The prior examples have assumed one line per unique subject/variable combination. This is not a typical way to enter data. A more typical way (found ., in Systat) is to have one row/subject. We need to "stack" the data to go from the standard input to the form preferred by the analysis of variance. Consider the following analyses of 27 subjects doing a memory study of the effect on recall of two presentation rates and two recall intervals. Each subject has two replications per condition. The first 8 columns are the raw data, the last 4 columns collapse across replications. The data are found in a file on the personality project server.
Kempthorne uses the randomization-distribution and the assumption of unit treatment additivity to produce a derived linear model , very similar to the textbook model discussed previously.  The test statistics of this derived linear model are closely approximated by the test statistics of an appropriate normal linear model, according to approximation theorems and simulation studies.  However, there are differences. For example, the randomization-based analysis results in a small but (strictly) negative correlation between the observations.   In the randomization-based analysis, there is no assumption of a normal distribution and certainly no assumption of independence . On the contrary, the observations are dependent !
Let’s make up a little story: let’s say we have three types of wine (A, B and C), and we would like to know which one is the best one (in a scale of 1 to 7). We asked 22 friends to taste each of the three wines (in a blind fold fashion), and then to give a grade of 1 till 7 (for example sake, let’s say we asked them to rate the wines 5 times each, and then averaged their results to give a number for a persons preference for each wine. This number which is now an average of several numbers, will not necessarily be an integer).