Hi, just the briefest of replies.
On Tue, 11 Sep 2001, Frank E Harrell Jr wrote: > I would be very interested to hear ideas about computing > combined summary statistics (R-square, ROC area, etc.) Any of these that can be viewed as estimating some popln quantity can be combined in the usual way, i.e., by taking the ave value across the data sets, although subasymptotically, it's better to do the averaging on a scale that makes things more likely to be normal, such as Fisher Z for correlations. and > whether there is any hope of getting a combined > imputation-adjusted likelihood ratio test. You can directly combine the underlying chi-squared statistics (not extremely accurate but OK for most purposes -- see Li,..., 1991, Statistica Sinica, pp. 65-92), or do the really correct thing using the LR computations ( Meng and Rubin, 1990, Biometrika, pp. 78-92). Both are summarized in Rubin and Schenker, 1991, Stat in Medicine). Also, > has anyone seen a reference where model diagnostics > are developed in the multiple imputation framework? Great question -- very little has been actively done on this topic. However, on each completed data set, the diagnostics will have too much precision, so if nothing wierd shows up on any of them, you know you are OK. > For example, would it be a good idea to make 5 > partial residual plots for 5 completed datasets, > and is there a reasonable way to combine these? To combine them using the standard things, you have to think of what each is estimating and the accociated standard error -- not all that natural with many diagnostics. Good topic for some brilliant insight! Best wishes, DBR > > Frank Harrell > > > "Raab, Gillian" wrote: > > > > Its all in Shafer's book - or various other places. Just combine the within > > and between imputation variance > > with simple formulae. What are you using to do computations? If it is SAS I > > have a pretty basic macro I have written that > > produces tables of oods-ratios and 95% confidence intervals from a SAS data > > set that contains all the imputed > > data. Most willing top pass on if it would help. > > > > I'd also be interested to hearing from anyone else who has been trying out > > the new SAS imputation procedures. > > > > Gillian Raab, Napier University, Edinburgh, Scotland > > > > -----Original Message----- > > From: [email protected] [mailto:[email protected]] > > Sent: 08 September 2001 10:19 > > To: [email protected] > > Subject: IMPUTE: (no subject) > > > > Dear all > > > > I have a logisitic regression model with continious and categorical > > variables. I carried out multiple imputation for missing values in most of > > the > > variables and redone the logisitic regression. > > > > Now I have 5 results my question is how to apply rubin's rules to combine > > the > > results, I found out it is straight forward for the odds ratios, the > > coeffecients and the standard deviations, I am only stuck with the p-values > > and the 95% confidence intervals. I would like to mention that my data > > consist of 23000 records so assumption of normality is quite feasiable. > > > > Ula Nur > >
