What a Mean-Difference Interval Says
A confidence interval for a difference of population means is useful only when its subtraction order and units are clear. If an interval is for “Group A minus Group B,” positive values indicate a larger population mean for A; negative values indicate a larger mean for B. The interval then describes effect sizes compatible with the data and the method’s assumptions, not just whether a calculation has crossed zero.
For two populations, define the parameter as \(\mu_A-\mu_B\). It is a fixed, unknown number: the difference between the two population means. The order matters. Reversing it gives \(\mu_B-\mu_A=-(\mu_A-\mu_B)\), so the interval’s endpoints change sign and reverse order as well.
Each number in the interval represents a possible value for the population mean difference under the inferential procedure. Interpret the numbers in the original measurement units. For example, an interval from 2 to 7 minutes for \(\mu_A-\mu_B\) represents plausible average differences of 2 to 7 minutes, with Group A’s population mean higher. It does not say that every observation in A exceeds every observation in B.
How to Read Direction, Size, and Zero
Start by restating the parameter in context, including the order of subtraction. Then describe the interval’s range and units. Finally, compare the interval with zero. Zero is the value that would mean equal population means, because \(\mu_A-\mu_B=0\) exactly when \(\mu_A=\mu_B\).
| Interval for \(\mu_A-\mu_B\) | Interpretation |
|---|---|
| Entirely positive | The interval supports a higher population mean for A, with the size of the difference ranging across the interval. |
| Entirely negative | The interval supports a lower population mean for A, or equivalently a higher population mean for B. |
| Contains zero | Equal population means are among the values not ruled out by the interval; values on either side may also be compatible with the data. |
If a two-sided 95% interval excludes zero, that corresponds to rejecting the null value \(\mu_A-\mu_B=0\) in a matching two-sided test at the 5% level, provided the same model and method are used. If it contains zero, the data do not establish a difference at that level. Neither result proves equality or inequality: the interval is about what values remain compatible with the data and assumptions.
A confidence interval also does not measure practical importance by itself. A narrow interval can exclude zero but contain only small effects. A wider interval can include both negligible and substantial effects. If a decision depends on a practically meaningful threshold, compare the whole interval with that threshold rather than relying only on whether it contains zero.
Why the Parameter Is Not Random
The conventional frequentist interpretation assigns probability to the interval procedure before the sample is observed. In repeated samples taken under the same model, a 95% confidence procedure produces intervals that cover the fixed parameter in 95% of repetitions in the long run. After one particular interval has been computed, its endpoints are fixed too. The parameter is not assigned a 95% probability of lying between those observed endpoints.
The central 95% probability statement for a standard normal variable gives a concise justification. Suppose independent samples come from normal populations, the population variances are known, and the standard error is therefore \(SE=\sqrt{\sigma_A^2/n_A+\sigma_B^2/n_B}\). Then the standardized difference of sample means has a standard normal distribution.
Rearranging the event between the two critical values gives an interval for the fixed mean difference:
Before observing the data, the sample means—and hence the interval endpoints—vary across repetitions, so the coverage probability describes the procedure. Once the data are observed, the interval is fixed. For t intervals, the corresponding procedure uses a t statistic and an appropriate critical value under its assumptions; the interpretation of confidence remains a repeated-sampling one.
Worked Interpretations
Worked Example: A Positive Interval for Two Independent Means
A hypothetical operations team compares the mean time, in minutes, needed to complete a task under Process A and Process B. The samples are independent, and the two population distributions are assumed normal with unequal, unknown variances. For A, \(n_A=36\), \(\bar x_A=42.0\), and \(s_A=6.0\); for B, \(n_B=64\), \(\bar x_B=37.2\), and \(s_B=8.0\). Construct and interpret a 95% Welch interval for \(\mu_A-\mu_B\). This is not a cash-flow or life-contingency calculation, so no interest rate or mortality assumption is used; the assumed distributions and sample summaries are stated above.
Method. Use the Welch two-sample t interval, which does not require equal population variances. The estimated difference is
The estimated standard error is
The Welch degrees-of-freedom estimate is
For a 95% two-sided interval with 90 degrees of freedom, use \(t_{0.975,90}\approx1.9867\). The margin of error is \(1.9867(1.4142)\approx2.8096\), so the interval is
Interpretation. The interval estimates that Process A’s population mean task time is between 1.99 and 7.61 minutes greater than Process B’s. All values are positive, so zero is excluded: under the stated model and method, the interval supports a higher mean time for A. The interval does not say that every task takes longer under A, nor does it establish that the difference is important for an operational decision.
Common mistake avoided. Reporting only “there is a difference” discards the size and direction information. The interval gives both: a positive direction and a range of plausible mean differences in minutes.
Worked Example: A Paired Interval and the Meaning of Its Sign
A hypothetical team records each of 16 employees’ task times before and after a revised workflow. Define each paired difference as \(D=\text{before time}-\text{after time}\), in minutes. Assume the 16 differences are an independent random sample from a normal population, with \(\bar d=3.0\) minutes and \(s_D=4.0\) minutes. Find and interpret a 95% paired t interval for the population mean difference \(\mu_D\). No interest rate or mortality assumption applies; the difference distribution is normal with unknown mean \(\mu_D\) and unknown variance \(\sigma_D^2\).
Method. Treat the paired differences as one sample and use a one-sample t interval. There are \(n=16\) differences, so the degrees of freedom are \(16-1=15\). The standard error is
The 95% critical value is \(t_{0.975,15}\approx2.131\). Thus the margin of error is \(2.131(1.0)=2.131\), and the interval is
Interpretation. Because the difference was defined as before minus after, a positive value represents a shorter time after the workflow change. The interval estimates an average reduction of 0.87 to 5.13 minutes. It excludes zero, so the data support a positive mean before-minus-after difference under the model.
Common mistake avoided. If the difference had instead been defined as after minus before, the estimate and interval would have been \(-3.0\) and approximately \([-5.13,-0.87]\). The substantive conclusion would be the same, but the sign and wording would reverse. Always check the defined order before describing direction.
Worked Example: An Interval That Contains Zero
A hypothetical laboratory compares the mean measurement under Method A with that under Method B. Independent samples each contain 25 observations. Assume normal populations with known standard deviations \(\sigma_A=4\) units and \(\sigma_B=3\) units. The sample means are \(\bar x_A=18.4\) and \(\bar x_B=19.4\). Construct and interpret a 95% z interval for \(\mu_A-\mu_B\). No interest rate or mortality assumption is relevant; the distributions are normal with means \(\mu_A,\mu_B\) and the stated known standard deviations.
Method. Since the population standard deviations are known, use the independent-sample z interval. The estimated difference is
The standard error and margin of error are
Therefore the 95% interval is
Interpretation. The interval includes zero and values on both sides. It is compatible with Method A having a population mean up to 2.96 units lower than Method B, equal means, or a population mean up to 0.96 units higher. The sample estimate is negative, but the interval does not establish that A’s population mean is lower.
Common mistake avoided. The sample difference of \(-1.0\) is not enough to conclude that A has a lower population mean. Directional interpretation must account for the uncertainty represented by the interval. Containing zero also does not prove that the means are equal.
Practical Importance and Careful Conclusions
A useful interpretation separates three questions: what is the estimated direction, what effect sizes are compatible with the interval, and whether those effect sizes matter in context. For example, an interval entirely above zero can show evidence of a positive difference while still including values too small to matter operationally. Conversely, an interval containing zero may still include effects large enough to matter, indicating that the data are inconclusive about the direction or size rather than showing that the effect is negligible.
The interval also does not by itself establish causation. A causal conclusion depends on how the groups or conditions were assigned and on the study design, not merely on whether zero lies in the interval. Keep the population, measurement units, subtraction order, and assumptions attached to the conclusion.
Check Your Understanding
For each problem, identify the direction and size of effects represented by the interval, and explain what zero implies. Values are hypothetical.
- Independent normal populations have known standard deviations \(\sigma_A=4\) and \(\sigma_B=3\). Two samples each have size 16, with means 12.5 and 10.5. Construct a 90% z interval for \(\mu_A-\mu_B\), using \(z_{0.95}=1.645\), and interpret it. (Answer: \(SE=1.25\); interval \([{-0.06},\,4.06]\) units to two decimals; zero is included, so the direction is not established.)
- For 25 paired measurements, the defined difference is after minus before. Assume normal differences, with \(\bar d=-2.4\) and \(s_D=3.0\). Construct a 95% interval using \(t_{0.975,24}=2.064\) and state the interpretation. (Answer: \(SE=0.60\), margin \(1.2384\); interval \([-3.64,\,-1.16]\) to two decimals, supporting a mean decrease of 1.16 to 3.64 units after the change.)
- A 95% Welch interval for \(\mu_A-\mu_B\), with degrees of freedom treated as 77 and critical value 1.99, uses independent samples with \(n_A=25\), \(s_A=5\), \(n_B=100\), and \(s_B=10\). The sample means are 31 and 26. Calculate and interpret the interval. (Answer: \(SE=\sqrt{25/25+100/100}=\sqrt{2}\approx1.4142\); interval \(5\pm1.99(1.4142)\approx[2.19,\,7.81]\) units, entirely positive.)
- A study reports a 95% interval of \([1.2,4.8]\) for \(\mu_A-\mu_B\). A difference of at least 2 units is considered practically important in the positive direction. What can be concluded about direction and practical importance? (Answer: The interval excludes zero and supports a positive difference, but it includes values below 2; therefore, it does not establish that the difference is at least practically important.)