Six moves, always in this sequence
Analysis feels arbitrary because most people meet it in the middle, at the software. It is not arbitrary. Every quantitative result any candidate has ever defended was built by the same six moves, and doing them out of order is what produces the panic.
- ◆State the question as a comparison, a relationship or a description. Anything that cannot be phrased as one of those three is not yet a question a number can answer.
- ◆List every variable and say what kind of number it is. This step takes ten minutes and decides almost everything downstream.
- ◆Write the design down: who is measured, how many times, and in how many groups.
- ◆Choose the test, which by now the first three steps have chosen for you.
- ◆Check the assumptions that test carries, and record what you found rather than what you hoped.
- ◆Read the output in a fixed order and convert it into one sentence a committee member can quote.
The variables pick the test, not the researcher
This is the part candidates believe is a matter of preference, and it is not. The kind of number your outcome is, plus the shape of your design, narrows the field to one or two options long before software is opened. Four questions settle it, and none of them requires arithmetic.
First, are you describing, comparing or relating? Second, how many groups are involved, and are they different people or the same people measured more than once? Two independent groups, one group measured twice, and three or more groups are three different procedures, and confusing them is the most common error the desk sees in a fourth chapter. Third, what kind of outcome is it: a measured quantity, an ordered category, or a count of people falling into boxes? Fourth, does something else have to be held constant while you look? The moment the answer to that is yes, you have moved from comparison into regression, and you have done so because of the question rather than because of the software.
What to read in the output, and in what order
Whatever produced it, SPSS or R or Python or a well-built sheet in Excel, the reading order is the same and it does not begin where most candidates begin. Start with the direction and the size of the difference or relationship, in the units of the thing you measured. Then the interval around that estimate. Then the number of cases the estimate rests on. Only then the p-value.
That last item carries more mythology than the rest of the chapter combined, so be exact about it. A p-value is the probability of seeing a result at least this extreme supposing there were genuinely nothing there. That is all it is. It does not tell you how likely your hypothesis is to be correct, it does not measure how much the finding matters, and the threshold most fields use is a convention those fields adopted rather than a law of nature. A confidence interval is usually the more informative line: it shows the range of values your data are compatible with, and a very wide one is an honest admission that the study was small.
Assumptions are a checklist, not a verdict
Every test carries conditions: independence of observations, a distribution that behaves reasonably, comparable spread between groups, a relationship that is roughly straight where the method assumes a straight one. Candidates treat a failed check as a catastrophe. It is a fork in the road, and all three branches are respectable.
You may transform the variable and say so. You may switch to the equivalent procedure that does not require the condition, which exists for nearly every common test. Or you may proceed, report the violation openly, and say which direction it is likely to have pushed the result. What you may never do is quietly delete the cases that spoil the check. Removing awkward observations without a stated rule is the one act in this chapter that a committee is entitled to treat as a finding about you rather than about your data, and it is the sort of thing that surfaces long after a quiet stretch of supervision has ended.
One sentence, and then the chapter writes itself
Every result becomes the same sentence: among this many participants, this group showed this much more or less of the outcome than that group, with this interval around the estimate, tested against the threshold set in the proposal. Fill that in for each hypothesis, in the order the hypotheses were stated, and you have the spine of a results chapter. Interpretation, which is a different act, comes afterward and belongs in the next chapter.
Two habits protect all of it. Keep a log of every decision, including the ones you reversed, because a defense is largely a conversation about choices rather than about numbers. And run the analysis twice, on different days, from the raw file. Where a candidate reaches this point with a deadline already tight and a design that has changed twice, the desk usually recommends a statistician sitting inside the project rather than a tutorial, since the difficulty by then is the design rather than the arithmetic.
Questions put to the house
Do I need to learn SPSS, R and Python?
No. Learn the reasoning once and one tool well. The reasoning transfers between them completely, because they run the same procedures on the same assumptions and disagree only about menus and syntax. Candidates who spend a term collecting software rarely gain anything a committee can see, whereas one clean, repeatable script or output file is defensible in any room.
What does a non-significant result mean for my study?
That your data did not provide strong evidence against the null hypothesis, which is not the same as showing nothing is there. Small samples produce wide intervals and wide intervals hide real effects. Report the estimate and its interval, discuss what size of effect your study could realistically have detected, and treat the finding as information rather than as failure.
How large does my sample need to be?
It is decided before collection, not after, by the smallest effect that would matter in practice, the variability you expect, and the certainty you want. That calculation belongs in the proposal and is one of the first things a committee will look for. If data are already collected, report what the study can detect rather than adjusting the target to match what arrived.
Can somebody else run my analysis for me?
A statistician may run and explain it, and in most doctoral programs that is ordinary practice rather than a concession. What must remain yours is the understanding: the design, why each test was chosen, what the assumptions were, and what the numbers mean for practice. You will be asked all four in a defense, by people who know the answers.
