treating measurement occasions as a numerical or as a factor predictor

On 15-12-12 11:48 AM, Rich Shepard wrote:

On Sun, 13 Dec 2015, K Imran M wrote:

I would like to ask a question about treating measurement occasions in a
longitudinal analysis specifically when using linear mixed model. In my
study, I have taken data on 3 separate occasions (at baseline, at 1 month
and at 3 months post baseline). I am not sure what is best approach treat
these measurement occasions in my analysis using lmer or lme functions.
Should I treat them as a numeric or as a factor variable. My feeling says
that I should treat such measurement occasions as a factor but I do not
have strong theoretical reasons for that.

Kamarul,

  What question do you want to answer with your data?

Rich

   A few thoughts to consider:

* with only 3 measurement occasions you won't really be able to treat
them as a random effect (not enough distinct levels to estimate
among-occasion variance reliably)
* if you treat measurement occasion as numeric (i.e., a linear effect of
time) you will assume that the change per month is identical throughout
the observation period (i.e. you will expect 2 times as much change from
1 month to 3 months post baseline as from baseline to 1 month post baseline)
* if you treat measurement occasion as categorical (factor) you will
estimate 2 parameters for the effect of occasion rather than 1.  There
are various ways you can break this up, depending on the contrasts you
choose (default 'contr.treatment': baseline vs 1 month, baseline vs. 3
months.  MASS::contr.sdif() gives you successive differences, making the
variable an ordered factor gives you contr.poly() (linear, quadratic
contrasts) by default.

   I would *generally* say that you're not complicating the model
much/spending very many parameters(degrees of freedom) by using a
categorical rather than a numeric input for time, and otherwise you're
making a fairly strong assumption, so I would recommend categorical.

  Ben Bolker