WOLFRAM|DEMONSTRATIONS PROJECT

Comparing Binomial Generalized Linear Models

​
link functions
logit
probit
cauchit
identity
log-log
complementary log-log
log
log-complement
odds power
linear predictor
linear
quadratic
cubic
quartic
link function
residual deviance
logit
1.51743
Generalized linear models are models of the form

y
i
=
-1
g
(
β
0
+
β
1
f
1i
+
β
2
f
2i
+…)
, where
g
is an invertible function called the link function and the
f
j
are basis functions of one or more predictor variables. The term
β
0
+
β
1
f
1i
+
β
2
f
2i
+…
is linear in the
β
i
and is referred to as the linear predictor. The value

y
i
is the predicted response for the
th
i
observed response
y
i
, and the
y
i
are assumed to be independent observations from the same exponential family of distributions. When the exponential family is the binomial family, the success probability
p
is modeled.
This Demonstration fits binomial models with various common link functions. Check the boxes next to the named link functions to fit models with those links. Select a linear predictor to choose the argument of
-1
g
in the model. The linear predictors are taken to be polynomials in a single predictor variable
x
, so for instance, with a quadratic linear predictor, the model is
-1
g

β
0
+
β
1
x+
β
2
2
x

.
Mouse over a fitted curve to see the functional form of the model. The residual deviances for the models are included in a table for comparison.