Presumably about Finite Population Correction factor -- On 9 Oct 2003 11:17:46 -0700, [EMAIL PROTECTED] (Mats Lingblad) wrote:
> I have data on 70 out of 130 projects completed in one firm. What is > the correct way to think about the the standard errors in regression > analysis, if I want to generalise to the firm? > 1. Adjust with the FPC factor since a large portion of the population > is covered. > 2. Do not adjust with the FPC factor since it is the potential number > of projects started by the firm that constitute the "true" population. > > Alternative 2 is safer and more conservative, but it also seems a bit > too hard since the large n/N ratio should count for something. *Should* the n-in-hand count for much, about the other ones? What conclusion are you trying to draw? - Are you in the sort of situation that is like "taking a vote" -- where, if you had 100% counted, you would look at some number and say, "Aha! -- this is what <wins> <the count is> ! " The basic, simple notion for tests and estimation is that you take what is in-hand as fixed and given, as estimates of counts and means; and you project the observed means and SDs as representing those who are not measured, with sampling variability. (For more realism, you would consider how the ones that are collected *differ* from the ones yet to be measured.) > > I would also appreciate it if somebody has references on this subject. > > Finally does anyone have the formula for adjusting the F-test in > standard OLS regression models for a finite population? > I have responded, in some fashion, to most questions posted on the subject in the stats-groups, over the last several years. You can check my stats-FAQ for a cursory summary, from a few years ago; you can use groups.google.org to search (advanced search) of sci.stat.* to get other comments. The assumptions for *simple* testing are tough, since there is that *strong* assumption that the ones 'collected' up to now are not different, in any relevant respect, from the ones that are still pending. If you are going to hope to convince any colleagues with an FPC-based analysis, I think you are going to need a model to work from, where someone has done it before. Else you should be a statistician who has figured out all these details long ago. Hope this helps. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html "Taxes are the price we pay for civilization." . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
