Dear all,
may I ask your opinion?
I am trying to evaluate the effect, treatment effect on the treated (TOT), of the labor market reform on the duration until exit to a job in the regression discontinuity design settings. The policy change affected people born after the introduction of the reform. I form the sample by selecting people born within 3 month before and after the reform.
****First****
The first problem I am facing is that before and after the reform groups are different. Hence my results are compromised by the selection which is not random. Due to the data restriction I observe less people after the reform. To get around this problem and make the groups homogeneous I use propensity score reweighting as suggested by e.g Dinardo (2002) and explained by Nichols (2008). I define weights to be one for the after reform group and p(x)/(1-p(x)) for before the reform group, p(x) - predicted probability of belonging to the after the reform group. I also tried rescaling the weights without apparent difference to the results. The reweighting is useful in that it allows me to make the before reform group comparable in observables to the after reform group. Does this approach make sense to you?
After I derived the weights I stset my data and specify a piecewise constant exponential model that includes after the reform dummy interacted with duration dependence parameters with streg in STATA. I interpret resulting hazard ratios of interaction of the reform dummy with duration dependence parameters as local TOT.
****Second****
The issue I am facing is that stset does not allow analytic weights, so that I am using probability weights. If I understand correctly this results in incorrect standard errors. I cannot bootstrap as it does not work with streg in the presence of weights. Is there any way around, glamm or R?
****Third****
I still see a need to control for the for unobserved heterogeneity in the model to account for the fact that some observation are more likely to fail than the others. To account for that I need a shared frailty model which is also not allowed in the presence of weights in STATA.
Maybe someone has an idea how to get around the above mentioned problems?
Thank you in advance.
DiNardo, J. 2002. Propensity score reweighting and changes in wage distributions.
Working Paper, University of Michigan. http://www-personal.umich.edu/˜jdinardo/bztalk5.pdf
Austin Nichols (2008) Erratum and discussion of propensity-score
reweighting. The Stata Journal8(4)
may I ask your opinion?
I am trying to evaluate the effect, treatment effect on the treated (TOT), of the labor market reform on the duration until exit to a job in the regression discontinuity design settings. The policy change affected people born after the introduction of the reform. I form the sample by selecting people born within 3 month before and after the reform.
****First****
The first problem I am facing is that before and after the reform groups are different. Hence my results are compromised by the selection which is not random. Due to the data restriction I observe less people after the reform. To get around this problem and make the groups homogeneous I use propensity score reweighting as suggested by e.g Dinardo (2002) and explained by Nichols (2008). I define weights to be one for the after reform group and p(x)/(1-p(x)) for before the reform group, p(x) - predicted probability of belonging to the after the reform group. I also tried rescaling the weights without apparent difference to the results. The reweighting is useful in that it allows me to make the before reform group comparable in observables to the after reform group. Does this approach make sense to you?
After I derived the weights I stset my data and specify a piecewise constant exponential model that includes after the reform dummy interacted with duration dependence parameters with streg in STATA. I interpret resulting hazard ratios of interaction of the reform dummy with duration dependence parameters as local TOT.
****Second****
The issue I am facing is that stset does not allow analytic weights, so that I am using probability weights. If I understand correctly this results in incorrect standard errors. I cannot bootstrap as it does not work with streg in the presence of weights. Is there any way around, glamm or R?
****Third****
I still see a need to control for the for unobserved heterogeneity in the model to account for the fact that some observation are more likely to fail than the others. To account for that I need a shared frailty model which is also not allowed in the presence of weights in STATA.
Maybe someone has an idea how to get around the above mentioned problems?
Thank you in advance.
DiNardo, J. 2002. Propensity score reweighting and changes in wage distributions.
Working Paper, University of Michigan. http://www-personal.umich.edu/˜jdinardo/bztalk5.pdf
Austin Nichols (2008) Erratum and discussion of propensity-score
reweighting. The Stata Journal8(4)
Comment