Comparing two proportions asks whether two populations have different rates of a characteristic. The estimate is a difference of sample proportions. Confidence intervals and tests use related formulas, but their standard errors differ because they answer different questions.
Define the difference and preserve its order
Let and be the population proportions, and estimate with Define which group is first before calculating. Reversing the order reverses the sign of the estimate and interval.
The groups must be independent for the usual two-sample procedure. If the same people provide before-and-after binary responses, the measurements are paired and this independent-samples method is not appropriate.
Check random sampling or random assignment as relevant, independence within groups, and the 10% condition for samples without replacement. Random assignment supports treatment causation; random sampling supports generalization to populations.
Confidence intervals use separate estimates
A large-sample confidence interval is
The two variance contributions add because the sample estimates are independent. For an interval, check at least 10 observed successes and failures in each group. Do not pool the estimates: the interval is designed to allow the two true proportions to differ.
Example: estimate a difference
In independent random samples, 120 of 200 people in group 1 and 90 of 200 in group 2 favor a policy. The estimates are 0.60 and 0.45, so the observed difference is 0.15, or 15 percentage points.
The unpooled standard error is A 95% interval is approximately
We are 95% confident that the population support proportion in group 1 exceeds that in group 2 by between about 5.3 and 24.7 percentage points. This is an absolute difference in proportions, not a relative percent increase.
An equality test pools under the null
To test the null asserts a common proportion. Estimate it by pooling successes and sample sizes:
The test statistic is
Use the pooled estimate to check expected successes and failures in each group under the null. The pooling step belongs to the equality test, not to the ordinary confidence interval. A test of a nonzero hypothesized difference requires a procedure appropriate to that null rather than blindly using this equality formula.
Example: test the same data
For the policy samples, The pooled standard error is approximately 0.04994, giving For a two-sided alternative, the P-value is about 0.0027.
At 5%, reject the equality null. The data provide strong evidence of a difference in population support rates, subject to the sampling conditions. The positive estimate indicates the observed direction favors group 1; the two-sided test allowed either direction.
The P-value is not the probability that the groups truly have equal proportions. It describes the extremeness of the sample difference under the equality model.
Interpret design and practical size
If groups are observational categories, a statistically significant difference does not by itself establish causation. Other characteristics may explain the association. In a randomized experiment, a treatment-effect interpretation can be justified by the design, with generalization limited by recruitment.
Also report the estimated difference and interval. A tiny difference can be statistically significant in a huge sample, while a practically important difference may be uncertain in a small one.
Try it yourself
Two groups have estimates 0.30 and 0.40. If an interval for is interpret its direction and rewrite it for
Answer: Group 1's population proportion is estimated to be 2 to 18 percentage points lower. Reversing the difference negates and reverses the endpoints, giving
What to remember
Define group order, verify independence and counts, use unpooled variability for intervals, and pool only for the usual equality test. Interpret differences in percentage points and respect the study design.