線性與指數模型的整合工具
Generalized Linear Model, GLM/GLZ
An Integrated Modeling Tool for
Linear and Exponential Analysis
神掌打通任督二脈‧易筋經以簡馭繁
廣義線性模型特色
廣義線性模式 (generalized linear model),縮寫早期同樣是GLM,這個術語,是統計史上最容易混淆的案例之一,所以,近來已有將縮寫改為GLZ的趨勢。
廣義線性模型的定義,簡單說: 範圍比一般線性模型更大,可包括指數、與對數迴歸,處理更多元的機率分配問題。
廣義線性模型與一般線性模型比較 GLM vs. GLZ
|
General linear model |
Generalized linear model |
Typical estimation method |
Least squares, best linear unbiased prediction |
Maximum likelihood or Bayesian |
Special cases |
ANOVA, ANCOVA, MANOVA, MANCOVA, ordinary linear regression, mixed model, t-test, F-test |
linear regression, logistic regression, Poisson regression |
統計模型的發展與統計史觀的分界
GLZ是由Nelder & Wedderburn (1972)所提出的,名稱還叫做線性模型,但已經能夠處理非線性的問題了;同時,(Nelder, 1966)也提出了Gamma 分配,以整合卡方分配、與F分配…等,在知識論上,提供了更整合性基礎(脫離原始框架)的觀照。
(其實,一般線性模型也能處理二次曲線,此細節在此不論。)
統雄老師-似乎零星西文文獻也有此議-感到,似可作為畫分「古典統計」與當代統計的界限。
同時,統雄老師也認為非線性分析,才是人類行為研究計量法的方向,不過,與GLZ 在基礎思想上又大不相同了。
統計模型的重要發展點
以下是由Lindsey, McCullagh, Nelder, Stiegler所建議的統計模型的重要發展點:
‧Multiple linear regression — normal distribution & identity
link (Legendre, Guass: early 19th century).
‧ANOVA — normal distribution & identity link (Fisher: 1920』s
– 1935).
‧Likelihood function — a general approach to inference about
any statistical model (Fisher, 1922).
‧Dilution assays — a binomial distribution with
complementary log-log link (Fisher, 1922).
‧Exponential family — class of distributions with sufficient
statistics for parameters (Fisher, 1934).
‧Probit analysis — binomial distribution & probit link (Bliss,
1935).
‧Logit for proportions — binomial distribution & logit link
(Berkson, 1944; Dyke & Patterson, 1952)
Item analysis — Bernoulli distribution & logit link (Rasch,
1960).
‧Log linear models for counts — Poisson distribution & log
link (Birch, 1963).
‧Regressions for survival data — exponential distribution &
reciprocal or log link (Feigl & Zelen, 1965; Zippin & Armitage,
1966; Glasser, 1967).
‧Inverse polynomials — Gamma distribution & reciprocal link
(Nelder, 1966).
‧Nelder & Wedderburn (1972): provided unification. They
showed "All the previously mentioned models are special cases of
a general model, 「Generalized Linear Models」 The MLE for all these models could be obtained using
same algorithm.
‧All of the models listed have distributions in the
"Exponential Dispersion Family」