Auto-correlation vs. number of observation periods I've just read an excellent post
mix model
I've a question connected to that. Can you recommend any reference to a comment that if one have not enough observation periods then it is difficult to model properly auto-correlation? Did I understand it properly, that four observation periods are probably not enough to successfully account on auto-correlation.
 A: 4 time points is not necessarily a problem for handling autocorrelation. The lme function in the nlme package has several methods which may be appropriate such as corAR1 for fitting an AR(1) process
Edit to adddress @whuber's comment:
I can see why the question could seem a bit vague. I took it to mean that, in the context of a mixed effects model used for a longitudinal study, a number of participants say, $N$, are measured 4 times each. Of course there are many more than 4 observations in total ($4N$ obviously, if there are no missing data). I think that is where the OP was making a distinction. A single time series of length 4 would be hopelessly inadequate for estimating an AR(1) model, but with many participants, this is a common model to fit, either in a mixed model framework or an SEM (for example. latent growth curve models and extensions thereof) - at least it is common in lifecouse epidemiology where the previous measurement often has a large influence on the next measure, compared to the other covariates, even with just 4 time points. BMI is a good example of this. Perhaps the estimated AR(1) coefficient still does not have a desirable level of precision but I have seen such models where they offer an improved interpretation as well as a better fit. Without the AR(1) process in the model, the covariances between residuals at each time point can be unrealistic.
I prefaced my answer with "not necessarily" because, of course it might be too few time points particularly if the level of autocorrlation is small (and hence my example in lifecourse epidemiology where it can be quite large). I may be able to find a toy dataset to illustrate this, if interested.
