Evaluation of the Australian Wage Subsidy Special Youth ...
Evaluation of the Australian Wage Subsidy Special Youth ... Evaluation of the Australian Wage Subsidy Special Youth ...
98 Φ = cumulative distribution function for the standard bivariate normal This approach has been applied in several evaluation problems for solving the issue of selection bias, where programme participants are not randomly selected, as discussed in the brief review of Maddala (1983). It allows for selection into participation in the wage subsidy on the basis of variables that are unobservable, or more usually variables that are unobserved in the data. The model is made operational by assuming a bivariate normal distribution for the errors in the participation and employment equations. This assumption makes the unobservable characteristics that jointly influence participation and employment follow the bivariate normal distribution. The participation equation is explicitly modelled to provide a variable that is then used to control for the part of unobserved variation of the employment equation that is correlated with the unobserved variation in the participation decision. With full information maximum likelihood methods, the selection and employment equations are estimated jointly, in which case the correlation between the unobservables is estimated directly. The sets of exogenous variables x i , z i can be overlapping, in which case only the functional form assumption provides identification. However, in practice this is not recommended as identification based solely on functional form is likely to be empirically fragile. The model is more appropriately identified if at least one variable in the set z i is not in the set x i , as the exclusion restriction improves the identification of the parameters of interest. Thus, practical implementation usually requires a variable included in the participation equation estimated that is excluded from the employment equation. 60 This exogenous variable should be suitable in that it influences selection but not employability. An important aspect of this estimation approach is the identification of a credible instrument for the exclusion restriction. Also, the results of estimation rest upon the 60 Wilde (2000) points out that earlier references presented contradictory opinions on this matter, so that Maddala (1983) held that a model with overlapping sets of exogenous variables x i , z i was only identified if at least one variable in the set z i is not in the set x i , while Heckman (1978) maintained that the functional form sufficed for identification. However Wilde (2000) clarifies that Maddala’s argument is only valid for his particular example, and is not the case generally due to the nonlinear relationships between z and the probability P(d=1 | z), and further concluded it could be avoided by assuming each equation contains at least one varying exogenous regressor. Wilde (2000) points out that this is a rather weak assumption in economic applications. Hence, Maddala’s argument commonly holds in econometric practice.
99 suitability of the underlying assumption of the bivariate normal distribution for the errors in the participation and employment equations. 3.3 Data and variables used for estimation The first stage of the replication, whereby the variables used in the analysis are reconstructed, is made easier by access to the final data set used as the basis for the Richardson (1998) analysis. 61 It was pointed out by Dewald et al. (1986) that access to the author’s data, with the variables in the final form used for the analysis, is an invaluable assistance in replication, but observed that this was an unusual occurrence. They attributed this to angst held by authors as they “…may interpret the very act of replication as a challenge to their professional competence and integrity” (Dewald et al. (1986): 601). In light of this, the access granted by the author James Richardson is strongly appreciated. As a result of access to the transformed data, replicating the exact construction of the variables is simplified. This would normally be an exacting task, as the construction of complex variables such as those derived from the work history is extremely difficult to repeat without the exact details of construction, and even the exact syntax coding. Any variation in construction would make replication far more difficult, and overcoming this hurdle also allows interpretation of the replication results to be more precise. It should not be overlooked that this stage of a replication would normally be most difficult. The variables are described in the data appendix. A brief introduction to the data is informative. The Australian Longitudinal Survey List Sample (Mcrae et al. 1984-1987) is used. This was a sample drawn from an administrative sample frame. The ALS list sample was a nationally representative sample of Australian youths aged 15-24 who had been registered as unemployed with the Commonwealth Employment Service for at least 3 months in June 1984. The 1984 survey took place with interviews between September 1984 and November 1984. The survey was repeated in each of the later years 1985, 1986 and 1987, following up where 61 The data were transformed from the ALS original SPSS files in extensive STATA processing by Lorraine Dearden, Alex Heath, Henry Overman, and James Richardson.
- Page 63 and 64: 47 In January 1979, variations were
- Page 65 and 66: 49 benefits were paid at a slightly
- Page 67 and 68: 51 2.2.3 SYETP operation Earlier re
- Page 69 and 70: 53 ceiling constraints’ applied t
- Page 71 and 72: 55 Award Conditions for employment
- Page 73 and 74: 57 Harris (2001) claims that during
- Page 75 and 76: 59 display boards listed details of
- Page 77 and 78: 61 restriction was used. If there w
- Page 79 and 80: 63 to the end of the 1980’s. An o
- Page 81 and 82: 65 for teens overall had risen, emp
- Page 83 and 84: 67 for Australia using data from th
- Page 85 and 86: 69 training, can provide a form of
- Page 87 and 88: 71 employer survey estimates were t
- Page 89 and 90: 73 provisions for SYETP and extende
- Page 91 and 92: 75 withdrawals occurred at similar
- Page 93 and 94: 77 Table 2.17 State usage of progra
- Page 95 and 96: 79 2.3.1 Stretton (1982, 1984) 53 S
- Page 97 and 98: 81 Stretton attributed the success
- Page 99 and 100: 83 included in the employment model
- Page 101 and 102: 85 completers. Their argument was t
- Page 103 and 104: 87 was an issue for the data. Unlik
- Page 105 and 106: 89 Table 2.21 Richardson (1998) Est
- Page 107 and 108: 91 2.3.5 General discussion Some ge
- Page 109 and 110: 93 Controlling for differences in i
- Page 111 and 112: 95 taken by a previous researcher a
- Page 113: 97 If employability is assumed to b
- Page 117 and 118: 101 Heckman, Lalonde and Smith (199
- Page 119 and 120: 103 effect on employment relative t
- Page 121 and 122: 105 Table 3.1, Part A Employment eq
- Page 123 and 124: 107 (-1.80) (-1.80) Tradesperson -0
- Page 125 and 126: 109 duration of Pre-June 1984 unemp
- Page 127 and 128: 111 4: Study 2 Propensity score mat
- Page 129 and 130: 113 Propensity score matching provi
- Page 131 and 132: 115 4.2 Propensity score matching m
- Page 133 and 134: 117 covariates that influence the a
- Page 135 and 136: 119 (7) E(Y c | D=1) = E P(X) {E[Y
- Page 137 and 138: 121 For CIA to be plausible, a ‘r
- Page 139 and 140: 123 employment and programme partic
- Page 141 and 142: Highest qualification in 1984 (1.56
- Page 143 and 144: 127 4.6 Distribution of the propens
- Page 145 and 146: 129 Figure 4.3 Histograms of estima
- Page 147 and 148: 131 Table 4.5 Summary statistics fo
- Page 149 and 150: 133 Table 4.5, that the variance of
- Page 151 and 152: 135 Table 4.6 Matching results, Sin
- Page 153 and 154: 137 Table 6.3 using Swedish data wi
- Page 155 and 156: 139 matching is the ability to weed
- Page 157 and 158: 141 Table 4.7 Matching results, All
- Page 159 and 160: 143 the unobserved component. If th
- Page 161 and 162: 145 5: Study 3 Attrition and non-re
- Page 163 and 164: 147 occur by design, because the mi
98<br />
Φ = cumulative distribution function for <strong>the</strong> standard bivariate normal<br />
This approach has been applied in several evaluation problems for solving <strong>the</strong> issue <strong>of</strong><br />
selection bias, where programme participants are not randomly selected, as discussed in<br />
<strong>the</strong> brief review <strong>of</strong> Maddala (1983). It allows for selection into participation in <strong>the</strong> wage<br />
subsidy on <strong>the</strong> basis <strong>of</strong> variables that are unobservable, or more usually variables that are<br />
unobserved in <strong>the</strong> data. The model is made operational by assuming a bivariate normal<br />
distribution for <strong>the</strong> errors in <strong>the</strong> participation and employment equations. This assumption<br />
makes <strong>the</strong> unobservable characteristics that jointly influence participation and<br />
employment follow <strong>the</strong> bivariate normal distribution. The participation equation is<br />
explicitly modelled to provide a variable that is <strong>the</strong>n used to control for <strong>the</strong> part <strong>of</strong><br />
unobserved variation <strong>of</strong> <strong>the</strong> employment equation that is correlated with <strong>the</strong> unobserved<br />
variation in <strong>the</strong> participation decision. With full information maximum likelihood<br />
methods, <strong>the</strong> selection and employment equations are estimated jointly, in which case <strong>the</strong><br />
correlation between <strong>the</strong> unobservables is estimated directly.<br />
The sets <strong>of</strong> exogenous variables x i , z i can be overlapping, in which case only <strong>the</strong><br />
functional form assumption provides identification. However, in practice this is not<br />
recommended as identification based solely on functional form is likely to be empirically<br />
fragile. The model is more appropriately identified if at least one variable in <strong>the</strong> set z i is<br />
not in <strong>the</strong> set x i , as <strong>the</strong> exclusion restriction improves <strong>the</strong> identification <strong>of</strong> <strong>the</strong> parameters<br />
<strong>of</strong> interest. Thus, practical implementation usually requires a variable included in <strong>the</strong><br />
participation equation estimated that is excluded from <strong>the</strong> employment equation. 60 This<br />
exogenous variable should be suitable in that it influences selection but not employability.<br />
An important aspect <strong>of</strong> this estimation approach is <strong>the</strong> identification <strong>of</strong> a credible<br />
instrument for <strong>the</strong> exclusion restriction. Also, <strong>the</strong> results <strong>of</strong> estimation rest upon <strong>the</strong><br />
60 Wilde (2000) points out that earlier references presented contradictory opinions on this matter, so that<br />
Maddala (1983) held that a model with overlapping sets <strong>of</strong> exogenous variables x i , z i was only identified if<br />
at least one variable in <strong>the</strong> set z i is not in <strong>the</strong> set x i , while Heckman (1978) maintained that <strong>the</strong> functional<br />
form sufficed for identification. However Wilde (2000) clarifies that Maddala’s argument is only valid for<br />
his particular example, and is not <strong>the</strong> case generally due to <strong>the</strong> nonlinear relationships between z and <strong>the</strong><br />
probability P(d=1 | z), and fur<strong>the</strong>r concluded it could be avoided by assuming each equation contains at<br />
least one varying exogenous regressor. Wilde (2000) points out that this is a ra<strong>the</strong>r weak assumption in<br />
economic applications. Hence, Maddala’s argument commonly holds in econometric practice.