tag:blogger.com,1999:blog-2198942534740642384.post5707355133746426939..comments2023-10-24T03:16:41.009-07:00Comments on Econometrics Beat: Dave Giles' Blog: Curious RegressionsDave Gileshttp://www.blogger.com/profile/05389606956062019445noreply@blogger.comBlogger12125tag:blogger.com,1999:blog-2198942534740642384.post-76610286546888689642012-10-31T16:46:53.432-07:002012-10-31T16:46:53.432-07:00Thanks for replying! Someone could write an algori...Thanks for replying! Someone could write an algorithm for that. :)Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-18812437546362093162012-10-29T08:01:01.930-07:002012-10-29T08:01:01.930-07:00Just good old-fashioned trial and error!Just good old-fashioned trial and error!Dave Gileshttps://www.blogger.com/profile/05389606956062019445noreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-52689839151428063162012-10-29T04:46:45.107-07:002012-10-29T04:46:45.107-07:00How do you generate this type of data? I mean, dat...How do you generate this type of data? I mean, data that display words when plotted.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-3266187977498381592011-04-17T08:28:48.677-07:002011-04-17T08:28:48.677-07:00Anonymous 1: Thanks for following this up. First, ...Anonymous 1: Thanks for following this up. First, I don't think it's controversial that we need a theory first - I TOTALLY subscribe to that position. Having got that in place, my concern is simply that you see people then diving into the estimation of their model before "looking" at the sample characteristics of their data: be it structural breaks; outliers, etc. <br /><br />Take a situation where we are using time-series data, for example, and we decide, correctly, to test for unit roots and cointegration. If structural breaks are present we need to modify our tests, or else we're likely to draw spurious inferences about the stationarity/non-stationarity/cointegration of our data. If that happened we'd move ahead with an invalid econometric specification of the model for estimatrion purposes. For instance, wwe might fail to use an error-correcion formulation and simply diference the data; or we might fail to difference certain variables. A theoretical model may suggest a relationship between Y and X1 and X2, say.<br /><br />The theory will give us an interesting model to estimate and test, but it doesn't always tell us about functional form, for instance. Should the dependent variable be in levels or logs? Data-plotting may asist in such cases too.<br /><br />Again, thanks for your interest in the postings!Dave Gileshttps://www.blogger.com/profile/05389606956062019445noreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-31547401650278995092011-04-16T22:04:54.243-07:002011-04-16T22:04:54.243-07:00Anonymous n. 1 here. Let me clarify: all this data...Anonymous n. 1 here. Let me clarify: all this data-plotting is a metaphore for the quest for a theory. Is it controversial that we should have a theory before we do any data-work, data plotting included?<br /><br />BTW thanks for your postsAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-33389637970220446902011-04-15T15:58:16.537-07:002011-04-15T15:58:16.537-07:00That seems like the best formal way to handle it, ...That seems like the best formal way to handle it, but I can't help feeling that your priors on finding messages written in the data are probably zero. Something funny is going on in such examples--you probably need to reason backward from the data to identify prior weights that would make your posteriors come to different conclusions, a la extreme bounds analysis.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-55906071454191874642011-04-15T11:01:11.463-07:002011-04-15T11:01:11.463-07:00Anonymous 3: Good question. Yep - I'd go all-o...Anonymous 3: Good question. Yep - I'd go all-out Bayesian, put a prior mass function on the model space, and go from there. In fact I'm a big fan of Bayesian model averaging.Dave Gileshttps://www.blogger.com/profile/05389606956062019445noreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-67802475642208151002011-04-13T17:28:19.750-07:002011-04-13T17:28:19.750-07:00How do you address the criticism that "peekin...How do you address the criticism that "peeking" at the data before building your model invalidates all your subsequent test statistics, because your model has been pre-adjusted to accommodate what might be random features? Do you just go all-out Bayesian?Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-57658217422346882732011-04-12T19:56:40.408-07:002011-04-12T19:56:40.408-07:00Anonymous 2: Thanks very much. I usually begin my ...Anonymous 2: Thanks very much. I usually begin my first intro. grad. econometrics each Fall by telling the students that if I instill some skepticism, then I've done my job!Dave Gileshttps://www.blogger.com/profile/05389606956062019445noreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-41786331499661952822011-04-12T19:54:23.848-07:002011-04-12T19:54:23.848-07:00Anonymous 1: I don't think there are any serio...Anonymous 1: I don't think there are any serious econometricians out there who would find data-plotting at all controversial. Regrettably, there are lots of "empirical economists", and students, who haven't had these points drummed into them. I see some scary things that could have been avoided with a little graphical analysis. Thanks for you comment!Dave Gileshttps://www.blogger.com/profile/05389606956062019445noreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-55637368486259954382011-04-12T18:25:08.214-07:002011-04-12T18:25:08.214-07:00Thanks a lot for these posts Dave! As someone who ...Thanks a lot for these posts Dave! As someone who is practicing econometrics in the private sector now, I appreciate your well written and informative posts as quick reminders not to become complacent!Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-2198942534740642384.post-15591001791355918392011-04-12T14:03:03.441-07:002011-04-12T14:03:03.441-07:00All you're saying is that it's best to hav...All you're saying is that it's best to have a theory before running a regression. The theory would have something to say about those "visual plots" that you recommend. I'm curious. Is this somewhat controversial among econometricians?Anonymousnoreply@blogger.com