Generalized Linear Models I

Generalized Linear Models I

Statistics 203: Introduction to Regression and Analysis of Variance Generalized Linear Models I Jonathan Taylor - p. 1/18 Today's class ● Today's class ■ Logistic regression. ● Generalized linear models ● Binary regression example ■ ● Binary outcomes Generalized linear models. ● Logit transform ■ ● Binary regression Deviance. ● Link functions: binary regression ● Link function inverses: binary regression ● Odds ratios & logistic regression ● Link & variance fns. of a GLM ● Binary (again) ● Fitting a binary regression GLM: IRLS ● Other common examples of GLMs ● Deviance ● Binary deviance ● Partial deviance tests 2 ● Wald χ tests - p. 2/18 Generalized linear models ● Today's class ■ All models we have seen so far deal with continuous ● Generalized linear models ● Binary regression example outcome variables with no restriction on their expectations, ● Binary outcomes ● Logit transform and (most) have assumed that mean and variance are ● Binary regression ● Link functions: binary unrelated (i.e. variance is constant). regression ● Link function inverses: binary ■ Many outcomes of interest do not satisfy this. regression ● Odds ratios & logistic ■ regression Examples: binary outcomes, Poisson count outcomes. ● Link & variance fns. of a GLM ● Binary (again) ■ A Generalized Linear Model (GLM) is a model with two ● Fitting a binary regression GLM: IRLS ingredients: a link function and a variance function. ● Other common examples of GLMs ◆ The link relates the means of the observations to ● Deviance ● Binary deviance predictors: linearization ● Partial deviance tests 2 ◆ ● Wald χ tests The variance function relates the means to the variances. - p. 3/18 Binary regression example ● Today's class ■ A local health clinic sent fliers to its clients to encourage ● Generalized linear models ● Binary regression example everyone, but especially older persons at high risk of ● Binary outcomes ● Logit transform complications, to get a flu shot in time for protection against ● Binary regression ● Link functions: binary an expected flu epidemic. regression ● Link function inverses: binary ■ In a pilot follow-up study, 50 clients were randomly selected regression ● Odds ratios & logistic and asked whether they actually received a flu shot. regression ● Link & variance fns. of a GLM ■ ● Binary (again) In addition, data were collected on their age and their health ● Fitting a binary regression GLM: IRLS awareness. ● Other common examples of GLMs ■ Here is the data. ● Deviance ● Binary deviance ● Partial deviance tests 2 ● Wald χ tests - p. 4/18 Binary outcomes ● Today's class ■ Suppose outcome Y is a 0-1 random variable, then ● Generalized linear models i ● Binary regression example ● Binary outcomes µ E Y π : ● Logit transform i = ( i) = i ● Binary regression ● Link functions: binary ■ regression Variance function ● Link function inverses: binary regression ● Odds ratios & logistic Var(Yi) = πi(1 − πi) = µi(1 − µi) regression ● Link & variance fns. of a GLM ● Binary (again) Variance is related to mean! ● Fitting a binary regression GLM: IRLS ■ A convenient way to model the dependence of Y on ● Other common examples of i GLMs ● Deviance covariates Xi1; : : : ; Xi;p−1 is through the logit transform. ● Binary deviance ● Partial deviance tests 2 p−1 ● Wald χ tests logit(πi) = β0 + βj Xij j=1 X - p. 5/18 Logit transform ● Today's class ■ Logit transform: ● Generalized linear models ● Binary regression example ● Binary outcomes π ● Logit transform logit π ; ● Binary regression ( ) = log 2 (−∞ +1) ● Link functions: binary 1 − π regression ● Link function inverses: binary ■ Inverse: regression x ● Odds ratios & logistic −1 e regression logit (x) = 2 (0; 1): ● Link & variance fns. of a GLM x ● Binary (again) 1 + e ● Fitting a binary regression ■ GLM: IRLS Derivative: ● Other common examples of GLMs ● Deviance d 1 1 ● Binary deviance logit(π) = = : ● Partial deviance tests dπ π(1 − π) V (π) 2 ● Wald χ tests Note: special relation between derivatives and variance function – more on this next lecture. - p. 6/18 Binary regression ● Today's class ■ Logistic regression model: ● Generalized linear models ● Binary regression example ● Binary outcomes logit(E(Yi)) = logit(πi) ● Logit transform ● Binary regression p−1 ● Link functions: binary regression ● Link function inverses: binary = β0 + βj Xij regression ● Odds ratios & logistic j=1 regression X ● Link & variance fns. of a GLM ■ ● Binary (again) Probit regression model: ● Fitting a binary regression GLM: IRLS ● Other common examples of p−1 GLMs −1 ● Deviance Φ (E(Yi)) = β0 + βj Xij ● Binary deviance ● Partial deviance tests j=1 2 ● Wald χ tests X where Φ is CDF of N(0; 1), i.e. Φ(t) = pnorm(t). ■ In each case, Var(Yi) = πi(1 − πi) but the regression model is different. ■ Here is an example - p. 7/18 Link functions: binary regression 6 4 logit probit 2 0 f(pi) −2 −4 −6 0.0 0.2 0.4 0.6 0.8 1.0 pi - p. 8/18 Link function inverses: binary regression 1.0 0.8 logit−inv probit−inv 0.6 f(pi) 0.4 0.2 0.0 −4 −2 0 2 4 pi - p. 9/18 Odds ratios & logistic regression ● Today's class ■ For any event A and any probability P ● Generalized linear models ● Binary regression example ● Binary outcomes P(A) ● Logit transform ODDS(A) = : ● Binary regression P A ● Link functions: binary 1 − ( ) regression ● Link function inverses: binary ■ regression In the logistic regression model with outcome Y ● Odds ratios & logistic regression ● ODDS Y : : : ; X x ; : : : Link & variance fns. of a GLM ( = 1j j = j + 1 ) βj ● Binary (again) = e ● Fitting a binary regression ODDS(Y = 1j : : : ; Xj = xj ; : : : ) GLM: IRLS ● Other common examples of GLMs is the (multiplicative) change in odds if variable X increases ● Deviance j ● Binary deviance by 1: βj is known as the ODDS RATIO for . ● Partial deviance tests e Xj 2 ● Wald χ tests ■ If Xj 2 f0; 1g is dichotomous, and P(Y = 1j : : : ) is small (rare event hypothesis) then group with Xj = 1 are approximately eβj more likely to have event, all other parameters being the same. - p. 10/18 Link & variance fns. of a GLM ● Today's class ■ If ● Generalized linear models ● Binary regression example k ● Binary outcomes E ● Logit transform ηi = g( (Yi)) = g(µi) = β0 + βj Xij ● Binary regression j=1 ● Link functions: binary regression X ● Link function inverses: binary then g is called the link function for the model. regression ● Odds ratios & logistic ■ regression If ● Link & variance fns. of a GLM E ● Binary (again) Var(Yi) = φ · V ( (Yi)) = φ · V (µi) ● Fitting a binary regression GLM: IRLS for φ > 0 and some function V , then V is the called variance ● Other common examples of GLMs ● Deviance function for the model. ● Binary deviance ■ ● Partial deviance tests “Canonical”reference: Generalized Linear Models, 2 ● Wald χ tests McCullagh and Nelder. - p. 11/18 Binary (again) ● Today's class ■ For a logistic model, ● Generalized linear models ● Binary regression example ● Binary outcomes g(µ) = logit(µ); V (µ) = µ(1 − µ): ● Logit transform ● Binary regression ● Link functions: binary ■ For a probit model, regression ● Link function inverses: binary regression −1 ● Odds ratios & logistic g(µ) = Φ (µ); V (µ) = µ(1 − µ): regression ● Link & variance fns. of a GLM ● Binary (again) ● Fitting a binary regression GLM: IRLS ● Other common examples of GLMs ● Deviance ● Binary deviance ● Partial deviance tests 2 ● Wald χ tests - p. 12/18 Fitting a binary regression GLM: IRLS ● Today's class ■ Algorithm ● Generalized linear models ● Binary regression example 1. Initialize: set µi = 0:999 or 0.001 depending on whether ● Binary outcomes ● Logit transform Yi = 1 or 0. ● Binary regression 0 ● Link functions: binary 2. Compute Z = g(µ ) + g (µ )(Y − µ ): regression i b i i i i ● Link function inverses: binary −1 0 2 regression 3. Use weights Wi = g (µi) V (µi) to regress Z onto X’s to ● Odds ratios & logistic regression get β using WLS.b b b ● Link & variance fns. of a GLM ● Binary (again) −1 b p b ● Fitting a binary regression 4. Compute µi = g β0 + Xijβj : GLM: IRLS j=1 ● Other common examples of b GLMs 5. Repeat steps 2-4 until convergence. ● Deviance b P b ● Binary deviance ■ b ● Partial deviance tests Approximate distribution 2 ● Wald χ tests β ∼ N(β; φ(XtW X)−1): ■ If φ has to be estimated, a simple choice is Pearson’s X 2: b c n 1 (Y − µ )2 φ = i i : n − p V (µ ) i=1 i X b b - p. 13/18 b Other common examples of GLMs ● Today's class ■ Standard multiple linear regression: g µ µ, Var µ ● Generalized linear models ( ) = ( ) = 1 ● Binary regression example ■ ● Binary outcomes Linear regression with variance tied to mean, for example: ● Logit transform 2 ● Binary regression g(µ) = µ, Var(µ) = µ . ● Link functions: binary ■ regression Poisson log-linear models: g(µ) = log(µ); Var(µ) = µ. ● Link function inverses: binary regression ● Odds ratios & logistic regression ● Link & variance fns. of a GLM ● Binary (again) ● Fitting a binary regression GLM: IRLS ● Other common examples of GLMs ● Deviance ● Binary deviance ● Partial deviance tests 2 ● Wald χ tests - p. 14/18 Deviance ● Today's class ■ Instead of least squares, models are fit on the basis of scaled ● Generalized linear models ● Binary regression example deviance, analogous to SSE when errors are not Gaussian. ● Binary outcomes ● Logit transform ■ ● Binary regression This is not completely clear from the IRLS algorithm, but it is ● Link functions: binary very close to the Newton-Raphson algorithm. Algorithm regression ● Link function inverses: binary often called Fisher scoring. regression ● Odds ratios & logistic ■ regression ● Link & variance fns. of a GLM ● Binary (again) DEV (µ, Y ) = −2 log L(µ, Y ) + −2 log L(Y; Y ) ● Fitting a binary regression GLM: IRLS ● Other common examples of where µ is a location estimator for Y (usually in an GLMs ● Deviance exponential family – more next lecture). ● Binary deviance ■ 2 ● Partial deviance tests If Y is Gaussian with independent N µ ; σ entries 2 ( i ) ● Wald χ tests n 1 DEV (µ, Y ) = (Y − µ )2 σ2 i i i=1 X - p.

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    18 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us