Paired versus Unpaired Data
If statistical methods are applied to analyze data from multiple
groups and each individual data point comes from a separate and distinct source,
the mathematics behind the statistics assumes an independence and randomness that
would not be the same if the sources are highly related. As an example, consider
a study of blood pressure in two groups of subjects, one treated with medication
and the other untreated. Because blood pressure measurements can be widely scattered
in distribution, depending on the subjects' previous medical condition, a small drug
effect might be difficult to notice. However, if the same individuals were studied
before and after therapy, the wide initial variations might cancel out and small
effects would be noticed. In this latter study, the data points can be paired, one
in each group, to detect these small effects. A study in which the data can be paired
can be very sensitive in picking out small changes. Conversely, the erroneous use
of paired statistics might suggest a statistically significant result when one does
not exist.
TABLE 23-3 -- Choice of tests
Comparison Goal |
Normally Distributed (Parametric) Interval
Data |
Nonparametric Interval Data |
Categorical Data |
Describe group |
Mean ± SD |
Median, mode, percentile |
Counts and proportions |
Compare group with value |
t-Test |
Wilcoxon |
Chi-square |
Compare two groups |
Unpaired t-test |
Mann-Whitney |
Chi-square, Fisher |
Two paired groups |
Paired t-test |
Wilcoxon |
McNemar |
Three or more groups |
ANOVA |
Kruskal-Wallis |
Chi-square |
Multiple matched groups |
Repeated-measures ANOVA |
Friedman |
— |
Regression |
Linear regression |
Nonparametric regression |
Logistic regression |
ANOVA, analysis of variance; SD, standard deviation. |