Independent vs Dependent Variables Explained: What Is an Independent Variable?

Say a team ships a checkout test. The new variant has a shorter form, a different button color, and a trust badge under the total. Conversion goes up. Someone asks which of the three did it, and the experiment has no answer to give — it was never built to separate them.

The failure happened long before the statistics, at the point where somebody decided what was being changed and what was being measured. Those two things have names, and getting them straight is most of what separates a test that settles an argument from a test that starts a new one.

What is an independent variable?

The independent variable is the thing you change on purpose. The dependent variable is the thing you measure to see whether the change mattered.

That's the whole distinction. In the checkout test, the independent variable is which version of the page a visitor sees. The dependent variable is whether they complete the purchase. You control the first. You observe the second, and you're hoping it depends on the first — which is where the name comes from and, once you've heard it, is hard to forget.

The trouble is the other name. "Independent" sounds like it means unrelated or standing alone, so people reach for it when describing whichever variable feels more like background. It doesn't mean that. It means independent of the outcome: free to be set by the researcher rather than determined by what happens in the study. If you find yourself deciding which variable is "more independent," you've already left the definition behind. Ask instead: which one did I set, and which one did I record?

The two-column test

When a study is confusing, the fastest way through it is to write two columns and force every variable into one of them.

Independent (you set it)	Dependent (you record it)
Which onboarding flow a user is assigned	Days to first successful export
Font size: 14px or 18px	Reading comprehension score
Search results as list or grid	Time to locate a target item
Number of form fields: 4, 7, or 11	Completion rate

Two things fall out of the exercise. First, the independent variable has levels: values you chose in advance. Four fields, seven, eleven. List or grid. If you can't list the levels, you don't have an independent variable yet.

Second, the dependent variable has to be something you can actually write down as a number or a category. "Whether the experience felt better" isn't a dependent variable. "Rating on a 1–7 ease-of-use item, collected immediately after the task" is. Turning a fuzzy intention into a recordable measure is called operationalizing it. NN/g's overview of quantitative research methodologies treats this as the defining move of quantitative work: the measure has to be fixed before collection, not chosen afterward from whatever the data happened to offer.

What is an independent variable when you didn't change anything?

Often, nothing you're studying was set by you at all. You compare users on mobile against users on desktop. Or new accounts against accounts older than a year. Nobody assigned anyone to those groups.

In that situation the honest answer to what is an independent variable here is: there isn't one. You have a predictor variable, something that varies across your data and may relate to the outcome, but you didn't manipulate it, so you can't claim it caused anything. Mobile users convert worse. Fine. Mobile users are also more often first-time visitors, browsing on transit, arriving from a different ad set. The device didn't necessarily cause the drop. It might just be the most visible thing about a group that differs in a dozen ways.

This is the line between experimental and correlational work, and product settings blur it constantly, because a dashboard hands you group comparisons that look exactly like experimental results. Deciding which of the two you're holding is a methods question, and the broader map of UX research methods is largely a set of answers to it. Manipulation plus random assignment is what buys the causal claim.

Everything else you're not measuring

Between the variable you set and the variable you record sits everything else about the situation, and mixing up two kinds of it is expensive.

Control variables are the things you deliberately hold still. Same task wording, same starting screen, same time of day, same device. They aren't interesting in themselves; they're interesting because holding them constant means they can't explain your result.

Confounding variables are the ones that moved when your independent variable moved, and now you can't tell them apart. This is the checkout test from the opening. Form length, button color, and the trust badge all changed together, so the lift belongs to all three jointly and to none of them individually. The experiment answers a much vaguer question than the team thought they were asking: does this bundle beat that bundle?

Sometimes a bundle is genuinely what you want to test. You're shipping a redesign. Isolating every factor would cost more than the answer is worth, and that's a legitimate call. The mistake is making it by accident and then reading the result as though you'd isolated something.

If you do want per-factor answers, the tool is a factorial design: two independent variables varied at once, in every combination. Short form with badge, short form without, long form with, long form without. That buys you what each factor does on its own plus whether they interact — whether the badge helps on the long form and does nothing on the short one. It also doubles your cells, and every cell needs enough participants to say anything about it. Budget accordingly.

Where it usually breaks

Four failure patterns show up repeatedly, and all of them are definition problems wearing statistical clothes.

Measuring twelve things and reporting the one that moved. Run enough dependent variables against one independent variable and something will clear significance by chance alone — the multiple-comparisons problem, and the reason NN/g argues for putting A/B testing in its place rather than treating a moving number as a finding. Name the primary dependent variable before you collect data, and label the rest exploratory when you report them.

Changing the independent variable mid-flight. Adjust the variant while the test runs and the group you measured in week one saw something different from the group you measured in week three. There is no coherent level left to attribute the result to.

Letting assignment correlate with something else. If the new flow went to users who signed up after the pricing change, the flow and the pricing cohort are welded together. Random assignment exists to prevent exactly this, and it is the first thing traded away when groups get split by cohort or by ship date.

A dependent variable that drifts on its own. Conversion rate has a seasonal shape. Compare November against January and the calendar is doing work you'll credit to the design. A concurrent control group is the only fix here; a before/after comparison inherits the calendar.

The questions worth asking before you run anything

Four, in order, and none of them require a statistician:

1. What exactly am I changing, and what are its levels?
2. What exactly am I recording, in what unit, at what moment?
3. What else differs between my groups, and did I hold it still or just hope?
4. If the number moves, what's the second-most-likely explanation after my hypothesis?

Question four is the one that earns its keep. There is always a second explanation. Naming it before the data arrives is also, in practice, when someone notices that the independent variable was never as clean as the plan said.

Frequently asked questions

What is an independent variable in an A/B test?

The variant assignment itself: which version of the page or flow a given user is served. Its levels are A and B, or A, B and C. Everything else about the two experiences should be identical, and anything that isn't is a confound riding along with your independent variable (A/B testing 101).

Which variable goes on which axis?

Independent on the x-axis, dependent on the y-axis. The convention matches the reading — the outcome plotted as a function of the thing you set — and it holds across nearly every field, so a reader can orient on your chart before reading the labels.

Can a study have more than one independent variable?

Yes, and it's often the better design. Varying two factors across all their combinations tells you what each one does and whether they interact. The cost is sample size: each additional cell needs enough participants to say anything about it.

Is a dependent variable always a number?

No. It can be categorical — completed the task or didn't, chose plan A or plan B. What it can't be is undefined at collection time. If two people watching the same session would record different values, the variable isn't specified tightly enough yet.

What's the difference between a control variable and a constant?

A control variable could have varied and you chose to hold it fixed — screen size, task order, time of day. A constant never varies in your context at all. The distinction matters when you write up limitations: a held-fixed control variable is also a boundary on how far your result generalizes.