Scoring algorithm explained

Scoring algorithm, also known as Fisher's scoring,[1] is a form of Newton's method used in statistics to solve maximum likelihood equations numerically, named after Ronald Fisher.

Sketch of derivation

Let

Y1,\ldots,Yn

be random variables, independent and identically distributed with twice differentiable p.d.f.

f(y;\theta)

, and we wish to calculate the maximum likelihood estimator (M.L.E.)

\theta*

of

\theta

. First, suppose we have a starting point for our algorithm

\theta0

, and consider a Taylor expansion of the score function,

V(\theta)

, about

\theta0

:

V(\theta)V(\theta0)-l{J}(\theta0)(\theta-\theta0),

where

l{J}(\theta0)=-

n
\sum
i=1

\left.\nabla\nabla\top

\right|
\theta=\theta0

logf(Yi;\theta)

is the observed information matrix at

\theta0

. Now, setting

\theta=\theta*

, using that

V(\theta*)=0

and rearranging gives us:

\theta*\theta0+l{J}-1(\theta0)V(\theta0).

We therefore use the algorithm

\thetam+1=\thetam+l{J}-1(\thetam)V(\thetam),

and under certain regularity conditions, it can be shown that

\thetam\theta*

.

Fisher scoring

In practice,

l{J}(\theta)

is usually replaced by

l{I}(\theta)=E[l{J}(\theta)]

, the Fisher information, thus giving us the Fisher Scoring Algorithm:

\thetam+1=\thetam+l{I}-1(\thetam)V(\thetam)

..

Under some regularity conditions, if

\thetam

is a consistent estimator, then

\thetam+1

(the correction after a single step) is 'optimal' in the sense that its error distribution is asymptotically identical to that of the true max-likelihood estimate.

See also

Further reading

Notes and References

  1. Nicholas T. . Longford . A fast scoring algorithm for maximum likelihood estimation in unbalanced mixed models with nested random effects . Biometrika . 74 . 4 . 1987 . 817–827 . 10.1093/biomet/74.4.817 .