I have classic block design of the experiment. There are blocks and treatments, and several observation within. The experiment is unbalanced thus I want to use mix models to analyze if the treatment has an effect on several parameters. I plan to have block as a random factor and treatment as a fix. The thing I do not get is: Shall I use all the data in my data set in the model or shall I calculate means of treatments per block and then use the data in the analysis. I tried both options in R (lmer) and I received slightly different results in estimation of parameters and different variance (which is obviously not surprising). May I get an explanation which approach shall I use in this case with mean or with raw data? Maybe I should specify the model in a way that it will account on the several measurements per block and treatment? Best regards
If 'observations within' are independent experimental units to which treatment level was randomly assigned, and applied, then use all data. If those observations come from the same experimental unit (e.g. scores from several students in a class with the same teacher, with teacher being the trt, or height of several plants in a pot with the same fertilizer trt) use averages, or model them as subsamples. But you'd still need several observations per treatment combination.