modelling time as continuous vs. discrete I am writing an analysis plan for data that is collected on approximately 30 people at approximately 5 unevenly spaced time points.  I am planning to analyze the data via a repeated measures mixed model, but I am unsure as to whether time should be treated as a continuous or discrete parameter in the model.  I was thinking it made more sense to model time as continuous due to the uneven spacing of the visits, but a colleague noted that this assumes a linear relationship between the outcome variable and time.  What are others' opinions on this?  How is power affected by the choice of continuous vs. discrete? Unfortunately, I do not have the benefit of having the actual data to determine the relationship with the best fit.  The model needs to be specified in advance.
 A: Time as continuous uses one degree of freedom (unless you include polynomials of course) - if it is treated as discrete, each dummy uses a degree of freedom. It may not be a big deal if you have lots of observations.
With 5 points of data, you may want to treat it as continuous; my experience is the more fixed time effects you have it gets harder to interpret their meaning, and I always end up looking for some sort of time trend anyways. Any nonlinearities can be treated with a quadratic or higher order term. Continuous time in a mixed model should be able to handle uneven spacing. That said, I'm not sure why you have to commit now. You can model both, and do a LR or Wald test to see which fit the data better. 
If you really cannot touch the data before making the decision, and once its decided you have to run with it, then I would recommend you rely on theory (which often trumps anything the data will "tell" you). But what kind of situation are you in that you cannot make adjustments to the model after modeling? That kind of goes against all principles of modeling I've ever followed. You define your initial model, test its fit, test your various hypotheses, and adjust model. I personally would never feel comfortably fully specifying a finalized model before analyzing data.
