Building and Checking a One-Way Model Specification

From a Group Description to a Model

A one-way model gives a precise account of how a quantitative response is expected to vary across levels of one treatment or grouping factor. Its value is not just that it provides an equation. A well-specified equation makes clear which groups are being compared, what the intercept represents, and what unexplained variation is left to the errors. A mismatch in any of these can make an otherwise familiar equation describe the wrong question.

Suppose a hypothetical laboratory compares a numerical biomarker across several fixed treatment levels. There is one quantitative response, each experimental unit belongs to one treatment level, and the aim is to describe how the mean response differs among those levels. The one-way model is appropriate as a starting point when those features match the study design.

Definition: In the fixed-effects one-way model, $Y_{ij}$ is the response for experimental unit $j$ at treatment level $i$, where $i=1,\ldots,a$ and $j=1,\ldots,n_i$. The model is $$ Y_{ij}=\mu+\tau_i+\varepsilon_{ij}. $$ Here $\mu$ is the grand mean under the chosen sum-to-zero convention, $\tau_i$ is the fixed effect of treatment level $i$, and $\varepsilon_{ij}$ is the error for that unit. The effects satisfy $\sum_{i=1}^{a}\tau_i=0$.

The corresponding population mean at level $i$ is $\mu_i=E(Y_{ij})=\mu+\tau_i$. With the stated constraint, $\mu$ is the equally weighted average of the $a$ treatment-level means, $\mu=(1/a)\sum_{i=1}^{a}\mu_i$. Thus each $\tau_i=\mu_i-\mu$ is a deviation from that average. This definition gives the intercept and effects a unique interpretation; without a constraint, many combinations of intercept and effects could produce the same group means.

The model’s standard error assumptions are that errors have mean zero, are independent across experimental units, are normally distributed, and have a common variance at every treatment level: $$ \varepsilon_{ij}\sim N(0,\sigma^2). $$ These statements concern the random deviations around the treatment means, not the fixed treatment effects. In an actual study, independence is principally justified from the design and knowledge of the experimental units. Residual plots can help assess patterns and differences in spread; a small data set cannot establish normality merely because no striking departure is visible.

Check the Parts of the Specification

Before calculating anything from a model, check that each symbol has a role consistent with the study. The $a$ levels are the particular treatment levels being compared, not a second factor hidden inside the equation. The $n_i$ units at level $i$ are the units that provide the observations. If several measurements come from the same unit, those measurements are not automatically independent experimental units.

The treatment effects are fixed quantities for the levels in the question. They describe how those levels’ population means differ from the grand mean; they are not the random noise around each individual response. Conversely, the error $\varepsilon_{ij}$ is an unobserved random deviation, not another treatment effect. Once data are observed, the fitted value for a unit at level $i$ is the fitted mean for that level, and the residual is the observed response minus that fitted value. A residual is a data-based estimate of an error, not the error itself.

Name the response and units.
Confirm that the response is quantitative and that each indexed observation corresponds to the experimental unit identified by the design.

Name the fixed levels.
Check that one treatment or grouping factor accounts for the group labels and that the levels in the model are the levels the question concerns.

Check the mean structure.
Make sure $\mu+\tau_i$ gives the intended mean at level $i$, and that the effects obey the stated constraint.

Check the errors against the design and data.
Consider independence, normality, and common variance. A model equation does not make these assumptions true.

Worked Example: Converting Three Group Means to Effects

In a hypothetical laboratory comparison, biomarker concentrations (in units per litre) are recorded for two independent experimental units at each of three fixed treatment levels. The observations are 10 and 12 for A, 14 and 16 for B, and 18 and 20 for C. Write the model’s mean structure and check its intercept and effects.

The observed group means are $$ \bar y_A=(10+12)/2=11,\qquad \bar y_B=(14+16)/2=15,\qquad \bar y_C=(18+20)/2=19. $$ Under the sum-to-zero convention, the grand mean of the three level means is $$ \mu=(11+15+19)/3=15. $$ The treatment effects are the group means minus this grand mean: $$ \tau_A=11-15=-4,\qquad \tau_B=15-15=0,\qquad \tau_C=19-15=4. $$ Their sum is $-4+0+4=0$, as required. The fitted mean structure is $Y_{ij}=15+\tau_i+\varepsilon_{ij}$: it gives means 11, 15, and 19 units per litre for A, B, and C. The negative effect for A means its population mean is 4 units per litre below the average of the three treatment-level means; it does not mean that every A observation is below every B observation.

For example, the observed value 10 at level A is 1 unit per litre below that group’s observed mean of 11. That difference is a residual when the group mean is fitted; the corresponding model error is unobserved. The design description supports treating the two units at each level as independent only if they are genuinely separate experimental units. These six observations alone do not establish normal errors or equal error variance.

Why the Constraint and the Group Sizes Matter

The equation for a group mean is $\mu_i=\mu+\tau_i$. If no rule fixes the intercept and effects, adding a constant to $\mu$ and subtracting it from every $\tau_i$ leaves all the group means unchanged. The sum-to-zero constraint removes this ambiguity. It also gives a direct way to check a proposed set of parameters.

Justification: If $\mu_i=\mu+\tau_i$ and $\sum_{i=1}^{a}\tau_i=0$, then $$ \frac{1}{a}\sum_{i=1}^{a}\mu_i =\frac{1}{a}\sum_{i=1}^{a}(\mu+\tau_i) =\mu+\frac{1}{a}\sum_{i=1}^{a}\tau_i =\mu. $$ Therefore $\tau_i=\mu_i-\mu$, and the effects sum to zero.

This convention gives every treatment level equal weight in defining the model intercept, whether or not the sample sizes are equal. In an unbalanced data set, the mean of all observed responses weights each level according to its sample size and need not equal the model intercept. Confusing those two averages can make a correctly calculated group mean look like an inconsistent model parameter.

Worked Example: Checking an Unbalanced Design

A hypothetical study records responses from 2 units at level A, 4 at level B, and 6 at level C. The recorded values are 8 and 12 for A; 9, 11, 10, and 10 for B; and 12, 14, 14, 16, 14, and 14 for C. Use the sum-to-zero convention to find the group means, intercept, and effects.

The group means are $$ \bar y_A=(8+12)/2=10,\quad \bar y_B=(9+11+10+10)/4=10,\quad \bar y_C=(12+14+14+16+14+14)/6=14. $$ There are three fixed treatment levels, so the model intercept is $$ \mu=(10+10+14)/3=34/3\approx11.33. $$ The effects are $$ \tau_A=10-34/3=-4/3,\quad \tau_B=10-34/3=-4/3,\quad \tau_C=14-34/3=8/3. $$ Their sum is $-4/3-4/3+8/3=0$, and adding each effect to $34/3$ recovers the corresponding group mean.

The mean of all 12 recorded responses is instead $$ (20+40+84)/12=144/12=12. $$ It differs from $34/3$ because the all-observation average weights the levels by their sample sizes, while the model intercept averages the three level means equally. This is not a contradiction: they answer different averaging questions. The specification should state which convention defines the intercept, then use it consistently.

Correcting Common Specification Mistakes

Some proposed equations reproduce the group means but do not use the chosen effects convention. Others attach the right-looking equation to the wrong units or treat a fixed effect as if it were random error. The following example checks all three distinctions using a small hypothetical data set.

Worked Example: Repairing an Intercept and an Error Term

Suppose a hypothetical experiment records one response in milligrams for each of three independent units at each of three fixed treatment levels. The observations are 48, 50, 52 at A; 53, 55, 57 at B; and 58, 60, 62 at C. A proposed specification uses $\mu=50$, $(\tau_A,\tau_B,\tau_C)=(0,5,10)$, and calls the treatment effects “the errors.” Check and correct it.

The group means are $$ \bar y_A=(48+50+52)/3=50,\quad \bar y_B=(53+55+57)/3=55,\quad \bar y_C=(58+60+62)/3=60. $$ The proposed parameters reproduce these means, but the effects sum to $0+5+10=15$, not zero. They are a reference-level description, not effects under the sum-to-zero convention. The grand mean and effects under the required convention are $$ \mu=(50+55+60)/3=55,\qquad (\tau_A,\tau_B,\tau_C)=(50-55,55-55,60-55)=(-5,0,5). $$ These effects sum to zero, and $\mu+\tau_i$ gives 50, 55, and 60 milligrams.

The errors are not the treatment effects. If the group means are fitted, the observed deviations from those means are $(-2,0,2)$ milligrams at each of A, B, and C. These are residuals; the model errors are the corresponding unobserved random deviations. The fixed effects describe between-level mean differences, while the errors describe unit-to-unit departures around each level’s mean.

The design states that the nine experimental units are independent, which supports the independence assumption if no units share an unmodeled source of dependence. The residuals have the same observed spread at each level, but this small pattern is not proof of equal population error variance or normality. Those assumptions must be considered using the design and appropriate residual diagnostics; they cannot be guaranteed by choosing parameter values.

Common Pitfalls and Why the Checks Matter

Using the wrong unit as an observation. If repeated readings are taken from one unit, counting them as independent units can misstate the design. Identify the unit to which the treatment was assigned before indexing $j$.
Calling effects errors. Effects are fixed level-specific deviations in the expected response. Errors are random departures for individual units.
Checking the sum incorrectly. Under the sum-to-zero convention, check $\sum_i\tau_i=0$. Do not substitute a sample-size-weighted sum unless a different constraint has explicitly been chosen.
Assuming an equation verifies assumptions. Writing independent, normal, equal-variance errors states a model condition; it does not demonstrate that the study or observations satisfy it.
Interpreting a treatment effect as a universal individual difference. An effect describes a difference between population means relative to the grand mean, not a guaranteed difference for every pair of units.

These checks matter because a one-way model is a description tied to a design, not merely an equation that can be fitted to any list of numbers. Confirming the response, levels, units, mean constraint, and error assumptions makes the model interpretable and clarifies what conclusions it could support. The model alone does not show that treatment caused a difference; that interpretation also depends on how units were assigned and on the study design.

Check Your Understanding

For each question, state the reasoning that supports your answer.

Under the sum-to-zero convention, group means are 6, 9, and 12 seconds. Find the intercept and the three treatment effects, and check the constraint.
There are two treatment levels with unequal sample sizes. Explain why the average of all observations need not equal the intercept under the equally weighted treatment-level convention.
A researcher labels the difference between a group mean and the grand mean as the error. What is wrong with this description?
A treatment is assigned to each patient, and three readings are taken from each patient. What must be clarified before treating all readings as independent observations in a one-way model?
In a small experiment, residual spreads look similar across groups. Does this establish equal error variance? Explain what the plot can and cannot show.