home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!uvaarpa!darwin.sura.net!zaphod.mps.ohio-state.edu!menudo.uh.edu!UHUPVM1.UH.EDU!KATTAN
- From: KATTAN@UHUPVM1.UH.EDU (Mike Kattan)
- Newsgroups: bit.listserv.stat-l
- Subject: Measurement Error vs. Bias
- Date: Thu, 28 Jan 93 07:55:55 CDT
- Organization: University of Houston
- Lines: 45
- Message-ID: <1k8qmvINN5eg@menudo.uh.edu>
- NNTP-Posting-Host: uhupvm1.uh.edu
-
- I'd greatly appreciate feedback concerning the following arguments
- concerning the terms "measurement error" and "bias". In particular,
- references supporting or refuting these arguments would be helpful.
-
- Say one is following the typical steps in prediction (1) determining a
- variable list, (2) collecting a sample, (3) modeling the sample, and (4)
- applying the model to new data. In the sample of step (2), the values
- of the predictor variables do not correspond appropriately with the
- values of the response variable. This would be called measurement
- error, which may occur due to misunderstood questions, approximations,
- incorrect observations, etc. If the values of the response variable are
- inappropriate (e.g., sample mean does not equal population mean), bias
- has occurred. Another example of bias is if the observed distribution
- of the response variable does not follow the true distribution of the
- response variable.
-
- [So basically the argument here is that X problems are due to
- measurement error, and Y problems are due to bias -- is this a
- reasonable interpretation of the terms "measurement error" and
- "bias"?]
-
-
-
- Now consider application of a regression model to new data [step (4)
- from above]. The betas have already been estimated. If it is not con-
- venient to measure the predictor variables from step (1), and one measures
- surrogate variables perfectly, measurement error (rather than bias) has
- occurred, even though the surrogate variables were measured perfectly.
- In fact, when making a single prediction with a regression equation, if
- the wrong values for the predictor variables are used, the predicted
- value for the response variable would be incorrect due to measurement
- error -- bias has no meaning in this context.
-
- [When making a single prediction with a regression equation, bias
- cannot occur in the *new* data (not the original sample data from which
- the regression coefficients were estimated) -- only measurement error
- could occur. Is this interpretation reasonable?]
-
- References for or against these interpretations would be most helpful.
- Should I receive many e-mail response, I'll post a summary. Thanks in
- advance.
- --
- Mike Kattan / Decision and Information Sciences / University of Houston
- Houston, TX 77204-6282 / 713-743-4734 / Fax x4693 / kattan@uhupvm1.uh.edu
- ==
-