1.6.9.1 The data🔗ℹ

The counts grow roughly like exp(0.3 x₁); x₂ is noise:

> (define X '((1.0 2.0) (2.0 1.0) (3.0 2.0) (4.0 1.0)               (5.0 2.0) (6.0 1.0) (7.0 2.0) (8.0 1.0)))

> (define y '(1 2 2 3 4 6 8 11))

1.6.9.2 Fitting🔗ℹ

> (define fit (poisson-fit X y #:lambda 0.2))

> fit

#<glmnet:poisson λ=0.2 dev=0.9859 nz=1/2>

> (coef fit)

'#(-0.026590920441019072 0.29703115403010283 0.0)

After the intercept, β₁ is close to the 0.3 the counts were built around, and the noise coefficient is exactly 0.0. The coefficients are on the log scale; exponentiated, β₁ is a rate ratio, the factor by which one more unit of x₁ multiplies the expected count:

> (exp (vector-ref (poisson-result-coefficients fit) 0))

1.3458572276715626

1.6.9.3 Predicting🔗ℹ

poisson-predict-mean applies the log link and returns the fitted mean for each row; it is predict with #:type 'response. Next to the observed counts:

> (for/list ([mu (in-list (poisson-predict-mean fit X))]              [count (in-list y)])     (list count mu))

'((1 1.3105412680428272)   (2 1.7638014377572937)   (2 2.3738249131831473)   (3 3.1948294166343585)   (4 4.299784261555074)   (6 5.786895725842328)   (8 7.788335438406571)   (11 10.481987541310053))

1.6.9.4 The response🔗ℹ

The response must be non-negative, but it need not be an integer, so rates such as events per hour fit too:

> (poisson-result-coefficients    (poisson-fit X '(0.5 1.0 1.2 1.5 2.0 3.1 4.0 5.6) #:lambda 0.1))

'#(0.29217287530369 0.0)

When observations cover different exposures (different lengths of time, different population sizes), R’s glmnet takes an offset of log(exposure). The bindings do not support offsets yet (#12).

1.6.9.5 When to use it🔗ℹ

Use the Poisson family for counts: defects per batch, visits per day, claims per policy. The model assumes a count’s variance equals its mean; counts that vary much more than that (overdispersion) are a poor fit for it.