Is it OK to use a mean score for a repeated measure as a predictor in a model? I have a single dependent score for each of my subjects. Most of my independent measures I took multiple times from each subject across multiple trials. In order to compare the independent measure to the dependent measure in a regression model I have compressed the repeated measures into mean scores for each subject. 
I'm aware that I'm losing a lot of variation in the data this way, so I also checked for significant variation between subjects before compressing scores to means. I.e. making sure the means were meaningful representations of each subject's performance. Is is absolutely necessary to do this before using mean scores as predictors in my model? 
I'm aware I could use some kind of multi-level or hierarchical model that includes all the trial-by-trial level measures, but I'm not that versed in those types of models and would prefer to keep it simple if I can avoid using them. 
 A: IMO parameterizations or summary measures, such as the mean of these measurements, are not wrong. As long as you understand they are not capturing all the possible variations, nor the characteristics of the possible underlying trends. 
Take the following example. Imagine you want to quantify the relation between disease severity in hospital patients and their survival, after the patients have been in hospital for, say, 5 days. You have daily measurements of disease severity. We assume/know the more severely ill you are, the higher your risk to die (sounds logical right?). What we do not know however, is whether the total duration of some level of disease severity is what kills you (i.e. a high mean disease severity or sum of the disease severity scores of one patient), or it is having a (single) high peak of illness (i.e. the maximum/highest value during the measurement period, or amount of days above a certain level). Moreover, there are actually a ton of other summary measures to think of which might have a different effect (e.g. mean, max, min, sum, weighted mean, difference between two scores, amount of days above a certain level, etc...). 
What I have seen in previous research is that whichever parameterization you choose indeed does matter. In a study I am working on different parameterizations of the same data were entered in a logistic regression model, and compared based on different measures of model performance. For the c-index for discrimination for example, a weighted mean performed better than the maximum score, but interestingly, just using the first, middle and last measurement without summarizing these into one score, performed even better. (sorry to say this study is still a work in progress and not yet published)
Concluding, it is ok to summarize your repeatedly measured independent variables data, but it does quite matter how you handle it. So if you want to stay away from intricate models which do not require you to 'arbitrarily' summarize/parameterize the data (e.g. Joint Models), I would recommend you to include parameterizations which have some (biological; in the biomedical field that is) plausibility while taking into account which parameterizations are easy to use (a mean weighted in such a way more more current days weigh more heavily might provide better performance than a regular mean, but does require more elaborate processing for the user to grasp).
