Same subjects or different subjects?
Are you comparing the same patients at two time points? Or two different groups of patients? The answer determines which statistical test is valid.
Use the wrong one and your p-values are meaningless.
This is one of the most common mistakes in clinical research: using an independent samples test when the data are paired, or vice versa.
The good news? Once you understand the concept, it's easy to get right.
Every comparison falls into one of two categories.
Each observation in one group has a specific, matched observation in the other group. The measurements are linked.
The observations in one group have no specific connection to observations in the other group. Different people entirely.
Paired and independent data require different statistical tests.
Paired data has built-in correlation. The same person's pre-op value is related to their post-op value. Patient A's high baseline means they'll probably have a high follow-up too.
Paired tests account for this correlation. Independent tests don't. Use the wrong one and you'll get the wrong answer.
Using an independent t-test on pre/post data. This ignores the pairing and usually gives you LESS statistical power. You're more likely to miss a real effect.
Paired tests are often more powerful because each subject serves as their own control. You remove between-subject variability from the equation.
Some situations aren't immediately obvious. Here's how to think through them.
"We measured patients at baseline, 6 months, and 12 months. Paired or independent?"
Paired (repeated measures). Same patients at multiple time points. But with 3+ time points, you need repeated measures ANOVA or a mixed model, not just a paired t-test.
"We compared the left and right leg in each patient. Paired or independent?"
Paired. Even though it's two different legs, they belong to the same person. Left and right measurements from the same patient are correlated.
"We matched cases to controls by age and sex. Paired or independent?"
Paired. Matched case-control studies create pairs. Each case has a specific matched control. Use McNemar's test for binary outcomes, not chi-square.
"Patients chose whether to get surgery or medical management. Paired or independent?"
Independent. Different patients in each group. Even though they're from the same population, there's no 1-to-1 matching. (Note: this design has selection bias issues, but that's a different lesson.)
"We randomized patients to Drug A vs Drug B. Paired or independent?"
Independent. Classic RCT design. Different patients in each arm. No pairing unless you specifically matched patients before randomization.
Ask yourself: "Can I draw a line connecting each observation in Group A to exactly one observation in Group B?"
If yes → Paired. If no → Independent.
For each scenario, identify whether the data are paired or independent.
How to identify whether your data are paired or independent. This determines which version of each statistical test you should use.
Next lesson: Number of Groups. Are you comparing two groups or more than two?