Why the Subtraction Order Matters
Before-and-after measurements are made on the same individuals, so the quantity of interest is usually their average change. To estimate that change, first calculate one difference for each individual. The subtraction order determines what a positive or negative difference means, and it must be stated before the interval is calculated.
For example, if a score rises from 72 before an intervention to 76 after it, defining the difference as after minus before gives \(4\) points. Defining it as before minus after gives \(-4\) points. Both are valid conventions, but the interval’s signs and interpretation must match the chosen convention.
Let \(\bar D\) be the sample mean of the differences and \(s_D\) their sample standard deviation. For a confidence level of \(100(1-\alpha)\%\), the paired t interval is
where \(t_{1-\alpha/2,\ n-1}\) is the Student t critical value with \(n-1\) degrees of freedom. Equivalently, the interval endpoints are \(\bar D-t_{1-\alpha/2,\ n-1}s_D/\sqrt n\) and \(\bar D+t_{1-\alpha/2,\ n-1}s_D/\sqrt n\).
Model and Conditions
The paired procedure reduces the data to a sample of \(n\) differences. For an exact t interval, assume the pairs are independently sampled from the population of interest and that the differences \(D_1,\ldots,D_n\) are normally distributed with mean \(\mu_D\) and variance \(\sigma_D^2\). The before and after measurements within a pair need not be independent; they are paired precisely because they belong to the same individual. It is the distribution of the differences that matters for the exact interval.
For a large sample, the procedure is often useful as an approximation when the differences are not extremely skewed and do not contain influential outliers. A small sample with strongly nonnormal differences does not meet the exact normal-difference condition; in that case, the stated exact t confidence level is not guaranteed.
The interval is calculated from the difference scores alone. It does not require separate standard errors for the before and after measurements, and it is not constructed by treating the two columns as unrelated samples.
Worked Example: Estimating an Average Increase in Test Scores
A hypothetical training program records scores for five participants before and after training. Scores are measured in points. Use after minus before, and calculate a 95% paired t confidence interval for the population mean change. Assume the five pairs are independent and the differences are normally distributed.
Method and differences. For each participant, subtract the before score from the after score:
The difference convention is after minus before, so a positive value represents an increase. There are \(n=5\) pairs and \(n-1=4\) degrees of freedom.
Sample mean. The sum of the differences is \(2+4+3+5+1=15\), so \(\bar D=15/5=3\) points.
Sample standard deviation and standard error. The deviations from the mean are \(-1,1,0,2,-2\), whose squared values sum to \(1+1+0+4+4=10\). Therefore,
Critical value and interval. For 95% confidence with 4 degrees of freedom, \(t_{0.975,4}=2.776\). The margin of error is \(2.776(0.7071)=1.9629\). Thus the endpoints are \(3-1.9629=1.0371\) and \(3+1.9629=4.9629\). Rounded to two decimal places, the interval is \((1.04,4.96)\) points.
Interpretation and check. Under the stated model, the interval estimates the population mean increase, because the differences were defined as after minus before. Both endpoints are positive, consistent with an estimated average increase. The interval is centered at \(3\), since \((1.04+4.96)/2=3.00\). A common mistake would be to describe this as an interval for the mean score after training; it is instead an interval for the mean of the paired changes.
Why the Paired t Formula Works
The justification comes from applying the one-sample t result to the difference scores. Under the model, the \(D_i\) are independent normal observations with mean \(\mu_D\) and variance \(\sigma_D^2\). Consequently, \(\bar D\) is normal with mean \(\mu_D\) and variance \(\sigma_D^2/n\), while the sample variance \(s_D^2\) estimates \(\sigma_D^2\).
Because \(\bar D\sim N(\mu_D,\sigma_D^2/n)\), the standardized quantity \((\bar D-\mu_D)/(\sigma_D/\sqrt n)\) has a standard normal distribution.
For a normal sample, \((n-1)s_D^2/\sigma_D^2\) has a chi-square distribution with \(n-1\) degrees of freedom and is independent of the standardized sample mean.
Replacing \(\sigma_D\) by \(s_D\) gives \((\bar D-\mu_D)/(s_D/\sqrt n)\), which has a Student t distribution with \(n-1\) degrees of freedom. The central probability statement is therefore \(P(-t_{1-\alpha/2,n-1}\leq(\bar D-\mu_D)/(s_D/\sqrt n)\leq t_{1-\alpha/2,n-1})=1-\alpha\).
Multiplying through by the positive standard error and isolating \(\mu_D\) gives \(\bar D-t_{1-\alpha/2,n-1}s_D/\sqrt n\leq\mu_D\leq\bar D+t_{1-\alpha/2,n-1}s_D/\sqrt n\), which is the paired t interval.
This argument also explains why the degrees of freedom are \(n-1\): the standard deviation is computed from \(n\) differences after estimating their mean.
Worked Example: Estimating an Average Reduction in Processing Time
A hypothetical operations team records the processing time of five matched cases before and after a new procedure. Times are in minutes. Define \(D_i=\text{before}_i-\text{after}_i\), so a positive difference represents a reduction. The observed differences are \(2,2,6,10,10\). Assume the pairs are independent and the differences are normally distributed. Find a 95% paired t confidence interval for the population mean reduction.
Method and mean. The sum is \(2+2+6+10+10=30\), so \(n=5\) and \(\bar D=30/5=6\) minutes. The deviations from 6 are \(-4,-4,0,4,4\), and their squared values sum to \(16+16+0+16+16=64\).
Standard deviation and standard error. The sample variance is \(s_D^2=64/(5-1)=16\), hence \(s_D=4\) minutes. The standard error is \(4/\sqrt5=1.7889\) minutes.
Critical value and interval. With 4 degrees of freedom, \(t_{0.975,4}=2.776\). The margin of error is \(2.776(1.7889)=4.9660\) minutes. The endpoints are \(6-4.9660=1.0340\) and \(6+4.9660=10.9660\). Rounded to two decimal places, the interval is \((1.03,10.97)\) minutes.
Interpretation and check. This interval estimates the population mean reduction, not the mean change in the opposite direction. Its midpoint is \(6\), and its half-width is \(4.97\) to the stated precision. Reversing the subtraction order would reverse both endpoints’ signs and would require describing the parameter as after minus before.
Reading the Interval in Context
The interval’s units are the units of the original measurements, and its direction is determined by the subtraction order. With after minus before, positive values indicate increases and negative values indicate decreases. With before minus after, positive values indicate decreases and negative values indicate increases. State the order and the meaning of a positive difference when reporting the result.
A confidence interval estimates a population mean difference; it does not say that every individual’s change lies between the endpoints. It also does not establish a cause for the observed change. The procedure quantifies uncertainty about an average change under its sampling and distribution assumptions.
Worked Example: Detecting a Reversed Interpretation
A hypothetical clinic measures a symptom score before and after a program. An analyst defines \(D_i=\text{after}_i-\text{before}_i\), obtains \(\bar D=-2.5\), \(s_D=2.0\), and \(n=9\), and assumes independent pairs with normally distributed differences. Construct a 90% interval and state what a negative difference means.
Method. The degrees of freedom are \(9-1=8\). For a two-sided 90% interval, \(t_{0.95,8}=1.860\). The standard error is \(2/\sqrt9=2/3=0.6667\).
Margin and endpoints. The margin of error is \(1.860(0.6667)=1.2400\). Therefore the interval is \(-2.5\pm1.2400\), with endpoints \(-3.7400\) and \(-1.2600\). To two decimal places, the interval is \((-3.74,-1.26)\) score points.
Interpretation and check. Because the order is after minus before, negative differences indicate that the score decreased after the program. The interval estimates a mean decrease between 1.26 and 3.74 points in magnitude. The common sign trap is to call \(-2.5\) an average increase; that reverses the stated subtraction order. The interval’s midpoint is \(-2.50\), as required.
Common Calculation Traps
- Changing the subtraction order mid-calculation. Compute every \(D_i\) using one stated order, then retain that order when describing \(\bar D\) and the interval.
- Using the wrong sample size. The sample size is the number of complete pairs, not the total number of individual measurements. With \(n\) pairs, use \(n-1\) degrees of freedom.
- Calculating variability from the original columns separately. Find \(s_D\) from the individual differences. The paired t formula uses \(s_D/\sqrt n\).
- Using a z critical value or the wrong t degrees of freedom. When \(\sigma_D\) is unknown, use the Student t critical value with \(n-1\) degrees of freedom.
- Claiming an exact interval without the needed conditions. For an exact small-sample interval, the differences must be normally distributed and the pairs independently sampled.
Check Your Understanding
For each problem, keep the stated subtraction order throughout. Give interval endpoints to two decimal places where needed.
- A researcher records four paired measurements and defines \(D=\text{after}-\text{before}\). The differences are \(1,3,2,2\). Assuming independent pairs and normally distributed differences, calculate a 95% paired t interval for \(\mu_D\). (Answer: \((0.70,3.30)\))
- A device’s operating time is measured before and after a hypothetical adjustment. Define \(D=\text{after}-\text{before}\); the sample has \(n=7\), \(\bar D=-1.5\) hours, and \(s_D=2.1\) hours. Assuming independent, normally distributed differences, calculate a 95% interval. (Answer: \((-3.44,0.44)\) hours)
- For eight matched cases, a hypothetical service team defines \(D=\text{before}-\text{after}\), with \(\bar D=4\) minutes and \(s_D=3\) minutes. Assuming independent, normally distributed differences, calculate a 90% interval for mean reduction. (Answer: \((1.99,6.01)\) minutes)
- A hypothetical sensor’s paired error magnitudes before and after recalibration produce differences \(D=\text{before}-\text{after}\) of \(3,4,4,5,4\), in units. Assuming independent, normally distributed differences, calculate a 95% interval for mean reduction. (Answer: \((3.12,4.88)\) units)
- A study defines \(D=\text{after}-\text{before}\) for nine matched observations and reports \(\bar D=-2.0\) and \(s_D=3.0\). Assuming independent, normally distributed differences, calculate a 95% interval and state whether its endpoints establish a nonzero mean change at this confidence level. (Answer: \((-4.31,0.31)\); no, because zero lies in the interval.)