Académique Documents
Professionnel Documents
Culture Documents
= =
= =
y J j X b b
X X y y p
X X y y p
k jk
K
k
j
K j j
K j j
For this model, the LRPC hypothesis is that:
j
k j
jk
k j
jk
b
b
b
b
= =
' , 1
'
, 1
. (j = 2,, J) (2)
Because the true variances of the underlying dependent variables in equations (1) are not
identified, however, the true coefficients for the effects of the X
k
are not identified. Equivalently
stated, the constants of proportionality,
j
, are estimable, but their values incorporate both
differences across equations in the effects of the regressors and also differences in the variances
of the underlying dependent variables.
One can rewrite the logit model as a set of equations for latent continuous variables,
say
*
j
y , as
) 3 (
1
0
*
j
j
k
j
jk
K
k j
j
j
j
X
y
+
|
|
.
|
\
|
+ =
=
where
jk
denotes the true coefficient for the effects of the kth independent variable on the log
odds of making the jth school transition and
j
is the standard deviation of the random
disturbance
j
for the equation for the jth transition. Neither the
jk
nor the
j
are estimable from
the data without additional information. They are, however, linked to the estimable quantities in
equation (1) because
j
jk
jk
b
= (4)
Mare
Hauser and Andrew's Model
May 7, 2006
7
If the LRPC constraints (2) hold for equation (1), then the LRPC model implies that
) ,..., 2 (
1
' , 1
'
1
, 1
J j
j
j
k j
j
jk
j
k j
j
jk
= =
|
|
.
|
\
|
|
|
.
|
\
|
=
|
|
.
|
\
|
|
|
.
|
\
|
. (5)
But this expression simplifies to
*
1 ' , 1
'
, 1
j
j
j
j
k j
jk
k j
jk
=
|
|
.
|
\
|
= =
, (6)
which shows that if the LRPC model holds for the estimable quantities b
jk
, it also holds for the
true underlying parameters
jk.
Although the constant of proportionality that links the true
parameters for the different transitions for a given regressor (6) is not, in general, the same as for
the estimable parameters (2), the LRPC hypothesis for the true parameters holds if and only if it
holds for the estimable parameters as well.
Although this is an appealing result inasmuch as it implies that the LRPC hypothesis is
not bedeviled by the identification problem that plagues the logistic response model itself, it may
also be construed as a hollow one because the coefficients of proportionality ) (
*
j
that link the
equations for the different transitions are not identified. Fortunately, however, it is possible to
draw stronger and more useful conclusions when the LRPC or LRPPC model holds. As noted
above, although the coefficients of binary response models are not identified without an arbitrary
scale restriction, other functions of the data are estimable, including functions of the predicted
probabilities under the models. In particular, the effects of the regressors as represented by the
partial derivatives of the probabilities with respect to the regressors (when they exist) or
differences in adjusted predicted probabilities between different values of the regressors are
Mare
Hauser and Andrew's Model
May 7, 2006
8
estimable. This raises the question of whether the restrictions of the LRPC model on the
coefficients of the logistic response model imply similar restrictions on the estimable functions
of the predicted probabilities. In fact, this is the case. To see this, consider the logistic response
model in equation (1) and assume, without loss of generality, that the derivatives of the
probabilities with respect to the X
k
exist.
3
These derivatives are
| |
| |
j jk
K j j K j j jk
k
K j j
f b
X X y y p X X y y p b
X
X X y y p
=
= = = = =
= =
) ,..., , 1 | 1 ( 1 ) ,..., , 1 | 1 (
) ,..., , 1 | 1 (
1 1 1 1
1 1
(7)
where f
j
, the density function of the standard logistic distribution, depends on the estimated
parameters for the jth transition and a selected vector of values for X
1
, , X
K
but not on the
specific regressor X
k
under consideration. Thus, the ratio of effects of a given regressor across
two adjacent transitions evaluated at a common vector of values of the regressors is
simply
1 , 1
/
j k j j jk
f b f b . If the LRPC restriction (2) holds for the estimated coefficients, then it
holds for the derivatives too. That is,
|
|
.
|
\
|
= =
1 1 ' , 1
'
1 , 1 j
j
j
j k j
j jk
j k j
j jk
f
f
f b
f b
f b
f b
(8)
Conversely, if (8) holds, (2) holds as well.
Thus the LRPC model implies a common hypothesis about the estimable coefficients, the
unidentified true coefficients, and the estimable effects of the independent variables evaluated at
a chosen vector of values of the regressors. This is an especially attractive property of the model,
which may not be shared by other restricted versions of the model for school transitions. When a
Mare
Hauser and Andrew's Model
May 7, 2006
9
LRPC or LRPPC model fits the data, therefore, it is not only a pleasingly parsimonious model as
Hauser and Andrew emphasize, but also deeply and robustly informative about the structure of
association in school transition models.
Despite the great potential of the LRPC and LRPPC models, Hauser and Andrew
overstate the case when they say that one ought to estimate the LRPC before proposing more
complex explanations of change in the effects of socioeconomic background. Whether or not the
LRPC model fits the data may or may not rule out specific explanations and, depending on
which covariates obey the proportionality assumptions, the model may or may not yield an
interpretable result. The LRPC and LRPPC models are elegant and sometimes adequate
summaries of multivariate relationships. But explanations and interpretations of variation in the
effects of social background may occur to an investigator before, during, or after fitting less or
more parsimonious specifications of transition-specific effects and may be tested by a variety of
alternative models. My own early efforts were influenced by the OCG educational stratification
studies of Featherman and Hauser (Hauser and Featherman 1976; Featherman and Hauser 1978),
which used relatively unparsimonious specifications of trends in social backgrounds effects.
Within the constraints of additive, linear models, these studies described changes in the effects of
specific background variables and, in yielding key descriptive statistics about trends in
stratification, set the stage for subsequent thinking about the causes of these trends. When one
turns to more parsimonious models, LRPC and LRPPC are just one of many possible classes of
models. Another, as Hauser and Andrew indicate, is the model with no variation across
transitions in the effects of some or all social background variables. The family of probability
models for cumulative educational attainment is another (e.g., Breen 2005; Cameron and
Heckman 1998). Yet another is models in which any or all background effects vary
Mare
Hauser and Andrew's Model
May 7, 2006
10
systematically with the known values of macro-level covariates, such as the proportion of a
cohort at risk to a given transition or characteristics of school systems and the economy (Mare
1977; 1981b; Rijken 1999). The LRPC and LRPPC models are very useful devices for studying
educational stratification. But in a model-rich world, the investigator needs a flexible toolkit and
some guiding substantive hypotheses to see what is going on. No single class of models or
sequence of model tests is likely to do the trick.
Mare
Hauser and Andrew's Model
May 7, 2006
11
ENDNOTES
1
When the error variance is fixed it is also inappropriate to make within sample comparisons
among the coefficients for a given covariate across equations with varying subsets of covariates.
In this case, the total variance of the latent dependent variable and thus the scale of the estimated
coefficients vary from model to model as a function of the different regressors that are included.
Fixing the variance of the latent dependent variable avoids this problem. It does not, however,
avoid the problems of comparison across samples and across dependent variables.
2
My discussion rules out the case discussed by Allison (1999), who shows that one can compare
the coefficients of binary response models across samples when one assumes that the coefficient
for at least one regressor is invariant across samples. This identifying restriction may be
defensible in many substantive investigations, but, without additional information, is hard to
maintain a priori in studies of school transitions where one of the main analytic objectives is to
describe variation in social background effects across school transitions.
3
A similar result to the one presented here for the effects of continuous X
k
holds for the adjusted
predicted probabilities for different values of discrete regressors.
Mare
Hauser and Andrew's Model
May 7, 2006
12
REFERENCES
Allison, Paul D. 1999. "Comparing Logit and Probit Coefficients Across Groups." Sociological
Methods and Research 28:186-208.
Allison, Paul D. 1982. "Discrete Time Methods for the Analysis of Event Histories."
Sociological Methodology 1982 (Vol. 13): 61-98.
Bollen, Kenneth A. 1989. Structural Equations with Latent Variables. New York: John Wiley
and Sons.
Boudon, Raymond. 1974. Education, Opportunity, and Social Inequality: Changing Prospects in
Western Society. New York John Wiley and Sons.
Breen, Richard. 2005. "Statistical Models of Educational Careers." Unpublished paper. Nuffield
College. Oxford.
Breen, Richard, and John H. Goldthorpe. 1997. "Exploring Educational Differentials: Towards a
Formal Rational Action Theory." Rationality and Society 9: 275-305.
Breen, Richard, and Jan O. Jonsson. 2000. "Analyzing Educational Careers: A Multinomial
Transition Model." American Sociological Review 65:754-72.
Cameron, Stephen V. and James J. Heckman. 1998. "Life Cycle Schooling and Dynamic
Selection Bias: Models and Evidence for Five Cohorts of American Males." Journal of
Political Economy 106:262-333.
Cox, David R. 1970. The Analysis of Binary Data. London: Methuen and Co. Ltd.
Duncan, Beverly. 1965. Family Factors and School Dropout: 1920-1960. Cooperative
Research Project No. 2258, Office of Education. Ann Arbor: The University of Michigan.
Duncan, Beverly. 1967. "Education and Social Background." American Journal of Sociology 72:
363-72.
Mare
Hauser and Andrew's Model
May 7, 2006
13
Duncan, Beverly. 1968. "Trends in Output and Distribution of Schooling." Pp. 601-72 in
Eleanor B. Sheldon and Wilbert E. Moore (eds.), Indicators of Social Change. New York:;
Russell Sage Foundation.
Gamoran, Adam, and Robert D. Mare. 1989. "Secondary School Tracking and Stratification:
Compensation, Reinforcement, or Neutrality?" American Journal of Sociology 94: 1146-
1183.
Featherman, David L., and Robert M. Hauser. 1975. "Design for a Replicate Study of Social
Mobility in the United States." Pp. 219-51 in Kenneth C. Land and Seymour Spilerman
(eds.), Social Indicator Models. New York: Russell Sage Foundation.
Featherman, David L., and Robert M. Hauser. 1978. Opportunity and Change. New York:
Academic Press.
Hauser, Robert M. 1973. "Disaggregating a Social-Psychological Model of Educational
Attainment." Pp. 255-84 in Structural Equation Models in the Social Sciences. edited by A.
S. Goldberger and O. D. Duncan. New York: Seminar Press.
Hauser, Robert M. 1976. "Review Essay: On Boudon's Model of Social Mobility." American
Journal of Sociology 81: 911-28.
Hauser, Robert M. and David L. Featherman. 1976. "Equality of Schooling: Trends and
Prospects." Sociology of Education 49(2):99-120.
Hauser, Robert M. and Arthur S. Goldberger. 1971. "The Treatment of Unobservable Variables
in Path Analysis." Pp. 81-117 in Sociological Methodology 1971, edited by Herbert L.
Costner. San Francisco: Jossey-Bass.
Long, J. Scott. 1997. Regression Models for Categorical and Limited Dependent Variables.
Thousand Oaks, CA: Sage Publications.
Mare
Hauser and Andrew's Model
May 7, 2006
14
Lucas, Samuel R. 2001. "Effectively Maintained Inequality: Education Transitions, Track
Mobility, and Social Background Effects." American Journal of Sociology 106:1642-90.
Manski, Charles F., and David A. Wise. 1983. College Choice in America. Cambridge, MA.:
Harvard University Press.
Mare, Robert D. 1977. Growth and Distribution of Schooling in White Male American Cohorts:
1907-1952. Unpublished Ph.D. Dissertation. Ann Arbor. University of Michigan.
Mare, Robert D. 1979. "Social Background Composition and Educational Growth." Demography
16: 55-71.
Mare, Robert D. 1980. "Social Background and School Continuation Decisions." Journal of the
American Statistical Association 75: 295-305.
Mare, Robert D. 1981a. "Change and Stability in Educational Stratification." American
Sociological Review 46: 72-87.
Mare, Robert D. 1981b. "Market and Institutional Sources of Educational Growth." in Donald J.
Treiman and Robert V. Robinson (eds.), Research in Social Stratification and Mobility 1,
Greenwich, Connecticut: JAI Press, 205-245.
Mare, Robert D., and Huey-Chi Chang. 2006. Family Attainment Norms and Educational
Stratification in the United States and Taiwan: The Effects of Parents' School Transitions.
Pp. 195-231 in S. L. Morgan, D. B. Grusky, and G. Fields (eds.) Mobility and Inequality:
Frontiers of Research in Economics and Sociology. Stanford, CA: Stanford University
Press.
Morgan, Stephen L. 2005. On the Edge of Commitment. Stanford, CA: Stanford University
Press.
Mare
Hauser and Andrew's Model
May 7, 2006
15
Rijken, Susanne. 1999. Educational Expansion and Status Attainment: A Cross-National and
Over-Time Comparison. Utrecht: Interuniversity Center for Social Science Theory and
Methodology.
Shavit, Yossi and Hans-Peter Blossfeld. 1993. Persistent Inequality: Changing Educational
Attainment in Thirteen Countries. Boulder, CO: Westview Press.
Spady, William. 1967. "Educational Mobility and Access: Growth and Paradoxes." American
Journal of Sociology73: 273-86.
Winship, Christopher, and Robert D. Mare. 1983. "Structural Equations and Path Analysis for
Discrete Data." American Journal of Sociology 88: 54-110.
Winship, Christopher, and Robert D. Mare. 1984. "Regression Models with Ordinal Variables."
American Sociological Review 49: 512-525.