Scribd
Upload
124 views

Panel Data Problem Set 6

This document provides information and assignments for an econometrics course on analyzing panel data. It covers nonlinear exponential regression models with fixed effects, binary fixed effects panel probit models, and common effects probit models. Students are asked to derive maximum likelihood estimators, solve likelihood equations, and estimate models using sample health insurance data to examine issues like the incidental parameters problem.

Uploaded by

mb001h2002
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
124 views

Panel Data Problem Set 6

This document provides information and assignments for an econometrics course on analyzing panel data. It covers nonlinear exponential regression models with fixed effects, binary fixed effects panel probit models, and common effects probit models. Students are asked to derive maximum likelihood estimators, solve likelihood equations, and estimate models using sample health insurance data to examine issues like the incidental parameters problem.

Uploaded by

mb001h2002
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

Department of Economics

Econometric Analysis of Panel Data


Professor William Greene Phone: 212.998.0876 Office: KMC 7-78 Home page:www.stern.nyu.edu/~wgreene Email: [email protected] URL for course web page: www.stern.nyu.edu/~wgreene/Econometrics/PanelDataEconometrics.htm

Assignment 6 Nonlinear Models for Panel Data


Part I. A Concentrated Log Likelihood
Consider an exponential regression model with fixed effects, The density is f(yit|xit) = [1/it] exp(-yit / it), yit > 0, where it = exp(i + xit), i = 1,,n; t = 1,,T. It will prove convenient to let i = exp(i) so it = i exp(xit) = i it. The log likelihood for this exponential regression model with fixed effects is log L = i=1 t =1 ( log it y it / it )
n T

(a) Obtain the first order condition for maximizing logL with respect to i. Note, there is one of these for each i, so you need only differentiate logL i = t =1 ( log it y it / it )
T

with respect to i and equate it to zero. You will gain some convenience by defining ait = yit/it. (b) Now, treating as if it were known, show that the implicit solution of this likelihood equation for i in terms of is

tT=1 y it / exp( x ) tT=1ait it = i = = ai = i () T T

(c) It follows that at the solution for the MLE, it will be true that i() = ai where ai is the
c sample mean of ait. Denote it = aiit . Insert this solution back into the log likelihood function, to obtain the concentrated log likelihood function c c logLc = i=1 t =1 ( log it y it / it ) . n T

Note that this is a function of but not of i. To obtain the maximum likelihood estimator of , we can now maximize this function with respect to . This is equivalent to maximizing the whole log likelihood function, while considering only the solutions for i that satisfy i = i() as shown above. When we find , we can then compute i. (No assignment for this part.) (d) With this in hand, it is now possible to maximize the function with respect to . Show that the likelihood equation will be
y c n T n T logLc 1 1 = i=1 t =1 c + cit 2 it = i=1 t =1 c it it (it ) y it c 1 it . c it

You now need the derivative, itc/ . Continuing, show that itc/ = ai it xit
1 T ait it xit T t =1

Hint: it/ = itxit. Insert your result in the log

likelihood equation to obtain the implicit solution for ,


logLc 1 n T = i=1 t =1 c it y it 1 T c 1 it ai x it t =1it ait x it = 0. T it

Part II. Solving for FE in Panel Probit


For the binary fixed effects panel probit model, Prob(yit = 1 | xit) = (i + xit), Prob(yit = 0 | xit) = 1 - Prob(yit = 1|xit) = (-i - xit). a. Write out the full log likelihood function. b. Write out the first order condition for maximizng the function with respect to i, taking as known. Take this derivation as far as possible you will ultimately find that unlike the exponential model we examined in class, in this model, there is no explicit solution for i in terms of and the data. c. Show that regardless of the finding in b, there is no solution for i when yit is always 1 or always 0 within a given group (i).

Part III. The Incidental Parameters Problem.


This is a purely empirical exercise. It will involve some computations using the German health care data. As we discussed in class, for the binary logit model, there are two estimators for the fixed effects model Prob(yit = 1 | xit) = (i + xit), i = 1,...,n, t = 1,...,T. The brute force approach maximizes the whole log likelihood for i,i=1,...,n and . This estimator is known to suffer from the incidental parameters problem; when T is small, the estimator is biased away from zero. The best known result is that when T = 2, there is a 100% bias. The other approach is the Rasch/Andersen/Chamberlain method, which computes a conditional MLE using the probabilities conditioned on the sum of the yits for each group. This estimator is known to be consistent. For this exercise, we will see if the effect is visible in a sample, using precisely the estimators described. a. We first see if we can observe Hsiao/Abrevayas finding when T = 2. The following commands compute the estimates of the logit model both ways. Estimate the equations, and report your results. Do the empirical results seem to conform to the theory?
Sample ; All $ Reject ; _groupti # 2 $ Namelist ; x = hhninc,age,married,working$ Create ; y = doctor $ Logit ; Lhs = y ; Rhs = x ; pds = 2 ; Fixed$ Logit ; Lhs = y ; Rhs = x ; pds = 2 $

b. A second result that seems intuitively reasonable is that the IP bias diminishes as T increases. Is this the case? Change the three 2s in the command set above to 3s and redo the experiment. What do you find? Now, change the 2s to 7s and repeat the experiment. In each case, report your findings and your conclusions. c. What do you conclude about the fixed effects model?

Part IV. A Common Effects Probit Model


In this exercise, you will fit a probit model with common effects, and develop the appropriate model based on your findings. The probit model we will use is Prob(yit = 1 | xit) = (ci + xit) yit = Publicit = whether or not the individual chose public health insurance in that year.

xit = one,age,educ,hhninc,handper,working,hsat
1. Suppose, for the moment, we ignore the heterogeneity, ci and just pool the data and fit a simple probit model. Is the estimator consistent? What assumptions are necessary for the pooled estimator to be a consistent estimator of ? 2. All of the suggested covariates in the model are time varying. Fit a random effects model and a fixed effects model (this can only be done by brute force). Report your results. 3. We are interested in deciding which is preferred, fixed or random effects. I propose to use a variable addition test. Add the group means to the model, then carry out a likelihood ratio test of the hypothesis that the coefficients on the group means are all zero. What do you find? What do you conclude is the preferred model? 4. Suppose it were hypothesized that the previous years choice of whether or not to choose public insurance were on the right hand side of the equation. That is, Prob(yit = 1 | xit) = (ci + xit + yi,t-1) What would this imply for how one (you) should go about estimating the parameters of the model. What issues should you be concerned with for a dynamic model? Hint: These commands can be used for parts 1. 3.
Sample ; All $ Probit ; Lhs = public ; Rhs = one,age,educ,hhninc,handper,working,newhsat ; Pds = _groupti ; Random $ Probit ; Lhs = public ; Rhs = one,age,educ,hhninc,handper,working,newhsat ; Pds = _groupti ; FEM $ Create ; ageb = Group Mean(age,pds=_groupti) $ Create ; educb = Group Mean(educ,pds=_groupti) $ Create ; handb = Group Mean(handper,pds=_groupti) $ Create ; workb = Group Mean(working,pds=_groupti) $ Create ; hsatb = Group Mean(newhsat,pds=_groupti) $ Namelist ; Means = ageb,educb,handb,workb,hsatb $ Probit ; Lhs = public ; Rhs = means,one,age,educ,hhninc,handper,working,newhsat ; Pds = _groupti ; Random $

You might also like