Вы находитесь на странице: 1из 12

Thesis Proposal: Research in Education Using

Structural Equation Modeling

Susan E. Stockdale
November 11, 1994

Until around the middle of the twentieth century, gender di erences in intel-
lectual abilities were considered \natural" or inherently dependent on biological
sex di erences (Sadker, Sadker, and Klein 1991). The landmark publication of
Maccoby and Jacklin's The Psychology of Sex Di erences in 1974 exposed this
belief as a myth and not supported by the research. Although they did nd ev-
idence of a gender-related di erence favoring males in spatial and quantitative
abilities, and a di erence favoring females in verbal abilities, they argue that
there is insucient evidence to attribute these di erences to biological char-
acteristics of males and females. In subsequent research, both biological and
environmental variables have been posited as \causes" of these gender-related
di erences. Biological explanations focus on the relationship between exposure
to testoserone in utero and hemispheric activity, while environmental explana-
tions are centered on childhood socialization and child-rearing practices.
Research over the past twenty- ve years indicates that gender di erences in
math achievement favoring males are not typically found prior to high school
(Sadker, Sadker, and Klein 1991; Kimball 1989; Friedman 1989). In high school,
di erences favoring males are common, particularly in the areas of problem solv-
ing and applications. These gender-related di erences have been attributed to
a number of variables, most notably, di erential course-taking patterns and
exposure to math, attitudes toward math, teacher behavior and learning envi-
ronment, parental attitudes and expectations, and socioeconomic status as well
as other background characteristics of students. Meta-analyses examining the
e ect size in terms of central tendency and variability indicate that the dif-
ferences are not large and that the gender gap may be diminishing over time
(Feingold 1992; Kimball 1989). However, some have suggested that the decrease
in gender di erences over time may be due to changes in the achievement tests
that reduce the amount of gender bias (Feingold 1992; Jacklin 1989).
In a review of the literature, Kimball (1989) notes that while evidence sug-
gests that males may have the advantage on math achievement tests, women
may receive higher grades in math courses. In fact, the gender di erence favor-
ing females in terms of grades in math increases with more advanced courses.
The di erence between males and females on grades in math courses is frequent-
ly ignored or down-played in the literature, however. It is dicult to assess the

1
size of this e ect, according to Kimball (1989) because very few of the studies
examining grades reported the descriptive statistics necessary to calculate e ect
size.
One factor that may contribute to the gender-related di erences is exposure
to math; males enroll in more math courses than females, and it has been posited
that they have more exposure to mathematics outside of the classroom (Kimball
1989). Controlling for the number of advanced math courses does diminish
the di erences between males and females on standardized math achievement
tests, but it does not completely eliminate them. Gender-related di erences
in math exposure outside of the class may contribute to gender di erences as
well. Males have more math-related experiences outside of the classroom and
are more likely to take advanced math courses, which may contribute to more
con dence and better performance on standardized tests. It has also been found
that earlier achievement in math is not related to future math course-taking
patterns for females; high and low female achievers are equally likely to enroll
in more advanced math courses. This would suggest that the females taking
more advanced math courses are not a select sample, and therefore, higher
grades for females in advanced math courses cannot be attributed to a more
select sample of females.
In addition, some research has found that academic and non-academic inter-
actions in the classroom occur more frequently for males, and males receive more
encouragement, praise and discipline while females receive more discouragement
(Kimball 1989). High achievers received more attention than low achievers in
one study, and high-achieving males and low-achieving females received more
attention that high-achieving females in another study. One study found that fe-
males receive better grades in a cooperative atmosphere as opposed to a compet-
itive one. Other studies have found that patterns of interaction in study groups
within the classroom have detrimental e ects on females' math self-concept but
positively a ect achievement for both males and females. Few studies have
attempted to correlate classroom environment with grades or achievement on
standardized tests, but some research indicates that teacher behavior and class-
room environment may be related to gender di erences in math achievement.
Gender-related learning styles may be partially responsible for the sex d-
i erences favoring females in the classroom and the di erences favoring males
on achievement tests (Kimball 1989). Speci cally, some evidence suggests that
females have a rote memorization learning style while that of males may be
characterized as autonomous. This rote learning style disadvantages females in
unfamiliar or novel situations, such as taking standardized math achievement
tests involving materials that may not have been covered in the classroom. Fur-
thermore, the rote learning style may be related to a lack of self-con dence in
their math abilities for females.
Some research suggests that parental attitudes and expectations may be
related to the gender gap in math achievement (Kimball 1989; Jacklin 1989).
It has been found that parental attitudes and expectations di er for sons and
daughters; parents report less con dence in daughter's math abilities given equal
performance. They also have lower expectations for daughters, both in terms of

2
math course-taking and future occupation, than for sons. These attitudes could
be transmitted from the parents to the students, resulting in females' perception
of mathematics as less valuable, useful, and important for them. Usefulness of
math and perceived value of math accounts for a portion of the gender di erences
in some studies, along with gender-stereotyped parental beliefs and math anxiety
(Jacklin 1989).
Researchers in education have increasingly called for the use of longitudinal
designs in order to examine the process of education. Previous attempts to mod-
el the learning processes of K-12 students have neglected the processual nature
of schooling, and use statistical models that are incapable of fully exploiting the
information provided in longitudinal educational data. A handful of researchers
(including Parkerson, Lomax, Schiller, and Walberg 1984; Ethington and Wol e
1986; Reynolds 1991) have examined the processes of math and science achieve-
ment using structural equation modeling (SEM) to establish causal relationships
among student characteristics, latent variables, and outcomes. However, there
is no continuity among the models used in this research or their ndings, and
some social statisticians (Freedman 1987; Berk 1988; Sobel 1993) have criticized
the interpretation of SEMs as used for modeling \causal" processes.
Only one study by Ethington and Wol e (1986) has used structural equation
modeling to examine gender di erences in math achievement speci cally. They
conclude that the process of math achievement is di erent for males and females
and may not be simply additive, meaning that interactions between gender
and other explanatory variables may a ect the process of math achievement.
Speci cally, they note that di erences in ability and attitudes were less exible
among women than among men. The factors that had a positive e ect on
math achievement, greater ability and more favorable attitudes, were stronger
for males than for females and the reverse was true for variables that had a
negative e ect on achievement for women (verbal ability on attitudes toward
math). Attitudes toward math were more negatively in uenced by verbal ability
for females than for males, and attitudes had a signi cant e ect on achievement
that was stronger for males than for females.
The model used in the Ethington and Wol e (1986) study exploited data
from high school students at only two time-points: math ability was measured
as the math achievement score on a standardized test taken in 10th grade and
math achievement was the score on a standardized test, composed of some of the
same items as the 10th grade test, taken in the 12th grade. Previous research
indicates that gender di erences begin to appear in math achievement around
the eighth or ninth grade, and thus the Ethington and Wol e model would
not be able to fully capture the process of di erential math achievement. In
addition, no controls for background characteristics of students were employed,
and besides mathematics ability, the only other explanatory factors in the model
were 10th-grade verbal ability, mathematics attitudes, and exposure to math as
measured by enrollment in advanced math courses.
In contrast, the model that this I proposes to use will exploit data from
three time-points beginning with eighth grade and will employ a measure of
math ability from the 6th grade, maximizing the possibility of capturing the

3
di erential achievement processes of males and females. A host of variables will
be used as controls, including background characteristics that are conspicuously
missing from the Ethington and Wol e model as well as indicators of math ex-
posure, parental and teacher attitudes, classroom environment, peer in uence,
educational expectations, and attitudes toward mathematics. A propensity s-
core (speci ed by Rosenbaum and Rubin 1983 and 1984) for gender will be
computed using these controls, and this propensity score will be used as an
input variable in the model.
As Freedman (1987) argues, the validity of the causal interpretations of path
models, as well as SEMs, depends on whether or not the assumptions hold.
The validity of the assumptions can be threatened from three sources: 1) mea-
surement error in the predictors, 2) nonlinearity, or the functional relationship
between the predictors and the outcome is incorrectly speci ed, and 3) omitted
variables. Non-experimental data, such as that used by researchers who employ
SEM with educational data, almost never approximate the conditions necessary
to ful ll the assumptions of the linear model. However, even if these conditions
are met, causal inferences may not be valid and one may be left instead with
associations. Speci cally, the experimental model of causality that most of the
causal inferences are based on requires that assignment to levels of the \cause"
be random. Sobel (1993) suggests that more plausible inferences may be drawn
from research using non-experimental data if a set of covariates, or control vari-
ables, are used in the model. Assignment to the causal variable is conditional
upon this set of covariates, and the parameter for the casual variable can be
estimated more accurately if this set of covariates are included in the model.
Sobel (1993) notes several general instances in which researchers have in-
validly used non-experimental data to make causal inferences that are based
on the experimental model. First, they commonly fail to distinguish between
variables that are causes and those that are concomitants. Some variables that
are treated as causes should be treated as covariates, because it is unrealistic
to assume that a change in the causal variable would result in a change in the
outcome. Second, the direct e ect of a predictor on an outcome is a param-
eter estimate that is based on the assumption that the indirect e ects due to
intervening variables can be controlled for. Researchers make causal inferences
about the direct e ects of causal variables although the intervening variables
cannot be held constant. Third, researchers frequently fail to recognize possible
relations between the exogenous variables when interpreting the total e ects of
the exogenous variables. The exogenous variables may be related to (depen-
dent upon) each other, thus, it may not be legitimate to treat some variables
as determined outside of the model. Fourth, parameters between endogenous
variables are often interpreted as causal e ects. If causal inference assumes that
the researcher has the ability to manipulate the causal variables (if only through
the thought experiment), it does not seem valid to interpret these parameters
in terms of causal e ects, as the causal variable is determined within the model
and is not subject to manipulation by the researcher.
The proposed research will include a critique of the use of structural equa-
tion modeling in educational research and will examine the learning process for

4
mathematics using a state space model with panel data from the National Edu-
cational Longitudinal Study of 1988 (NELS:88). In addition to the substantive
issue of gender di erences in mathematics achievement, this study will address
two methodological concerns: 1) Is this type of modeling (state space model
or, more generally, SEM) appropriate for examining the process of achievement,
and 2) What, if any, valid inferences can be drawn from the results? Causality
cannot be \proven" with statistics, rather it is inferred. Consequently, some
or even most of the issues regarding causality and causal inference may not be
solved with the proposed model and data.
NELS:88 includes survey and achievement data for a national sample of s-
tudents with a base year in the eighth grade and two follow-ups at two year
intervals (10th and 12th grades). A two-stage strati ed probability design was
used in the base year to select the sample of eighth grade students and schools,
with the schools being the primary sampling unit. In the rst stage, a strati ed
systematic sample of 1,734 schools were selected, 1,052 of which actually par-
ticipated. The schools were strati ed according to region, urbanicity, percent
minority students, and school sector, and the probability of selection was pro-
portional to the size of the eighth grade, with oversampling of private schools.
In the second stage, a random sample of eighth grade students were selected
from within the schools. Of the 26,432 students selected to participate in the
base year, 24,599 actually participated. For the rst follow-up, 21,474 of the
students who participated in the base year were selected; this included both
drop-outs and students who were still in school. Each of these students were
again included in the sample for the second follow-up. The panel sample con-
sists of a total of 16,489 students who participated in all three waves. Sampling
weights are provided with the data to adjust for design e ects and will be used,
if appropriate, in the analysis.
The basic structure of the model is given in the gure below. No lagged
e ects are speci ed in the model because they are can be speci ed by including
additional state variables. In other word, the lagged e ects of inputs at timet
are expressed as a component of the state at timet . Input variables are at the
top of the gure, output variables are at the bottom. Unobserved state variables
mediate the relations between input and output, with time going from the left
to the right.

xi1 xi2 xit



-
?-

?- -

?-

zi0 zi1 zi2    zit
? ? ?
yi1 yi2 yit

In this analysis, the input variables will include the propensity score for gen-

5
der, computed from measured variables for each time point, and the respondent's
sex and the output variables will include the IRT number answered correctly on
the mathematics achievement test for each time point. The propensity score,
as described by Rosenbaum and Rubin (1983) \is the conditional probability of
assignment to a particular treatment given a vector of observed covariates" (41).
In this case, the \treatment" would be the respondent's sex and the covariates
would include all of the input variables (described below and contained in the
Appendix) for each time point. More speci cally, the gender propensity score
will be the predicted probability of \being" a male or a female, as estimated from
a logit regression of sex on the input (control) variables. One advantage of the
gender propensity score is that it will control for the associations between the
input variables (including background characteristics, math exposured, parental
and teacher attitudes, classroom environment, peer in uence, educational ex-
pectations, and attitudes toward math). The literature indicates that gender
may be related to a number of the variables that have an e ect on gender dif-
ferences in math achievement. If the gender di erence in achievement is in fact
a function the di erent learning experiences of males and females rather than
gender, then the e ects due to the respondent's sex should disappear when the
gender propensity score is included in the model. In addition, the propensity
score reduces the dimensionality of the problem to be examined by the state
space model.
Although the model described above does not exactly correspond with theirs,
MacCallum and Ashby (1986) have demonstrated that linear systems theory, or
state space models, are actually a special case of structural equation modeling:
\: : :the covariances of the inputs and outputs can be expressed as functions of
the LST model parameters" (p 12). The LST model can be expressed in terms
of the following equations1 :
Vt+1 = A Vt + B  Ut (1)

Wt = C Vt + D Ut (2)
Input variables are contained in U, state variables in V , and output variables
in W. If i , k , and j  are the number of input, state, and output variables,
respectively, the parameters, as MacCallum and Ashby explain:
1 The model to be used in this analysis only corresponds roughly to the one decribed by
McCullum and Ashby, with x corresponding to U , z corresponding to V , and y corresponding
to W . The left portion of the model above can be speci ed, using the terminology of McCullum
and Ashby, by the following equations:
Vt1 = A Vt0 + B Ut1
W t1 = C  V t1
The primary di erence between this model and the McCullum and Ashby model is that no
direct e ects are speci ed between inputs and outputs (parameter D ), and instead of a lagged
e ect between inputs at time0 and the state at time1 , a direct e ect is speci ed between inputs
at time1 and the state at time1 (parameter B ).

6
In these equations A is a k  k matrix of coecients representing
the e ects of the state variables at time t on the state variables at
time (t + 1); B  is a k  i matrix of coecients representing the
e ects of the inputs at time t on the state variables at time (t + 1);
C  is a j   k matrix of coecients representing the e ect of the
state variables at time t on the outputs at time t; and D is a j   i
matrix of coecients representing the e ects of the inputs at time t
on the outputs at time t. (pg 3)
The covariance structural model, or structural equation model, is composed
of a measurement model and a structural model. The measurement model,
which speci es the relationships among the measured variables can be expressed
as:
x = x +  (3)

y = y  + " (4)
In the above equations, x represents the independent measured variables, y
represents the measured outcome variables,  is the matrix of independent latent
variables,  is the matrix of errors in the measurement of x,  is the matrix of
dependent latent variables, " is the matrix of errors in the measurement of y, x
is the matrix of coecients relating the independent latent variables with the
independent measured variables, and y is the matrix of coecients relating the
dependent latent variables to the measured outcome variables. The structural
model, de ning the relationships among the latent variables, is:
 = B +  +  (5)
In this equation, B is the matrix of coecients representing the relationships
between the dependent latent variables and will have zeros on the diagonols, is
the matrix of coecients representing the relationships between the independent
and the dependent latent variables, and  contains the error terms representing
the variation in the dependent latent variables that is not explained by the
independent latent variables and the other dependent latent variables. The
covariances among the measured variables can be expressed as parameters of
the structural model, with the speci cation of the following covariance matrices:
 = E( 0 ) (6)

 = E( 0 ) (7)

" = E(""0 ) (8)

= E( 0 ) (9)

7
, the (p+q)(p+q) population covariance matrix for the measured variables
can be expressed as a function of the eight parameter matrices, x , y , B, ,
,  , " , and :
 
=    yy  yx
(10)
xy xx
 0 ) E(yx0 ) 
E(yy
= E(xy0 ) E(xx0) (11)

 
= (I B) 1 ( 0 + )(I B 0 ) 1 y (I B) 1 0x (12)
x 0 (I B) 1 0y x 0x + 
The eight parameter matrices in the structural equation model are reduced
to three in the linear systems theory model, B, , and .  is a covariance
matrix of the inputs, and B and contain copies of the A , B  , C  , and D
matrices from the state space model for each time point. The matrix containing
the errors (for the structural portion) in structural equation modeling, , are
null, as are the matrices for the measurement portion of the model.
However, some researchers have pointed out that such state space models
are not identi ed and therefore have no unique solution. MacCallum and Ashby
(1986) suggest that this can be corrected by adding error terms ev +1 and ew ,
t
which are (respectively) the variance in the state at timet+1 that is not explained
by the state and the inputs at timet , and the variance in the output at timet
that is not explained by the inputs and states at timet .
Vt+1 = A + B  Ut + ev +1
t (13)

Wt = C Vt + D Ut + ew t (14)
This can be accomplished by specifying  as a diagonal matrix with all
diagonal elements free, that is, assuming all errors across and within time points
are uncorrelated. Correlations among errors can be speci ed by freeing the
corresponding o -diagonal elements of , as will probably be necessary with
these data.
The proposed model will incorporate some of the speci cations proposed by
MacCallum and Ashby (1986), however one parameter matrix, that specifying
the direct e ects of the inputs on the outputs within time points (D ), will be
eliminated, as has been proposed by Meditch (1969). The primary goal of this
model will not be description or causal explanation of a process that occurs,
rather the goal is to achieve the best possible prediction. Some method of cross
validation will be used in addition to the goodness of t statistics to evaluate
the performance of the model with regard to prediction.
A total of 154 variables have been selected for inclusion in this analysis. A
complete listing of the variables and frequency distributions for each appear in
the Appendix. The three output variables used in this model will be the IRT

8
(Item Response Theory) number correct on the math achievement tests for the
8th, 10th, and 12th grades. The IRT measure, as opposed to formula scores, is
desirable because students were administered di erent test depending on ability
level in the 10th and 12th grades. The IRT scores provide comparable measures
of achievement for all students between time periods.
The questionnaires administered to the students in the 8th, 10th, and 12th
grades did not contain all of the same questions; some of the questions in the
base year survey were not included in the rst or second follow-ups, and some
questions not appearing in the base year survey were added to the follow-ups.
Thus, the items used to measure some of the input variables were not identical
over time. However, in most cases, comparable measures are available. Items
measuring the background characteristics at each time period are composed of
identical items from the questionnaires. These include the following:
Sex SEX, F1SEX, and F2SEX are composite variables composed of question-
naire items or information from school records.
Race/ethnicity RACE, F1RACE, and F2RACE1 are composites variables
composed of information from student and/or parent questionnaires and/or
school rosters.
Socioeconomic status BYSES, F1SES, and F2SES are composite variables
composed of items from the parent questionnaires, including father's edu-
cation, mother's education, father's occupation, mother's occupation, and
family income.
School sector G8CTRL, G10CRTL1, and, G12CTRL1 describe the type of
school, public, private, Catholic private, etc., and are composed of items
in the school administrators questionnaire.
Parent education BYPARED, F1PARED, and F2PARED are the highest
level of education attained by either parent, composed of items from the
parent questionnaires.
Urbanicity G8URBAN, G10URBAN, and G12URBN3 describe the urbanici-
ty of the student's school. These were taken from the Quality of Education
Data (QED), which was used to create the sampling frame.
Ed aspirations BYS45, F1S49, and F2S43 are items from the student ques-
tionnaires asking how far in school the respondent thinks he/she will get.
Mother's aspirations BYS48B, F1S48B, and F2S42B are items from the
student questionnaire asking how far in school the respondent's mother
wants him/her to go.
Father's aspirations BYS48A, F1S48A, and F2S42A are items from the
student questionnaire asking how far in school the respondent's father
wants him/her to go.

9
Program F1HSPROG and F2S12A measure the high school program, academ-
ic, general, or vocational, that the respondent is enrolled in. A comparable
item did not appear on the base year survey.
The remainder of the input variables were not measured consistently across
administrations of the questionnaire. A description of those found in the base
year questionnaire follows. More information about each item is contained in
the Appendix.
Classroom/learning environment BYS59F: The teaching is good; BYS59G:
Teachers are interested in students; BYS59H: Teachers praise my e orts;
BYS59I: I feel put down by my teachers; BYS59J: Most of my teachers
listen to what I say; BYS59L: Student disruptions inhibit learning.
Home learning environment BYS35A: R's family has speci c place for
study; BYS35B: R's family has a daily newspaper; BYS35C: R's family
has regularly received magazine; BYS35D: R's family has an encyclopedia;
BYS35E: R's family has an atlas; BYS35F: R's family has a dictionary;
BYS35M: R's family has more than 50 books.
Parental involvement BYS36A: Discuss programs at school with parents;
BYS36B: Discuss school activities with parents; BYS36C: Discuss things
studied in class with parents; BYS37A: R's Parents attended a school
meeting; BYS37B: R's parents spoke to teacher/counselor.
Signi cant other in uence BYS50A: Talk to father about planning high
school program; BYS50B: Talk to mother about planning high school
program; BYS50C: Talk to counselor about planning high school pro-
gram; BYS50D: Talk to teacher about planning high school program;
BYS50F: Talk to friends about planning high school program; BYS51DA:
Talk to counselor about courses at school; BYS51DB: Talk to teachers
about courses at school; BYS51DC: Talk to other adult about courses at
school.
Prior ability BYS60A: R's ability group for mathematics; BYS81B: Math
grades from grade 6 until now.
Math in uence BYS61: Talk to teacher/counselor about taking algebra;
BYS62: Did parents/guardians want R to take algebra; BYS63: Friends
encourage/ discourage R from taking algebra; BYS65: Who had the most
to say about R taking algebra.
Math courses BYS67A: Attend remedial math at least once a week; BYS67B:
Attend regular math at least once a week; BYS67C: Attend algebra at least
once per week.
Other math exposure BYS79A: Time spent on math homework each week;
BYS82J: Participated in math club.

10
Attitudes toward math BYS69A: Usually look forward to math class; BYS69:
Afraid to ask questions in math class; BYS69C: Math will be useful to my
future.
The rst and second follow-ups contain similar items that can be used as
indicators of the above listed variables. In addition to those listed, the rst and
second follow-ups also contain questions about teaching methods in mathematics
classes. The rst follow-up contains more detailed information about math
course work.
The data are currently available on CD-Rom with an electronic codebook
that can be used to select variables and automatically writes the code to produce
an SPSS or SAS system le. One variable, IRT math number answered correctly
in the second follow-up, was not available on the current version of the CD-Rom,
but will be available on the new release (which is due out this fall). The data can
be written into a raw ASCII data le to be used with other statistical software.
The software that will most likely be used for this study is EQS by Peter Bentler
or XLISP-STAT.
The utility of this model for prediction is to aid educators and policy makers
in their decision-making. The implication of the model is that all of the relevant
information for predicting math achievement is contained in the state variables.
The parameters between states contain all of the necessary information and
summarize the information contained in all of the parameters that could be
speci ed between background, intervening, and outcome variables. No direct
e ects are speci ed between math achievement at di erent points in time; only
the state variables directly a ect math achievement. With fewer parameters
than the typical structural equation model, the model may not t as well, but
it will be more stable, i.e. yield smaller standard errors. Thus, one purpose of
this project will be to produce stable predictions of math achievement.
A second, and equally important, goal of this study is to examine gender
di erences in math achievement over time while controlling for factors that con-
tribute to the di erences and are associated with gender, such as math course-
taking patterns. By computing propensity scores using the many factors related
to gender di erences in math achievement, it is anticipated that the di erences
will be greatly diminished, in comparison to past research.

11
References
Berk, R.A. 1988. \Causal Inference for Sociological Data". pps 155-172 in
(ed.) Smelser, N.J. Handbook of Sociology Newbury Park, CA: Sage Publica-
tions.
Ethington, C.A. and L.M. Wol e. 1986. \A Structural Model of Mathemat-
ics Achievement for Men and Women". American Educational Research Journal
23: 65-75.
Freedman, D.A. 1987. \As Others See Us: A Case Study in Path Analysis".
Journal of Educational Statistics 12: 101-128.
Friedman, L. 1989. \Mathematics and the Gender Gap: A Meta-Analysis
of Recent Studies on Sex Di erences in Mathematical Tasks". Review of Edu-
cational Research 59: 185-213.
Feingold, A. 1992. \Sex Variability in Intellectual Abilities: A New Look at
an Old Controversy". Review of Educational Research 62: 61-84.
Jacklin, C.N. 1989. \Female and Male: Issues of Gender". American Psy-
chologist 44: 127-133.
Kimball, M.M. 1989. \A New Perspective on Women's Math Achievement".
Psychological Bulletin 105: 198-214.
MacCallum, R. and R.G. Ashby. 1986. \Relationships between Linear Sys-
tems Theory and Covariance Structure Modeling". Journal of Mathematical
Psychology 30: 1-27.
Meditch, J.S. 1969. Stochastic Optimal Linear Estimation and Control New
York: MacGraw-Hill.
Parkerson, J.A., R.G. Lomax, D.P. Schiller, and H.J. Walberg. 1984. \Ex-
ploring Causal Models of Educational Achievement". Journal of Educational
Psychology 76: 638-646.
Reynolds, A.J. 1991. \The Middle Schooling Process: In uences of Sci-
ence and Mathematics Achievement from the Longitudinal Study of American
Youth". Adolescence 26: 133-158.
Rosenbaum, P.R. and D.B. Rubin. 1983. \The Central Role of the Propen-
sity Score in Observational Studies for Causal E ects". Biometrika 70: 41-55.
Rosenbaum, P.R. and D.B. Rubin. 1984. \Reducing Bias in Observation-
al Studies Using Subclassi cations on the Propensity Score". Journal of the
American Statistical Association 79: 516-524.
Sadker, M., D. Sadker, and S. Klein. 1991. \The Issue of Gender in Elemen-
tary and Secondary Education". Review of Educational Research 17: 269-334.
Sobel, M.E. 1993. \Causal Inferences in the Social and Behavioral Sciences".
pps1-38 in (eds) G. Arminger, C.C. Clogg, and M.E. Sobel A Handbook for
Statistical Modeling in the Social and Behavioral Sciences New York: Plenum
Press.

12

Вам также может понравиться