Hi Stata Heroes,
I'm running a Difference-in-difference model to check whether a reform in the Dutch medical malpractice system in 2016 had a different effect on C-section rate of expert patients (patients who had a medical training) than on non-expert patients. I use delivery level data of first births 14 weeks before and after the reform to estimate the following equation (number of observations = 31 320):
in which the dependent variable is a dummy equal to one if the delivery d of mother m in institution i at time t was by C-section. Post is a dummy equal to one if the delivery was after the tort reform, Expert is a dummy equal to one if the patient was an expert-patient. I control for some mother and delivery characterstics, institution and biweek fixed effects.
I have two questions related to this analysis:
1. Should I cluster standard errors on insitution level? I'm considering clustering since the tort reform concerns every Dutch clinical institution at the same time (so no staggered DiD) but every institution had the opportunity to implement the reform differently (e.g. they had to establish a safe reporting point for medical incidents by themselves and to join a self-chosen certified dispute resolution authority within a year). Therefore I guess the reform may have had a different impact on different institutions so that the standard errors of patients in the same hospital might be correlated. What do you guys think?
2. I also run a dynamic specification with leads and lags of the biweeks before and after the reform with the expert dummy in which I omitted the interaction term of experts with the first biweek before the torm reform as reference category (I replaced my difference-in-difference estimator in the main analysis with these leads and lags). While my difference-in-difference estimate in my main analysis is significant at the 5% significance level, the lags aren't in the dynamic specification. Although the coefficients of the first lags after the reform are greater in absolute terms than my main effects. Did I do something wrong or how can you explain this phenonemon? The leads are also insignificant, supporting the common trends assumption.
Thank you in advance for your expertise.
Kind regards,
Lotte
I'm running a Difference-in-difference model to check whether a reform in the Dutch medical malpractice system in 2016 had a different effect on C-section rate of expert patients (patients who had a medical training) than on non-expert patients. I use delivery level data of first births 14 weeks before and after the reform to estimate the following equation (number of observations = 31 320):
in which the dependent variable is a dummy equal to one if the delivery d of mother m in institution i at time t was by C-section. Post is a dummy equal to one if the delivery was after the tort reform, Expert is a dummy equal to one if the patient was an expert-patient. I control for some mother and delivery characterstics, institution and biweek fixed effects.
I have two questions related to this analysis:
1. Should I cluster standard errors on insitution level? I'm considering clustering since the tort reform concerns every Dutch clinical institution at the same time (so no staggered DiD) but every institution had the opportunity to implement the reform differently (e.g. they had to establish a safe reporting point for medical incidents by themselves and to join a self-chosen certified dispute resolution authority within a year). Therefore I guess the reform may have had a different impact on different institutions so that the standard errors of patients in the same hospital might be correlated. What do you guys think?
2. I also run a dynamic specification with leads and lags of the biweeks before and after the reform with the expert dummy in which I omitted the interaction term of experts with the first biweek before the torm reform as reference category (I replaced my difference-in-difference estimator in the main analysis with these leads and lags). While my difference-in-difference estimate in my main analysis is significant at the 5% significance level, the lags aren't in the dynamic specification. Although the coefficients of the first lags after the reform are greater in absolute terms than my main effects. Did I do something wrong or how can you explain this phenonemon? The leads are also insignificant, supporting the common trends assumption.
Thank you in advance for your expertise.
Kind regards,
Lotte
