Model predictors included main effects and interactions of group with linear and
quadratic effects of time (i.e., time and time squared, respectively), and the three - way interactions of 1 - year environmental factors × group × linear time and environmental factors × group × quadratic time.
Not exact matches
Individual growth curve models were developed for multilevel analysis and specifically designed for exploring longitudinal data on individual changes over
time.23 Using this approach, we applied the MIXED procedure in SAS (SAS Institute) to account for the random
effects of repeated measurements.24 To specify the correct model for our individual growth curves, we compared a series
of MIXED models by evaluating the difference in deviance between nested models.23 Both fixed
quadratic and cubic MIXED models fit our data well, but we selected the fixed
quadratic MIXED model because the addition
of a cubic
time term was not statistically significant based on a log - likelihood ratio test.
But at the very moment when one considers joined piecewise linear approximations [fig 4], without giving the reasons why the system would change its characteristics at that very
time, perhaps it is more sensible to look for mechanisms (or
effects, if we are unsure
of the mechanisms) that describe the changing rate
of change (leading to
quadratic description) or perhaps even other functional forms: periodic, logistic, cubic.
Because initial RR analyses revealed both
quadratic and linear
effects of time on treatment outcomes, we computed the log
of the number
of days since randomization for each assessment point and used these log values in all RR analyses.