Ordered Factors and Ordered Contrasts

Why Order Is Not the Same as a Numeric Predictor

Categories such as policy tiers, clinical severity bands, or service ratings can have a meaningful order. That order contains information, but it does not automatically tell us that the gaps between categories are equal. A model must distinguish “low, medium, high” from unrelated labels such as region names, and it must also distinguish an ordered category from a numeric variable whose one-unit changes have a defined meaning.

In R, an ordered factor records a sequence of levels. Polynomial contrasts provide a set of columns that can describe linear and higher-order patterns across that sequence. These columns are a choice of model-matrix basis: their coefficients describe components of the fitted pattern, not necessarily changes from one category to the next.

For a factor with \(k\) levels, a full set of polynomial contrasts has \(k-1\) columns, in addition to the model intercept. R labels these columns with suffixes such as `.L`, `.Q`, and `.C`: linear, quadratic, and cubic. For a factor with more than four levels, further polynomial columns are available. The column values depend on the level order and the scores supplied to R.

A nominal factor should not be assigned an order merely because its labels can be sorted alphabetically. Conversely, an ordered factor should not be treated as numeric unless a numerical scale and the meaning of its unit steps are justified. The earlier material on factor contrasts establishes how a model matrix represents categories; here the key question is what the ordered contrast columns mean.

Creating and Inspecting Ordered Contrasts in R

Set the level sequence explicitly. For example, the R expression `ordered(x, levels = c("Basic", "Enhanced", "Premier"))` creates an ordered factor with Basic below Enhanced and Enhanced below Premier. Explicit levels protect against accidental alphabetical ordering and make the intended order visible in the analysis.

R commonly uses polynomial contrasts by default for ordered factors, but contrast settings can be changed. To make the choice reproducible, assign them explicitly: `contrasts(tier) <- contr.poly(3)`. The function `contr.poly(3)` creates contrasts for three ordered levels scored equally as 1, 2, and 3. The model matrix can be inspected using `model.matrix(~ tier, data = dat)`; its columns reveal exactly which values enter the model.

What the Trend Coefficients Mean

With a Gaussian identity-link model, a factor’s fitted mean at level \(j\) is formed by adding the intercept and the relevant contrast contributions. The `.L` coefficient describes the linear component in the chosen contrast basis. A `.Q` coefficient describes a quadratic component: a departure from a straight-line pattern across the scores. Higher-order columns represent further departures. A coefficient’s sign indicates the direction of that component relative to the signs in R’s contrast column.

For a generalized linear model, the same contrast columns enter the linear predictor. If the response has mean \(\mu_j\) at a level and link function \(g\), then \(g(\mu_j)\) is represented by the intercept plus the contrast contributions. The coefficient is therefore on the link scale. For example, in a binomial model with a logit link, it describes a component of the log odds, not a direct change in probability. Translating a coefficient into a probability difference requires evaluating the inverse link at specified predictor values.

The columns from `contr.poly(k)` are normalized and mutually orthogonal across the \(k\) levels when levels are equally weighted. For three levels, for example, the `.L` values sum to zero, the `.Q` values sum to zero, and the sum of their pairwise products is zero. Normalization matters: it sets the numerical scale of coefficients. Using a different contrast basis to represent the same fitted values changes coefficient values, even though fitted means may remain unchanged.

Why the Balanced-Data Formula Works

The projection interpretation can be derived directly for a balanced Gaussian identity-link model. Let \(c_{jr}\) be the value of contrast column \(r\) at level \(j\), and let \(\bar y_j\) be the response mean at that level. Suppose there are \(m\) observations at every level, the contrast columns are centered and orthonormal across levels, and the model includes the intercept and all \(k-1\) contrast columns. Thus, \(\sum_j c_{jr}=0\) and \(\sum_j c_{jr}c_{js}=1\) when \(r=s\), or 0 otherwise.

The model predicts \(\bar y_j = \alpha+\sum_r\beta_r c_{jr}\). Minimizing the sum of squared residuals gives the normal equation for coefficient \(r\): the sum of the residuals multiplied by \(c_{jr}\), across all observations, equals zero. Since there are \(m\) observations at each level, this equation is

The intercept term vanishes because the contrast is centered. Orthogonality makes every cross-term vanish except the one involving \(\beta_r\). Dividing by \(m\) leaves

This is why, in the balanced example, each coefficient was calculated by multiplying level means by contrast values and summing. The result does not apply in this simple form to unequal replication, non-identity links, or a model containing other predictors: the estimation then uses the model’s actual design and, for a generalized linear model, its likelihood and link.

Unequal Spacing and the Choice of Scores

Default polynomial contrasts treat successive levels as equally spaced. That may be appropriate for ordered bands deliberately defined as consecutive steps, but not for every ordered variable. Suppose ordered categories correspond to actual scores 1, 2, and 5. The order is clear, but the gap from 2 to 5 is larger than the gap from 1 to 2. R can construct polynomial contrasts using those scores: `contr.poly(3, scores = c(1, 2, 5))`. This changes the contrast columns and therefore the coefficient scale and interpretation.

Using scores is a modeling decision, not a way to make the data automatically more informative. If a variable genuinely has measured numerical values, fitting it as numeric may be more direct. If categories only have an order and no defensible distances, a numeric spacing assumption should not be smuggled in. A full ordered-factor model with all \(k-1\) columns can represent any set of level means; a reduced trend model makes stronger assumptions about the shape of the pattern.

Unbalanced Data, Model Scope, and Practical Checks

Orthogonality describes the contrast columns across levels, not necessarily the columns as they appear in the observed sample. When level counts differ, the sample model-matrix columns need not be orthogonal under the observation-level inner product. In a Gaussian model with unequal counts, coefficient estimates are not generally obtained by applying the simple balanced-data formula to unweighted level means. A generalized linear model also estimates coefficients on its link scale, using the response distribution and the model’s fitting procedure.

Before interpreting a trend, check four things: the factor’s level order; the scores used to build the contrasts; which contrast columns are included in the fitted model; and the response scale on which the coefficients operate. In R, inspecting `levels(tier)`, `contrasts(tier)`, and `model.matrix(...)` helps confirm the first three. Then identify the model family and link to establish the fourth.

Ordered contrasts are useful when the ordering is meaningful and the question concerns a trend across levels. They can summarize a pattern compactly, or impose a trend structure when only selected columns are included. They are not a substitute for checking the data, and they do not establish that a trend is practically important or statistically reliable. In particular, a positive linear component does not guarantee that every adjacent-level difference is positive when higher-order components are also present.