Turning Observations Into Signs

Why Signs Begin With a Comparison

A sign procedure reduces numerical observations to their direction relative to a reference value. That makes it essential to identify the reference before counting anything. If the question concerns whether a population median differs from a proposed value, the relevant signs come from comparing each observation with that proposed median—not merely from checking whether the observation itself is positive or negative.

For paired data, the first step is to calculate a change using a clearly stated order, such as after minus before. The signs are then assigned by comparing each change with the hypothesized median change. A change equal to the target has sign zero and does not contribute to either directional count.

The sign records direction, not magnitude. For example, differences of \(1\) and \(100\) both contribute one positive sign. The procedure deliberately discards the size of each nonzero difference; its input is whether the observation lies above, below, or exactly at the hypothesized median.

One-Sample Signs Relative to a Hypothesized Median

Suppose a study records \(n\) observations and asks about a specified median \(m_0\). For each observation, subtract \(m_0\), or equivalently compare the observation directly with \(m_0\). Do not compare with zero unless zero is itself the hypothesized median. Once every sign has been assigned, count the positive signs, negative signs, and zeros separately.

The number of observations used in the sign calculation is the effective sample size \(n^*\), not necessarily the original sample size \(n\). Observations exactly equal to the target have no direction relative to it, so they contribute to neither \(S_+\) nor \(S_-\).

Here \(I(A)\) equals \(1\) when statement \(A\) is true and \(0\) otherwise. Each observation either lies above the target, below it, or equals it. These three possibilities are mutually exclusive and exhaustive, so the following identity is a useful arithmetic check.

Paired Signs and the Direction of Change

For paired observations, calculate one difference for each pair. The subtraction order determines what a positive change means, so state it before assigning signs. For example, if \(D_i=\text{after}_i-\text{before}_i\), then a positive \(D_i\) means the measured quantity increased after the intervention. If the proposed median change is \(\Delta_0\), compare each calculated \(D_i\) with \(\Delta_0\). The sign is the sign of \(D_i-\Delta_0\), not necessarily the sign of \(D_i\).

This paired construction turns each pair into one change and then treats those changes as the observations to be compared with the target. It does not treat the two measurements in a pair as separate signs. When the target change is zero, a positive sign means an increase under the chosen subtraction order; with a nonzero target, a positive sign instead means a change greater than that target.

What the Sign Counts Do—and Do Not—Establish

The counts \(S_+\) and \(S_-\) summarize how many nonzero differences fall on either side of the reference value. The sign sequence preserves the order of observations if that is useful for checking the work, while the counts provide the basic numerical summary. Neither count retains the magnitude of a difference: a value just above the target contributes the same one positive sign as a value far above it.

The sign construction itself is a data transformation. To use the counts for a formal test, an inferential null model and its assumptions are also needed. In the standard continuous one-sample setting, a null median \(m_0\) implies equal probabilities of falling above and below \(m_0\). If the distribution has a point mass at \(m_0\), the treatment of ties and the assumptions supporting the null calculation require care. Forming the signs correctly does not, by itself, establish that a particular null distribution applies.

For paired data, the corresponding signs are based on the paired changes and their hypothesized median \(\Delta_0\). The pairs must be matched as described, and the subtraction order must be consistent. These requirements ensure that a positive sign has the same interpretation across all pairs.

Common Errors to Avoid

• Using zero as the reference automatically. Compare \(X_i\) with \(m_0\), or \(D_i\) with \(\Delta_0\). Zero is the reference only when it is the stated target.
• Signing the raw paired measurements. First calculate one change per pair in the stated order; then compare each change with the target change.
• Counting target matches as positive or negative. An observation equal to the hypothesized value has a zero difference and contributes to neither sign count.
• Using the original sample size after removing zeros. The effective sample size is \(n^*=S_++S_-\), not the number of observations before zero differences are excluded.
• Reversing the paired subtraction order midway. If change is defined as after minus before, use that order for every pair. Reversing it changes the interpretation of every nonzero sign.
• Confusing direction with size. Sign counts use only whether a difference is positive, negative, or zero relative to the target; they do not measure how large the difference is.