missing individual covariates

[brouwerl]

I have a data set with natal territory quality as an individual covariate, however, this ind cov is not available for all years.
Can I just include it for the years available and for the years not available just include a '1' in my design matrix? Or do I have to reduce my dataset to one in which the ind cov is available for all years?


Lyanne Brouwer
University of Groningen
The Netherlands

[gwhite]

Lyanne:
You can use missing values in individual covariates, but you have to be very careful how you do it. First, use the z transformation to replace all the values that you do have with their z-transformed value (i.e., (x - mean)/sd). Then, for the missing values, give them the value of zero, so that they correspond to the mean. Because the new variate is scaled to have a zero mean, a zero value does not affect the slope, and represents the pivot of the line relating the covariate to the real parameter.

I only suggest this option for when the individual covariate is missing in a few places, and not because you failed to catch the animals. If you are missing the value because you failed to catch the animal, then you need to convert the covariate to a categorical variable, and use the multistrata model to handle the missing values. That is, each of the categories now represents a stratum, and animals transition between the strata, regardless of whether you captured them or not.

Gary

[rlong]

Is there a similar way to treat missing categorical covariate data that is dummy coded? The only approach I can think of is to include an extra dummy variable for "missing", and model this as one of the categories.

Thanks for any thoughts.

Robert

Robert Long
USGS Vermont Cooperative Fish & Wildlife Research Unit
University of Vermont

[cooch]

Is there a similar way to treat missing categorical covariate data that is dummy coded? The only approach I can think of is to include an extra dummy variable for "missing", and model this as one of the categories.

Thanks for any thoughts.

Robert

Robert Long
USGS Vermont Cooperative Fish & Wildlife Research Unit
University of Vermont

Gary's suggestion (previous posting in this thread) on using a multi-state approach is, in fact, a natural solution for categorical covariates.

[brouwerl]

So the missing ind cov do not influence the slope in this way, but what about the p-value?
I would like to test the effect of an ind cov that is only available for 3 years, but to get the best survival estimates I would like to use my complete data set which includes data from over 10 years. However, in this way I would have a lot of missing values and I cannot test for the significance of the ind cov anymore? Is there a way to do this?

Lyanne Brouwer
University of Groningen
The Netherlands

[cooch]

So the missing ind cov do not influence the slope in this way, but what about the p-value?
I would like to test the effect of an ind cov that is only available for 3 years, but to get the best survival estimates I would like to use my complete data set which includes data from over 10 years. However, in this way I would have a lot of missing values and I cannot test for the significance of the ind cov anymore? Is there a way to do this?

Lyanne Brouwer
University of Groningen
The Netherlands

No...and with only 3 years of data, there will be significant limits to how strong an inference you can make about the influence of an individual covariate anyway.

[Bernd Gruber]

Dear Markers,

I am looking for a reference in the literature that can be used to cite the idea that if the individual covariate is missing for some animals the mean of the covariate can be used to still include these animals for the survival estimate.

I found this idea mentioned here and in the gentle introduction, but nowhere else. If anyone could send me a reference (as far as I recall the "Gentle introduction should not be cited), it would be very helpful.

Thanks in advance,

Bernd Gruber