Median chat 95% CIs

[peteM]

Dear MARK forum
I am trying to figure out if my median chat estimate is 'acceptable'.
I ran the median chat test on my global model which is phi(site*time),p(site*time), with 12 parameters.
In the window that appears just before you run the median chat test it says observed deviance/df = 0.98
Which sounds like an OK result, if this is the chat value RELEASE would generate and is meant to be near 1. I tried to set the lower bound to 0.5 but only 1 is allowed - why is this? Because anything below 1 doesn't need adjustment? So I left the bounds at the defaults which appeared (1 to 1.3) but increased the number of replicates from the default of 10 to 30 to try and improve confidence in the median chat estimate.
The result is a median chat value of 0.83 with a SE of 0.17. This doesn't sound too bad but the 95% CI on the median chat estimate sounds very wide (-0.49 to 2.15). Is this a problem? What does the 95% CI on a median chat estimate actually mean?
Thanks!

[abreton]

See the side bar at the top of page 5-6 (Chapter 5, Page 6). At the moment those of us (like myself) that need guidance from the statisticians are being told to set c = 1 when our estimates of c are < 1 (a condition known as underdispersion). I suspect that the median c-hat tool was developed with this rule in mind, hence, you're not allowed to drop the lower boundary below 1. The c-hat that was reported is the deviance/deviance df from the model selected in your results browser when you requested the median c-hat option in MARK (see bottom of page 5-29).

I'm surprised the defaults bounds were 1-1.3, I would have expected 1-5.5. If the upper boundary limit was 1.3 then I suggest increasing this initially to 3 or 4. If again, your estimate is close to 1 then decrease it to 2...and the less than 2 if the result continues to be close to one. This is the advice given in the MARK book (chapter 5)...it will help you refine your estimate of c-hat using the 'median' approach.

If my estimate of c-hat was < 1 then I wouldn't be concerned with the SE or CI. However, I would have taken care when extracting the dataset (from a larger pool of available data)...always thinking about heterogeneity. If I thought I done this well, and c-hat was < 1....I'd leave work early that day and celebrate.

Just some thoughts...

andre

[cooch]

See the side bar at the top of page 5-6 (Chapter 5, Page 6). At the moment those of us (like myself) that need guidance from the statisticians are being told to set c = 1 when our estimates of c are < 1 (a condition known as underdispersion). I suspect that the median c-hat tool was developed with this rule in mind, hence, you're not allowed to drop the lower boundary below 1.

Andre is entirely correct. As he surmised, MARK has a 'hard lower limit' of 1 for the median c-hat approach.