Link functions, Advanced Statistics

Link functions:

The link function relates the linear predictor ηi to the expected value of the data. In classical linear models the mean and the linear predictor are identical. However, when dealing with counts and the distribution is Poisson, we must have the Poisson distribution parameter satisfy µi > 0 so that the identity link is less attractive, partly because ηi may be negative while µi > 0 must not be. It is advisable to utilize the log link η = log(µ) in this case. Similarly, when dealing with binomial distribution, the parameter p of probability of success in a single trial is restricted to be in (0,1) and the link function serves the purpose to map the interval (0,1) on to R1 . Therefore, links like the following (here µ is replaced by p):

614_Link functions6.png

have been suggested and widely used.

Let us illustrate the most commonly met examples of GLM together with describing the types of response variable, distribution, and the link function:

- Traditional Linear Model:

i) response variable: continuous

ii) distribution: normal

iii) link function: identity : η = µ

- Logistic Regression:

i) response variable: probability ( µ)

ii) distribution: binomial

iii) link function: logit: η = log( µ/1-µ)

- Poisson Regression in Log Linear Model:

i) response variable: count

ii) distribution: Poisson
iii) link function: η = log(µ)
- Gamma model with Log Link:
i) response variable: a positive continuous variable
ii) distribution: Gamma
iii) link function: η = log(µ)

Intermezzo and history. If you read di?erent references, you may get confused about the terminology. You may have already come across the term "general linear model" in your introductory Statistics courses or in some reference books. Note, however, that this term refers to a conventional linear regression model for a continuous response variables given continuous and/or categorical predictors. It includes multiple linear regression, as well as ANOVA and ANCOVA. In SAS, such models are ?t by least squares and weighted least squares using (typically) proc glm. HOWEVER, the "generalized linear model" we are speaking about here, refers to the larger class discussed in this section. The ?rst widely used software package for ?tting these models was called GLIM. Because of this program, "GLIM" became a well-accepted abbreviation for generalized linear models, as opposed to "GLM". Since we clari?ed the confusion though, we will continue using "GLM" for generalized linear models since many recent references use it. Today, generalized linear models are ?t by many packages, notably by the SAS proc genmod. (End of intermezzo).

One of the advantages of the full probabilistic speci?cation of the GLM model is that ML Estimation suggests itself as a natural general estimation method. We have to maximize the log-likelihood

2487_Link functions2.png

where β is linked to θ through the link function. Recall that the main parameter- vector of interest is β, the vector of regression coeffcients in the relation ηi = g(µi) = x0

1971_Link functions3.png

There is nowadays, with the availability of modern computing power, seldom any reason to consider estimators of β that are di?erent from the MLE. By using the chain rule, we get for the components of the score function:

1533_Link functions4.png

The (expected) Fisher information matrix is given then by

687_Link functions5.png

The ML Estimator is de?ned by equating the score function to zero. Numerically, the equation is solved by applying iterative procedures which we discuss next.

Posted Date: 2/27/2013 1:11:48 AM | Location : United States







Related Discussions:- Link functions, Assignment Help, Ask Question on Link functions, Get Answer, Expert's Help, Link functions Discussions

Write discussion on Link functions
Your posts are moderated
Related Questions
Per-experiment error rate is the possibility of the incorrectly rejecting at least one null hypothesis or assumption in the experiment including one or more tests or comparisons,

Artificial neural network : A mathematical arrangement modelled on the human neural network and designed to attack various statistical problems, particularly in the region of patte

Graphical deception : Statistical graphics which are not as honest as they should be. It is relatively simple. To mislead the unwary with the graphical material. For instance, c

Law of likelihood : Within framework of the statistical model, a particular set of data supports one statistical hypothesis or assumption better than another if the likelihood of t

A directed graph is simple if each ordered pair of vertices is the head and tail of at most one edge; one loop may be present at each vertex. For each n ≥ 1, prove or disprove the

A test for equality of the variances of the two populations having normal distributions, based on the ratio of the variances of the sample of observations taken from each. Most fre

Minimum volume ellipsoid is a term for ellipsoid of the minimum volume which covers some specified proportion of the set of multivariate data. It is commonly used to construct rob

Hello! I am currently in graduate school earning a masters in mental health counseling. I am in a stats course at current and we are reviewing z-scores. I am a little lost because

Genstat: The basic purpose piece of statistical software for the management and the analysis of data. The package incorporates the wide variety of data handling events and a wi

The Null Hypothesis - H0: β 1 = 0 i.e. there is homoscedasticity errors and no heteroscedasticity exists The Alternative Hypothesis - H1: β 1 ≠ 0 i.e. there is no homoscedasti