Nerdish.Org

By Chapter 14, statistics reaches a really beautiful point.

We already learned:

• GLMs → Different distributions
• GAMs → Nonlinear relationships
• Mixed Models → Random effects
• GEE → Population averages

But real data often has both:

• Nonlinear behavior
• Grouped behavior

Example:

Sales may:

• Increase nonlinearly with customer count
• Vary by retailer

That is where Generalized Additive Mixed Models (GAMMs) come in.

What Is a GAMM?

A GAMM combines a GAM with a Mixed Model:

$$
\text{GAMM}=\text{GAM}+\text{Mixed Effects}
$$

One of the most practical modern modeling frameworks.

The Problem GAM Solves — But Not Completely

Suppose we model:

$$
\log(\mu)=\beta_0+f(\text{CustomerCount})
$$

Great.

Customer count can affect sales nonlinearly.

But now suppose:

• Retailer A consistently sells more.
• Retailer B consistently sells less.

The smooth curve cannot explain those retailer-specific differences.

We need random effects.

The Problem Mixed Models Solve — But Not Completely

Suppose we model:

$$
\log(\mu)=\beta_0+\beta_1\text{CustomerCount}+u_j
$$

Now retailers can differ.

However, the relationship between customer count and sales is still assumed to be linear.

That assumption may fail.

GAMM Combines Both

The general GAMM model is:

$$
g(\mu)=X\beta+f(X)+Zu
$$

Interpretation:

• $$X\beta$$ → Fixed effects
• $$f(X)$$ → Smooth nonlinear effects
• $$Zu$$ → Random effects

Everything together in one model.

Breaking Down the Model

Fixed Effects

$$
X\beta
$$

Represent overall trends.

Smooth Effects

$$
f(X)
$$

Represent nonlinear relationships.

Random Effects

$$
Zu
$$

Represent group-level variation.

Example: Customer Count and Retailers

Suppose sales depend on:

• Customer count
• Retailer

Model:

$$
\log(\mu)=\beta_0+f(\text{CustomerCount})+u_{\text{Retailer}}
$$

Interpretation:

• Customer count → nonlinear effect
• Retailer → individual adjustment

Why Not Just Use Polynomial Terms?

You could use:

$$
x+x^2+x^3
$$

However, polynomials can create:

• Instability
• Oscillation
• Difficult interpretation

Splines are usually smoother and more stable.

Random Intercepts

The most common GAMM specification:

$$
g(\mu)=\beta_0+f(X)+u_j
$$

Meaning:

• Same curve for everyone
• Different baseline levels

Example:

The shape of the curve stays the same.

Random Slopes

A more flexible model:

$$
g(\mu)=\beta_0+f(X)+(\beta_1+u_j)X
$$

Now both:

• Baseline level
• Predictor effect

can vary by retailer.

Example:

Smooth Interactions

Now GAMMs become very powerful.

Suppose month changes the effect of customer count.

Model:

$$
g(\mu)=\beta_0+f(\text{CustomerCount},\text{Month})
$$

Now the shape of the curve changes over time.

Tensor Product Smooths

Suppose:

• Month is nonlinear
• Customer count is nonlinear

Model:

$$
f(\text{CustomerCount},\text{Month})
$$

This creates smooth surfaces rather than simple curves.

Visualization

Instead of a 2D curve:

Sales
 ^
 |
 |
 +-------------------->

You obtain a 3D response surface:

• Customer Count
• Month
• Sales

all modeled simultaneously.

Example: Inventory Forecasting

Suppose you model:

• Month
• Customer Count
• Year
• SKU

Model:

$$ \log(\mu) \beta_0+ \beta_1\text{Month} +f(\text{CustomerCount})+\beta_2\text{Year}+u_{\text{SKU}}$$

Interpretation:

This is essentially a GAMM.

Inventory Forecast Workflow

Step 1

Fit a GAMM.

↓

Step 2

Forecast future periods.

↓

Step 3

Aggregate demand.

↓

Step 4

Adjust for lead time.

↓

Step 5

Add safety stock.

This closely resembles enterprise forecasting systems.

Why Random Effects Help Forecasting

Without random effects:

all SKUs share information equally.

With random effects:

weak SKUs borrow information from stronger SKUs.

This is known as:

partial pooling.

One of the most powerful ideas in modern statistics.

Shrinkage Happens Again

Suppose observed sales are:

$$
50
$$

The model believes this month is unusually high.

The GAMM estimate may become:

$$
35
$$

Random effects help stabilize estimates.

GAMM vs GAM

GAMM vs GLMM

Estimation

Fitting GAMMs is harder than fitting GAMs.

We must estimate:

• Smoothing parameters
• Random effects
• Fixed effects

simultaneously.

A common approach is:

REML

REML

REML stands for:

Restricted Maximum Likelihood

It generally provides better variance estimation than ordinary maximum likelihood.

Very common in mixed-model applications.

When Should You Use GAMMs?

Use GAMMs when:

• Measurements are repeated
• Predictors are nonlinear
• Data is hierarchical
• Random effects matter

Examples:

• Inventory forecasting
• Retailer sales analysis
• Customer behavior modeling
• Demand forecasting

When Should You Avoid GAMMs?

Avoid GAMMs when:

• Datasets are very small
• Relationships are clearly linear
• Simple interpretation is the primary goal

Real Business Applications

Examples include:

• Customer growth modeling
• SKU demand forecasting
• Seasonal inventory planning
• Price elasticity analysis
• Energy demand forecasting
• Website engagement modeling

Inventory Problem Revisited

Suppose your model already contains:

• Month
• Spline(Customer Count)
• Year
• Lead Time

You are already close to a GAM.

Add:

$$
u_{\text{SKU}}
$$

and the model becomes a full GAMM.

Very realistic for inventory forecasting.

Chapter 14’s Big Lesson

Reality is both:

• Nonlinear
• Heterogeneous

Good models must handle:

• Curves
• Groups
• Uncertainty

simultaneously.

Final Thought

Before Chapter 14, you ask:

“What curve fits the data?”

After Chapter 14, you ask:

“What curve fits each group?”

That shift moves statistics from:

one-size-fits-all modeling

to

adaptive learning across populations.