The counter argument to this is that many mappers are doing bad science anyway, with "sample sizes" of less than 5 rounds. I'll include myself on that list. Sometimes you learn something in half an hour and you don't want to see people banging on about it for another half before you fix it.
Let me explain why this is a problem with more science. If you were examining a disease, you'd like to do controlled double blind experiments with hundreds of people with the disease and hundreds without. This will average out differences from patient to patient and the results will be more accurate or true to the effect of the disease on the average man (or woman). You want the random factors - old, young, male, female, with other conditions and so on - to be more representative of the whole population of the earth.
But when the sample sizes get really small, you have to throw this randomness out the window. With diseases, this is the cases with really rare diseases where only a handful of people have what you are looking for. In pharmaceuticals, you can't have a larger sample size because literally nobody else suffers from the disease. In mapping, your sample size is tiny because you don't want anyone else to suffer your poorly made alpha map with glaring design flaws. This time, doctors would really want to eliminate all randomness between patients - identical twins are ideal in that they can show whether the disease is genetic or, if only one has the disease, tests can be performed on the both of them to uncover the exact effect of the disease.
Let's take this bloated metaphor and apply it in mapping. Each round is a patient, with random characteristics and showing various symptoms, such as "difficult to push C" in the same way a patient might have restricted breathing. If you have a large sample size, a huge number of patients, because you are testing snowplow or a Valve map or whatever, you want rounds with a huge variety of features - unbalanced teams, scout rushes, 6v6,
critically timed critical hits and so on - so that you can construct an impression of the "average game". Then you can see what symptoms it has e.g. "difficult to push C" and act on it.
Alternatively, if you just want to act quickly, you want your rounds to be as uniform as possible, so you can make the
assumption that they accurately represent the map. You'd like every round to be perfectly balanced, with great team composition and no
surprise factors like random crits so that there are the minimal number of symptoms not caused by the actual map. In other words, we want perfect team composition so that we know "difficult to push C" isn't caused by red having three engineers.
TL;DR:
WITH SCIENCE