Polychoric and Tetrachoric Correlation

Two bond issuers' credit ratings each move between just a few discrete letter grades, say AAA, AA, A, BBB, BB, over ten years. You compute an ordinary correlation between the two letter-grade sequences (after converting letters to numbers) and get 0.35. But ratings are coarse snapshots of an underlying continuous variable, the issuer's true, continuously varying creditworthiness, that only crosses your attention when it moves far enough to flip a discrete letter grade. Ordinary correlation on the coarse categories understates the underlying continuous correlation, because genuine co-movement within a grade band never shows up as a category change. Polychoric correlation (for multiple ordered categories) and tetrachoric correlation (its two-category special case) estimate what the correlation would be if the smooth underlying variable were visible directly.

An analogy: judging rainfall correlation from "did it rain" alone

Imagine you only know, for two nearby cities on each day, whether it rained (yes/no), not how much. Ordinary correlation on the yes/no data will show some relationship, but it throws away enormous information: a light drizzle and a torrential downpour both just register as "yes," so days when one city gets a light sprinkle and the other gets nothing look the same as days when one gets a downpour and the other gets nothing, even though the underlying rainfall amounts might actually track each other closely. Tetrachoric correlation assumes there's a smooth, continuous "rainfall tendency" underneath the yes/no threshold in each city, and works backward from the pattern of joint yes/no outcomes to estimate how correlated those smooth underlying tendencies actually are, recovering information the coarse yes/no data alone destroyed.

The mechanics, one piece at a time

The model assumes each observed categorical variable is really a coarsened version of a continuous, normally distributed latent variable, cut into categories by unobserved thresholds. For two binary variables (the tetrachoric case), you observe a 2×2 table of joint frequencies, how often both are "yes," both "no," and each mixed combination. Tetrachoric correlation finds the value $\rho$ of a bivariate normal distribution's correlation parameter that would generate that exact 2×2 table of frequencies if you sliced the bivariate normal at the thresholds implied by each variable's marginal yes/no rate. In plain English: it asks "if two smooth, correlated bell-curve variables were hiding underneath these two yes/no flags, what correlation $\rho$ would best reproduce the agreement pattern we actually observed?" Polychoric correlation generalizes this to more than two ordered categories (like the five-letter-grade rating example), estimating both the category thresholds and the underlying correlation jointly, typically by maximum likelihood.

Worked example 1: a simplified tetrachoric estimate

Two analysts each flag a stock as "overvalued" or not on 100 stocks. Both flag 30 stocks as overvalued; they agree (both flag, or both don't) on 76 of the 100. Ordinary phi correlation (the binary analogue of Pearson correlation) on this 2×2 table works out to about 0.30. But because each analyst is only flagging the top 30% as overvalued, a fairly extreme cutoff on their continuous judgment, a lot of near-miss agreement (stocks just below one cutoff but just above the other) gets counted as pure disagreement. Solving for the bivariate-normal correlation that reproduces this 76% agreement rate at a 30% marginal cutoff gives a tetrachoric estimate around 0.55, nearly double the raw phi correlation, because it correctly accounts for how a strict cutoff mechanically suppresses agreement even when the underlying judgments are quite correlated.

Worked example 2: rating migrations, polychoric vs. naive

Two issuers' annual ratings (mapped AAA=1 through BB=5) over 15 years give an ordinary Pearson correlation, treating the numbers as literal quantities, of 0.35. A polychoric estimate, treating the ratings as coarse buckets of an underlying continuous credit-quality process, and using the actual observed frequency of each of the 5×5 joint rating combinations, comes out around 0.52. The gap matters directly for portfolio credit risk: default correlation models (see the asset threshold model) build on exactly this kind of latent continuous-variable correlation, and underestimating it by using naive ordinal correlation understates how much two issuers' defaults would cluster together in a downturn.

What this means in practice

Polychoric and tetrachoric correlation matter wherever ordinal, coarsely-bucketed data (credit ratings, analyst buy/hold/sell recommendations, survey Likert scales, binary event flags) is a proxy for something genuinely continuous underneath. Using ordinary correlation on such data systematically understates the true relationship, which matters for anything downstream: correlation matrices feeding a default-correlation model, factor analysis on ordinal survey or rating data, or any risk model that needs an honest estimate of co-movement rather than one artificially deflated by coarse measurement.