Why Pairing Can Change Precision
When two measurements come from the same person, item, or matched unit, they are usually related. A before-and-after measurement on one policyholder, for example, may share characteristics that make both values high or both values low. An analysis that uses this relationship can produce a more precise estimate of the mean difference than an analysis that treats the measurements as independent. But pairing does not always help: if the two measurements tend to move in opposite directions, it can increase the variability of their differences.
The key quantity is the within-pair covariance. This tutorial compares the standard error of a paired mean difference with the standard error for two independent samples having the same sample sizes and marginal variability. The independent-sample calculation is a useful precision benchmark, but it is not a valid substitute for paired analysis when the observations are dependent.
The Variance Relationship
Write the population variances of the two measurements as \(\sigma_X^2\) and \(\sigma_Y^2\), and their within-pair correlation as \(\rho\). The covariance is \(\operatorname{Cov}(X,Y)=\rho\sigma_X\sigma_Y\). Because the difference is \(D=X-Y\), its variance is
For a paired sample, the standard error of \(\bar{D}\) is estimated by \(s_D/\sqrt{n}\). If the two samples instead consisted of independent observations, the variance of their difference in sample means would be \(\sigma_X^2/n+\sigma_Y^2/n\), estimated by \(s_X^2/n+s_Y^2/n\). The paired variance includes the covariance term; the independent-sample variance does not.
Why the Covariance Term Appears
The relationship follows by expanding the squared deviation of a difference. At the population level, center \(X\) and \(Y\) by their respective means. Then \(D-E[D]=(X-E[X])-(Y-E[Y])\). Squaring and taking expectations gives
The cross-product is the covariance, so it cannot be discarded unless the covariance is zero. A positive covariance subtracts from the variance of the difference: when both measurements tend to be above or below their means together, subtracting them tends to cancel some shared variation. A negative covariance adds to that variance: the measurements tend to move in opposite directions, so their difference can vary more.
In a two-sided \(100(1-\alpha)\%\) paired t interval, the estimated standard error is multiplied by \(t_{1-\alpha/2,n-1}\), provided the \(D_i\) are a random sample of independent, identically distributed differences from a normal population. The interval is \(\bar{D}\pm t_{1-\alpha/2,n-1}s_D/\sqrt{n}\). For two independent samples, a Welch interval uses standard error \(\sqrt{s_X^2/n_X+s_Y^2/n_Y}\) and an approximate degrees of freedom. These assumptions matter: the paired method requires independence across pairs, not independence within a pair.
Worked Comparisons
Worked Example: Strong Positive Correlation Narrows the Interval
A hypothetical study measures two quantities on each of 16 matched devices. Let \(X\) be the first reading and \(Y\) the second; the reported mean difference is \(\bar{X}-\bar{Y}=2.500\). The sample standard deviations are \(s_X=10\) and \(s_Y=8\), and the sample correlation within pairs is \(r=0.75\). Compare the paired and independent-sample standard errors, then calculate the paired 95% t interval. Assume independent pairs and normally distributed paired differences.
Method. First obtain the sample covariance: \(s_{XY}=r s_Xs_Y=0.75(10)(8)=60\). Then use the variance of a difference:
The paired standard error is \(6.6332/\sqrt{16}=6.6332/4=1.6583\). For the independent-sample benchmark, omit the covariance term: \(\sqrt{10^2/16+8^2/16}=\sqrt{100/16+64/16}=\sqrt{10.25}=3.2016\). The paired standard error is much smaller because the within-pair correlation is strongly positive.
For a paired 95% interval, \(n-1=15\) degrees of freedom and \(t_{0.975,15}\approx2.131\). The margin is \(2.131(1.6583)=3.535\), to three decimals. Therefore the interval is \(2.500\pm3.535=(-1.035,6.035)\), to three decimals.
Check. The half-width, \(3.535\), is the stated critical value times the paired standard error, and the interval midpoint is \(2.500\). The independent standard error is a benchmark only; using it to analyze these correlated pairs would misrepresent their sampling variability.
Worked Example: Negative Correlation Can Make Pairing Less Precise
A hypothetical matched assessment records \(X\) and \(Y\) for each of 10 participants. The sample standard deviations are \(s_X=3\) and \(s_Y=4\), the sample correlation is \(r=-0.50\), and \(\bar{X}-\bar{Y}=-1.200\). Compare the two standard errors and calculate the paired 95% t interval. Assume the paired differences are independent across participants and normally distributed.
Method. The sample covariance is \((-0.50)(3)(4)=-6\). The difference variance estimate is
The paired standard error is \(\sqrt{37}/\sqrt{10}=\sqrt{3.7}=1.9235\). The independent-sample benchmark is \(\sqrt{3^2/10+4^2/10}=\sqrt{2.5}=1.5811\). Here pairing gives a larger standard error because the negative covariance increases the variability of the differences.
For a paired 95% interval, there are \(9\) degrees of freedom and \(t_{0.975,9}\approx2.262\). The margin is \(2.262(1.9235)=4.351\), to three decimals. The interval is \(-1.200\pm4.351=(-5.551,3.151)\), to three decimals.
Check. The interval midpoint is \(-1.200\), and its half-width is \(4.351\). A larger paired standard error is not a calculation failure: pairing describes the data structure, while its effect on precision depends on the direction and strength of association.
Worked Example: Recovering Correlation From the Difference Variability
In a hypothetical study of 20 matched service cases, \(s_X=5\), \(s_Y=4\), and the sample standard deviation of the within-pair differences \(D=X-Y\) is \(s_D=\sqrt{17}\). The observed mean difference is \(\bar{D}=1.100\). Find the sample correlation, compare standard errors, and calculate the paired 95% t interval. Assume independent, normally distributed differences.
Method. Rearrange the sample variance identity to find the covariance:
The sample correlation is \(r=s_{XY}/(s_Xs_Y)=12/(5\cdot4)=0.600\). The paired standard error is \(\sqrt{17}/\sqrt{20}=\sqrt{0.85}=0.9220\). The independent-sample benchmark is \(\sqrt{25/20+16/20}=\sqrt{2.05}=1.4318\). The positive correlation explains why the paired estimate is more precise.
For 19 degrees of freedom, \(t_{0.975,19}\approx2.093\). The margin is \(2.093(0.9220)=1.930\), to three decimals. The paired interval is \(1.100\pm1.930=(-0.830,3.030)\), to three decimals.
Check. Substituting the recovered covariance into the variance identity gives \(25+16-2(12)=17\), matching the stated difference variance. The correlation is positive and below 1, as expected for these valid summary values.
Interpreting Precision Without Overstating It
For equal sample sizes, the estimated standard-error ratio is
This ratio compares estimated standard errors, not necessarily the final interval widths. The critical values can differ because the paired t interval uses \(n-1\) degrees of freedom, whereas an independent-sample interval may use a pooled or Welch procedure with different degrees of freedom. Even when the paired standard error is smaller, the full interval-width comparison should use each method’s appropriate critical value.
If the two measurements have the same population standard deviation \(\sigma\), the population ratio simplifies to \(\sqrt{1-\rho}\). For example, a correlation of \(0.64\) gives a paired-to-independent standard-error ratio of \(\sqrt{0.36}=0.60\). A correlation of \(-0.25\) gives \(\sqrt{1.25}>1\), so the paired difference has a larger standard error than the independent benchmark.
Pairing is not an optional adjustment chosen only because it narrows an interval. It is determined by how the data were collected. If the same units contribute both measurements, their dependence must be accounted for. Conversely, if two samples are genuinely independent, inventing pairs does not justify using a paired standard error.
Common Mistake: Dropping the Covariance Term
A frequent error is to use \(\sqrt{s_X^2/n+s_Y^2/n}\) for matched observations simply because there are two sample means. That expression is the independent-sample standard error; it ignores the relationship within each pair. In the first example, it gives \(3.2016\), while the correct paired estimate is \(1.6583\). The correction is to form \(D_i=X_i-Y_i\) for every matched unit and calculate \(s_D/\sqrt{n}\), or equivalently use \(s_D^2=s_X^2+s_Y^2-2s_{XY}\).
Another trap is to assume that pairing always improves precision. The second example shows the opposite: its negative covariance increases the difference variance. The safe procedure is to identify the design, calculate the appropriate variability, and compare the resulting standard errors rather than relying on a rule of thumb.
Determine whether each \(X_i\) is naturally matched to a \(Y_i\), and whether different pairs can reasonably be treated as independent.
For matched observations, calculate the differences and use \(s_D/\sqrt{n}\); do not discard within-pair association.
Positive within-pair correlation tends to reduce the difference variance, while negative correlation tends to increase it. Confirm the conclusion from the calculated standard errors.
Check Your Understanding
For each question, compare standard errors using the stated design. Round standard errors to four decimals where needed.
- Twenty-five matched units have \(s_X=6\), \(s_Y=8\), and sample correlation \(r=0.50\). Find the paired and independent-sample benchmark standard errors. (Answer: paired \(=\sqrt{52}/5=1.4422\); independent \(=2.0000\).)
- Sixteen pairs have \(s_X=3\), \(s_Y=5\), and \(r=-0.40\). Find the estimated variance of the paired differences and compare the paired standard error with the independent benchmark. (Answer: \(s_D^2=46\); paired standard error \(=\sqrt{46}/4=1.6956\); independent benchmark \(=\sqrt{34}/4=1.4577\).)
- Twelve pairs have \(s_X=3\), \(s_Y=4\), and \(s_D=4\). Find the sample covariance, sample correlation, and paired standard error. (Answer: covariance \(=4.5\); correlation \(=0.375\); paired standard error \(=4/\sqrt{12}=1.1547\).)
- Two measurements have equal population standard deviations and within-pair correlation \(0.64\). What is the ratio of the paired standard error to the independent-sample standard error? (Answer: \(\sqrt{1-0.64}=0.6000\).)
- Nine matched units have \(s_X=5\), \(s_Y=12\), and \(s_D=9\). Find the sample covariance and correlation, and state whether the paired standard error is smaller than the independent benchmark. (Answer: covariance \(=(25+144-81)/2=44\); correlation \(=44/60=0.7333\); paired standard error \(=9/3=3.0000\), smaller than the independent benchmark \(\sqrt{25+144}/3=4.3333\).)