Because I primarily worked with official statistics from many small, highly skewed establishment surveys of diverse entities, using relatively simple regression to predict for missing/out-of-sample data, and some small area estimation, in a monthly and even weekly data production environment, there were no resources that would have allowed one to even consider the data collection effort needed for measurement error models. Further, heteroscedasticity was/is important, and the disproportionately larger measurement error apparent for data submitted by smaller establishments, which lacked the resources to have data collection departments, complicates this, and modifies the coefficient of heteroscedasticity as well. That modification produced results that much scrutiny showed useful.
Note in Measurement Error Models, Wayne A. Fuller, John Wiley & Sons, Inc., 1987, 2006, pages 13-14, that an estimate for the sigma for the measurement error of x is needed, and the suggestion given for obtaining it was to take many repeated "independent" measurements. This may often be beyond the resource limits for many projects. Further, I suggest that it may be likely that if measurement error is a substantial problem for x, that it may be problematic for y as well.
So, above I have noted an area of application where measurement error models may not be useful. You have to handle what is most important and can be handled in a practical manner. However, I would like to hear of applications where these models are useful. ... I would especially like to hear from those on ResearchGate who have actually used a measurement error model for some practical application.
Thank you.
An example of a 'real world' application, which was actually conducted, would be most interesting. In particular, what amount of resources were used to estimate measurement error of regressor variables, and measurement error of y, as opposed to basic data collection for the sample?
Has anyone tried a sensitivity analysis such as what would happen if estimates of measurement error were off, say by 10%?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression