[R-sig-ME] Linear mixed model - heterogeneity

[Ben Bolker]

  If you have 5 reps per patient, unless those reps vary in some
interesting way or they have structure you need to know about (e.g.
you're trying to estimate slope of improvement over time, or
patient/rep instances having differing values of covariates you need
to adjust for), it would definitely help to just aggregate (i.e., take
the mean over reps: see Murtauch "Simplicity and complexity in
ecological data analysis") ...

On Mon, Nov 2, 2015 at 5:58 AM, Etn bot <etnbot1 at gmail.com> wrote:
> @Ben many thanks or your response - with reference to the source of the
> zeros - the clinical data: patients force is recorded using a machine, this
> force reading is recorded 5 times for each patient at each time point (4
> different visiting times). Sometimes the machine has a reading of zero (for
> all 5 reps) and other times it has a zero reading for e.g. 1st rep, 3rd rep.
> If there is a full zero reading (for all 5 reps at each of the four time
> points), this is due to the patient having no force (true reading and this
> does not happen very often in the data). If there is zero reading (for some
> of the 5 reps) then this could be due to the patient not having ability to
> consistently push hard enough for that reading and the machine recorded
> zero.
>
> On 30 October 2015 at 01:10, Ben Bolker <bbolker at gmail.com> wrote:
>>
>> lme4 will run Gamma mixed models, but these don't accomodate zeros.  I
>> don't think Weibull will either.  You're also right that
>> transformation won't generally solve these problems. There are very
>> few positive distributions, not considering censored variants of
>> real-valued distributions, that will naively allow zeros.   You could
>> run a two-stage model (Bernoulli model for zero vs non-zero, then a
>> positive-distribution model for the conditional effects on the
>> non-zero values only).
>>
>> The cplm package allows tweedie mixed models, which might work for
>> you. AD Model Builder and Template Model Builder will allow you to fit
>> fixed models from any distribution you can specify (with a generic
>> Laplace approximation engine built in), but the learning curve is
>> pretty steep ...
>>
>> It's important in this case to consider the source of your zeros.  Are
>> they below minimal detection limits (in which case something like a
>> Tobit is appropriate)?  Do they represent a separate process (in which
>> case two-stage models are sensible)? Or ... ?
>>
>> On Fri, Oct 23, 2015 at 10:15 AM, Etn bot <etnbot1 at gmail.com> wrote:
>> > I have a run a linear mixed effects model in R to model clinical data,
>> > however this model is heteroscedastic (as there excess zeros in the
>> > response variable)....
>> >
>> > I have tried transforming the data (log transform) and (sqrt), however
>> > neither transformation resolve the issue (see residual versus fitted
>> > value
>> > plot). I have not used cox proportional hazards model as the data is not
>> > time-to-event data, the data measures force and there are a large number
>> > of
>> > observations have a reading of zero. I cannot exclude these readings as
>> > they are valid.
>> >
>> > I have found a R package that runs Tobit regression (AER), however this
>> > will not accommodate the random effects in the model. I cannot find any
>> > R
>> > packages that run Weibull mixed effects models (or gamma mixed effects
>> > models)...
>> >
>> > Does anyone know if there is a package to run these type of models? (or
>> > can
>> > they suggest any alternative approach).
>> >
>> > Many thanks
>> >
>> >
>> > Etn
>> >
>> >         [[alternative HTML version deleted]]
>> >
>> > _______________________________________________
>> > R-sig-mixed-models at r-project.org mailing list
>> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>