I am running a RD (23 primary occasions and variable secondary occasions) with two groups (female,male). I did run all the models in RMark, but then imported them into MARK for further tests. I have done model averaging for survival parameter for models with AICc weight>0, but the estimated 95% CI for some of the survival parameters gives UCI>1, which doesn't make sense.
Based on the MARK book (Chapter 4.5), MARK first calculates the 95% CI on the logit scale, before back-transforming to the real probability scale, to ensure a [0,1] bound, instead of using the stardard approach θ ± (1.96×SE). As long as I see it, MARK is actually using the stardard approach, see a bit of output from model averaging:
- Code: Select all
Survival Parameter (S) females Parameter 1
Model Weight Estimate Standard Error
---------------------------------------- ------- -------------- --------------
{ S(~sex * time)Gamma''(~constrain)Gamma 0.89083 0.9999982 0.0003608
{ S(~1)Gamma''(~constrain)Gamma'(~constr 0.06153 0.9774169 0.0047503
{ S(~sex)Gamma''(~constrain)Gamma'(~cons 0.04372 0.9754844 0.0064441
{ S(~time)Gamma''(~constrain)Gamma'(~con 0.00392 0.9999956 0.0006696
---------------------------------------- ------- -------------- --------------
Weighted Average 0.9975370 0.0008980
Unconditional SE 0.0074102
95% CI for Weighted Average Estimate is 0.9830130 to 1.0120611
Percent of Variation Attributable to Model Variation is 98.53%
In the example above, UCL for female, primary period 1 is 0.9975370± (1.96×0.0074102)=1.0120611
Any ideas why MARK is not estimating 95%CI to keep estimates within [0,1]?
Many thanks,
Monica