Dummy Individual Covariates and Real Parameter Estimates

[gwhite]

Todd, All:
As Evan points out, using a categorical individual covariate to describe group effects can be tricky, but not impossible to work with. Todd, you first have to convert your ordinal variable (1, 2, 3) into 2 dummy or indicator variables. You can do this manually, and recreate the MARK file, or you can use the design matrix functions that I have implemented in MARK -- see the Design Matrix Functions help document.

The EQ function could be used to create the 2 dummy variables. Suppose you called your ordinal individual covariate TREAT. Then the design matrix:
1 EQ(TREAT,1) EQ(TREAT,2)
1 EQ(TREAT,1) EQ(TREAT,2)
1 EQ(TREAT,1) EQ(TREAT,2)
...

would do the job. The EQ(TREAT,1) returns the value 1 when TREAT = 1, 0 otherwise. Likewise, EQ(TREAT,2) returns the value 1 when TREAT = 2, 0 otherwise. So, in effect, you have created the dummy variables with the EQ function.

Now the tricky part. The real parameter estimates generated will reflec the value of TREAT specified. See the help document "Individual Covariates and Real Parameter Values" for the details. You will have to run the model 3 different times for each of the TREAT values to get the estimates for each level of TREAT.

Gary

P.S. And we did talk about this problem in last summer's MARK workshop, like we do every time in the intermediate workshop. However, you are drinking out of a fire hose at these workshops, so some of the information splashes off.

[cooch]

Todd, All:
As Evan points out, using a categorical individual covariate to describe group effects can be tricky, but not impossible to work with. Todd, you first have to convert your ordinal variable (1, 2, 3) into 2 dummy or indicator variables. You can do this manually, and recreate the MARK file, or you can use the design matrix functions that I have implemented in MARK -- see the Design Matrix Functions help document.

Also discussed in some detail in Chapter 12 (the individual covariates chapter).

Now the tricky part. The real parameter estimates generated will reflec the value of TREAT specified. See the help document "Individual Covariates and Real Parameter Values" for the details. You will have to run the model 3 different times for each of the TREAT values to get the estimates for each level of TREAT.

Which is part of the reason for not including this in the book that I alluded to earlier. While you *can* get there from here coding using individual covariates to specify levels of a classification variable, its often non-trivial to implement, difficult to interpret (e.g., interaction effects?), and takes more time than it might save.

[Todd]

Thanks gentlemen for clarifying this issue. It now becomes obvious why it is preferred to use group attributes instead of covariates to handle non-binary, categorical assignments, in addition to the already obvious issue of not having to run your model twice when using a group instead of a "covariate" option. In other words, it requires a little more trickery to use a >2 level covariate than to use three groups. Again, the benefit is that you don't have a PIM chart with 3x the number of starting variables that makes your DM really long.