Psychometric scaling is the process of transforming qualitative judgments into quantitative scales. It is widely used in market research, psychology, and sensory analysis. The fundamental goal is to assign numerical values to stimuli based on paired comparisons or category judgments.
When subjects evaluate stimuli (products, brands, statements), their responses are inherently ordinal. Psychometric scaling methods provide a principled way to construct interval-level measurement from these ordinal observations, enabling meaningful arithmetic operations on the resulting scale values.
Given \(K\) stimuli evaluated by judges on ordinal categories, we want to construct an interval scale — a scale where differences between values are meaningful and comparable.
Stimuli can be of many kinds:
The key assumption underlying all scaling methods in this family is that there exists a latent continuous variable driving the observed categorical judgments. This is the foundation of Thurstone's model: each stimulus occupies a position on a psychological continuum, and observed judgments are noisy reflections of these true positions.
Thurstone (1927) proposed that each stimulus \(i\) evokes a discriminal process — a random variable representing its perceived value on any given occasion:
\[ X_i \sim N(S_i, \sigma_i^2) \]
where \(S_i\) is the true scale value (mean of the discriminal process) and \(\sigma_i^2\) is the discriminal dispersion.
When a judge compares stimuli \(j\) and \(i\), the probability that \(j\) is judged greater than \(i\) is:
\[ P(j > i) = \Phi\left( \frac{S_j - S_i}{\sqrt{\sigma_i^2 + \sigma_j^2 - 2\rho_{ij}\sigma_i\sigma_j}} \right) \]
where \(\Phi\) is the standard normal CDF and \(\rho_{ij}\) is the correlation between discriminal processes for stimuli \(i\) and \(j\).
Thurstone's Case V assumes equal discriminal dispersions (\(\sigma_i = \sigma\) for all \(i\)) and zero correlations (\(\rho_{ij} = 0\)). Under these assumptions:
\[ P(j > i) = \Phi\left( \frac{S_j - S_i}{\sigma\sqrt{2}} \right) \]
This is the most commonly used form because it requires the fewest parameters to estimate and leads directly to Torgerson's scaling method.
Torgerson provided a systematic procedure to recover scale values from observed comparison data under Case V assumptions. The method proceeds in four steps:
Collect paired comparison or category judgment data and organize into a frequency matrix \(F\), where \(F_{ij}\) is the number of times stimulus \(j\) was judged greater than stimulus \(i\).
Convert frequencies to proportions:
\[ P_{ij} = \frac{F_{ij}}{N} \]
where \(N\) is the total number of judges. \(P_{ij}\) represents the proportion of judges who rated stimulus \(j\) higher than stimulus \(i\).
Apply the inverse normal (probit) transformation to each proportion:
\[ Z_{ij} = \Phi^{-1}(P_{ij}) \]
Under Case V, we have \(Z_{ij} = \frac{S_j - S_i}{\sigma\sqrt{2}}\), which is the key relationship linking observed data to latent scale values.
The scale values are obtained as column means of the \(Z\) matrix:
\[ S_j = \frac{1}{K} \sum_{i=1}^{K} Z_{ij} \]
This works because summing over rows for a fixed column \(j\):
\[ \sum_{i=1}^{K} Z_{ij} = \frac{1}{\sigma\sqrt{2}} \sum_{i=1}^{K}(S_j - S_i) = \frac{K \cdot S_j - \sum_i S_i}{\sigma\sqrt{2}} \]
Setting the scale origin such that \(\sum_i S_i = 0\) and unit such that \(\sigma\sqrt{2} = 1\), the column mean directly yields \(S_j\).
Note: The resulting scale is an interval scale — the origin and unit of measurement are arbitrary. Only differences between scale values are meaningful.
When data comes from category ratings rather than direct paired comparisons, Torgerson's method can still be applied with an intermediate step:
Thresholds are estimated as row means and scale values as column means (with sign reversal) of the \(Z\) matrix, analogous to the paired comparison case.
Once scale values are estimated, we can assess how well the Case V model fits the data through an internal consistency check.
From the estimated scale values, compute predicted proportions:
\[ \hat{P}_{ij} = \Phi\left( \frac{S_j - S_i}{\sigma\sqrt{2}} \right) \]
Compare predicted \(\hat{P}_{ij}\) with observed \(P_{ij}\) using:
If the fit is poor, Case V assumptions may not hold. In such cases, one may relax assumptions (Case III: unequal variances) or consider non-metric alternatives such as multidimensional scaling.