Adjacent-Difference Coding for Category Means

Why Code Neighboring Differences?

When categories have a meaningful order, a comparison between neighboring categories can be useful: for example, comparing low, medium, and high exposure, or successive levels of a service. A usual reference-category coding compares every non-reference category with one selected baseline. Adjacent-difference coding instead makes each coefficient compare one category with the next. This can make a sequence of local changes easier to read.

The coding describes the fitted mean for each category. It does not require the changes between categories to be equal, and it does not treat category labels as measured numerical scores. The categories need an order so that “next” has a clear meaning.

The Model and Its Coding Columns

To specify the coding, define \(z_{ij}=1\) when observation \(i\) is in category \(j+1\) or any category after it, and define \(z_{ij}=0\) otherwise. There is one such column for each adjacent pair, so there are \(K-1\) columns. The model is

Here \(Y_i\) is the response for observation \(i\), \(\beta_0\) is the fitted mean for category 1, and \(\delta_j\) is the coefficient for the step from category \(j\) to category \(j+1\). For three categories, the coding columns are:

The columns are cumulative: once a category reaches a step, it has a value of 1 for that step and every earlier step. For instance, category 3 is past both the first and second transitions, so both indicators equal 1. The coefficient \(\delta_1\) is not the whole difference between category 1 and category 3; that overall difference is \(\delta_1+\delta_2\).

Why the Coefficients Are Adjacent Differences

The coding works because category \(r\) has value 1 for exactly the first \(r-1\) step columns. Its fitted mean is therefore the baseline plus the first \(r-1\) differences. In symbols, category 1 has no steps, while each later category accumulates the differences it has passed:

Subtract the expression for category \(r\) from the expression for category \(r+1\). All terms up to \(\delta_{r-1}\) cancel, leaving exactly the next step:

This also shows how to recover the coefficients from any set of fitted category means: set \(\beta_0=\mu_1\), then calculate each successive difference \(\delta_r=\mu_{r+1}-\mu_r\). Conversely, adding the differences back to the baseline recovers every category mean. There is no restriction that the differences have the same size or even the same sign.

Estimating the Category Means

In a one-factor model with one free fitted mean for each category, ordinary least squares estimates each category mean by that category’s sample mean. Adjacent-difference coding changes how those same fitted means are expressed as coefficients; it does not change the fitted mean for any category. Once the category means are estimated, the coding coefficients are their successive differences.

Reconstructing Means From a Coefficient Table

A fitted coefficient table may report a baseline and adjacent changes without listing the category means directly. To find the mean for a category, start at the baseline and add every step coefficient up to that category. A positive coefficient raises the fitted mean at that transition; a negative coefficient lowers it. The sum of several coefficients describes the total fitted-mean difference across those transitions.

What This Coding Does—and Does Not—Say

Adjacent-difference coding is useful when the scientific or practical question concerns transitions between neighboring ordered groups. Its coefficients retain the response’s units, so a coefficient can be described directly as a difference in fitted response units. It also gives a compact way to reconstruct all category means without treating the category labels as numerical measurements.

The coding itself does not impose a smooth trend, a constant increase, or a constant decrease. Each transition has its own coefficient. Nor does a positive or negative coefficient establish that the category change caused the response to change; the coefficient describes a modeled difference in fitted means. A causal interpretation requires an appropriate study design and additional justification.

Common Mistakes

• Reading a coefficient as a comparison with the baseline. The coefficient \(\delta_j\) compares category \(j+1\) with category \(j\), not category \(j+1\) with category 1. To compare with the baseline, add all intervening differences.
• Using a one-step indicator for only one category. In this coding, the column for step \(j\) equals 1 for category \(j+1\) and every later category. A later category accumulates the earlier steps too.
• Assuming every adjacent difference must be equal. No such condition is part of the coding. The values of \(\delta_1,\ldots,\delta_{K-1}\) can differ in size and sign.
• Reversing the subtraction order. With the convention here, \(\delta_j=\mu_{j+1}-\mu_j\). Reversing the order changes the sign and therefore changes the interpretation.