[R-sig-ME] estimating AR1 parameters of level one error

[Viechtbauer Wolfgang (STAT)]

Good point! In fact, besides the fact that the OP was not fitting exactly the same model, numerical instabilities may be another reason for why different values were obtained. I've been dealing with lots of analyses recently where we have 30-60 measurements per person, so fitting models with both random slopes and AR(1) structure is not an issue. But if there are only a handful of measurements, differences are more likely.

Best,
Wolfgang

> -----Original Message-----
> From: R-sig-mixed-models [mailto:r-sig-mixed-models-bounces at r-
> project.org] On Behalf Of Steve Candy
> Sent: Friday, May 22, 2015 02:55
> To: r-sig-mixed-models at r-project.org
> Subject: Re: [R-sig-ME] estimating AR1 parameters of level one error
> 
> Thanks Wolfgang
> 
> Yes I agree its possible and I have fitted such a model using random
> subject-splines. I found numerical instability but this may not be the
> case with other applications. So my statement "not sensible" is not
> justifiable. I should have said "can be problematic". The marginal
> covariance is much more complex and harder to describe graphically  than
> the random intercepts plus AR(1) error model which I have used
> extensively and combined with the log-transform of the response to allow
> for "splaying-out" of curves in time or adding extra variance parameters
> to model that or other types of variance heterogeneity. I have found this
> type of model more stable in practice.
> 
> Thanks for correcting me.
> 
> Regards
> 
> Steve
> 
> >It is certainly possible to have a model with random intercepts and
> slopes
> (for time) and also AR(1) correlated residuals over time within
> individuals.
> The random slopes model differences in the trend between individuals,
> while
> >the AR(1) structure models how the residuals are fluctuating around the
> person-specific slopes within individuals. Of course, you need to have a
> sufficient number of follow-up measurements within individuals to
> distinguish >those two elements. But this is certainly possible. And in
> fact, ignoring serial correlation in the residuals when it is present
> could
> lead to inflated Type I error rates for the mean trend effect.
> 
> >Best,
> >Wolfgang