Non-Linear Regression Models B 非线性回归.ppt

Fitting Non-Linear Regression Models VR 8.2 Oy Mae Louie Bessie Nguyen Gina Piscitelli When do we use a non-linear regression model? When at least one of the parameters is not linear The general form of a non-linear regression model is y=η(x,β) + ε where ε ~N(0, σ 2) Plot of Weight Loss vs Days for an Obese Patient oldpar-par(mar=c(5.1,4.1,4.1,4.1)) plot(Days,Weight,type=p,ylab=Weight(kg)) Wt.lbs-pretty(range(Weight*2.2025)) axis(side=4,at=Wt.lbs/2.205,lab=Wt.lbs,srt=90) mtext(Weight(lb),side=4,line=3) Weight Loss Dataset Model: y = βo + β1*2-t/θ + ε The parameters: βo: ultimate lean weight β1: total amount of weight to be lost θ: time taken to lose half the amount of weight remaining to be lost Method Used to Analyze Data Iterative procedure The number of iterations depend on how quickly the parameters converge. The converged parameters are close approximates for βo,β1,and θ. The likelihood for the non-linear regression model is maximized when residual sum of squares is minimized. Non-Linear Regression Methods 1. Direct Computation Method: y = βo + β1*2-t/θ Derivative Method: η(β) ≈ ω(0) + Z(0)β Self-Starting Method: y = βo + β1 exp(-x/θ) + ε 1. Direct Computation Method The direct computation method initializes values for b0, b1, and theta by inspection of the weight loss data set. R Code wtloss.st -c(b0 = 90, b1 = 95, th =120) wtloss.fm - nls(Weight ~ b0 + b1*2^(-Days/th), data = wtloss, start = wtloss.st, trace = T) wtloss.fm lines(Days,fitted.values(wtloss.fm)) 2. Derivative Method The equation η(β) ≈ ω(0) + Z(0)β defines (in vector terms) the tangent plane to the surface at the coordinate point β = β(0) where η(β) is the mean vector ω (0) is the offset vector Z (0) is the matrix which defines the tangent plane Each iteration gives a new approximation for β until convergence is reached. The derivative method uses numerical methods to compute a formula for the gradient unless first derivatives are supplied. 2. Derivative