They could've spent some time clearing up their differences. Why exactly does Deaton believe randomizing can't account for I observed heterogeneity?
- macro bro
I think Deaton's main point is that in the real world, RCTs always involve attrition, nonresponse, crossover and other deviations from the ideal. There are ways to adjust for these deviations, but they are not content-free: you are still making assumptions about how the heterogeneity you observe reflects or does not the unobserved heterogeneity of the underlying population.
There's a deeper issue that I think Deaton and his wife Anne Case's paper from last year illustrates well, too, that somewhat relates to his criticism of RCTs. Sometimes, economists in particular but social scientists at large are liable to think that well-identified Local Average Treatment Effects are all that is going on in the world. But that can make you lose the forest for the trees. As their paper points out using simple observational data, there are huge increases in early mortality due to suicide, overdose, and alcoholism among non-hispanic American whites that are sufficient to drive all-cause mortality up. (The 2015 data which is now available shows that this increase has gone up still further.) You really can't get at these overall trends by simply aggregating across all the different LATEs you measure from individual quasiexperimental or experimental studies, but they're clearly important, and reflect some of the overall drivers of collective behavior.
I think Deaton's main point is that in the real world, RCTs always involve attrition, nonresponse, crossover and other deviations from the ideal. There are ways to adjust for these deviations, but they are not content-free: you are still making assumptions about how the heterogeneity you observe reflects or does not the unobserved heterogeneity of the underlying population.
There's a deeper issue that I think Deaton and his wife Anne Case's paper from last year illustrates well, too, that somewhat relates to his criticism of RCTs. Sometimes, economists in particular but social scientists at large are liable to think that well-identified Local Average Treatment Effects are all that is going on in the world. But that can make you lose the forest for the trees. As their paper points out using simple observational data, there are huge increases in early mortality due to suicide, overdose, and alcoholism among non-hispanic American whites that are sufficient to drive all-cause mortality up. (The 2015 data which is now available shows that this increase has gone up still further.) You really can't get at these overall trends by simply aggregating across all the different LATEs you measure from individual quasiexperimental or experimental studies, but they're clearly important, and reflect some of the overall drivers of collective behavior.
Economist detected!
I think Deaton's main point is that in the real world, RCTs always involve attrition, nonresponse, crossover and other deviations from the ideal. There are ways to adjust for these deviations, but they are not content-free: you are still making assumptions about how the heterogeneity you observe reflects or does not the unobserved heterogeneity of the underlying population.
There's a deeper issue that I think Deaton and his wife Anne Case's paper from last year illustrates well, too, that somewhat relates to his criticism of RCTs. Sometimes, economists in particular but social scientists at large are liable to think that well-identified Local Average Treatment Effects are all that is going on in the world. But that can make you lose the forest for the trees. As their paper points out using simple observational data, there are huge increases in early mortality due to suicide, overdose, and alcoholism among non-hispanic American whites that are sufficient to drive all-cause mortality up. (The 2015 data which is now available shows that this increase has gone up still further.) You really can't get at these overall trends by simply aggregating across all the different LATEs you measure from individual quasiexperimental or experimental studies, but they're clearly important, and reflect some of the overall drivers of collective behavior.Economist detected!
You are a gentlelady and a scholar.