Redundancy Between Totals and Their Components

Why a Total and Its Parts Can Be Redundant

A data set may record both an overall quantity and the amounts that make it up. For example, a clinic might record total visit time as well as minutes spent in consultation, testing, and administration. If those component times add exactly to the recorded total for every visit, the total does not provide a separate piece of predictor information: it is already determined by the components.

That identity matters in a regression model. Including the total and every component as predictors creates exact collinearity. The issue is not simply that coefficient estimates might be imprecise. The model cannot determine unique values for all those coefficients, because different coefficient combinations give exactly the same fitted values. The practical decision is therefore not “Which redundant variable should always be removed?” It is “Which representation expresses the comparison that matters?”

Definition: Suppose a recorded total $T_i$ is the exact sum of $k$ recorded components $X_{i1},\ldots,X_{ik}$ for every observation $i$: $$ T_i=\sum_{j=1}^{k}X_{ij}. $$ A regression that includes $T$ and all $k$ components as predictors has exact collinearity: one predictor column is an exact linear combination of the others. The regression coefficients in that full representation are not uniquely identifiable.

The Model and the Algebra of Nonunique Coefficients

Consider the usual additive regression model for a quantitative response $Y_i$:

Y_i=\beta_0+\beta_TT_i+\sum_{j=1}^{k}\beta_jX_{ij}+\varepsilon_i.

Here $\beta_0$ is the intercept, $\beta_T$ is the coefficient for the total, and $\beta_j$ is the coefficient for component $j$. In the usual setting for regression inference, the errors have mean zero, are independent, and have a common variance; normal errors are additionally assumed for exact small-sample $t$- and $F$-inference. These error assumptions concern inference about the model. The collinearity conclusion below follows from the predictor identity alone.

Substituting $T_i=\sum_j X_{ij}$ into the linear predictor gives:

\beta_0+\beta_TT_i+\sum_{j=1}^{k}\beta_jX_{ij} = \beta_0+\sum_{j=1}^{k}(\beta_T+\beta_j)X_{ij}.

Thus the data can at most determine the combined coefficients $\beta_T+\beta_j$ in this expression, not $\beta_T$ and every $\beta_j$ separately. The same point can be shown by changing the coefficients without changing any fitted value. For any number $c$, set $\beta_T'=\beta_T+c$ and $\beta_j'=\beta_j-c$ for every component. Then:

\beta_T'T_i+\sum_{j=1}^{k}\beta_j'X_{ij} = (\beta_T+c)T_i+\sum_{j=1}^{k}(\beta_j-c)X_{ij} = \beta_TT_i+\sum_{j=1}^{k}\beta_jX_{ij} +c\left(T_i-\sum_{j=1}^{k}X_{ij}\right) = \beta_TT_i+\sum_{j=1}^{k}\beta_jX_{ij}.

The last equality holds because the quantity in parentheses is zero for every observation. Least-squares fitted values can still be determined, but there is no unique set of coefficients in this full parameterization. A software package may drop a predictor or report a coefficient solution, but that does not make all the original coefficients separately estimable.

Worked Example: Seeing the Same Fit From Different Coefficients

Suppose three measured components $A$, $B$, and $C$ make up a total $T=A+B+C$. A hypothetical data row has $A=2$, $B=3$, and $C=5$, so $T=10$. Consider the linear predictor:

\hat{Y}=12+1.5T+0.5A-0.5B+2C.

For this row, substitution gives:

\hat{Y}=12+1.5(10)+0.5(2)-0.5(3)+2(5) =12+15+1-1.5+10 =36.5.

Now choose $c=1$ in the coefficient change above. The total coefficient becomes $2.5$, and each component coefficient decreases by $1$. The alternative predictor is $12+2.5T-0.5A-1.5B+1C$. Its value for the same row is:

$$ 12+2.5(10)-0.5(2)-1.5(3)+1(5) =12+25-1-4.5+5 =36.5. $$

The two coefficient sets differ, but the fitted value is identical. In fact, they give the same fitted value for every row satisfying $T=A+B+C$, not just for this example. The data cannot select one set of these four coefficients as uniquely correct.

Choose the Representation That Matches the Question

If the total and all its components are exact, there are two common and defensible representations. A model with the total alone asks how the response is associated with the overall amount, without distinguishing how that amount is divided among components. A model with the components alone allows their associations with the response to differ. Neither representation is universally preferable: the choice depends on which distinction the analysis is intended to describe.

Predictors included	Question represented	What the model does not separate
Total only	How is the response associated with a larger or smaller overall amount?	Differences in composition at the same total
All components, without the total	How is the response associated with each component, conditional on the other components?	A distinct additional effect of the total, beyond its components
Total and every component	No unique separate coefficient question is available in this parameterization	The coefficients are not all separately identifiable

Worked Example: Modeling Total Visit Time or Its Components

In a hypothetical service setting, total visit time is exactly consultation time plus testing time plus administrative time. Suppose a larger analysis reports the illustrative total-only model $\hat{Y}=3.0+0.020T$, where $Y$ is a recorded follow-up interval in days and $T$ is visit time in minutes. A visit with 80 total minutes has predicted follow-up interval:

\hat{Y}=3.0+0.020(80)=4.6\text{ days}.

The slope in this model describes a 0.020-day difference in predicted follow-up interval per additional minute of total visit time, within the model. It does not distinguish which part of the visit accounts for the added time, and it does not by itself establish that changing visit time causes a change in follow-up.

If the question instead concerns the components, suppose an illustrative component-only model is $\hat{Y}=3.0+0.010C+0.040Q+0.005A$, where $C$, $Q$, and $A$ are consultation, testing, and administrative minutes. For a visit with $C=30$, $Q=20$, and $A=30$, the components sum to $80$ minutes and:

\hat{Y}=3.0+0.010(30)+0.040(20)+0.005(30) =3.0+0.30+0.80+0.15 =4.25\text{ days}.

Each component coefficient describes a comparison that holds the other listed components fixed. In this model, an additional testing minute is associated with a 0.040-day difference in predicted follow-up interval, conditional on consultation and administrative minutes. The total is already the sum of these predictors, so adding it as a fourth predictor would make the coefficient representation nonunique. These illustrative models need not have the same predictions: they answer different modeling questions and are not asserted to be fits to a particular data set.

A Total With Only Some Components

Including a total with only a subset of its components does not automatically create exact collinearity. The remaining components can supply variation that is not determined by the included subset. But coefficient interpretation changes: a coefficient for the total is conditional on the included component values, and therefore corresponds to changes in the part of the total not held fixed.

Worked Example: Total Spending and Grocery Spending

Suppose total household spending $T$ is the sum of grocery spending $G$ and all other spending $O$, so $T=G+O$. A model includes $T$ and $G$, but not $O$:

\hat{Y}=50+0.20T+0.50G,

where $Y$ is a hypothetical monthly outcome measured in points and spending is measured in dollars. Replacing $T$ with $G+O$ gives:

\hat{Y}=50+0.20(G+O)+0.50G =50+0.70G+0.20O.

For a household with $G=60$ dollars and $O=40$ dollars, $T=100$ dollars, and the predicted outcome is:

\hat{Y}=50+0.20(100)+0.50(60)=100\text{ points}.

In the original parameterization, increasing $T$ by one dollar while holding $G$ fixed means that other spending increases by one dollar; the predicted difference is 0.20 points. Increasing $G$ by one dollar while holding $T$ fixed means replacing one dollar of other spending with grocery spending; the predicted difference is 0.50 points. The second comparison is a change in composition at a fixed total, not an increase in overall spending. The algebra makes these interpretations explicit.

Check the Identity Before Choosing Predictors

The exact-collinearity conclusion relies on a genuine row-by-row identity. Check that the total and components use the same units, cover the same categories, and refer to the same observations and time period. If a total includes unrecorded components, then it is not exactly the sum of the listed predictors. If values are rounded, an identity may hold before rounding but fail in the recorded data. A small discrepancy does not automatically make the chosen representation scientifically sensible; it does mean that the exact algebraic argument no longer applies as written.

State the intended comparison before deciding what to include. If only overall magnitude matters, consider representing it with the total. If component-specific differences matter, consider using the components without the total. If the scientific question is about composition at a fixed total, a model with the total and a carefully chosen subset of components may express that comparison, provided it avoids an exact dependency. In every case, coefficient interpretations should be checked against what is held fixed.

Common Mistakes

Keeping the total because it “controls for overall amount” while also including every component. If the components sum exactly to the total, the full predictor set is redundant and cannot have unique separate coefficients. Choose a nonredundant representation that matches the question.
Calling every coefficient in the redundant model meaningless. The full set of coefficients is not separately identifiable, but the fitted values and certain coefficient combinations can still be determined. Do not confuse nonunique coefficients with an inability to make any predictions.
Interpreting a component coefficient as an effect of adding that component alone when the total is fixed. In a model with all components and no total, holding the other components fixed while increasing one component also increases the total. A fixed-total change instead involves replacing some other component.
Assuming the total must always be dropped. A total-only model can be appropriate for an overall-amount question. A component-only model can be appropriate for a composition or component-specific question. The intended comparison determines the representation.

Key takeaway: When a total is exactly the sum of all its recorded components, do not include the total and every component together as if their coefficients could be separated. Use a total-based or component-based representation—or a justified partial-component representation—whose conditional comparisons match the question.

Check Your Understanding

Use the definitions and interpretations from this tutorial to answer the following questions.

A recorded total is the exact sum of four component predictors. What is the relationship between the total predictor column and the four component columns?
Why can two different coefficient sets in a model containing the total and all four components produce exactly the same fitted values?
A researcher wants to describe association with overall expenditure, regardless of its allocation among categories. Which broad representation is more directly aligned with that question: total only or all components only? Explain.
In a model containing a total and one component, the total is the sum of that component and all remaining spending. What does increasing the total while holding the included component fixed mean?
Why should a data analyst verify units, categories, and rounding before concluding that a recorded total is exactly the sum of its components?