Introduction
A signed-rank procedure uses both the signs of observations’ deviations from a proposed location and the ranks of their absolute sizes. Its location interpretation depends on a symmetry assumption: the procedure is aimed at a center of symmetry, often described through the pseudomedian. That target is not, in general, the ordinary median.
For paired data, the observations to which the procedure is applied are typically the paired differences. For a one-sample question, they are the observations themselves. The calculation of the signed-rank statistic concerns their signed ranks; the pseudomedian gives a precise location quantity associated with those observations.
Why the Pseudomedian Is the Signed-Rank Location Target
For a hypothesized location \(\theta\), the signed-rank procedure is applied to deviations \(X_i-\theta\). It ranks their absolute values and uses the signs attached to those ranks. If the observations have a distribution symmetric about \(\theta\), positive and negative deviations are balanced in the specific sense required by the signed-rank null model. With a continuous distribution and no zero deviations, each sign is equally likely to be positive or negative and, under symmetry, the signs are independent of the absolute values.
The population pseudomedian matches this center. The reason is short but important. If \(X\) is symmetric about \(\theta\), then \(X-\theta\) is symmetric about zero. For independent \(X_1,X_2\), their centered sum is symmetric about zero as well. Therefore, \((X_1+X_2)/2\) is symmetric about \(\theta\), so \(\theta\) is its median. When the median is unique, the population pseudomedian is exactly \(\theta\).
This argument explains the role of symmetry: it connects the distributional center tested by signed ranks to the pseudomedian. Without symmetry, the signed-rank test should not automatically be described as a test of an ordinary median. The pseudomedian remains a defined quantity, but its value need not equal the ordinary median.
Worked Example: Computing Walsh Averages for Symmetric Data
A hypothetical analyst records four changes in a process measurement: \(-3,-1,1,3\). Compute the sample pseudomedian and compare it with the ordinary sample median.
Method. Form every average \( (x_i+x_j)/2 \) with \(i\leq j\). There must be \(4(5)/2=10\) Walsh averages.
The self-averages are \(-3,-1,1,3\). The six averages from distinct pairs are \((-3-1)/2=-2\), \((-3+1)/2=-1\), \((-3+3)/2=0\), \((-1+1)/2=0\), \((-1+3)/2=1\), and \((1+3)/2=2\).
Combining all ten gives the unsorted list \(-3,-2,-1,0,-1,0,1,2,1,3\). In order, the Walsh averages are \(-3,-2,-1,-1,0,0,1,1,2,3\). The fifth and sixth are both 0, so the sample pseudomedian is \((0+0)/2=0\).
The ordinary sample median is also \((-1+1)/2=0\). Here the two summaries agree, consistent with the data’s symmetry around zero.
Check. There are four self-averages and six distinct-pair averages, for ten total, as required. A common mistake is to calculate \((1+3)/2\) as 3; it is 2, and including that correct value gives the sorted list above.
Calculating a Sample Pseudomedian
The Walsh-average calculation is mechanical, but there are two details to keep straight: include the self-averages and sort the complete list before selecting its middle. The ordinary sample median, by contrast, is calculated directly from the ordered observations. These procedures use different sets of numbers and can give different answers.
Worked Example: A Pseudomedian Different From the Ordinary Median
Suppose a hypothetical study records four positive changes, in units, of \(2,4,7,11\). Find the sample pseudomedian and the ordinary sample median.
Method. There are \(4(5)/2=10\) Walsh averages. The four self-averages are \(2,4,7,11\). The six averages from distinct pairs are \((2+4)/2=3\), \((2+7)/2=4.5\), \((2+11)/2=6.5\), \((4+7)/2=5.5\), \((4+11)/2=7.5\), and \((7+11)/2=9\).
In sorted order the ten averages are \(2,3,4,4.5,5.5,6.5,7,7.5,9,11\). The fifth and sixth are \(5.5\) and \(6.5\), so the sample pseudomedian is \((5.5+6.5)/2=6\).
The ordinary sample median uses only the four observations: \((4+7)/2=5.5\). Thus the sample pseudomedian is 6, while the ordinary sample median is 5.5.
Check. The list contains all ten required averages, and the two middle Walsh averages average to 6. A common mistake is to call 5.5 the pseudomedian simply because it is the ordinary sample median.
Using Signed Ranks to Assess a Location Shift
For data \(x_1,\ldots,x_n\) and a proposed location \(\theta_0\), calculate the deviations \(x_i-\theta_0\), rank their absolute values, and attach each deviation’s sign to its rank. The sum of the positive ranks is commonly denoted \(T^+\). Under the signed-rank null model, the proposed location is the center of a symmetric distribution. The exact treatment of tied absolute deviations and zero deviations depends on the stated convention; those cases must be handled consistently with the procedure being used.
A large positive-rank sum indicates that, relative to \(\theta_0\), larger ranked deviations tend to be positive; a small sum indicates the opposite. This is evidence assessed under the symmetric-location model. A signed-rank result alone does not establish that the observations are symmetric, nor does it license a median interpretation when symmetry is implausible.
Worked Example: Centering the Data Before Ranking
A hypothetical paired study produces three differences, in minutes: \(1,4,10\). Assess the signed ranks relative to the proposed shift \(\theta_0=2\), and calculate the sample pseudomedian of the observed differences.
Method. First subtract the proposed shift. The deviations are \(1-2=-1\), \(4-2=2\), and \(10-2=8\). Their absolute values are \(1,2,8\), whose ranks are \(1,2,3\), respectively. There are no ties or zero deviations. The positive deviations have ranks 2 and 3, so \(T^+=2+3=5\).
For the sample pseudomedian of the original differences, the six Walsh averages are the three self-averages \(1,4,10\), together with \((1+4)/2=2.5\), \((1+10)/2=5.5\), and \((4+10)/2=7\). In order they are \(1,2.5,4,5.5,7,10\). Their middle two are 4 and 5.5, giving a sample pseudomedian of \((4+5.5)/2=4.75\).
The ordinary sample median of the differences is 4. The proposed shift used for ranking is 2; it is not the sample pseudomedian. The rank sum describes the observed signs and magnitudes relative to 2, while 4.75 is a separate sample location summary.
Check. The signed ranks are attached to deviations from 2, not to the uncentered observations. Ranking \(1,4,10\) without subtracting 2 would answer a different question.
Population Pseudomedian and Ordinary Median Can Differ
The distinction is not merely a matter of how a sample is calculated. The two population quantities can differ substantially when a distribution is asymmetric. The next example computes both from a simple discrete distribution.
Worked Example: An Asymmetric Distribution
Let \(X\) be a hypothetical measurement equal to 0 with probability 0.6 and 10 with probability 0.4. Find its ordinary population median and its population pseudomedian.
Ordinary median. The cumulative probability at 0 is 0.6, so at least half the distribution is at or below 0. Also, the probability at or above 0 is 1. Thus 0 is a median. No value greater than 0 can be a median: for any \(m\) with \(0<m\leq10\), the probability \(P(X\geq m)\) is 0.4, below one-half. The ordinary median is therefore 0.
Pseudomedian. Take two independent draws \(X_1,X_2\). Their average is 0 if both are 0, which has probability \(0.6(0.6)=0.36\). It is 5 if one is 0 and the other is 10, which has probability \(2(0.6)(0.4)=0.48\). It is 10 if both are 10, which has probability \(0.4(0.4)=0.16\).
The average has cumulative probabilities 0.36 at 0 and \(0.36+0.48=0.84\) at 5. Thus its median is 5, and the population pseudomedian is 5. It differs from the ordinary median, 0.
Check. The probabilities \(0.36+0.48+0.16=1\). This distribution is not symmetric around 0 or 5, so the example illustrates why the signed-rank symmetry assumption matters rather than demonstrating a valid symmetric-location model.
Common Interpretation Traps
- Equating the two medians. The ordinary median summarizes the distribution of \(X\); the pseudomedian summarizes the distribution of \((X_1+X_2)/2\). They may agree, but agreement is not guaranteed.
- Calling signed ranks a general test of the median. The location interpretation relies on symmetry about the proposed center. Without that assumption, the ordinary median need not be the parameter being assessed.
- Omitting self-averages. Walsh averages include \((x_i+x_i)/2=x_i\). Omitting them changes the statistic and the resulting sample pseudomedian.
- Forgetting to sort the full Walsh list. For an even number of averages, the sample pseudomedian is the average of the two central ordered Walsh averages—not necessarily a Walsh average itself.
- Ranking before centering. To assess a proposed location \(\theta_0\), rank \(|x_i-\theta_0|\) and retain the signs of \(x_i-\theta_0\), not the signs and ranks of uncentered observations.
Check Your Understanding
For each problem, distinguish the location summary requested from any ordinary median or signed-rank quantity.
- For observations \(1,3,9\), list the six Walsh averages, find the sample pseudomedian, and compare it with the ordinary sample median. (Answer: Walsh averages \(1,2,3,5,6,9\); sample pseudomedian \(4\); ordinary median \(3\).)
- A sample has a pseudomedian of 6.5. Each observation is increased by 3, with no other changes. What is the new pseudomedian, and why? (Answer: \(9.5\); every Walsh average increases by 3.)
- For observations \(0,2,8\), calculate all six Walsh averages and the sample pseudomedian. Then state the ordinary sample median. (Answer: Walsh averages \(0,1,2,4,5,8\); pseudomedian \(3\); ordinary median \(2\).)
- A random variable is 1 with probability 0.7 and 9 with probability 0.3. Find its ordinary median and pseudomedian, and state whether equality holds. (Answer: ordinary median \(1\); pseudomedian \(5\); they differ.)
- A signed-rank analysis is conducted on deviations from a proposed value \(\theta_0\). State the distributional condition needed to interpret \(\theta_0\) as the pseudomedian target of the procedure, and explain why a generic median interpretation is not automatic. (Answer: the observations should be symmetrically distributed about \(\theta_0\); without symmetry, the ordinary median and pseudomedian need not agree.)