2004, Vol. 15, No. 3, pp. 689∼694
A Note on Test for Model Adequacy in Nonlinear Regression 1)
Myung-Wook Kahng 2)
Abstract
We investigate the test for model adequacy in nonlinear regression. We can expect the usual likelihood ratio statistic to be unaffected by any parametric- effect curvature; only the effect of intrinsic curvature needs to be considered. Multiplicative correction factor is derived for the limiting distribution of test statistic, which is a function of the intrinsic curvature arrays.
Keywords : Correction factor, Intrinsic curvature, Intrinsic curvature array, Model adequacy, Parametric-effect curvature
1. Introduction
Once a nonlinear regression model has been chosen and the unknown parameters estimated, the question of the adequacy of the model to the data arises. Usually one would look at the residuals and carry out various plots. If the intrinsic curvature is negligible, as is often the case, then the usual residual plots and likelihood ratio test carried out for linear models are also appropriate in nonlinear regression analysis. However, if there is substantial intrinsic curvature, then the projected residuals suggested by Cook and Tsai(1985) can be used and its F-approximation can be seriously affected by severe curvature (Bates and Watts 1980, 1981; Hamilton, 1986).
When the model is replicated, the adequacy of the model can be tested directly.
We consider the likelihood ratio statistic for the test of lack of fit. This test is unaffect by any parametric-effect curvature, but the effect of intrinsic curvature 1) This research was supported by the Sookmyung Women's University Research Grants 2003.
2) Professor, Department of Statistics, Sookmyung Women's University, Seoul, 140-742, Korea
E-mail : [email protected]
distribution of test statistic, which is a function of the intrinsic curvature arrays.
If here are no replicates, then there are a number of test procedures available, generally based on the idea of near-replicates. A number of these, which are also suitable for nonlinear models, are given by Miller, Neill and Sherfey(1999).
2. Nonlinear Regression Model
Consider the standard nonlinear regression model
y
i= f ( x
i: θ ) + ε
i, i = 1, 2, …, n , (2.1)
in which the i -th response y
iis related to the q -dimensional vector of known explanatory variable x
ithrough the known nonlinear model function f , which depends on p -dimensional unknown parameter vector θ , and ε
iis error. We assume that f is twice continuously differentiable in θ , and errors ε
iare independent identically distributed normal random variables with mean 0 and variance σ
2. The least squares estimate of θ , denoted by ˆ , minimized the θ error sum of squares
S ( θ ) = ∑
n
i = 1
[ y
i- f ( x
i: θ ) ]
2.
Let the p -dimensional parameter vector θ be partitioned in the form θ = ( θ
T1, θ
T2)
Twhere θ
1is a p
1-dimensional nuisance vector and θ
2is a p
2-dimensional vector of primary interest with p
1+ p
2= p . Now we consider the testing H
0: θ
2= θ
20. To apply the likelihood ratio procedure we need to minimize S ( θ ), subject to θ
2= θ
20. Let S ( θ ˜ ) = S ( θ ˜
1
( θ
20) , θ
20) be the minimum value of S ( θ ) under H
0. Then the likelihood ratio statistics is
F
LR= S( θ ˜ ) - S( θ ˆ )
S( θ ˆ ) ⋅ n - p
p
2(2.2)
which is approximately distributed as F
p2, n - pwhen H
0is true.
3. Lack of Fit Test
Suppose a design is replicated, say n
itimes for point x
i, so that our nonlinear model (2.1) now becomes
y
ij= η
i+ e
ij= f ( x
i: θ ) + ε
ij, i = 1, 2, … , k, j = 1, 2, …, n
i(3.1)
where the errors ε
ijare assumed to be i.i.d. N ( 0, σ
2) and k is the number of the distinct setting of the explanatory variables. This model includes the genuine variation between different experiments with the same x observation as well as error of measurements.
Ignoring the structure on η
i, we have the usual decomposition
∑
i∑
j
( y
ij- η
i)
2= ∑
i
∑
j
( y
ij- y
i⋅+ y
i⋅- η
i)
2= ∑
i
∑
j
( y
ij- y
i⋅)
2+ ∑
i
n
i( y
i⋅- η
i)
2,
(3.2)
where ∑
i
∑
j
( y
ij- y
i⋅)
2is usually referred to as the pure error sum of squares.
Defining n = ∑ n
i, an unbiased estimate of σ
2is
s
2e= 1 n - k ∑
i
∑
j
( y
ij- y
i⋅)
2, and under the normality assumptions of ε
ij,
(n- k) s
2eσ
2∼ χ
2n - k.
To find the least squares estimate ˆ of θ we minimize ∑∑ (y θ
ij- η
i)
2with
respect to θ , which is equivalent to minimizing ∑ n
i( y
i⋅- η
i)
2, i.e. a
weighted least squares analysis with weight n
i. The normal equations for θ ˆ are
therefore
- 2 ∑
i
n
i( y
i) - f ( x
i: θ ) ∂ f ( x
i: θ )
∂ θ | θ = θˆ
= 0
If η
iis replaced by η ˆ
i
= f ( x
i: θ ˆ ) , the identity (3.2) still applies, so that
∑
i∑
j
( y
ij- η ˆ
i
)
2= ∑
i
∑
j
( y
ij- y
i⋅)
2+ ∑
i
n
i( y
i⋅- η ˆ
i
)
2SSE = SS
PE+ ( SSE - SS
PE)
= SS
PE+ SS
LOFThe left-hand side, called the residual sum of squares, is therefore split into a pure error sum of squares SS
PEand a lack of fit sum of squares SS
LOF. For the case when η
iis a linear function of θ , a test for the validity of this model is given by Seber(1977),
F = SS
LOF/ ( k - p) SS
PE/ ( n - k)
= ∑
i
n
i( y
i⋅- η ˆ
i
)
2∑
i∑
j
( y
ij- y
i⋅)
2⋅ n - k k - p
(3.3)
where F ∼ F
k - p , n - kwhen the model is valid. Because of asymptotic linearity, we find the above F-test is approximately valid for large n when the model is nonlinear.
4. Correction Factors
Since f ( x : θ ) and S ( θ ) are invariant under one-to-one transformations of θ , the likelihood ratio statistic, F
LRis unaffected by the parameter-effect curvature, however, the effect of intrinsic curvature need to be considered.
Hamilton and Wien(1987) studied second-order approximations for the distribution
of F
LRand developed correction terms which are functions of the intrinsic
curvature arrays. Assuming normality of the ε
i, they showed as σ
2→ 0 (rather
than n → ∞ ), the limiting null distribution of F
LRis, to order σ
3,
( 1 - γ σ
2) F
p2, n - p(4.1)
where γ = α
2/ ( n - p) - α
1/ p
2and
α
1= α
0- α
2α
2= - 1
2 ∑
n - p
i = 1
tr { ( A
Ni ..)
2} + 1 4 ∑
n - p
i = 1
{ tr ( A
Ni ..) }
2α
0= - 1
2 ∑
n - p1
i = 1
tr { ( A
N0 i ..)
2} + 1 4 ∑
n - p1
i = 1
{ tr ( A
N0 i ..) }
2.
Here the p×p matrix A
Ni .., i = 1,2,…, n - p is the i -th face of the intrinsic curvature array A
N..for the full model (2.1) (Seber and Wild, 1989, p. 142), and the p
1×p
1matrix A
N0 i .., i = 1, 2, …, n - p
1is the i -th face of the intrinsic curvature array for the restricted model obtained by fixing θ
2= θ
20and working with θ
1. To evaluate γ we use θ
2= θ
20and θ
1= θ ˜
1