(This is a follow-up of a previous question of mine, and similar to this unanswered question)
Scenario Say I have 30 subjects. On each subject, I have a number of categorical and continuous variables, like gender, height, weight, smoker, etc. I am interested in how a number of these variables influence how fast people run. Furthermore, I suspect height is a factor, although my research question is not about height. I let the subjects run on 20 days, giving me, in principle, 20 measurements of my dependent variable for each subject. Maybe some subjects didn't show up some of the time, so I might have some missing measurements.
Question Which statistical method maximizes the statistical power for this scenario (multiple categorical and continuous predictors on subjects, repeated measurement of single dependent variable of interest for each subject, some measurements missing), while ensuring that all necessary controls are done? It would be a plus if the method is straight forward and easy to understand for others (and myself).
Some Thoughts The problem is that the 600 measurements cannot be treated as independent. Per-subject averaging would be a solution, but you would loose statistical power. If you're dealing with a single (or multiple) categorical values, a repeated-measure anova seems to be the best solution (see my previous question). But this doesn't work if you want to control for a continous variable like height right? I read some things about multilevel analysis and panel-analysis, would one of these methods be appropriate in this case? I am afraid I don't really get the difference between those.
Current Solution (probably wrong) At the moment, I test the influences I am interested in separately using a repeated-measure anova. Any continous variables, I convert to categorical values by making a high/low marker. I try to control for height by doing my analysis three times, for all subjects, for 'low-height' subjects and for 'high-hight' subjects. I adjust my alpha for multiple comparison, and I have made sure that my predictors are only mildly correlated (largest r = 0.3), but otherwise do not check for any interactions between them.
