Time-varying covariates before time of capture

[Kenup17]

Hello,

I'm modelling survival as a function of age (through a binary time-varying covariate) on RDPNE (thanks bmcclintock for the tip ). Logically, I don't the covariate's value of a given individual before its first capture, so I have a string of missing values before the occasion of first capture.

Section 11.6 of the handbook recommends using the mean value of the covariate for missing values. However, it does not approach the case of missing values of time-varying covariates. Should I:

(a) Use the mean value of the covariate over all occasions?
(b) Use the mean value over each interval (i.e., the mean value of occasion i for missing values on i )?
(c) Substitute missing values for zeroes?
(d) MARK won't read covariate values for not-yet captured individuals; it won't make a difference?

I thought about for a while, and still think it's a silly question; however, I couldn't find any reference to it on the book nor the forum. Sorry if I overlooked it.

Thanks in advance!

[cooch]

Hello,

I'm modelling survival as a function of age (through a binary time-varying covariate) on RDPNE (thanks bmcclintock for the tip ). Logically, I don't the covariate's value of a given individual before its first capture, so I have a string of missing values before the occasion of first capture.

Section 11.6 of the handbook recommends using the mean value of the covariate for missing values. However, it does not approach the case of missing values of time-varying covariates.

Thats funny. I was pretty sure I wrote 11.6 to deal with this problem. Try reading it again. You either impute the value (say, by using the mean), or discretize, and use a multi-state approach.

Or, go Bayesian (which is large part isn't much different than using the mean of those individuals that are captured). If the covariate has a plausible generating model (like, say, growth), then if you're clever, you can possibly impute. But, for general problems, within MARK, best strategy is to discretize, and use a MS approach.

Should I:


(d) MARK won't read covariate values for not-yet captured individuals; it won't make a difference?

MARK has nothing to do with it. Whether or not values prior to first capture are used depends on the data type. For what you're working with, the models condition on first encounter. So, the covariate value prior to first encounter has no meaning.

[bootzies]

Hello all,

Following on from the thread above, I just wanted to clarify what this means in practical terms when setting up the raw data. Does this mean that the cell values for time-varying individual covariate values can be left blank for occasions prior to capture ? Or, does some numeric value need to be entered to avoid MARK throwing an error (I remember reading that NA values in time-varying individual covariates cause errors - but I am not sure whether this just refers to NA values for occasions after initial capture - i.e. missing data).

I have an age growth curve for my study species, so I am able to allocate each individual an estimated age at first capture. I am trying to run a simple CJS model with survival as a function of this estimated age. I want to include estimated age as a time-varying covariate (as opposed to creating age sturucture in the design data) so that I can later make use of the covariate.predictions function to generate a survival ~ age curve (I am using RMark).

I am now trying to complete the time-varying covariate columns in my raw data for the 'estimated-age' at each sampling occasion. I am doing this by simply adding (or subtracting) the time since initial capture to the initial age estimate. This obviously works fine for generating an age value at occasions after first capture, but it results in some negative age estimates for young individuals captured later in the survey. I want to know whether I can leave the 'estimated age' field blank in these cases where I have negative age estimates? Or, should I change them to zero (to prevent negative age estimates being included, but have some numeric value in the cell)? Or, shoudl I change them to NA? Or, should I just leave the negative values, as they won't be used in the modelling process anyway?

Thanks in advance.

Alice

[jlaake]

You can enter any value because for CJS model, any value before and including the first 1 are not used. But I don't understand why you are doing this. Age is already available in the design data and it is unlike environmental covariates like in my recent post in the sense that you wouldn't predict for values other than the ages represented in the data.

--jeff

[bootzies]

Hi Jeff,

Thanks very much for the reply. I am not sure why I am doing it this way! I thought that it would be easier in terms of creating the output that I am after - which is a curve of survival by age (actually a number of curves for different populations, so that I can compare them and look for differences). the reason I thought it would be easier to do with a time-varying covariate, rather than using the design data, is that I would be able to predict to regular intervals within the range of the ages of sampled individuals if I used a time-varying covariate (as this would enable me to use the covariate predictions function).

I have previously used the design data as you suggest; where I generated my own 'estimated age' column by summing my 'estimated age at capture' column (converted from factor to numeric) and the numeric 'Time' column (i.e. time since first capture). In these previous models I was really only after age-class estimates to look for differences in survival for hatchlings, juveniles and adults - so I binned the resulting 'estimated age' column into age classes.

So, what you are suggesting is that I could use the same method as I used above to generate a column of 'estimated age' (without the binning) within the design data, and that would enable me to get 'estimated age'-based survival estimates? And, that this would be more appropriate than using 'estimated age' as an individual time-vary covariate?

Sorry for the confusion, I just want to make sure that I fully understand the options I have, and that I choose the most appropriate one!

Coudl you please post the link to the recent comment that you mentioned in your previous response?

Cheers,

Alice

[jlaake]

We are starting to move over to RMark territory, so I'll post a message over there addressing your issue. The post I was referring to is at http://www.phidot.org/forum/viewtopic.php?f=21&t=3044