Module 11: Linear Regression. Rebecca C. Steorts

Transcription

1 Module 11: Linear Regression Rebecca C. Steorts

2 Announcements Today is the last class Homework 7 has been extended to Thursday, April 20, 11 PM. There will be no lab tomorrow. There will be office hours this week. Optional review class next Tuesday, April 25th. Module 9 has been updated to match the Hoff book s notation.

3 Agenda What is linear regression Motivating Example Application from Hoff

4 Setup X n p : regression features or covariates (design matrix) x p 1 : ith row vector of the regression covariates y n 1 : response variable (vector) β p 1 : vector of regression coefficients Goal: Estimation of p(y x). Dimensions: y i β T x i = (1 1) (1 p)(p 1) = (1 1).

5 Health Insurance Example We want to predict whether or not a patient has health insurance based upon one covariate or predictor variable, income. Typically, we have many predictor variables, such as income, age, education level, etc. We store the predictor variables in a matrix X n p.

6 Normal Regression Model The Normal regression model specifies that E[Y x] is linear and the sampling variability around the mean is independent and identically (iid) from a normal distribution Y i = β T x i + e i (1) e 1,..., e n iid Normal(0, σ 2 )

7 Normal Regression Model (continued) This allows us to write down p(y 1,..., y n x 1,... x n, β, σ 2 ) (2) n = p(y i x i, β, σ 2 ) (3) i=1 (2πσ 2 ) n/2 exp{ 1 2σ 2 n (y i β T x i ) 2 } (4) i=1

8 Multivariate Setup Let s assume that we have data points (x i, y i ) available for all i = 1,..., n. y is the response variable y = y 1 y 2. y n n 1 x i is the ith row of the design matrix X n p. Consider the regression coefficients β = β 1 β 2. β p p 1

9 Multivariate Setup y X, β, σ 2 MVN(Xβ, σ 2 I) β MVN(0, τ 2 I) The likelihood in the multivariate setting simpifies to p(y 1,..., y n x 1,... x n, β, σ 2 ) (5) (2πσ 2 ) n/2 exp{ 1 n 2σ 2 (y i β T x i ) 2 } (6) i=1 (2πσ 2 ) n/2 exp{ 1 2σ 2 (y Xβ)T (y Xβ)} (7)

10 Posterior computation Let a = 1/σ 2 and b = 1/τ 2. p(β y, X) p(y X, β)p(β) (8) exp{ a/2(y Xβ) T (y Xβ)} exp{ b/2β T β)} (9) Just like in the Multivariate modules, we just simplify. (Check these details on your own). p(β y, X) MVN(β y, X, Λ 1 ) where Λ = ax T X + bi and µ = aλ 1 X T y.

11 Posterior computation (details) p(β y, X) (10) exp{ a 2 (y Xβ)T (y Xβ)} exp{ b 2 βt β)} (11) exp{ a 2 [y T y 2β T X T y + β T X T Xβ] b 2 βt β} (12) exp{aβ T X T y a 2 βt X T Xβ b/2β T β} (13) exp{aβ T [X T y] 1/2β T (ax T X + bi)β} (14) Then Λ = ax T X + bi and µ = aλ 1 X T y.

12 Linear Regression Applied to Swimming We will consider Exercise 9.1 in Hoff very closely to illustrate linear regression. The data set we consider contains times (in seconds) of four high school swimmers swimming 50 yards. There are 6 times for each student, taken every two weeks. Each row corresponds to a swimmer and a higher column index indicates a later date.

13 Data set read.table(" ## V1 V2 V3 V4 V5 V6 ## ## ## ##

14 Full conditionals (Task 1) We will fit a separate linear regression model for each swimmer, with swimming time as the response and week as the explanatory variable. Let Y i R 6 be the 6 recorded times for swimmer i. Let X i = be the design matrix for swimmer i. Then we use the following linear regression model: Y i N 6 ( Xβ i, τ 1 i I 6 ) β i N 2 (β 0, Σ 0 ) τ i Gamma(a, b). Derive full conditionals for β i and τ i.

15 Solution (Task 1) The conditional posterior for β i is multivariate normal with while V[β i Y i, X i, τ i ] = (Σ τxi T X i ) 1 E[β i Y i, X i, τ i ] = (Σ τ i Xi T X i ) 1 (Σ 1 0 β 0 + τ i Xi T Y i ). τ i Y i, X i, β Gamma ( a + 3, b + (Y i X i β i ) T (Y i X i β i ) 2 These can be found in in Hoff in section ).

16 Task 2 Complete the prior specification by choosing a, b, β 0, and Σ 0. Let your choices be informed by the fact that times for this age group tend to be between 22 and 24 seconds.

17 Solution (Task 2) Choose a = b = 0.1 so as to be somewhat uninformative. Choose β 0 = [23 0] T with Σ 0 = [ ] This centers the intercept at 23 (the middle of the given range) and the slope at 0 (so we are assuming no increase) but we choose the variance to be a bit large to err on the side of being less informative.

18 Gibbs sampler (Task 3) Code a Gibbs sampler to fit each of the models. For each swimmer i, obtain draws from the posterior predictive distribution for yi, the time of swimmer i if they were to swim two weeks from the last recorded time.

19 Posterior Prediction (Task 4) The coach has to decide which swimmer should compete in a meet two weeks from the last recorded time. Using the posterior predictive distributions, compute Pr {yi = max (y1, y 2, y 3, y 4 )} for each swimmer i and use these probabilities to make a recommendation to the coach. This is left as an exercise.