*184*
*83*
*1MB*

*English*
*Pages 95*
*Year 1984*

- Author / Uploaded
- John H. Aldrich
- Forrest D. Nelson

*Table of contents : ContentsSeries IntroductionAcknowledgments1. The Linear Probability Model 1.0 Introduction 1.1 Review of the Multivariate, Linear Regression Model 1.2 A Dichotomous Dependent Variable and the Linear Probability Model 1.3 A Dichotomous Response Variable with Replicated Data 1.4 Polytomous or Multiple Category Dependent Variables 1.5 The Linearity Assumption 1.6 The Effect of an Incorrect Linearity Assumption2. Specification of Nonlinear Probability Models 2.0 Introduction 2.1 The General Problem of Specification 2.2 Alternative Nonlinear Functional Forms for the Dichotomous Case 2.3 Derivation of Nonlinear Transformations from a Behavioral Model 2.4 Nonlinear Probability Specifications for Polytomous Variables 2.5 Behavior of the Logit and Probit Specifications 2.6 Summary3. Estimation of Probit and Logit Models for Dichotomous Dependent Variables 3.0 Introduction 3.1 Assumptions of the Models 3.2 Maximum Likelihood Estimation 3.3. Properties of Estimates 3.4 Interpretation of and Inference from MLE Results 3.5 Conclusions4. Minimum Chi-Square Estimation and Polytomous Models 4.0 Introduction 4.1 Minimum Chi-Square Estimation for Replicated, Dichotomous Data 4.2. Polytomous Dependent Variables5. Summary and Extensions 5.0 Introduction 5.1 Summary 5.2 ExtensionsNotesReferencesAbout the Authors*

Series / Number 07-045

LINEAR PROBABILITY, LOGIT, AND PROBIT MODELS

JOHN H. ALDRICH University of Minnesota

aan ae

NELSON FORREST D. lowa of ity Univers

SAGE PUBLICATIONS Beverly Hills

London

New

Delhi

ay

£

a.

ye

PRE

Copyright © 1984 by ; Sage Publications, Inc. irs t. 4

*

ee

.

;

.

Printed in the United States of America ‘

All rights reserved. No part of this book may be reproduced or utilized in any form or by any means, electronic or mechanical,

including photocopying, recording, or by any information storage and retrieval system, without permission in writing from the publisher,

Loner

400 ‘

For information address:

5 4

| -

Lye a) %

SAGE Publications, Inc. 275 South Beverly Drive Beverly Hills, California 90212

SAGE Publications India Pvt. Ltd. C-236 Defence Colony New Delhi 110 024, India

6)

SAGE Publications Ltd 28 Banner Street London ECtY 8QE, England

Internaliona] Standard Book Number 0-8039-21 33-0 Library of Congress Catalog Card No. 84-051766

FIRST PRINTING When citing a professional paper, please use the proper form. Remember to cite the correct Sage University Paper series title and include the paper number. One of the following formats can be adapted (depending on the style manual used): (1) IVERSEN, GUDMUND R. and NORPOTH, HELMUT (1976) “Analysis of Variance.” Sage University Paper series on Quantitative Applications in the Social Sciences, 07-001. Beverly Hills and London: Sage Pubns. OR (2) Iversen, Gudmund R. and Norpoth, Helmut. 1976. Analysis of Variance. Sage University Paper series on Quantitative Applications in the Social Sciences, series no. 07-001, Beverly Hills and London: Sage Pubns.

CONTENTS

Series Introduction

Acknowledgments

5

7

1. The Linear Probability Model

1.0 1.1 1,2

9

13

Introduction 9 Review of the Multivariate, Linear Regression Model A Dichotomous Dependent Variable and the Linear Probability Model 12 A Dichotomous Response Variable with

1.4

Replicated Data 20 Polytomous or Multiple Category

1.5 1.6

Dependent Variables 22 The Linearity Assumption

24

The Effect of an Incorrect

Linearity Assumption

27

2. Specification of Nonlinear Probability Models 30 2.0 Introduction 30 2.1 The General Problem of Specification 30 2.2 Alternative Nonlinear Functional Forms for 2.3

the Dichotomous Case 31] Derivation of Nonlinear Transformations

from a Behavioral Model 2.4 2.5 2.6

35

Nonlinear Probability Specifications for Polytomous Variables 37 Behavior of the Logit and Probit Specifications

Summary

47

40

10

3. Estimation of Probit and Logit Models for Dichotomous Dependent Variables

48

3.0 3.1

Introduction 48 Assumptions of the Models

3.2 3.3. 3.4

Maximum Likelihood Estimation Properties of Estimates 52 Interpretation of and Inference

3.5

from MLE Results Conclusions 65

48

49

54

4. Minimum Chi-Square Estimation and Polytomous Models 66

4.0 4.1

Introduction 66 Minimum Chi-Square Estimation for Replicated, Dichotomous Data 67 Polytomous Dependent Variables 73

4.2.

5. Summary and Extensions

Notes

5.0

Introduction

5.1

Summary

5.2

Extensions

78

78 8]

85

References

93

About the Authors

95

78

Series Editor’s Introduction

Earlier in this series, we published papers on regression analysis by Lewis-Beck and by Achen. Picking up where these papers left off, Nelson and Aldrich develop the logit and probit models analyzing dependent variables which are not continuous. Although there has been

a tendency for many social scientists who analyze dichotomous dependent variables, such as vote choice, to simply proceed with ordinary regression analysis, Nelson and Aldrich show that this is really not an appropriate strategy. Probit analysis is more ideally suited to such

problems, and, in the general case of polytomous dependent variables, logit ought to be the method of choice. Aldrich and Nelson review the linear probability model, and then introduce the reader to several alternative specifications of nonlinear models, including the normal, logistic, Gompertz, Burr, and several others. The remainder of the monograph is devoted to the discussion of

the estimation of probit and logit models, the assumptions of these models, and problems of inference. The authors use several examples to demonstrate the differences among the linear, logit and probit models, and to illustrate the importance of various assumptions in these models. Although the material covered here is inherently difficult for many social scientists, the authors have done quite a good job of presenting formal statements followed by verbal translations followed by detailed

examples. This pedagogic strategy should serve the neophyte well, and this of course is the intention of our sertes of monographs. The topic— probit and logit models—is not an introductory one. The reader should have a good understanding of linear regression analysis and of statistical inference. But for such readers, the presentation of this topic is an introductory one, and we fully expect that this monograph will have the effect of increasing the usage of logit and probit models in social science disciplines such as sociology, psychology, political science, and perhaps others as well. It has been aimed directly at the economics market, of course, and should provide a readable and lucid introduction for economists who, as 2 group, already make considerable use of logit and

probit models. As such, this paper represents the first in a series of publications that we intend to present for the econometrics market. In so doing, we hope to provide a service supplementing the standard econometrics textbooks, and also provide a vehicle to upgrade statistical methodology in the other social science disciplines. Many graduate students in social sciences other than economics often have great difficulty mastering a standard econometrics text, and it is our hope that by presenting some topics in econometrics in our series, we can provide a

5

bridge between standard econometrics and other social science treat-

ments of statistical concepts. We find this manuscript to be an excellent choice as an initial move in this direction, because the presentation is so

well done and because it has received rave reviews from some of the best econometricians. —John L. Sullivan Series Co-Editor

Acknowledgments We would like to thank Professors Lee Spector and Michael Mazzeo for making their data available for our use, Professor Raymond Duvall and two referees for their useful comments, and Professors John Sullivan and Richard Niemi, editors, and Sage Publications for their sup-

port. We alone remain responsible for remaining errors.

LINEAR PROBABILITY, LOGIT, AND PROBIT MODELS JOHN H. ALDRICH University of Minnesota

FORREST D. NELSON University of lowa

1. THE LINEAR

PROBABILITY

MODEL

1.0 Introduction Regression analysis has become a standard statistical tool in the social sciences. Its popularity stems from several sources. It provides

much explanatory power, especially due to its multivariate nature. The Gauss-Markov Theorem (Johnston, 1984) proves that it has some very

desirable statistical properties. It is widely available in computer packages, it is easy to interpret, and there is a widespread belief that it remains a reasonable procedure even if some of the assumptions underlying it are not met in the data (a property statisticians refer to as “robustness”). Perhaps because of its widespread popularity, regression may be one of the most abused statistical techniques in the social sciences. While estimates derived from regression analysis may be robust against errors

in some assumptions, other assumptions are crucial, and their failure dependent variable is a qualitative measurere rather. than a continuous,

interval measure. After reviewing the linear regression model, we shall show that regression estimates with a qualitative dependent variable can lead to serious errors in inference. For example, we shal] show that regression estimates with a qualitative dependent variable may seriously

misestimate the magnitude of the effects of independent variables, that all of the standard statistical inferences such as hypothesis tests or the

9

10

construction of confidence intervals are unjustified, and that the regression estimates will be highly sensitive to the range of particular values

observed for the independent variables (thus making extrapolations or forecasts beyond the range of the data especially unjustified). The failure of multivariate regression with a qualitative dependent variable is particularly troubling in the behavioral sciences, for many of the behaviors of interest, as well as attitudes and preferences underlying them, are qualitative—or at least observed qualitatively. Voting, for

example, is one of the key variables in political science. In economics, individual choices to sell or to purchase a good, to enter a contract, and so on, are central problems. In sociology, one might be concerned with the decision to marry, to go to college, to have children, or any number

of other qualitative choices. All of these behaviors are qualitative, often representing even dichotomous decisions. Many attitudes may be measurable in only a smal! number of categories, Party affiliation, ideology, the importance of religion in one’s life, and many others may be potentially quantitative attitudes, but in practice are often measured qualitatively, especially in large surveys. What is needed, therefore, is some

set of statistical techniques that can do the work of multivariate regression but that are not subject to its liabilities in the presence of qualitative dependent variables. Fortunately, several such procedures exist, and the

purpose of this volume is to present some of them. In this chapter, we will build the case that they are needed. The first

section presents a quick review of the linear regression model, presenting notation, assumptions, and central results (see Lewis-Beck, 1980, and Achen, 1982 in this series for more details). The second section considers the use of regression with a dichotomous dependent variable. Section

three and four extends the analysis when the dichotomous dependent variable is repeatedly observed and when it is polytomous (i-e., qualitative but with more than two categories), respectively. Section five ex-

amines the linearity assumption, and section six covers some of the errors of inference that may arise from assuming it incorrectly.

1.1 Review of the Multivariate, Linear Regression Model A regression equation with, say, two independent (aka exogenous)' variables may be written; Y: = bi + beXi2 + br Xi3 + Ui or, in the general case, using summation notation:” Yi = TbyXe + ui

[1.1]

7 In 1,1, ¥ is the dependent variable, Xi for k = 2,..., K are the exogenous or independent variables, u is the random error (or disturbance term) which is, of course, unobserved, b; are unknown constants,’ and

the subscript i denotes the i" observation from the sample of size N. In 1.1, we observe only Y; and Xi, the values in the sample for the dependent and independent variables. The central problem is to estimate the unobserved constants or parameters, by. Let the parameter

estimates be denoted by bh. Replacing the by in 1.1 with by yields Yj = DbXu te;

[1.2]

This equation may be rearranged to solve for the error term: e= Yi - bX

[1.3]

Squaring equation 1.3 and summing over all N observations yields the sum of squared errors. The ordinary least squares (OLS) estimates of the

b; are those which minimize that sum.‘ Elementary calculus is used to find the minimum. So doing yields K equations, one for each of the unknown b,’s. To solve K equation for K unknowns, by, requires that

each equation not be an exact linear combination of any of the others. But what this means is that the independent variables are not perfectly

collinear, the first assumption of OLS regression. At the very least, that requires at least as many observations as there are parameters to estimate (i.e.,

N = K).

To this point, we have not discussed specification of the equation. Uniess otherwise noted, we assume throughout, as users of regression assume, that all relevant—and no irrelevant—X’s have been included in the equations 1.1 and 1.2 (see Lewis-Beck, 1980; Achen, 1982, for more

details). The

remaining

assumptions

concern

the disturbance

term

ui.’

Possibly the most important of these assumptions is that u; is not correlated with any of the independent variables, X,, and that it has a mean of zero (assumption 3a and 3d in Lewis-Beck, 1980, p. 26). This

implies that, given the X;, the mean, or expected value, of Y; is

E(Y¥il Xi, ... Xin) = DOeXix

[1.4]

If so, then the OLS solutions, by, will be unbiased estimators of the unknown parameters of interest, by. The last two assumptions made about u; are it has a constant variance, oi, across all observations, and that it is uncorrelated from one

12 observation to another. These two are called “homoscedasticity” and

“serial independence,” respectively. The set of three assumptions about ui are often called the Gauss-Markov assumptions, because they are those used in the Gauss-Markov Theorem. This theorem states that, given this set of assumptions, the OLS estimators are the best (most “efficient” in the sense of having the smallest sampling variance) esti-

mators of those which are linear (in Y;) and unbiased—so-called “BLUE” for “best,” linear, unbiased estimator. It is easy to construct an unbiased estimator of the variance of the disturbance term 2:

S2 = Ye} /(N-K) This statistic is closely related to the F statistic and R’ which provide checks on the overall fit of the model.

A final common assumption is that u; is normally distributed. That implies that b, are normally distributed. Finally, the sampling variances of the by’s (s? as estimators of oi) can be computed, and their distributions known. From these, hypothesis tests can be constructed, con-

fidence intervals computed, and so on. 1.2 A Dichotomous Dependent Variable and the Linear Probability Model The regression model places no restrictions on the values that the independent (exogenous) variables take on, except that they not be exact linear combinations of each other. They may be continuous, interval level (net worth of

a company), they may be only positive or

zero (percent of vote a party received), they may be integers (number of children in a family), or they may be dichotomous (so-called dummy variables, e.g., scored as a one if make, zero if female). The dependent variable, however, is assumed to be continuous. Since there are no restrictions on the X,’s, the b,'s, or the u;’s, then Y; must be free to take on any value from negative infinity to positive infinity. In practice, Y; will take on only a small set of values in the data. For

example, even if Y; is family income for the year, only a relatively small range of values will be observed. But, in this case, since any Xx’s will be similarly restricted, the assumption of continuous, interval measurement of Y: will not be a bad approximation. But, if Y; can take on only two values (say zero and one), the violation of this assumption is so egregious as to merit special attention.

13 Suppose Y; equals either zero or one. The expected value of Yi reduces to the probability that Y; equals one [symbolized as P(Y; = 1)J: E(Y;) = | + PCY, = 1) + 0- PCY; = 0) = P(Yi=

L)

[1.5]

If we then combine equations 1.4 and 1.5, we get E(Y)) = PCY: = Y= LoiXe

[1.6]

From this we conclude that the right-hand side of the regression equation must be interpretable as a probability, i.e., restricted to between zero and one. For this reason, a linear regression model with a de-

pendent variable that is either zero or one is called the linear probability model or LPM. If Y; takes on only two values, then u; can assume only two values

itself, for any given values of Xix. That is, by equation [.1 for ¥;equalling zero and one, respectively, we get:

If Y, = 0, then (O= 2b, X,, + u,) or (u, = - 2b, X,)

a7 1.7

If ¥,=1, then (1 = 2b, X,, + u,) or (u, = 1 - xb, X,) Now, it can be shown that the first key assumption about wu, that its

expectation be zero, can be maintained:

E(u,) = PCY, = 0) [-2b, X,.] + PCY, = 1)[1 - 2b, X,,] = -[1- PCY, = 1] PCY, = 1) + PCY, = Dl - PCY, =D] =0 As aresult, OLS estimates of by will be unbiased. However, the assumption that the u; have a constant variance cannot be maintained. In fact,

the variance of u; varies systematically with the values of the independent variables,

v(u,) = E(u?) = P(Y, = 0) [- 2b, X17 + PCY, = 1) {1 ~ 2b, X17

= (1 - PCY, = 1] (PCY, = D1? + PCY, = DL -POY, = DY? =P(Y, = 1) [1 -PCY, =1)] = [Eb X,,] [1 - 2b, X,,] Thus, the OLS estimate, bx, will be unbiased but not best (i.e., not have the smallest possible sampling variance). As a further result, estimates of the sampling variances will not be correct, and any hypothesis tests (e.g.,

14 the t and F

tests) or confidence

intervals based

on these sampling

variances will be invalid, even for very large samples. Thus, even in the best circumstances, OLS regression estimates of a dichotomous dependent variable are, although

unbiased,

not very desirable. There

is,

however, a solution to this problem that is a fairly simple modification of OLS regression. 1.2.1 WLS ESTIMATION TECHNIQUE THE LINEAR PROBABILITY MODEL

FOR

Goldberger (1964) proposed a two-step, weighted estimator to correct the problems of OLS regression of the linear probability model. The

first step is to do the usual OLS regression of Yi on the Xiu. So doing yields the unbiased estimates b,. From these estimates, construct the set of weights, one for each observation:

wi = [1/(Zb Xa

— VbeXi)J'?

[1.8]

These weights are just the reciprocals of the estimated standard errors of u;. Now multiply both sides of eq. 1.1 by wi:

(wi) = T(bewiXn) + (wins)

[1.9]

It can be shown (Goldberger, 1964), that (w,u;) has a constant variance, so that regressing (wiYi) on (wiXix) by OLS yields new estimates, say Bh, which are not only unbiased but also have the smallest possible sampling variances. The standard errors of the second set of estimates, by, are the correct ones for conducting hypothesis tests, and the like.* Finally, ui,

taking on bul two values, cannot be normally distributed. However, in large samples,

by turn out to be approximately normally distributed, so

that hypothesis tests and the like can be used in the usual manner. The only statistic resulting from the second stage estimation not of

much use is the R*. The one computed at this stage correctly reports the proportion of variance explained, but the variance referred to is the variance of the weighted dependent variable, not the original one. The R’ could be recomputed from the untransformed data according to the formula

R? = f - EY, - 36,%4)) 20, ~¥) where bi represents the WLS coefficient estimates, Yi and X», are the original data, and Y is the sample mean of the Y;. But even this quantity

[5

is of limited use for at least two reasons. First, since the Y; are inherently heteroscedastic, even computing a sample variance is like adding apples and oranges, and the statistical interpretation of the fraction of this ageregation explained is imprecise at best. Second, since OLS has the

property of maximizing the R’, the quantity computed as above is guaranteed to be smaller than the OLS R’, leading to another interpretive quandary—generally we prefer models with the highest R’, but here we know, on theoretical grounds, that a model with lower R? is preferred. Given these difficulties, use of the coefficient of determi-

nation as a summary statistic should be avoided in models with qualitative dependent variables. What we have done, therefore, is to describe a procedure for valid

estimation of a linear regression equation in the face of a dichotomous dependent variable—so long as the new assumption about the disturbance term (following from equation 1.7) can be made. Clearly, the implications of this assumption are very different from the usual

assumptions made in OLS regression, The u; are not assumed to be continuous, homoscedastic or normally distributed. Rather, they are assumed to be dichotomous and dependent upon the parameters, by, and values of the independent variables. 1.2.2 AN EXAMPLE

Spector and Mazzeo (1980) examined the effect of a teaching method

denoted as PSI onthe performance of students in a course, intermediate macro economics. The question was whether students exposed to the method scored higher on exams in the class. They collected data on students from two classes, one in which PSI was used and another in which traditional teaching methods were employed. Those data are

reproduced in Table | .1 and include for each of 32 students the entering grade point average (GPA), the score on an exam given at the beginning of the term to test entering knowledge of the material (TUCE), adummy variable indicating teaching method (PSI), and the final grade, recorded

here as | for an A and 0 for a B or C (GRADE). GRADE is the dependent variable, and of particular interest is whether PSI has a significant influence on GRADE. Casual inspection of the data reveals a high correlation between PSI and GRADE—eight of fourteen students taught by the PSI method earned A’s while only three of eighteen non-PSI students received A’s. But TUCE and GPA are also positively related to GRADE. Students earning an A had an average GPA of 3.43, while other students averaged only 2.95, forexample. And those taught by the PSI method had, on average, slightly higher GPAs and TUCE

scores. A multivariate analysis is cequired to ascertain whether the

16 TABLE

1.)

Data on the Effect of “Personalized System of Instruction” (PSI) on Course Grades OBS

GPA

TUCE

PS{

Grade

Grade

| 2.66 20 2 «2.89 22 3. 3.28 24 4 2.92 12 § 400 21 6 «2.86 17 7 2.76 17 8 2.87 21 9 3.03 25 10 3.92 29 41 2.63 20 12. 3.3223 13. 3.57 23 4143.26 25 15 3.53 26 162.74 9) SOURCE:

Letter|

OBS

PSI

Grade

Grade

| PSI

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 i 0 0 0 0 1 6 0 0 1 0 0

c B B B A B B B c A Cc B B A B B

GPA

TUCE

Grade

Grade

PSI

Grade

Grade

25 19 23 25 22 28 14 26 24 27 17 24 21 23 21 19

0 0 I 1 1 i 1 1 1 1 } 1 1 ! 1 1

0 0 0 1 9 1 9 0 1 1

Cc Cc B A Cc A Cc B A A A B A A Cc A

17. 2.75 18 2.83 9 3.12 20 «3.16 21. 2.06 22 #362 232.89 24 «3.51 25 3.54 262 —-.2.83 27) -3..39 28 = 2.67 293.65 30 4.00 31 3.10 32 —-2.39

Letter

0 I } 0 1

Spector and Mazzeo () 980).

PSI-GRADE relation is independent of or in addition to other factors,

or whether the apparent positive relation is spurious. Thus GPA and TUCE are included in the analysis to control for ability and background.

Results of a simple OLS regression on these data are presented in Table |.2a. Except for the coefficient estimates and descriptive statistics, the results appearing there are not of much use for the reason noted above—the

heteroscedastic residuals. To proceed with the weighted

least squares analysis, the weights must be constructed using the coefficient estimates of Table |.2a. Define the variable Y; according to Y, = -1.498 + .4638*GPA(i) + .3785*PSI(i) + .0105*TUCE(i) Then the weights (w) are determined according to

w= 1/¥) A - Yo” Each variable, including the intercept, is then multiplied by this weight,

and OLS is applied to the transformed variables.’ Results of the weighted least squares analysis are reported in Table 1.2b. (The descriptive statistics reported there are for the transformed

17 TABLE 1.2a

Ordinary Least Squares Results on PSI Data Descriptive Statistics Standard

Grade GPA

PSI TUCE

Sum

Mean

11.0 99.75

0.3437 3.117

0.4826 0.4667

0.4375 21.94

0.5040 3.902

14.0 702.0

Model:

MODELO1

SSE

Dependent Variable:

Grade

MSE

Weight:

%t/(P(1-P))

4.216474

DFE

28

Standard

Estimate

Error

F Ratio

6.65

R-Square

0.4159

Prob>F

0.150588

Parameter

Deviation

Variable

DF

Intercept GPA PSI

1 1 !

—1498 0.4639 0.3786

0.5239 0.1620 0.1392

—2.859 2.864 2.720

0.0079 0.0078 0.0111

TUCE

1

0.0105

0.0195

0.539

0.5944

TABLE

T Ratio

0.0016

Prob >|7|

1.2b

Weighted Least Squares Results on PSI Data Descriptive Statistics Standard

Sum Grade

97.377

GPA PSI

15051.41 122.56

TUCE

970184

Model: Weight: Dependent

MODELO02 1/(P (1-P)

SSE DFE

Variable:

Grade

MSE

Variable

Mean

DF

Intercept

1

GPA PSI TUCE

1 1 1

22.739 28

0.8121 Standard

Estimate

Error

0.3982 0.3878 0.0122

0.0179

1.756

2.767 0.0225

2.384 1.966

17.83

Parameter --1.309

Deviation

0.2885 0.08783 0.1052 0.0045

41.48 F Ratio Prob > F

29.92 0.0001

R-Square

0.7622

T Ratio

Prob >|T|

—4.536

0.0001

4.533 3.687 2.676

0.0001 0.0010 0.0123

18 variables, GPA*W, INTERCEPT*W, and so on. Comparing the OLS and WLS results, note that the coefficient estimates change but not by

much—the only problem with OLS estimates is high sampling variance. All other results, on the other hand, are markedly different. The coefficient of TUCE appears significant at a level of about 0.015, for example, while an incorrect inference drawn from the OLS results would have suggested TUCE to be insignificant. These WLS results can be used to draw the inferences sought. In particular, the t ratio for PSI is

3.69, which is highly significant, indicating that PSI does have a positive influence on the chance of earning an A even after controlling for ability

and background. (Of course 32 observations are too few for strict reliance on the distributional properties of these WLS estimates, so the inferences drawn should be taken as tentative. As noted above, nonnormality implies that the distributional properties hold only as an

approximation in large samples. Thirty-two is not large enough for this approximation to be particularly tight.) The coefficients can be interpreted as in regression with a continuous

dependent variable except that they refer to the probability of a grade of A, rather than to the level of the grade itself. They can be used, for example, to predict the chances of earning an A. According to these results, a student with a grade point of 3.0, taught by traditional methods, and scoring 20 on the TUCE exam would earn an A with probability of .1286.

Three points are worthy of note. First, the reported obtained directly from the computer printout, is highly explained above, it refers to the explained fraction of the transformed dependent variable (GRADE*W) and not

R’ of .7622, as misleading. As variance of the to the original

variable. Recomputation of the R’ using the original variables yields the number .3755. Note that this value is smaller than the one reported above for the OLS analysis. By no means does this suggest that the OLS resulis are better. Rather it demonstrates the inappropriateness of the

R’ statistic in analyses involving qualitative variables. Second, the sum of squared errors (SSE) has a useful interpretation beyond that in the standard regression model. The effect of the weight-

ing is to divide each observation by anestimate of the standard deviation of the error term in that observation. Thus the weighted residual has a variance that is approximately 1, and the sum of squared residuals from

the WLS analysis should behave approximately hike a Chi-Square random variable with degrees of freedom N - K (32 - 4 = 28 in this case).

Since a Chi-Square variable with degrees of freedom d has a mean of d and a variance of 2d, we can compute a statistic t = (SSE - (N-K))/

¥(2(N - K)) which should behave like a drawing from Student's t distri-

19 bution with N-K degrees of freedom when N -K is large. For the data used here we obtain t = (22.74 - 28)/sqrt(S6) = -.703, Comparing this value to tabulated values of the t distribution, we conclude that it is not an unlikely outcome and therefore does not raise questions about the appropriateness of the model. Had it been larger than say 2.1, we might

well reject the linear probability model as an adequate description of the data employed here. Finally, while the results of this problem make the linear probability model appear quite acceptable, there are difficulties with it to be explained below. 1.2.3 A DIFFICULTY WITH GOLDBERGER’S PROCEDURE AND A SECOND EXAMPLE

One difficulty that arises with some frequency is that the estimates of bi in the first OLS regression lead to values of LbyXix (or estimated values of the probability that Yi equals one) that are less than zero or greater than one." This problem is not necessarily very severe. Given the

LPM assumption (equation 1,6), the probability estimates are unbiased but will differ from the true value, being perhaps negative or preater than one simply due to sampling error. A practical solution is to truncate the estimates of Xb, Xx to values close to zero or one, say .001 or .999.

If there are very many values that fall out of the range 0 to 1, especially if they do so by very much, then we might be led to wonder

about the reasonableness of the assumption in equation 1.6. That is, we may be led to wonder whether or not our specification of the process as a linear probability model is in fact tenable. For example, Ostrom and Aldrich (1978), in an example that we will detail below, examined the relationship between the incidence of large-scale war and the number of

major power international dichotomous, independent

nation states or alliances involving major system from 1824 to 1938. The dependent | if there was a large-scale war in a year, variable was the number of major actors

powers in the variable was 0 if not. The in the inter-

national system that year. Since several of the hypotheses were approx-

imated by a linear probability model, Goldberger’s estimation technique was used with an LPM of X linearly related to Y. The first stage OLS regression yielded 12% of the cases having a negative “probability” of war (all the years with a bipolar system; the estimated “probability” was about -.07), too high a percentage to attribute solely to “sampling error.” The conclusion reached was that the assumption of the LPM, in

combination with the assumed model, was implausible. Rather, a dif-

ferent set of assumptions was necessary.’

20 1.3 A Dichotomous Response Variable with Replicated Data A random variable which can take on only two different values, designated as | and 0 for convenience, is known in statistics as a

“Bernoulli response variable” and has associated with it a parameter, say P, that represents the probability of observing a response Y = |. A single observation, say Yi, on the dichotomous dependent variable of the

previous section is an example of a Bernoulli random variable. In equation 1.6 we specified that the probability of a positive response, designated there as P(Y;= 1), depended on the values of K independent variables, that is,

P; = P(Y; = 1) = SbXx Since the Xi. are presumed to vary across observations, the parameter P;

also changes, so we would say that each Y; follows a different Bernoulli distribution. Thus in section [.3, a sample of size N represented drawings from N different Bernoulli distributions. The sum of N independent drawings (or N replications) from the same Bernoulli distribution (that

is with the same probability, P, of success) is known as a “binomial response variable” and has associated with it two parameters, N and P. Binomial variables also arise in social science contexts. In particular, it is not uncommon to encounter data sets in which observations can be grouped into subsets with the independent variables taking on the same values across all observations within a group. For example, in the Ostrom-Aldrich example in the last section, there were | 11 data points.

However, there were only five different values to the single independent variable. There were 13 cases of two major actors in the international system, 7 of three, 54 of four, 11 of five, and 16 of six actors. Since the only variable assumed relevant in that model was the number of major actors, the probability of war is the same for all years with the same

number of actors. We could (and implicitly do) treat all 54 years with four major actors as drawings, or replications, from the same distri-

bution, and the total number of wars begun during years with four actors as a binomial variable.

More generally, let the number of groups in a data sel be designated by M (in Ostrom-Aldrich, M = 5) and let Nj, j= 1,..., M) represent the number of observations within a group, (in Ostrom-Aldrich, N, = 13,

21 N2 = 17, N3 = 54, Na = 11, and Ns = 16). Identify the N, distinct obser-

vations on Y within group j as Yj, (i= 1,.-.., Nj), and define

Ny

Y,=2YiThe groups are defined so that each exogenous

variable is constant

across the Nj observations in a group; designate these constant values as Xix, (k= 1,...K). Since the probability of a positive response depends

only on the X; and parameters by, it too is constant across observations in a group. That is PCY;

= 1) = P; = Loi Xx

for alli=1,...,Nj, j= 1,...,M. Clearly the Y;’s are Bernoulli variables with parameter P;, and Yj, defined above, is binomial with parameters P; and Nj {so long as the N; observations Yj in group j are independent), From a theoretical point of view it matters little whether we treat as the focus of our analysis the binomial observations Yj, (j= 1,..., M), or the Bernoulli observations Yj, (i= 1,...,Nj,j=1,-.., M), since the infor-

mation content is the same in either set. But from a practical point of view, the binomial variables represent a much smaller set of numbers (M) than do the binomial variables (N, where N = EN). So it generally proves more economical to treat the binomial {or “grouped,” or “replicated”) responses Yj; as the unit of analysis. What appears to be an advantage for binomial variables turns out to be only illusory. Define the fraction of positive responses within group

jas f,= Y¥;/Nj. (The sample proportion fj is an estimate of Pj, the population proportion.) While the Bernoulli variable Y; of section 1.2 was limited to two values (0 or 1), the group proportion fj can assume Nj + | different values (0/.Nj, 1/Nj, . . ., Nj/Nj). It would seem that if we regress f, on the independent variables, then the difficulties encountered

in section 1.2 would not be so severe. That is, under the linear probability model assumption, E(f)) = P; = SbiXx, so we can write the regression equation f; = SOX

Since fis

+ uy;

[1.10]

acloser approximation to acontinuous dependent variable, it

would seem that OLS

regression of fj on Xx, would suffice. However,

22 little is changed from section 1.2. OLS regression would continue to yield

unbiased

estimates,

given

assumption

1.6. But

as before,

the

assumption of homoscedasticity would fail, as the variance of uj would depend on the by and Xj, values.'° Therefore, a weighted least squares

procedure, similar to Goldberger’s as presented above, is more appropriate. A one-step procedure is all that is required in this case. Insection 1.2, the first OLS regression was used to estimate the weights, wi. Here,

we can use the observed frequencies or proportions f, themselves to compute the weights:

w= [Ni/((Q)C - #1”

[ltl]

The weights can then be used as in the Goldberger procedure. In sum, the grouped or replicated data allow us (only) the shortcut of estimating

w) from the original data without recourse to the first OLS regression. Otherwise, this binomial variable case is the same as the Bernoulli case.‘'! The fina! similarity to the last section is that statistical properties of the estimates hold only as large sample approximations, though here we

might obtain them from the fact that normality of u; holds as a reasonable approximation for a large number of observations in each group. The only real difference in this regard is that with replicated or binomial data, “large samples” refers to a large number of replications, Nj, for each observation, not to a large number of groups or observations, M. 1.4 Polytomous or Multiple Category Dependent Variables

So far, we have examined only the dichotomous variable. This is a special case of the more general problem of categorical dependent

variables.’’ Party identification, for example, is usually a nominal or categorical variable. Choice of travel (say, car, bus, or subway), occupation, religion, and many other variables in the social sciences may be in principle nominal, while practical measurement problems may yield only nominal measures of many others. Let us consider the case of J Mutually exclusive and exhaustive categories, which can be labeled arbitrarily as 1,2,..., J, but the numberings can not be taken to indicate order, let alone magnitude. An obvious approach would be to divide the dependent variable, Yi, into a set of dichotomous variables. If there are J categories, there will be

J such variables, say Y\, which equal one if Yi equals j, zero otherwise. For party identification, for example, there are three categories, Republican, Independent, and Democrat (assuming them only categorical,

23 not ordinal). Thus, we will have three dichotomous dependent variables, say Yi, Yi2, and Yi3, respectively. Then Yi: will be one if individual i is identified as a Republican, zero if i is not a Republican, and so on. Since each Yj is a dichotomous, or Bernoulli, variable, we could

assume that each one of the Yj;’s may be modeled as a linear probability model: Yq = TbyXa

+ uy

j=1,2,.-.,J

[1.2]

The only difference between equations 1.12 and the LPM of section 1.2 is that 1.12 includes J equations instead of just one. Further, if the assumptions analogous to 1.6 can be assumed to hold for all J of the

equations, the analysis of section 1.2 applies directly to each of the J equations separately. That is, OLS estimates of each separate equation in 1.12 will be unbiased but the variance of u: will again be heteroscedastic. There is one conceivable problem. We are estimating probabilities. In

section 1.2, we faced the problem that the estimated probability that Y; equalled | might be negative or greater than 1. Here, we face the additional must sum individual And, it is

problem that the probabilities estimated for each observation to | across all equations. For example, the probability that iis a Republican, a Democrat, or an Independent must be |. not particularly obvious that estimating three, four, or even

more equations separately will yield the appropriate sum, By algebraic manipulation of the so-called “normal equations” of regression (Johnston,

1984) it turns out that least squares

estimation

of each

equation separately will automatically meet that condition—provided that each equation in 1.12 contains

an intercept term. This fact yields an

added bonus. There is no reason to estimate all J equations. Rather, we need only estimate J—1 of them, for the unestimated equation can be solved for algebraically. And so doing will yield the same answer that

you would get from running the regression of it. To see why, suppose that everyone is either a Republican or a Democrat. This variable is a dichotomous one, just as we studied in section 1.2. But, we could have, as in this section, constructed two dichotomous variables, one for Republican (say Yi) and one for Democratic (Yi2) identification. And, we could run regressions on both of these, just as we have suggested doing here. But we don’t have to, for the coefficients in estimating, say

Yi2, would be equal to the negative of their counterparts in Yu. (The intercept for Yi. would be one minus the intercept of Yi.) And, that is, in

24

effect, just what we would do in solving for P(Y; = 0) in the dichotomous case: knowing PCY; = 1) means that solving for P(Y, = 0) is straight-

forward, being ]-P(Y; = 1). So, in short, the extension to the polytomous case is straightforward

so far. The OLS estimates of each of the J (or J - 1) equations in 1.12 will be unbiased. The disturbance terms, uj, will be heteroscedastic as before.

Again, a weighted, two-step approach could be used.'” However, in Goidberger’s procedure, the intercept is dropped from the equations, being replaced by the weight, wi. What this means is that the estimated

probabilities might mer sum to | after the second round of estimation. Beyond that, however, the extension of the Goldberger procedure is rather straightforward. It is not used very often, however, since a

procedure to be discussed in Chapter 3 is more attractive. Everything we have considered from section |2 to this point rests on the linear probability assumption. We have suggested that assumption is problematic. It is time to consider why it is so. 1.5 The Linearity Assumption Throughout the last three sections, we have argued that, so long as the linearity assumption of equation 1.6 is maintained, least squares estimation with a correction for heteroscedasticity and some care in interpretation in small samples is viable. Here, we argue that, since Xb. Xue must be interpreted as a probability, the fact that it is linear in X makes it highly suspect. Probabilities are, of course, restricted to the interval from 0 to 1. Asa

result, the linear probability model imposes harsh and quite possibly arbitrary constraints on the values that the regression coefficients by may assume. Suppose we have a bivariate case: PCY: = 1) = E(Yi) = bo + bi Xi

and suppose that b, is positive. Then, the smallest value that X; takes on, say Xu), must yield a predicted probability that is greater than or equa!

to zero. The largest value of Xj, say X.xj), must yield a probability no larger than one: 0 =

det

biXp)