Previous: What computational details should

Up: Frequently Asked Questions

Next: What global values should

What statistical issues should I check?

It is essential to verify that the model fits the data. First, look at eigraph's fit: for the X by T graph on the left, verify that the $E(T_i\vert X_i)$ line passes through the middle of the points, and the 80% confidence intervals capture around 80% of the points, vertically for each value of $X$ on the horizontal axis. For the tomography plot of the right of the fit graph, verify that the contours capture the place from where the ``emissions'' come from (roughly speaking, where the lines are crossing). Then check eigraph's tomogS and verify that the maximum likelihood contours (on the top left, the same as the tomography plot in the fit graph) and mean posterior contours (on the top right) both fit the data in roughly the same way. If there are problems here, see the next question in this FAQ.

Other issues to check are whether there outliers or multiple modes. Is there aggregation bias? Check the results at the aggregate level (eigraph's post; eiread's Paggs) and the precinct level (e.g., eigraph's beta or eiread's betaB and betaW); are these consistent with your qualitative knowledge? Verify that the relationship between $\beta_i^b$ and $X_i$ and between $\beta_i^w$ and $X_i$ are consistent with your understanding of your substantive problem (see eigraph's boundX). Is there survey, qualitative, or other external information you could have used but didn't? Chapter 16 provides a complete checklist that should be used for every serious application.