[R-sig-ME] logistic model with exponential decay
Douglas Bates
bates at stat.wisc.edu
Mon Dec 22 10:59:51 CET 2008
On Sun, Dec 21, 2008 at 2:02 PM, Stijn Ruiter <s.ruiter at maw.ru.nl> wrote:
> Hi,
> I have official judicial data on criminal offending (dichotomous dependent
> variable=conviction(=Y)) of all (adult) children of fathers who differ with
> respect to their level of criminal behavior. These data were registered on a
> yearly basis. So, I am able to follow people over the course of their lives
> and model whether they get convicted. I intend to estimate a discrete-time
> logit model on a person-year file. Of course, because children are nested
> within their fathers, I need to take that into account. Furthermore, many
> subjects get convicted more than once during their lives, so I need to
> estimate a repeated events model.
> I have several time-constant variables (e.g., gender) and several
> time-varying variables (e.g., number of years since father committed a
> crime(=T)). I would like to estimate something like this:
>
> logit(Y) ~ alpha + beta1*GENDER + exp(-T/beta2) + ... + error term for
> nesting within fathers + error term for nesting within subject
At present nlmer does not allow families other than gaussian for the
conditional distribution of the response given the random effects. I
do plan to combine nonlinear model functions and glm families to allow
for generalized nonlinear mixed models but must finish a couple of
other projects first.
> Douglas Bates schreef:
>>
>> On Sun, Dec 21, 2008 at 4:00 AM, Stijn Ruiter <s.ruiter at maw.ru.nl> wrote:
>>
>>>
>>> Hi all,
>>> Frederik is right. Do you think such a model can be estimated using
>>> nlmer?
>>>
>>
>>
>>>
>>> How?
>>>
>>
>> We would need more detail about the data and the model to be able to
>> answer.
>>
>> Is the response on a continuous scale (i.e. not binary or a count)?
>>
>> What are the covariates?
>>
>> What is the model for the mean response?
>>
>> How many random effects would be defined and how would they enter the
>> model for the mean response?
>>
>
More information about the R-sig-mixed-models
mailing list