Mixed model fitted on time course data I have some time course data which plotted looks like this:

I have fitted a mixed model to my raw datapoints with R's lme4::lmer, as seen in this code.
In essence I get the following output for my model:
measurement~condition*time:
Random effects:
 Groups   Name        Variance Std.Dev.
 ID       (Intercept) 1118082  1057.4  
 Residual               58680   242.2  
Number of obs: 5784, groups: ID, 12

Fixed effects:
             Estimate Std. Error t value
(Intercept)  2480.379    305.375   8.122
CoIhard       -90.294     12.659  -7.133
Time          -17.326      3.884  -4.461
CoIhard:Time   62.931      5.492  11.458

Correlation of Fixed Effects:
            (Intr) CoIhrd Time  
CoIhard     -0.021              
Time        -0.025  0.611       
CoIhard:Tim  0.018 -0.864 -0.707

Now, what am I to make of these results? As I predicted in this question the factors only get one intercept value each, as if they would follow a steady linear increase/decrease. Obviously my time course does not.
What added information pertaining to the description of the difference between my conditions does fitting this model give me?
Cheers,
 A: (1) your constant = 2480! Where does that fit on the graph?
(2) Graph doesn't have axis labels /legend. Looks like fitted data.
(3) is time a linear term in this model? why?  Have you tried fitting quadratic terms, cubic terms, quartic terms...
- unless the change in scale is distorting the graph, the time trend is clearly non-linear and using a simple linear term lacks face validity. I'd build a main effects model first (time time2, time time2 time3 etc. where time2=time^2 and time3=time^3), find which one fitted best and then consider adding interaction terms. Your categorical would need to interact with each time term if you wanted to add an interaction. I would also test to see if you need the interactions
- how you represent that in R is not clear to me (I usually use Stata)
- There are other ways of handling non-linearity from piecewise models, to I think there is a mixed-effects GAM package. But this is a solid and traditional approach.
(4) have you considered time as a random slope?
(5) have you done residual diagnostics?
- simplest would be to generate histograms of your level-1 and level-2 standardized residuals. but I think the linear time assumption is the key.
(6) have you plotted predicted vs. actual?
- from the model you should be able to predict a mean/model averaged trajectory over time. If you do this with the above model it should be clear that it doesn't fit the data at all. Or show that it sort of does fit the data.      
I don't think you're at the "what does this model tell me" stage yet. It's always hard to tell from just a snapshot like this, sorry if I'm way off base and wasting your time, but does seem like you still have a lot of model refinement to do.            
