 |
 |
Dear Allstat ..... I have a problem with expressing a confidence/prediction interval around a regression line (model) that would make sense easily to medical clinician colleagues. The situation is as follows. The dataset is N=300,000 approx and I am using standard multiple regression techniques to contrast two models having the same dependent variable (birthweight at full term pregnancy) with firstly ... A: 2 co-variates (1 binary, 1 scale) and then ... B: 11 co-variates (a mixture of binary, categoric and scale). Particularly relevant statistics are given as follows: A: Constant = 3553; SEE = 442: R-square = 0.180. B: Constant = 3511; SEE = 411; R-square = 0.291 Now, clinicians understand the concept of R (correlation explaining the 'fit' of the model), R-square (coeff of determination describing the percentage of variation in birthweight explained by the co-variates present) and SEE (Std error of estimate = sd of model residuals as a way of describing variation around the model). However, they seem nervous about the fact that the increase in R-square (about 62%) is reflected only in a corresponding decrease in SEE (approx 7%). Even if I convert R-square to R, the increase is still large (27%) compared with the SEE decrease. When I attempt to convert these statistics to diagrams, the results are quite uninspiring, since the population sampled is so large and the corresponding intervals so small as a result. Question 1: Has anyone any suggestions on how I might explain the above statistics in a way that might make sense to my colleagues? SPSS has a facility for calculating 'individual confidence intervals' (LICI and UICI) which seem much wider than I personally obtain for standard confidence/prediction intervals. Question 2: I cannot find any information on these! Does anyone know (or can they refer to) how these are calculated? Thanks in advance. Andre Francis Perinatal Institute Birmingham United Kingdom You may leave the list at any time by sending the command SIGNOFF allstat to [log in to unmask], leaving the subject line blank.