Higher level variables not correctly treated in multiple response module of MLWin
I need help with the multiple response module in MLWin, which I can’t get to correctly specify higher-level variables, erroneously treating them as first-level ones.
1) My data is essentially three-level drawing on time-series for countries so that the first level is 3-year period (within country), the second level is country. Then there is a third level with one variable containing period-wise world market prices. The latter then varies according to 3-year periods but only within the third level.
2) Data are sorted correctly, with a vector of ones for world market level in the first column, with country codes in the second column and with 3-year periods in the third column. The fourth column contains a serial number.
3) When I specify a univariate model MLWin correctly specifies first and second-level variables, but not third level ones (see screen dump in attachment).
In the model in screen dump 1 in the attachment I have included GRPR_D1 which is the world market variable where the correct subscript should have been ‘k’, not ’ijk’. The three following terms refer to a variable REGTYPE with four categories. This is a country level variable, here correctly specified with subscript ‘jk’. The final term CELY_D1 is correctly specified with subscript ‘ijk’. Already here I get a problem with the first variable. When I specify a multiple response model, neither GRPR_D1 nor CELY_D1 get correctly specified.
4) When I specify the same model as part of multiple response equation, both the second- and third level variables come out as first level ones, as is clear from the second screen dump in the attachment:
In the above, resp3 contains the same independent variables as the model in screen dump 1. As can be seen from the attachment all three independent are treated as level 5 variables with subscripts 'ijklm' where ‘l’ stands for serial number and ‘m’ for response category.
5) I am aware that the treatment of GRPR_D1 may have to do with multiple classification problems, but there are no problems like that with the categorical variable REGTYPE.
6) Question: Why can’t I get the second level independent variables (GRPR_D1 and REGTYPE) correctly specified according to level? As far as I can judge it has nothing to do with the sorting of the file, nor with the definition of SERNO.
Göran
Higher level variables not correctly treated in multiple res
Re: Higher level variables not correctly treated in multiple
When you specify a multivariate model with several responses in MLwin , it creates a long vector with interleaved responses. In the User manual on page 213 you will see that the two responses of coursework and written have this pattern. You will also see that there is a dummy variable that picks out which variables is the particular responses. In a three level model (responses on students in schools) these dummies will have the ijk subscript. If you then add a school variable as a separate coefficient for each of the responses they will be multiplied by this dummy so that the will not have the k subscript but correctly the ijk subscript. I think that this is the answer to you question and that is why you higher level variables have a lower level subscript.
There are two more things that are difficult for me to understand about what you are doing and why
1 Your non - multivariate model : You say - " . Then there is a third level with one variable containing period-wise world market prices. The latter then varies according to 3-year periods but only within the third level." you would appear to be mixing up variables and levels here - what are the countries a repeated measures of? - I am sorry but I do not follow your logic. I think you have a two level model.
2 Your multivariate model: it is unusual to put predictors into only one of the response (yours are related to only response); multivariate models are usually used because you want to see how the predictors are differently related to each response, ( as well as getting the correlation/covariance at each level)
I would recommend that you work through Chapter 14 of the User manual.
Kelvyn Jones
There are two more things that are difficult for me to understand about what you are doing and why
1 Your non - multivariate model : You say - " . Then there is a third level with one variable containing period-wise world market prices. The latter then varies according to 3-year periods but only within the third level." you would appear to be mixing up variables and levels here - what are the countries a repeated measures of? - I am sorry but I do not follow your logic. I think you have a two level model.
2 Your multivariate model: it is unusual to put predictors into only one of the response (yours are related to only response); multivariate models are usually used because you want to see how the predictors are differently related to each response, ( as well as getting the correlation/covariance at each level)
I would recommend that you work through Chapter 14 of the User manual.
Kelvyn Jones