From Question to Hypothesis
Every study starts as a hunch: "I bet music helps people study." That's a fine question, but it isn't yet something you can test. Research design begins by sharpening a fuzzy idea into a prediction so specific that the data could prove it wrong.
The ladder from idea to prediction
A good hypothesis is built in stages. Each rung makes the claim more concrete, and more falsifiable.
- Research question. The broad curiosity: "Does music affect studying?"
- Conceptual hypothesis. A proposed relationship between ideas (constructs): "Background music influences learning."
- Testable prediction. The same claim pinned to a specific population, a manipulable cause, a measurable outcome, and a direction: "Among first-year students, studying with lyrical pop (vs. silence) lowers recall on a 20-item word test."
Falsifiability comes first. A hypothesis is scientific only if there's some result that would show it's wrong. "Music affects people somehow" can never be refuted (every outcome fits it), so it's useless. "Music lowers recall" is testable, because a study that finds no drop in recall would refute it.
🛠️ The Hypothesis Machine
Pick a vague claim, then sharpen each part. The falsifiability meter climbs as your prediction gets specific enough for a result to contradict it.
Your prediction so far
H₀
H₁
Directional or not?
Once you pick a direction, you're choosing between two flavors of alternative hypothesis:
- Non-directional (two-tailed): "music changes recall" — you predict a difference but not which way. Safer when you don't know the direction.
- Directional (one-tailed): "music lowers recall", a bolder, riskier bet that puts all your statistical power on one side. Justified only when theory or prior evidence points clearly one way.
Where H₀ and H₁ come from
Your prediction becomes the alternative hypothesis (H₁). Its mirror image, "nothing is going on, no effect, no difference," is the null hypothesis (H₀). You never test H₁ directly; instead you assume H₀ and ask how surprising your data would be under it. How that test works is the subject of hypothesis testing logic. It can only be run on a prediction sharp enough to be falsified.
Why it matters: a study is only as good as the prediction it started from. Fuzzy questions produce fuzzy results that can be spun to fit any story. Nail down the population, the variables, and the direction before you collect data, and you've already done the hardest part of the work. Preregistration records that prediction where you cannot later revise it. A preregistration also has to name the analysis, and the test chooser helps you pick one.
The complete worked project runs the whole sequence once, from a question this vague to the results paragraph. Once your own prediction names an outcome and a comparison, Plan My Analysis suggests a test and a sample size for it. Choosing Statistics for Your Dissertation covers the same decision in prose.
Common questions
My hypothesis has three parts. Is that one hypothesis or three?
Count the tests. "Music lowers recall, slows reading, and raises reported stress" is three predictions about one subject. Each will get its own statistical test, its own chance of a false positive, and its own line in your results, and packing them into one sentence doesn't change that. Decide in advance which ones are the confirmatory tests you are willing to be judged on, write that down in your preregistration, and correct across that set the way multiple-comparison corrections describe. A compound prediction is perfectly good science, as long as you say in advance which parts must hold for it to count as supported.
What makes a hypothesis falsifiable?
There has to be some possible result that would count as evidence against it. "Music changes recall" is falsifiable: a clear no-difference result contradicts it. "Music affects people somehow" is not, because any outcome at all can be squeezed to fit, so it can never be wrong and therefore never informative. Falsifiability, following Popper, is the line between a scientific claim and an empty one.
Should I use a one-tailed or a two-tailed hypothesis?
Default to two-tailed (non-directional) unless strong theory or prior evidence justifies predicting the direction. A one-tailed test is more powerful if you guessed the direction correctly, but it cannot detect a real effect in the opposite direction. Switching to one-tailed after peeking at the data is a form of p-hacking. Whichever you choose, choose it before you collect data.