Comments on: From Terence’s stuff: You want proof?
http://hea-www.harvard.edu/AstroStat/slog/2009/from-terences-stuff-you-want-proof/
Weaving together Astronomy+Statistics+Computer Science+Engineering+Intrumentation, far beyond the growing bordersFri, 01 Jun 2012 18:47:52 +0000hourly1http://wordpress.org/?v=3.4By: Raoul LePage
http://hea-www.harvard.edu/AstroStat/slog/2009/from-terences-stuff-you-want-proof/comment-page-1/#comment-1046
Raoul LePageFri, 12 Aug 2011 17:44:19 +0000http://hea-www.harvard.edu/AstroStat/slog/?p=4133#comment-1046I've long been cautious about any model employing hypotheses of randomness in model components, particularly the backbone method of multiple linear regression, popular accounts of which may have encouraged the use of such hypotheses elsewhere. One particular liability of models is they may suggest particular classes of sample points are all that are needed. Examples include regression models positing iid errors that are much the same no matter which sample points are included in the regression design. Something similar is seen in time series and elsewhere. If the model is wrong on that point it may have encouraged a laxity in data selection from which there is no recovery once the data are in. I call this "brittle" modeling. Statistics is yet young. Stay tuned for models that are more in the spirit of sampling theory where inference is rooted in randomness introduced by the experimenter rather than hypothesized (a different sample space) and no model is correct although some may be estimated to provide better population description than others. Simpler may for some purposes prove better in the long run.I’ve long been cautious about any model employing hypotheses of randomness in model components, particularly the backbone method of multiple linear regression, popular accounts of which may have encouraged the use of such hypotheses elsewhere. One particular liability of models is they may suggest particular classes of sample points are all that are needed. Examples include regression models positing iid errors that are much the same no matter which sample points are included in the regression design. Something similar is seen in time series and elsewhere. If the model is wrong on that point it may have encouraged a laxity in data selection from which there is no recovery once the data are in. I call this “brittle” modeling. Statistics is yet young. Stay tuned for models that are more in the spirit of sampling theory where inference is rooted in randomness introduced by the experimenter rather than hypothesized (a different sample space) and no model is correct although some may be estimated to provide better population description than others. Simpler may for some purposes prove better in the long run.
]]>