strong> term bias here is <strong>the</strong> statistical sense, and refers to <strong>the</strong> differences in <strong>the</strong> conditioning<br />

variables used for matching and not bias in an estimator for a parameter.

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

Saved successfully!

Ooh no, something went wrong!