Why the Representation Matters
Suppose a model uses two categorical predictors, such as service channel and contract type. There are two natural ways to represent them. A model can keep the factors separate and assign each its own effect, or it can treat each observed pairing—such as “online and annual”—as a single category. These representations do not make the same assumptions about the outcome.
The distinction matters for both interpretation and model size. Separate additive effects summarize how each predictor is associated with the outcome while holding the other fixed. A combined category can describe every pairing separately, but it uses more parameters and no longer summarizes the two factors through separate effects. The comparison below uses reference-level coding to make those differences explicit.
Models and Parameter Counts
First consider a model with an identity link, such as a linear model for a continuous response \(Y\). Let \(A=1\) and \(B=1\) denote the chosen reference levels. Under reference-level coding, the additive model is
The intercept \(\beta_0\) is the expected response for the reference pair. The terms \(\alpha_j\) and \(\gamma_k\) are differences from the reference level of their respective factors. Because the model has no interaction, the difference associated with changing \(A\) is the same at every level of \(B\), and vice versa.
If all levels are represented and the model matrix has full rank, the additive model has \(1+(J-1)+(K-1)=J+K-1\) coefficients, including the intercept. A full interaction model adds the \(A\)-by-\(B\) terms and has \(JK\) coefficients in total. It can assign a distinct fitted mean to every pairing. A combined-category factor with \(M\) observed pairs has \(M\) coefficients, including its intercept representation, when fitted with reference-level coding.
| Representation | What it describes | Number of coefficients under the stated conditions |
|---|---|---|
| Additive factors | Separate effects for \(A\) and \(B\); no interaction | \(J+K-1\) |
| Full interaction | A separate fitted mean for each of the \(JK\) possible pairs | \(JK\), when all pairs are represented |
| Combined category | A separate fitted mean for each observed pair | \(M\), where \(M\) is the number of observed pairs |
When all \(JK\) pairs are present and no other model terms differ, the full interaction model and the combined-category model describe the same set of fitted means. They are alternative parameterizations of a model with one freely estimated mean per pair. The additive model is more restrictive: it requires the pair means to be explainable as a sum of separate factor effects. The parameter counts assume the design is estimable; a lack of data for some levels or other rank deficiencies can alter which coefficients can be estimated.
Why Additivity Means a Zero Difference in Differences
For two factors with two levels each, write the four cell means as \(m_{11},m_{21},m_{12},m_{22}\). Under the additive identity-link model,
Subtract the first row’s factor-\(A\) difference from the second row’s factor-\(A\) difference. The result must be zero:
Equivalently, \(m_{22}=m_{21}+m_{12}-m_{11}\). This is the identity-link form of the no-interaction condition. It explains why an additive model can predict an unobserved pair from the separate effects, provided those effects are estimable. With a different link, additivity is instead a condition on the linear predictor, not necessarily on the response-scale means.
Worked Example: Finding the Departure From Additivity
A hypothetical insurer studies average claim-handling cost, in hundreds of dollars, by service channel and contract type. The channels are desk and mobile; the contract types are standard and extended. The four observed cell means are $42, $48, $50, and $64, respectively: desk-standard, mobile-standard, desk-extended, and mobile-extended. Use desk and standard as the reference levels. Find the additive prediction for mobile-extended, compare it with its observed mean, and count the coefficients in the additive and combined-category representations.
Method. Calculate the reference mean and each separate factor difference, then add those differences to predict the joint pair. This avoids treating the observed mobile-extended mean as if it were forced to equal an additive prediction.
The reference mean is \(\beta_0=42\). The mobile difference at standard contracts is \(48-42=6\), so \(\alpha_{\text{mobile}}=6\). The extended-contract difference at the desk channel is \(50-42=8\), so \(\gamma_{\text{extended}}=8\).
The observed mobile-extended mean is 64, so the departure from the additive prediction is \(64-56=8\) hundreds of dollars, or $800. The same result follows from the difference in differences: \((64-50)-(48-42)=14-6=8\). The additive representation has \(1+(2-1)+(2-1)=3\) coefficients. A combined factor has four observed pairs and therefore four fitted cell means. Check: the additive formula gives \(48+50-42=56\), matching the prediction above. The common mistake is to report 64 as the additive prediction; 64 is the observed cell mean, while 56 is the prediction constrained by the separate effects.
When a Combined Category Is Useful
A combined category is often easy to explain when the joint groups are the substantive units of interest. For example, a report might need a distinct estimate for each product-and-channel pairing rather than separate statements about product and channel. The combined factor does not assume that a channel difference is constant across products. Its cost is that each observed pair gets its own parameter, so estimates are not linked by an additive structure.
By contrast, the additive representation is concise and directly describes separate effects. It can be useful when that structure is scientifically or operationally plausible. It also determines fitted values for pairs through the sum of the factor effects rather than estimating every pair independently. That is a structural assumption, not merely a shorter way to write the same unrestricted model.
If some possible pairs are absent, the distinction deserves particular care. A combined factor has levels only for pairs present in the data and cannot assign a fitted cell mean to an unobserved pair as a level of that factor. An additive model may imply a prediction for that pair if its factor effects are estimable, but that prediction relies on the additive assumption. It is not an observed cell mean.
Worked Example: An Unobserved Pair and the Parameter Count
A hypothetical service operation records average resolution time in minutes for three request types—account, payment, and delivery—and two service modes—self-service and assisted. The observed means are: account, 10 minutes in self-service and 14 in assisted; payment, 13 in self-service and 17 in assisted; delivery, 16 in self-service, with no delivery-assisted observations. Use account and self-service as references. Compare the additive and combined-category parameter counts, and calculate the additive prediction for delivery-assisted.
Method. Estimate the reference mean, the two request-type differences at self-service, and the assisted-mode difference for the observed reference request type. Then apply those effects to the unobserved pair. The key assumption is that the assisted-mode difference is common across request types.
The reference mean is 10. The payment difference is \(13-10=3\), and the delivery difference is \(16-10=6\). The assisted-mode difference for account requests is \(14-10=4\). The additive model therefore predicts
There are \(1+(3-1)+(2-1)=4\) additive coefficients. There are five observed pairs, so a combined-category factor has five coefficients, one per observed pair; it has no delivery-assisted level in this data. The additive model supplies the 20-minute prediction only by imposing its common-mode-difference assumption. Check: the observed assisted-minus-self-service differences are \(14-10=4\) minutes for account and \(17-13=4\) minutes for payment, consistent with the additive pattern among the observed pairs. The common mistake is to claim that the combined-category model also estimates a delivery-assisted mean of 20; it does not, because that pair is absent.
The Same Comparison on a GLM Linear-Predictor Scale
For a generalized linear model, the representation comparison applies to the linear predictor. If \(g\) is the link function and \(\mu_{jk}=E[Y\mid A=j,B=k]\), an additive GLM has
Thus, the no-interaction condition is additivity of \(g(\mu_{jk})\), not necessarily additivity of \(\mu_{jk}\). For a logistic model, for instance, the log odds are additive under this model; probabilities need not satisfy the zero difference-in-differences equation. A combined-category GLM can instead assign each observed pair its own linear-predictor value. It still uses the same response distribution and link; only the factor representation changes.
Worked Example: Additive Effects in a Logistic Model
A hypothetical team models whether a customer renews a subscription using a logistic regression with two binary factors: reminder status (no reminder or reminder) and plan type (basic or premium). For a no-reminder, basic-plan customer, the fitted renewal probability is 0.20. The fitted odds ratio for reminder versus no reminder is 2, and the fitted odds ratio for premium versus basic is 3. Under a model with additive effects on the log-odds scale and no interaction, calculate the predicted probability for a reminded premium-plan customer. Compare the parameter counts with a combined category.
Method. Convert the reference probability to odds, multiply by the two odds ratios, and convert the resulting odds back to probability. This avoids incorrectly adding probability differences in a logistic model.
The reference odds are \(0.20/(1-0.20)=0.20/0.80=0.25\). Additivity on the log-odds scale means the two effects multiply the odds, so the predicted joint odds are \(0.25\times2\times3=1.50\). Converting odds \(o\) to probability \(o/(1+o)\) gives
The additive logistic model has \(1+1+1=3\) coefficients. A combined factor for the four observed pairs has four coefficients. Check: the reference log odds are \(\log(0.25)\); adding \(\log(2)\) and \(\log(3)\) gives \(\log(1.50)\), whose corresponding probability is 0.60. The common mistake is to add probability changes of 0.20 and 0.40 to the reference probability. Those changes do not represent the stated odds ratios, and the logistic model specifies additivity on the log-odds scale.
A Practical Decision Sequence
For an identity-link model, interpret additive effects on the mean scale. For a GLM, identify the link and interpret additivity on the linear-predictor scale.
Choose separate factors when separate effects are meaningful and a common effect across the other factor’s levels is a defensible assumption. Choose a combined category when the observed pairings are the intended units of comparison.
For full-rank additive factors, count \(J+K-1\). For a combined factor, count its \(M\) observed pairs. If all pairs are present, a full interaction also has \(JK\) coefficients.
Separate an observed cell mean from a model-implied prediction. If a pair is absent, a combined factor has no level for it; any additive prediction depends on the additive structure.
A parameter count describes the flexibility built into a representation; it does not decide which representation is appropriate. The substantive question is whether separate effects adequately describe the relevant pairings. When reporting results, state both the representation and the scale on which its effects are additive. This is aligned with the SOA Exam PA (Predictive Analytics) syllabus.
Check Your Understanding
For each problem, identify the model structure as well as the requested calculation.
- A two-by-two identity-link model has cell means 20 for \(A_1B_1\), 24 for \(A_2B_1\), and 25 for \(A_1B_2\). Under additivity, predict \(A_2B_2\). If its observed mean is 32, calculate the difference from the additive prediction and state the additive and combined-category parameter counts. (Answer: additive prediction \(=24+25-20=29\); observed-minus-predicted difference \(=3\); counts \(=3\) and \(4\).)
- Three request classes have two service modes. The observed means are 7 and 9 for class 1, 10 and 12 for class 2, and 13 for class 3 in the first mode; class 3 in the second mode is unobserved. Under an additive identity-link model, predict the missing mean and give the additive and combined-category coefficient counts. (Answer: \(13+(9-7)=15\); counts \(=1+(3-1)+(2-1)=4\) and \(5\) observed pairs.)
- A logistic model has reference odds \(1/4\), and two factor effects have odds ratios 3 and 2. Under log-odds additivity, find the joint probability and compare the number of coefficients with a combined factor when all four pairs are observed. (Answer: joint odds \(=(1/4)(3)(2)=3/2\), probability \(=(3/2)/(1+3/2)=0.6000\); coefficient counts \(=3\) and \(4\).)
- Two factors have three levels each, and all nine pairs are observed. How many coefficients are in the additive model and in the full interaction or equivalent combined-category model? What is the modeling distinction? (Answer: additive \(=1+2+2=5\); full interaction and combined category \(=9\); the additive model constrains pair effects to sums of separate effects, while the latter allows a separate fitted mean for each pair.)
- Two factors each have two levels, but only three pairs are observed. A combined-category factor is used. How many coefficients does it have, and can it estimate a mean for the unobserved pair as a category level? (Answer: \(3\) coefficients, one for each observed pair; no, it has no level for the unobserved pair.)