Trend of a continuous variable in a repeated model

[tka726]

I am trying to see if there is a trend in a continuous lab value over time. There are 3 timepoints (1=baseline, 2=month 1, 3=month 2). I structured the data such that each patient has 3 records. Most patients have data for all 3 timepoints, but not all.

Is a mixed model appropriate, with repeated time? Or do I use proc reg?

I am relatively new with proc mixed; this is the mixed model I currently have.

proc mixed data=lab covtest;

      class time ptno;

      model labval = time;

      repeated time/subject=ptno type=un;

        lsmeans time / cl pdiff adjust=tukey;

      ods output LSMeans=lsmeans1;

run;

Type 3 Tests of Fixed Effects
Effect Num DF Den DF F Value Pr > F
time 2 207 137.04 <.0001

Do I use the type 3 tests of fixed effects to determine if there is a trend over time?

IF I want to see the differences between any 2 timepoints, should I use the p-values from the differences of least square  means section, or should I do paired t-tests/wilcoxon signed rank?

[SteveDenham · Accepted solution]

EDIT: I forgot to say that your PROC MIXED approach is quite good.  I wouldn't shift to REG.  The Type3 tests are whether at least one time point mean differs from all the rest.  That may not be a trend, so...

If you want to look at a trend in time, use an LSMESTIMATE statement.

lsmestimate 'linear time' time -1 0 1;
lsmestimate 'quadratic time' time 1 -2 1;

Note that with 3 time points, for linear time the middle time point has no effect, which is why looking at the quadratic time (or deviation from linearity) is useful.

SteveDenham

[tka726]

Thank you!
The p-value for 'linear time' and 'quadratic time' are both significant. I understand that the significant p-value for linear time means that there is a difference in lab value between time 1 and time 3, but I don't exactly understand what the significant p-value for quadratic time means...?

[SteveDenham]

Here is an interpretation:  Significant linear means that a straight line fit through the 3 means will have a slope that is significantly different from 0.  Significant quadratic tests whether there is a significant deviation from linearity in the means - that while a straight line fits, there is more that should be considered. 

In this case, without seeing the means, I would expect that the month1 and month2 means differ from baseline by approximately the same amount.  A hockey-stick graph of some sort.

SteveDenham

[tka726]

Thank you so much, this is perfect! And you are absolutely right - the graph looks like a boomerang.