Hello. Thanks a lot for the explanation, it was so clear and useful. I have a question: I didn't get the point why for the proc mixed using time as a repeated measurement, you classified islt as a subject and not a model, can you please clarify when should we chose it as a subject, model or group? And especially for time, because I always thought a subject should be the subject from the experiment. Thanks in advance.
Great question! Where there is a time variable, it should ALWAYS be modeled. The question is how? How time should be factored in to a model (fixed, random or repeated) depends on the goal of the modeler. In the video we considered time as a nuisance factor and therefore put it in the Repeated statement. This reflects the common situation where repeated measurements are necessary to achieve precision of the estimate, but those repetitions are not considered individually important. If understanding an effect at a particular date or time is important, it should be included as a fixed effect in the Model statement. For example, mixed modelers often model a time by treatment interaction to see if the treatment effect changes at different time points. This is parameterized as: Model response= treatment time treatment*time; While it's less common to include time as a factor in the random statement, it could be a valid intermediate scenario between including it as a fixed effect in the model statement or a residual effect in the repeated statement.
Many thanks for the detailed explanation that is very useful, however, I keep wondering about the classification made #during minute 16:00 of this video# regarding the #year# that has been classified as a repeated measurement but as a subject effect and not as a model effect, and this is the basis for my question: according to what we decide if the effect should be subject or model for repeated measurements.
Thanks for the clarification and great question. Valid scenarios could have been 1) to include Year either as “Repeated Year,” or 2) as “Repeated / group=year” (with Year also in the Class statement). In the first scenario we would assume residuals to be independent and of similar variance among the years. By using scenario 2 in the video we requested residuals be estimated within each year. Doing that relaxes the assumption of common variance (homoscedasticity) among all years, where we allow each year’s residual distribution to differ from the other years. In effect we’re saying “a year is a long time for this biological system, and its variance components should change year to year”. It's often hard to know in advance whether a time component will have equal variance among its time points. So, trying both scenarios and evaluating the characteristics of the residual plots (e.g. Pearson panel) and the Fit Statistics (e.g. AIC) can help determine which of the two scenarios is best. I hope this answers your question.
