Nonlinear Methods in Econometrics

Table of Contents

Readings and References:


Nonlinear Model Estimation

Nonlinear least squares and maximum log-likelihood are the most common methods for parameter estimation of nonlinear models in econometrics. Assume that the model, linear or nonlinear, is expressed as:

F(z,β) = ε

F(z,β) defines the functional form of the model, where z = [y x] is the data matrix which includes both dependent (endogenous) y and independent (exogenous) x variables, β is the vector of unknown parameters and ε is the model error. A typical nonlinear model in econometrics takes a separable form (between y and x) like this:

ε = F(z,β) = F(y,x,β) = y - f(x,β), or

y = f(x,β) + ε

For a general nonseparable (between y and x) nonlinear model F(z,β), the asymptotic theory of nonlinear least squares does not apply. Maximum likelihood or generalized method of moments should be considered instead.

Nonlinear Least Squares

The model is estimated by minimizing the sum of squared errors:

S(β|y,x) = ε'ε = (y-f(x,β))' (y-f(x,β))

Least squares estimates of the parameters are computed from the first-order condition for minimization (zero gradient):

∂S/∂β = 2ε'(∂ε/∂β) = - 2ε'(∂f(x,β)/∂β) = 0.

Finally, the following Hessian matrix must be checked for the positive definiteness:

2S/∂β∂β' = 2[(∂ε/∂β)'(∂ε/∂β) + ∑i=1,2,...,Nεi (∂2εi/∂β'∂β)]
= 2[(∂f(x,β)/∂β)'(∂f(x,β)/∂β) - ∑i=1,2,...,Nεi (∂2f(xi,β)/∂β'∂β)]

Given E(∂S/∂β) = 0, following from Taylor approximation of ∂S/∂β at the NLS estimator b of β, the asymptotic theory implies that

√N(b-β) →d N(0,H-1VH-1)

where V = Var(∂S/∂β) = E((∂S/∂β)'(∂S/∂β)), and H = E(∂2S/∂β∂β').

Evaluated at the NLS estimator b of β, the sample analogy of H and V is respectively:

H = 2[(∂ε/∂β)'(∂ε/∂β)]/N
V = 4[(∂ε/∂β)'εε'(∂ε/∂β)]/N.

Therefore, b ~a N(β,[(∂ε/∂β)'(∂ε/∂β)]-1 [(∂ε/∂β)'εε'(∂ε/∂β)] [(∂ε/∂β)'(∂ε/∂β)]-1)

Under the assumption of homschedasticity, E(εε') = σ2I, the estimated variance-covariance matrix of the parameters b is simplified as follows:

Var(b) = s2[(∂ε/∂β)'(∂ε/∂β)]-1

where s2 is the estimated model variance σ2. That is, s2 = e'e/N, and e = y-f(x,b) is the estimated errors or residuals.

If there are equality or inequality parameter constraints (e.g., non-negativity) expressed in terms of a continuous transformation β = φ(α) where α is an unconstrained parameter vector. Then from the estimator of α and Var(α), we have

β = φ(α)
Var(β) = (∂φ/∂α) [Var(α)] (∂φ/∂α)'

Nonlinear Weighted Least Squares

The technique of nonlinear least squares can be generalized straightforwardly to consider the weighted model errors. Denote the weighting scheme w = w(β|y,x), a scalar or a vector, which may be linearly or nonlinearly dependent on part or all of the parameters. Define the weighted error terms as ε* = wε. The model can be estimated by minimizing the sum of weighted squared errors:

S*(β|y,x) = ε**

Since the weighting function w may depend on the unknown parameters β, the consistency condition is not satisfied in general for the weighted least squares model. Weighted least squares estimator may be inconsistent.

Maximum Normal Likelihood

Assuming the normal probability distribution of the model error, or ε ~ normal(0,σ2I), then the log-likelihood function for each data observation i is

ll(β,σ2|yi,xi) = -½ [ln(2π)+ln2) + εi22] + ln(Ji(β))

where εi = F(yi,xi,β), and Ji(β) = |∂εi/∂yi| is the Jacobian of transformation from εi to yi. The model is estimated by maximizing the sum of log-likelihood over a sample of N observations as follows:

ll(β,σ2|y,x) = -½N [ln(2π)+ln2)] -½ (ε'ε/σ2) + ∑i=1,2,...,Nln(Ji(β))

The solution is obtained from the system of first-order condition as follows:

ll/∂β = - ε'/σ2 (∂ε/∂β) + ∑i=1,2,...,N[1/Ji(β)](∂Ji/∂β) = 0.
ll/∂σ2 = - N/(2σ2) - ε'ε/(2σ4) = 0.

Usually the maximum likelihood estimation is performed by substituting out the asymptotic variance estimate σ2. That is, σ2 = ε'ε/N. Then the following concentrated log-likelihood function is maximized to find the parameter estimates β:

ll*(β|y,x) = -½N [1+ln(2π)-ln(N)] -½N ln(ε'ε) + ∑i=1,2,...,Nln(Ji(β))

Let ε* = ε/[(J1...JN)1/N]. Then the last two terms of the above concentrated log-likelihood function can be combined and the function is re-written as

ll*(β|y,x) = -½N [1+ln(2π)-ln(N)] -½N ln**)

Therefore, maximizing the concentrated log-likelihood function ll*(β|y,x) is equivalent to minimizing the sum of squared weighted errors:

S*(β|y,x) = ε**

where ε* = wε, with the weight w = 1/[(J1...JN)1/N] (inverse of the geometric mean of Jacobians) applied to each observation of the error terms ε.

Solving from the first-order condition or zero-gradient condition:

ll*/∂β = -½N ∂ln(S*)/∂β = -½(N/S*)(∂S*/∂β) = -(N/S*)[ε*'(∂ε*/∂β)] = 0,

the solution must be checked for the negative definiteness of the Hessian matrix (the second-order condition):

2ll*/∂β∂β' = ½N ∂2ln(S*)/∂β∂β'
= ½(N/S*)[(1/S*)(∂S*/∂β)'(∂S*/∂β) -(∂2S*/∂β∂β')]

Since ∂S*/∂β = 0 from the first-order condition for the maximum likelihood solution, the corresponding negative definite Hessian matrix is simply

2ll*/∂β∂β' = -(N/S*)[½(∂2S*/∂β∂β')]
= -(N/S*)[(∂ε*/∂β)'(∂ε*/∂β) + ∑i=1,2,...,Nεi* (∂2εi*/∂β∂β')]

Given E(∂ll*/∂β) = 0, following from Taylor approximation of ∂ll*/∂β at the ML estimator b of β, the asymptotic theory implies that

√N(b-β) →d N(0,H-1VH-1)

where V = Var(∂ll*/∂β) = E((∂ll*/∂β)'(∂ll*/∂β)), and H = E(∂2ll*/∂β∂β').

Evaluated at the ML estimator b of β, the sample analogy of H and V is respectively:

H = (-1/σ2*)[(∂ε*/∂β)'(∂ε*/∂β)]/N
V = (-1/σ2*)2[(∂ε*/∂β)'ε*ε*'(∂ε*/∂β)]/N, where σ2* = S2/N.

For a class of models satisfying regularity assumptions, the Information Matrix Equality holds as - H = V or

- E(∂2ll*/∂β∂β') = E((∂ll*/∂β)'(∂ll*/∂β))

Therefore, √N(b-β) →d N(0,-H-1). In other words, b ~a N(β,σ2*[(∂ε*/∂β)'(∂ε*/∂β)]-1)

The estimated variance-covariance matrix of the parameters b is:

Var(b) = s2*[(∂ε*/∂β)'(∂ε*/∂β)]-1, where s2* is the sample estimate of σ2*.

Further, as in the case of nonlinear least squares, if there are equality or inequality parameter constraints (e.g., non-negativity) expressed in terms of a continuous transformation β = φ(α) where α is an unconstrained parameter vector. Then from the estimator of α and Var(α), we have

β = φ(α)
Var(β) = (∂φ/∂α) [Var(α)] (∂φ/∂α)'

A Special Case

If the Jacobian Ji(β) = |∂εi/∂yi| = 1 for all observation i, then the log-Jacobian terms in the above concentrated log-likelihood function vanish. Therefore,

ll*(β|y,x) = -½N [1+ln(2π)-ln(N)] -½N ln(ε'ε)

This is exactly the special case of classical nonlinear model in which ε = F(y,x,β) = y - f(x,β). For this special case, maximizing the concentrated log-likelihood function ll*(β|y,x) is the same as minimizing the sum of squared errors S(β|y,x).

Example 1: Generalized Production Functions

First, we fit the following two classical production functions based on 30 data observations of labor L, capital K, and output Q given in the file JUDGE.TXT (The data of this example is taken from Judge, et. al. [1988], Chapter 12, p. 512):

  1. Cobb-Douglas Production Function
    ln(Q) = β1 + β2ln(L) + β3ln(K) + ε

  2. CES Production Function
    ln(Q) = β1 + β4ln2Lβ3 + (1-β2)Kβ3) + ε

Based on the least squares and maximum likelihood criteria, estimate and compare the Cobb-Douglas and CES production function, respectively.

According to Zellner and Revankar [1970], the classical production functions may be generalized to consider variable rate of returns to scale as follows:

  1. Generalized Cobb-Douglas Production Function
    ln(Q) + θ Q = β1 + β2ln(L) + β3ln(K) + ε

  2. Generalized CES Production Function
    ln(Q) + θ Q = β1 + β4ln2Lβ3 + (1-β2)Kβ3) + ε

Modify and estimate the Generalized version of Cobb-Doulass and CES production function, repectively.


Statistical Inferences in Nonlinear Models

The classical assumption of statistical inferences is the normal probability distribution of the model error:

ε = F(y,x,β) ~ normal(0,σ2I).

Thus the estimated least squares or maximum likelihood parameters b of β is normally distributed:

b ~ normal(β,Var(b))

where the estimated variance-covariance matrix is

Var(b) = [-E(∂2ll(b)/∂β∂β')]-1 = s2[½ E(∂2S(b)/∂β∂β')]-1
= s2[(∂ε(b)/∂β)'(∂ε(b)/∂β)]-1

and the estimated asymptotic variance of the model is s2 = S(b)/N.

Confidence location of the true parameter vector β is derived from the estimated b based on the following familiar F statistic:

F = (S(β)-S(b))/J/s2

where J is the degrees of freedom associated with the testing hypotheses.

By approximating the sum of squares function S(β) at b up to the second order and set ∂S(b)/∂β = 0,

S(β) - S(b) = ½ (β-b)' [∂2S(b)/∂β∂β'] (β-b)

Therefore, the test statistic for testing β = b:

JF = (β-b)' [Var(b)]-1 (β-b)

follows a Chi-Square distribution with J degrees of freedom.

Wald Test

Consider J active constraints of parameters, linear or nonlinear (continuous and differentiable), expressed as the equation:

c(β) = 0.

If the constraints were true, without estimating the constrained model, the unconstrained parameter estimator b is expected to satisfy the constraint equation closely. That is, c(b) = 0. The test statistic

W = c(b)'[Var(c(b)]-1c(b)

has a Chi-square distribution with J degrees of freedom. With the first-order linear approximation of the constraint function c(β) at b,

W = c(b)' {(∂c(b)/∂β) [Var(b)] (∂c(b)/∂β)'}-1 c(b)

Note that this test statistic does not require the computation of the constrained parameter estimator.

Lagrangian Multiplier Test

Given the J-element constraint equation c(β) = 0, let b* denote the constrained maximum likelihood estimator of the parameters β. The Lagrangian multiplier test is based on the score vector ∂ll(b*)/∂β of the original parameterization. If the constraints hold, then ∂ll(b*)/∂β should be close to ∂ll(b)/∂β for the unconstrained parameter estimater b, which is of course zero.

The test statistic is written as:

LM = (∂ll(b*)/∂β) [Var(∂ll(b*)/∂β]-1 (∂ll(b*)/∂β)'
= (∂ll(b*)/∂β) [Var(b*)] (∂ll(b*)/∂β)'

The estimated variance-covariance matrix of the constrained estimator b* is computed as follows:

Var(b*) = H-1 [I - G'(G H-1G')-1GH-1]

where H = [-∂ll(b*)2/∂β∂β'] and G = [∂c(b*)/∂β].

LM test statistic is easily approximated with the following formula:

LM = {[ε(b*)'(∂ε(b*)/∂β)] [(∂ε(b*)/∂β)'(∂ε(b*)/∂β)]-1 [ε(b*)'(∂ε(b*)/∂β)]'}/(ε(b*)'ε(b*)/N)

Note that the maximum likelihood estimates of errors ε(b*) may be properly weighted, and this test statistic is based on the constrained parameters alone.

Likelihood Ratio Test

If both the constrained and unconstrained maximum likelihood solutions are available, then the test statistic

LR = -2(ll(b*)-ll(b))

follows a Chi-square distribution with J degrees of freedom, in which there are J constraints in the equation c(β) = 0. In terms of sum of squares, it is

LR = N ln(S(b*)/S(b)).

Figure: Three Bases for Hypothsis Tests

Example 2

Returning to the previous example of Generalized CES production function:

ln(Q) + θ Q = β1 + β4ln2Lβ3 + (1-β2)Kβ3) + ε

Using Wald test, Largrangian multiplier test, and likelihood ratio test to verify the nonlinear equality constraints (or classical CES): β4 = 1/β3, and θ = 0.


Copyright© Kuan-Pin Lin
Last updated: 10/12/2012