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.

117<br />

covariates that influence <strong>the</strong> assignment to treatment as well as potential outcomes. The<br />

precise form <strong>of</strong> CIA depends on <strong>the</strong> parameter being estimated. For <strong>the</strong> treatment on <strong>the</strong><br />

treated parameter, <strong>the</strong> CIA requires that, conditional on observable characteristics,<br />

potential non-treatment outcomes are independent <strong>of</strong> treatment participation.<br />

Formally, CIA can be written as<br />

(3) Y c ╨ D | X = x, ∀x∈ χ<br />

where ╨ denotes independence and χ denotes <strong>the</strong> part <strong>of</strong> <strong>the</strong> attribute space for which<br />

<strong>the</strong> treatment effect is defined. This assumption is sufficient for <strong>the</strong> treatment on <strong>the</strong><br />

treated parameter. The fuller CIA assumes that<br />

(3a) Y t ,Y c ╨ D | X.<br />

However, Heckman et al. (1997) show that for estimating <strong>the</strong> mean effect <strong>of</strong> treatment on<br />

<strong>the</strong> treated, CIA for Y c is sufficient. This is because inference <strong>of</strong> Y c for persons where<br />

D=1 is estimated from data on persons where D=0. Hence, after adjusting for observable<br />

differences, <strong>the</strong> mean <strong>of</strong> <strong>the</strong> no-treatment (potential) outcome is <strong>the</strong> same for those<br />

receiving treatment as for those not receiving treatment. This allows non-participants’<br />

outcomes to be used to infer participants’ counterfactual outcomes.<br />

However, this is only valid if <strong>the</strong>re are non-participants for all participants’ values <strong>of</strong> X<br />

(<strong>the</strong> support condition):<br />

(4) 0 < Pr ( D = 1 | X = x ) < 1<br />

where “Pr” is <strong>the</strong> probability <strong>of</strong> participating in <strong>the</strong> programme <strong>of</strong> treatment. This ensures<br />

that <strong>the</strong>re are no values <strong>of</strong> <strong>the</strong> characteristics in X for which <strong>the</strong> propensity score is zero<br />

or one. Heckman, Ichimura, Smith, and Todd (1998) show that failure to ensure common<br />

support will give biased estimates. Rosenbaum and Rubin (1983) showed that when this<br />

assumption <strong>of</strong> common support holds toge<strong>the</strong>r with <strong>the</strong> CIA, <strong>the</strong>n if treatment assignment<br />

is strongly ignorable when X is given, <strong>the</strong>n it is also strongly ignorable for any balancing<br />

score, such as <strong>the</strong> propensity score or probability <strong>of</strong> participating in <strong>the</strong> programme.

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

Saved successfully!

Ooh no, something went wrong!