Cell-Means Coding Without an Intercept

Introduction

A categorical predictor with several levels can be represented in more than one way. In reference-category coding, one category is represented by the intercept and the remaining categories by indicators. In cell-means coding, the model instead uses one indicator for every category and has no intercept. The result is direct: each regression coefficient is the fitted mean for one category.

This is a way to write a factor-only model, not a claim that the categories have a numerical order or that their labels are quantitative. The observations are still grouped by category; the coding simply assigns one coefficient to each group.

The No-Intercept Indicator Model

Suppose each observation belongs to exactly one of \(K\) categories. Let \(y_i\) be the response for observation \(i\), and define \(D_{ij}=1\) if observation \(i\) belongs to category \(j\), and \(D_{ij}=0\) otherwise. The cell-means model is

There is no separate constant or intercept term. Because every observation belongs to exactly one category, exactly one of its indicators is 1 and all the others are 0. Thus the model's mean for an observation in category \(j\) is \(\beta_j\). The error \(\epsilon_i\) describes the difference between the observed response and that category's model mean.

The corresponding design row has a 1 in the column for the observation's category and 0 in every other category column. For three categories, the rows are \((1,0,0)\), \((0,1,0)\), or \((0,0,1)\). Multiplying any such row by the coefficient vector selects just one coefficient as the fitted value.

Assumptions and What Is Being Estimated

The model describes the response mean using category membership alone. In particular, observations in a category share the same model mean, and no other predictor is included. The least-squares procedure chooses the coefficients to minimize the sum of squared differences between observed and fitted responses. This calculation does not require normally distributed errors.

For the usual regression model used in standard inference, errors are assumed to have mean zero, to be independent across observations, and to have a common variance. Normal errors are additionally assumed when exact small-sample \(t\)- and \(F\)-inference is required. These conditions concern the error structure, not the indicator coding itself. The examples below estimate fitted means, not significance tests, so they do not need a normality assumption.

The coefficients are estimable only for categories represented by observations in the data: a category with no observations has no data-based mean to estimate. In applying the model, also consider whether grouping observations by the factor alone is a useful description of their response means. If relevant systematic differences have been left out, the fitted category means may not answer the scientific question of interest.

Why the Coefficients Are the Category Means

The direct interpretation follows from the least-squares calculation. Let category \(j\) contain \(n_j\) observations, with response values \(y_{1j},\ldots,y_{n_jj}\). Each of those observations has fitted value \(\beta_j\). The residual sum of squares for the whole model can therefore be grouped by category:

Each \(\beta_j\) appears only in the part of the sum for category \(j\), so minimizing the full sum amounts to minimizing each category's sum separately. Differentiate that part with respect to \(\beta_j\), set the derivative to zero, and solve:

This stationary point is a minimum: the second derivative of the category's sum of squares is \(2n_j\), which is positive when the category contains at least one observation. Consequently, for every observed category, the least-squares coefficient is its sample mean. The fitted values within that category are all this mean.

The calculation also explains why unequal category sizes cause no difficulty for estimating the category means. Each coefficient minimizes the squared residuals for its own category. A category with more observations contributes more residuals to the overall sum, but it does not change the formula for any other category's fitted mean.

Reading a Coefficient and a Fitted Value

Once the coefficients have been estimated, use the indicator row to obtain a fitted value. An observation in category \(j\) has fitted value \(\hat{y}_i=\hat{\beta}_j\). Its observed response may be above or below that value; the difference \(y_i-\hat{y}_i\) is its residual. A coefficient is therefore a fitted category mean, not an individual response and not a difference from another category.

For instance, if the response is blood pressure measured in mmHg, a coefficient of 122 means a fitted mean of 122 mmHg for the category named by that coefficient. It does not, by itself, mean that the category is 122 mmHg higher than a comparison category. To discuss differences between categories requires comparing their fitted means; the coefficients themselves are on the response scale.

Why This Coding Matters

Cell-means coding is useful when the main quantities of interest are the fitted means for each category. It gives one coefficient per category, keeps the coefficients in the response's units, and makes a category's fitted value easy to read directly from the model. It can also make a model's output easier to connect to a table of group means.

The absence of an intercept is essential to that interpretation. If a separate intercept were added while retaining every category indicator, the intercept and indicators would not be uniquely identifiable: adding a constant to the intercept and subtracting it from every category coefficient would leave every fitted value unchanged. With no intercept, each category's indicator selects its own coefficient, and the coefficients are separately estimated as long as each category appears in the data.

This is not the same as fitting a numerical predictor with its line constrained to pass through zero. In cell-means coding, the indicators are category-membership variables. Since one and only one is 1 for each observation, each category has its own freely estimated mean; the coding does not impose a scientifically meaningful zero response for any category.

Common Mistakes

• Calling a coefficient a difference. A no-intercept coefficient is the fitted mean for its category. A difference requires subtracting one category mean from another.
• Leaving out a category indicator. With no intercept, every modeled category needs its own indicator. Otherwise, observations in the omitted category would have no category coefficient selecting their fitted mean.
• Adding an intercept while keeping all indicators. That makes the parameters redundant because the indicators sum to 1 for every observation. Use the specified no-intercept model for cell-means coding.
• Interpreting the fitted mean as every individual's response. The coefficient is the fitted value for the category; observations can differ from it, and their differences are residuals.
• Using a coefficient for a category with no observations. Its mean is not determined by the data, so that category-specific coefficient cannot be estimated from this model.