Why use a saturated model?

It gives a reference for assessing simpler restrictions.

Is it automatically the best forecasting model?

No; fitting the observed table does not establish forecasting performance.

Can zero-count cells cause difficulties?

Yes; log-scale parameter estimates may become infinite.

AI Glossary / Statistics and probability / Advanced

Saturated model

Definition

A saturated model has enough free parameters to reproduce all the independently variable features of the observed data being modeled. In the supplied glossary's log-linear context, it includes the main effects and interactions needed to fit every cell count of a contingency table. It provides a benchmark for assessing restrictions imposed by simpler models; exact fit does not establish predictive accuracy or causal explanation.

What it means

Suppose a table classifies requests by channel and resolution status. A simpler model might assume that channel and status are independent. A saturated model allows the complete observed pattern of association between them.

The word "saturated" concerns the model relative to the data representation and sampling constraints. Four cell counts from one hundred requests do not necessarily require one hundred free parameters. Conversely, writing many coefficient names does not guarantee saturation if constraints or redundant terms prevent independent adjustment of the fitted cells.

How it works

In a two-way log-linear model, express the logarithm of an expected cell count using an intercept, row effects, column effects and row-by-column interactions. With suitable identifiability constraints, the full model can reproduce the cell counts. Omitting the interactions gives the independence model.

Residual deviance compares a fitted model's maximized likelihood with that of the saturated model. A saturated fit has zero residual deviance relative to itself and no remaining goodness-of-fit degrees of freedom for that comparison.

Zero counts require care: fitted means at zero can correspond to infinite log-scale coefficients. Structural zeros also change the possible table and its dimension. Distinguish these issues from positive-count examples.

Three practical examples

1. Service analytics: compare a channel-by-outcome table with a model that assumes independence. The saturated model supplies the unrestricted reference for the likelihood comparison.

2. AI error analysis: classify errors by model version, task category and error type. A saturated table model permits all interactions among those factors, although a simpler model may communicate patterns more economically.

3. Survey analysis: fit a table of responses across demographic categories. Record which combinations are possible and which totals were fixed by the sampling design before counting free parameters.

Think of it this way

Think of adjusting each displayed cell separately until it matches the observed table. That does not explain what the next table will contain.

Worked example

Consider a fictional table of one hundred requests. Channel A has thirty resolved and ten unresolved requests. Channel B has twenty resolved and forty unresolved requests. The row totals are forty and sixty; both column totals are fifty.

For independent Poisson cell counts, a saturated model can estimate four cell means separately. Their fitted values are 30, 10, 20 and 40. One reference-coded log-linear representation is log(mu) = b0 + bR R + bC C + bRC RC, where R is zero for A and one for B, and C is zero for resolved and one for unresolved.

Set b0 = log(30), bR = log(2/3), bC = log(1/3) and bRC = log(6). For the B-unresolved cell, exponentiating the linear predictor gives 30 × (2/3) × (1/3) × 6 = 40. The other three cells similarly reproduce 30, 10 and 20.

The interaction is not optional if these particular counts are to be fitted exactly under this representation. The observed cross-product odds ratio is (30 × 40)/(10 × 20) = 6, whereas independence would require an odds ratio of one.

Under independence, the fitted count is row total times column total divided by the grand total. The four fitted counts become 20, 20, 30 and 30. They preserve the margins while failing to reproduce the observed cells.

The likelihood-ratio deviance is 2 times the sum of observed count multiplied by log(observed/fitted). Here it is 2[30 log(30/20) + 10 log(10/20) + 20 log(20/30) + 40 log(40/30)], approximately 17.2609. Natural logarithms are used. The independence model has three Poisson mean parameters, compared with four for saturation, leaving one degree of freedom for their comparison under regular conditions.

If instead the grand total of one hundred is fixed and the table is modeled multinomially, the four cell probabilities sum to one. The saturated probabilities have three free parameters, and independence has two. The comparison still has one degree of freedom. The sampling formulation changes the parameter count, not the observed table.

Illustrative case study

Fictional Birch Support fits the worked table and announces that its saturated model has "perfect accuracy."

A reviewer replaces that description with "reproduces the observed cell counts." Birch retains the saturated fit as a reference and reports how the independence model differs. It separately examines whether the association could reflect different request types entering each channel.

The team collects later data to assess whether patterns persist. Matching the original table cannot demonstrate performance on that later table, and the association does not by itself show that changing a request's channel would change its resolution. This fictional case illustrates interpretation rather than a real service result.

Common misconceptions and mistakes

"Zero deviance proves the model is true." Saturation produces exact in-sample fit by construction in this example.

"Saturation always means one parameter per individual." The table has four cells, despite summarizing one hundred requests.

"Every model with interactions is saturated." A multi-way table may require higher-order interactions that a pairwise-only model omits.

"Zero counts and impossible cells are equivalent." An unobserved but possible combination differs from a structural impossibility.

Business use

Specify the table, sampling assumptions and permitted cell combinations before interpreting model output. Report whether a result concerns fitted counts, an association parameter or prediction on later observations.

For AI-assisted reporting, prevent generated explanations from presenting zero residual deviance as an independent quality certificate. Keep the simpler model's assumptions visible. An independence comparison can reveal a pattern worth investigating, but operational decisions still need context about how requests entered the table and what outcomes matter.

Check your understanding

Why does independence predict twenty A-resolved requests?

Answer: Its row total forty times column total fifty is divided by one hundred.

What is the observed odds ratio?

Answer: Six.

Why are there only three free multinomial probabilities?

Answer: The fourth is determined by their sum being one.

Related terms

Log-linear models
Contingency table
Deviance (statistics)
Log-linear model
deviance
likelihood-ratio test
interaction
structural zero
model degrees of freedom

Learn next

Replace the table with its independence-fitted counts and recalculate the odds ratio and deviance. Then distinguish a lack of detected association from a causal conclusion.

Odds ratio

Sources

David C. Howell, Statistical Methods for Psychology glossary, Saturated model: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

Maarten Speekenbrink, Statistics: Data analysis and modelling, Chapter 12, model fit and log-linear models: https://mspeekenbrink.github.io/sdam-book/ch-generalized-linear-models.html

MASS package documentation, loglm, deviance and zero-cell limitations: https://stat.ethz.ch/R-manual/R-devel/library/MASS/html/loglm.html

Previous: Saturated array Next: Saturation (point operation)