www.cmm.bristol.ac.uk/forum

Hi.

I am trying to estimate a 2-level multinomial model. The dependent variable has 4 categories and I have about 35 explanatory variables. I am working on with a data set with 83,123 records. Unfortunately, everytime I use runmlwin to estimate the model, I keep on receiving an error that says, "Unable to allocate 888379 k. Worksheet size unchanged." I tried adding mlwinsettings(size(900000)) as an additional option but I still get a similar error. I also tried keeping the variables that are needed but it didn't solve the problem... Any help will be greatly appreciated....


Thanks a lot.

Have a nice week ahead,
Arturo

900,000Kcells works out as about 3.4Gb which it is very unlikely that MLwiN can allocate in one contiguous block. It is possible that the calculation runmlwin is doing to work out the space needed is wrong, in which case you can try using the mlwinsettings(size()) option with lower values. The highest I have been able to get on a 64bit machine is 500,000 (~1.9Gb). If it really needs a worksheet of 900,000 then unfortunately I don't think that MLwiN will be able to provide it.

Hi Arturo

MLwiN wants to expand your 80,000 records to 3*80,000 = 240,000 records (the response is modelled as three log-odds contrasts)

However, even so 240,000 obs and 34 vars doesn't sound impossible to me so am surprised you get hit by the error message.

If you can find no way around the warning message then you could

(1) work with a random sub sample of your data. You might take a random sample of clusters and then a random sample of observations within clusters.
(2) Collapse the response from 4 to 3 categories (not really ideal) since then you will only be needing 160,000 records in MLwiN

George