If at first you don’t succeed, try a different specification
| Gabriel |
Cristobal Young (with whom I overlapped at Princeton for a few years) has an article in the last ASR on model uncertainty, with an empirical application to religion and development. This is similar to the issue of publication bias but more complicated and harder to formally model. (You can simulate the model uncertainty problem as to control variables but beyond that it gets intractable).
In classic publication bias, the assumption is that the model is always the same and it is applied to multiple datasets. This is somewhat realistic in fields like psychology where many studies are analyses of original experimental data. However in macro-economics and macro-sociology there is just one world and so to a first approximation what happens is that there is basically just one big dataset that people just keep analyzing over and over. To a lesser extent this is true of micro literatures that rely heavily on secondary analyses of a few standard datasets (e.g., GSS and NES for public opinion; PSID and ADD-health for certain kinds of demography; SPPA for cultural consumption). What changes between these analyses is the models, most notably assumptions about the basic structure (distribution of dependent variable, error term, etc), the inclusion of control variables, and the inclusion of interaction terms.
Although Cristobal doesn’t put it like this, my interpretation is that if there were no measurement error, this wouldn’t be a bad thing as it would just involve people groping towards better specifications. However if there is error then these specifications may just be fitting the error rather than fitting the model. Cristobal shows this pretty convincingly by showing that the analysis is sensitive to the inclusion of data points suspected to be of low quality.
I think it’s also worth honoring Robert Barro for being willing to cooperate with a young unknown researcher seeking to debunk one of his findings. A lot of established scientists are complete assholes about this kind of thing and not only won’t cooperate but will do all sorts of power plays to prevent publication.
Finally, see this poli sci paper which does a meta-analysis of their two flagship journals and finds a suspicious number of papers that are just barely significant. Although, they describe the issue as “publication bias,” I think the issue is really model uncertainty.