A person holding a stack of survey results faces a problem before they face any question about what the results mean. The stack is not organized. Some items in it speak directly to whatever the person wants to know. Some speak to it sideways. Some do not speak to it at all, even though they concern the same general subject. Before any of it can be used, it has to be sorted, and how it gets sorted determines what conclusions are available afterward.

The sorting categories are called buckets, and choosing them is the real work. A bucket is not a topic. It is a kind of evidence. Two items can both be about school attendance and still belong in different buckets, because one asks a student how often they come to school and the other asks a teacher how often that student comes to school. Same subject, different kind of evidence, different bucket.

This matters because the buckets are what make disagreement visible. If every item goes into one undifferentiated pile labeled "attendance," and the pile contains both student answers and teacher answers, then a researcher who summarizes the pile will produce an average that hides the fact that the two groups said different things. Sorted into separate buckets, the disagreement becomes the finding rather than being smoothed away.

Every good bucket system needs one bucket that most people forget to build. Some items in the stack will not answer the question at all. They are related, they are interesting, and they are irrelevant to the specific thing being asked. An item recording which brand of shoes students wear is genuinely about students and genuinely useless for a question about attendance. That item needs somewhere to go, and the somewhere cannot be a bucket that implies it counts as evidence. Researchers who do not build a reject bucket end up quietly stretching a good bucket to hold something that does not belong in it, and the stretch is invisible in the final report.

Deciding an item does not answer the question is a judgment, and it is a judgment that has to be defended. The defense usually comes down to a difference between what the item asked and what the question asks. A question about how often something happens is not answered by an item about whether it happens at all. A question about how much is not answered by an item about which. Naming that mismatch out loud is the difference between rejecting an item and dismissing it.

So far, everything described here counts one item at a time. A researcher looks at an item, reads the counts under each answer, and decides what that item shows. This is called univariate work, from a root meaning one variable. Almost all reporting on surveys is univariate. Sixty-two percent said this. Forty said that. Each number describes the answers to a single question.

Univariate work is not a lesser form of analysis. It is the foundation, and most published survey findings never go beyond it. But it has a specific limit that is worth naming early. A univariate count tells you how many people gave each answer. It cannot tell you anything about which people. Two items counted separately show two distributions and nothing about the relationship between them, because counting one item at a time discards the information about who said what.

Once items are sorted and counted, the last step is to say something. That statement is a claim, and a claim is not the same thing as a guess or an opinion. A claim is a statement that specific evidence supports and that could be shown wrong by other evidence. Both halves matter. A statement nothing supports is a guess. A statement nothing could contradict is not a claim either, because a statement that survives every possible piece of evidence is not saying anything about the world.

A well-built claim carries its evidence with it. Instead of asserting that attendance is a problem, it names which items support that and what those items actually asked. Doing so makes the claim checkable, which is the entire point. A reader who disagrees can go look at the same items and argue about them. A reader facing an unsupported assertion can only agree or refuse.

The order is fixed and it does not reverse. Sort first, count second, claim third. A researcher who forms the claim first will sort the evidence toward it without noticing, and will build the buckets that produce the answer they already have. This is not usually dishonesty. It is what happens when the conclusion arrives before the categories do.