Académique Documents
Professionnel Documents
Culture Documents
Olaf Koller
Jurgen Baumert
Reciprocal effects models of longitudinal data show that academic self-concept is both a cause and an effect of
achievement. In this study this model was extended to juxtapose self-concept with academic interest. Based on
longitudinal data from 2 nationally representative samples of German 7th-grade students (Study 1: N 5 5,649, M
age 5 13.4; Study 2: N 5 2,264, M age 5 13.7 years), prior self-concept significantly affected subsequent math
interest, school grades, and standardized test scores, whereas prior math interest had only a small effect on
subsequent math self-concept. Despite stereotypic gender differences in means, linkages relating these constructs were invariant over gender. These results demonstrate the positive effects of academic self-concept on a
variety of academic outcomes and integrate self-concept with the developmental motivation literature.
398
theory. Consistent themes have emerged from reviews of the development of competence self-beliefs
(Harter, 1990, 1992, 1998; Jacobs, Lanza, Osgood,
Eccles, & Wigfield, 2002; Marsh, 1989; Marsh, Craven,
& Debus, 1991, 1998; Marsh, Debus, & Bornholt, 2005;
Watt, 2004; Wigfield, 1994; Wigfield & Eccles, 1992).
With improved methodology (better measurement,
stronger applications of confirmatory factor analyses
[CFA]), researchers have demonstrated that even very
young children are able to differentiate between
different domains of self-concept (e.g., verbal, mathematics, physical ability, physical appearance, peer
relations, relations with parents). There is clear evidence for increasing differentiation among these domains through age 12 (Marsh, 1989; Marsh & Ayotte,
2003), but not for older children (Marsh, 1989).
Age and gender differences in mean levels of selfconcept are generally small but systematic. Selfconcept declines from a young age through adolescence, levels out, and then increases at least through
early adulthood (Marsh, 1989, 1993b; see also Crain,
1996; Jacobs et al., 2002; Marsh & Craven, 1997; Wigfield et al., 1997). There are also counterbalancing
gender differences consistent with gender stereotypes.
Consistent across preadolescent, adolescent, late-adolescent/young adult periods, males report higher physical ability, physical appearance, and math
self-concepts, whereas females report higher verbal
self-concepts (Marsh, 1989; see also Crain, 1996;
Wigfield et al., 1997). Contrary to gender intensification hypotheses, gender differences did not vary
substantially with age. Based on longitudinal growth
trajectories of children in Grades 1 through 12, Jacobs
et al. (2002) reported gender stereotypic differences
and age-related declines in competence perceptions
but concluded that their results were broadly consistent with Marshs (1993b) findings of no age-related changes in gender differences in self-concept.
Most self-concept studies have focused on gender
and age differences in mean levels of self-concept but
not on factor structure differences, including relations among key constructs. Byrne and Shavelson
(1987), for example, concluded, Clearly, interpretations of mean differences in SC [self-concept] between males and females are problematic unless the
underlying construct has the same structure in the
two groups (p. 369). Hattie (1992) also emphasized
that the differences in means may not be as critical
in the development of self-concept as changes in
factor structure (pp. 177 178). Testing how relations among these constructs vary with gender and
age is even more complicated. Thus, for example,
Marsh (1993b) tested the gender-stereotypic model
that hypothesized that: (a) math self-concept would
be more highly correlated with academic self-concept and global self-esteem for boys than for girls, (b)
verbal self-concept would be more highly correlated
with academic self-concept and global self-esteem
for girls than for boys, and (c) the contrasting pattern
of results would intensify and increase with age.
Instead, however, he found support for the genderinvariant model in which relations among math,
verbal, academic, and general self-concepts did not
vary as a function of gender or age. More recently,
Watt (2004; see also Jacobs et al., 2002) demonstrated
that gender differences favoring boys for math and
English for girls showed little support for either
gender-intensification or -convergence hypotheses.
Academic self-concept and achievement: A reciprocal
effects model. The causal ordering of academic selfconcept and academic achievement has important
theoretical and practical implications, and has been
the focus of considerable research. Byrne (1996) emphasized that much of the interest in the self-concept/achievement relation stems from the belief that
academic self-concept has motivational properties
such that changes in academic self-concept will lead
to changes in subsequent academic achievement.
Calsyn and Kenny (1977) contrasted self-enhancement and skill development models. According to
the self-enhancement model, academic self-concept
is a primary determinant of academic achievement
(ASC ! ACH), whereas the skill development model
implies that academic self-concept emerges principally as a consequence of academic achievement
(ACH ! ASC). However, Marsh and colleagues
(Marsh, 1990, 1993a; Marsh, Byrne, & Yeung, 1999;
Marsh & Craven, in press) argued that much of the
early research was methodologically unsound and
inconsistent with the academic self-concept theory.
Based on theory, a review of empirical research, and
methodological advances in SEM, he argued for a
reciprocal effects model in which prior self-concept
affects subsequent achievement and prior achievement affects subsequent self-concept. In their metaanalysis of self-belief measures, Valentine, Dubois,
and Cooper (2004) also found clear support for a reciprocal effects model. They concluded that the effects of self-beliefs on subsequent performance were
stronger when the measure of self-belief was based
on domain-specific measures rather than global
measures, such as self-esteem, and when self-belief
and achievement measures were matched in terms of
subject area (e.g., mathematics achievement and math
self-concept) as is typical in self-concept research.
Academic achievement: School grades and standardized
test scores. Academic self-concept, interest, and related motivation constructs should be substantially
399
400
401
402
403
groups so that item and scale variances are comparable across groups. Although latent mean differences are typically tested with CFAs, it is also
possible to test for differences with an SEM in which
subsequent mean differences are corrected for differences in variables occurring earlier in the causal
ordering of longitudinal data (Marsh & Grayson,
1990). In the present investigation, for example, we
tested for latent mean differences between responses
by boys and girls in four constructs (self-concept,
interest, grades, test scores) at T1 and T2 in a CFA
model, and then evaluated differences at T2 controlling for differences at T1 in an SEM.
Results and Discussion
We began by evaluating the complete SEM that
included T1 and T2 measures of math self-concept,
interest, school grades, and test scores (see Table 1).
Math self-concept and interest were both positively
correlated with math grades and test scores (Table 1).
However, consistent with a priori predictions, both
self-concept and interest were systematically more
highly correlated with school grades (about .40 for
self-concept and .22 for interest) than with standardized test scores (about .30 for self-concept and .15
for interest). It is also important to note that correlations of self-concept with achievement were more
positive than the corresponding correlations between interest and achievement. Finally, the pattern
of these results and actual sizes of the correlations
were consistent across T1 and T2 responses.
In the evaluation of path coefficients relating T1
and T2 constructs, we juxtaposed the results of several models. In each case we began with the overall
model that contained all four constructs (self-concept, interest, grades, test scores). However, because
there were positive correlations among all four constructs, multicollinearity might obscure the pattern
of results. Thus, for example, it would be possible for
both self-concept and interest to have significant effects on achievement when considered separately
but for the effects of neither construct to be statistically significant when both are considered simultaneously (i.e., the unique effect of neither is significant
when the effects of each are controlled for the effects
of the other). For this reason, we also conducted a
series of supplemental models in which we evaluated the causal ordering among various pairs of
constructs (see Table 2 and Figure 1) that are more
like traditional causal models used in previous reciprocal effects research.
Math self-concept and achievement. Consistent with
a priori predictions and previous research, reciprocal
404
Table 1
Factor Solution Relating Academic Self-Concept, Interest, School Grades, and Test Scores at Times 1 and 2 in Study 1: Full Information Maximum
Likelihood Estimation
Time 1 constructs
MASC
Factor loadings
T1MASC1
.63a
T1MASC2
.77
T1MASC3
.80
T1MASC4
.63
T1MASC5
.80
T1MINT1
T1MINT2
T1MINT3
T1MINT4
T1MGrade
T1Mtest
T2MASC1
T2MASC2
T2MASC3
T2MASC4
T2MASC5
T2MINT1
T2MINT2
T2MINT3
T2MINT4
T2MGrade
T2MTst
Path coefficients
T2MASC
.57
T2Mint
.07
T2MGrd
.24
T2MTst
.09
Residual variances/covariances
T1MASC
1.00
T1Mint
.56
T1MGrd
.41
T1MTst
.32
T2MASC
T2Mint
T2MGrd
T2MTst
Correlationsb
T1MASC
1.00
T1Mint
.56
T1MGrd
.41
T1MTst
.32
T2MASC
.62
T2Mint
.39
T2MGrd
.44
T2MTst
.29
MInt
MGrd
Time 2 constructs
MTst
MASC
Mint
MGrd
MTst
.62a
.75
.80
.62
1.00a
1.00a
.72a
.84
.86
.73
.84
.65a
.78
.82
.65
1.00a
1.00a
.04
.55
.01
.02
.03
.00
.35
.17
1.00
.22
.15
1.00
.35
1.00
.22
.15
.37
.59
.25
.17
1.00
.35
.29
.16
.50
.35
.06
.02
.15
.40
1.00
1.00
.25
.12
.35
.49
.61
.16
.09
.07
.65
.06
.03
.66
.10
1.00
.41
.38
.28
1.00
.23
.15
1.00
.37
.71
1.00
Note. All variables were given a label that identifies the Time (T1 or T2), the construct (MASC 5 math self-concept, MINT 5 math interest,
Mgrd 5 math grade, or MTst 5 math test), and, for the multiple indicators of each latent construct, the item number. All parameter estimates are presented in completely standardized form. Not presented are the uniquenesses and correlated uniquenesses. Although the
full information maximum likelihood chi-square of 3516.3 (df 5 176) was highly significant because of the large sample size, the root mean
square error of approximation (RMSEA) of .058 demonstrated that the model was able to fit the data well. For comparison purposes, the
same model was estimated, in which missing values were imputed using the expectation maximization algorithm. The parameter estimates were nearly identical to those presented here and the goodness-of-fit statistics indicated a good fit to the data (normal theory
weighted least squares w2 5 4250.681, df 5 176, RMSEA 5 .064, non-normed fit index 5 .965, comparative fit index 5 .973, standardized root
mean square residual 5 .0383).
a
In the unstandardized model, the first indicator of each construct was fixed to 1.0 to fix the metric of the factor.
b
Factor correlations are based on the equivalent confirmatory factor analysis model in which all constructs are correlated. Time 1 correlations are the same as in the structural equation model (see residual variance and covariance estimates) but differ from Time 2 estimates in
that the effects of Time 1 constructs are not partialed out of the correlations, whereas they are partialed out for the Time 2 residual
variances and covariances.
p o .05. p o .001.
Time 1
T1MASC
T1MInt
1. Study 1
T2MASC
.57
.04
T2MInt
.07
.55
T2MGrd
.24
.01
T2MTst
.09
.02
2. Study 2, subject-specific interest
T2MASC
.55
.07
T2MInt
.12
.51
T2MGrd
.26
.01
T2MTst
.16
.01
3. Study 2, domain-specific interest
T2MASC
.55
.09
T2MInt
.16
.51
T2MGrd
.26
.01
T2MTst
.16
.01
4. Study 2, combined interest
T2MASC
.54
.09
T2MInt
.13
.52
T2MGrd
.26
.00
T2MTst
.16
.01
T1MGRd
Time 2
.57/.55
MSC
.07/.12
MSC
.04/.07
.09/.16
Time 1 constructs
Time 1 constructs
405
T1MTst
MInt
.03
.00
.35
.17
.02
.15
.40
.08
.04
.37
.14
.03
.03
.06
.35
.08
.01
.37
.14
.03
.00
.07
.35
.08
.04
.37
.14
.03
.02
.06
.35
MInt
.55/.51
.06
.24/.26
.03*/.08
MGrd
.35/.37
.06/.03*
.15/.06
Note. All variables were given a label that identifies the Time (T1 or
T2), the construct (MASC 5 math self-concept, MInt 5 math interest, MGrd 5 math grade, or MTst 5 math test). All models were
based on full information maximum likelihood. All parameter
estimates are presented in completely standardized form. Results
for Model 1 (see Table 1) and Model 3 (see Table 3) are presented in
more detail and are the main focus of the present investigation,
whereas critical parameter estimates for the alternative models are
presented here for comparison purposes.
p o .05. p o .001.
MGrd
MTest
.17/.14
.40/.35
MTest
406
Time 2
Time 1
MSC
MSC
.60/.60
MSC
.28/.27
MGrd
MGrd
MGrd
.60/.63
MTest
.06/.05
.45/.39
Time 2
Time 1
MInt
MInt
.03*/.08
MTest
Time 2
.59/.58
MInt
.09/.10
MGrd
.47/.50
MSC
.59/.56
.15/.14
Time 2
.15/.21
.04/.09
.39/.39
MInt
MTest
.48/.45
MTest
Time 2
.60/.60
MSC
.08/.08
MInt
MSC
.04/.10
.55/.60
MInt
Figure 2. Structural equation model paths relating Time 1 (T1) to Time 2 (T2) constructs. Separate models were fitted to selected pairs of
constructs. Stability (horizontal) paths are presented in gray and all statistically significant paths between different constructs are presented in black. The first coefficient in each box is based on Study 1 results (see also Table 1) and the second is based on Study 2 results (see
also Table 5). Only statistically significant paths are presented (except where a path was significant in only one of the two studies, in which
case the nonsignificant path is presented with an asterisk). MSC 5 math self-concept, MInt 5 math interest, MGrd 5 math grade,
MTest 5 math test scores.
Math interest and achievement. Although math interest was correlated with math achievement, there
was no support for any reciprocal effects between the
two constructs based on the overall model (Table 1).
Whereas T1 math interest had a substantial effect on
T2 math interest, it had no significant effects on either T2 math test scores (.02) or T2 math grades (.01).
Similarly, influences on T2 math interest were not
statistically significant for either T1 math test scores
(.02) or T2 math grades (.00). However, in different
supplemental analyses that considered only pairs of
the four constructs, some of these effects were statistically significant. In particular, there were statistically significant effects of T1 math interest on T2
407
gender differences in mean levels of the four constructs? As emphasized earlier, advances in SEM
allow us to incorporate both questions into a latent
factor model such that inferences about means are
based on latent mean differences derived from an
appropriate factor structure.
First, to evaluate the invariance of the SEM across
gender, we pursued a traditional two-group analysis
in which we constrained various sets of parameter
estimates to be invariant over gender. The invariance
of factor loadings is typically considered the minimal
condition for factorial invariance. In the present investigation, we compared RMSEA indexes for models
with a variety of different sets of invariance constraints ranging from model MG1 (no invariance
constraints for any parameter estimates) to the most
restrictive model MG9 (all parameter estimates
Ffactor loading, factor variances and covariances,
factor path coefficients, and measured variable
uniquenessesFconstrained to be the same in solutions for males and females). Although many models
were considered, the results are easy to summarize.
RMSEA values improved progressively for each set of
invariance constraints such that the best model (with
lowest RMSEA value) was the model imposing complete invariance across gender (MG9 in Table 3).
Table 3
Structural Equation Models of Gender Invariance of Factor Structure and Latent Means: Fit of Alternative Models
Study 1
Model
Invariance
MG1
MG2
MG3
MG4
MG5
MG6
MG7
MG8
MG9
Invariance
MG10
df
w2
RMSEA
of factor structure
352
3759
.060
366
3825
.059
382
3882
.058
386
4018
.059
393
3978
.058
402
4071
.058
409
4040
.057
413
4188
.058
429
4244
.057
of latent means
429
4244
.057
Study 2
95% CI
w2
RMSEA
95% CI
Invariance constraints
(.058 .061)
(.057 .061)
(.057 .060)
(.057 .061)
(.056 .060)
(.056 .060)
(.056 .059)
(.057 .060)
(.056 .059)
1387
1432
1467
1548
1517
1582
1553
1650
1684
.051
.051
.050
.051
.050
.051
.050
.051
.051
(.048 .054)
(.048 .054)
(.047 .053)
(.049 .054)
(.048 .053)
(.048 .054)
(.047 .052)
(.049 .054)
(.048 .054)
(.056 .059)
1684
.051
(.048 .054)
MG11
443
4286
.056
(.055 .058)
1737
.051
(.048 .053)
MG12
451
4665
.059
(.057 .060)
1919
.054
(.051 .056)
Note. In each model, factor structures for responses by males and females were compared subject to constraints that some parameter
estimates were the same (Inv 5 invariant) in the two solutions or were unconstrained and freely estimated in the two solutions (Free).
w2 5 full information chi-square; RMSEA 5 root mean square error of approximation; 95% CI 5 95% confidence interval about the RMSEA;
FL 5 factor loading; FV/CV 5 factor variance/covariance matrix; uniq 5 measured variable uniqueness; PC 5 path coefficient;
LFMD 5 latent factor mean differences (between males and females).
408
Hence, consistent with a priori predictions, the results provide strong support for the generalizability
of results of Study 1 over gender.
Second, although tangential to our main focus, we
evaluated gender differences in the latent means for
factors representing our four constructs at T1 and T2.
In pursuing this issue, we began with the model of
complete invariance of the latent factor structure and
evaluated whether item intercepts were invariant
over gender. Item intercepts reflect the difficulty of
an item in the sense that students at a given level of
the underlying latent construct (e.g., math self-concept) will give systematically more positive responses to easy items and systematically less
positive response to difficult items. At least reasonable support for the invariance of item intercepts
across males and females is typically taken to be the
minimal condition for valid comparisons of mean
differences. Unless there is such support, differences
on latent means are not consistent across items used
to define the factor. Fortunately, a comparison of the
RMSEAs for models MG10 and MG11 provided
good support for the invariance of item intercepts.
Next, we compared models in which latent means
are constrained to be equal across responses by
males and females. In contrast to all other tests of
invariance over gender, comparison of these models
(MG11 and MG12 in Table 3) indicated that the latent
means differed for males and females.
To evaluate the nature of these latent mean differences (Table 4), we compared responses for males
and females on the four T1 constructs, the corresponding four T2 constructs without controlling for
T1 constructs, and the corresponding four T2 constructs after controlling for T1 constructs. In general,
males had substantially higher math self-concepts
and interest at T1 and T2. Controlling for T1 constructs largely, but not completely, eliminated differences in these two constructs at T2. For math test
scores, males scored higher than did females at T1,
but their advantage was much smaller at T2 so that
after controlling for T1 scores females actually did
slightly better than males at T2. For math grades,
there were no significant differences between males
and females at T1, and girls performed marginally
better than did males at T2.
In summary, tests of gender differences largely
supported a priori predictions that males would have
substantially higher math self-concept and interest
scores and moderately higher math test scores.
However, also consistent with a priori predictions
based on the gender-invariant model, the pattern of
path coefficients linking T1 constructs to T2 constructs
was remarkably similar for male and female students.
Table 4
Latent Mean Differences for Males and Females in Four Math
Constructs (Positive Values Reflect Higher Scores for Males)
Study 1
Math self-concept
Math interest
Math Grades
Math test scores
Study 2
Math self-concept
Math interest
Math grades
Math test scores
Time 1
Time 2
Time 2
No control
No control
Control
Mn
SE
Mn
SE
Mn
SE
.46
.29
.00
.26
.03
.03
.03
.03
.39
.29
.08
.10
.03
.03
.03
.03
.12
.10
.08
.06
.03
.03
.02
.02
.46
.27
.16
.33
.05
.05
.04
.04
.34
.35
.06
.27
.04
.05
.05
.04
.06
.21
.17
.06
.04
.05
.04
.02
Study 2
The purpose of Study 2 was to evaluate the replicability of results from Study 1 and the generalizability
of results across responses by students in two different school years to two different interest measures
based on data from the German component of the
Third International Mathematics and Science Study
(TIMSS; Baumert et al., 1997). Unlike the international TIMSS study, the German component was
longitudinal in that math achievement, interest, and
self-concept were collected in both Grades 7 and 8.
Study 2 contained the same math self-concept and
interest items used in Study 1 and similar measures
of math achievement (standardized test scores and
school grades). It differed in that data were based on
responses by students in Grades 7 and 8 so that the
data collection waves were separated by a full academic year (Study 1 was based on responses from
two occasions in Grade 7) and data in Study 2 were
collected about 3 years after those in Study 1. In
addition, two measures of interest were used in
Study 2. To maintain comparability with Study 1, the
same interest items were included in Study 2.
However, a separate measure of interest was also
included based on subsequent theoretical work by
Krapp, Schiefele, and colleagues (e.g., Krapp, 2000;
Method
Data. Study 2 was based on a sample of German
Grade 7 students who participated in the TIMSS
(Baumert et al., 1997; Beaton et al., 1996). The sample
was nationally representative with respect to region,
school type, and gender. The German TIMSS study
in Grades 7 and 8 contained some national extensions compared to the international design. The
longitudinal, nationally representative sample consisted of 128 randomly selected schools in which one
class per school was sampled randomly. Most students (90%) were born in Germany, including 94%
with German citizenship (including 3.4% with dual
citizenship). Students spoke German at home always
or almost always (88%) or sometimes (10%), and
most were Caucasian (95%). Because the sample was
representative, it was heterogeneous in relation to
socioeconomic status. The final sample in the present
investigation included a total of 2,264 students (50%
female, M age in Grade 7 5 13.7 years) who were
tested at two points. Excluded from this final sample
were schools that participated in the study on only
one of the two occasions, or for which there were less
than 10 students who responded.
Measures. The math self-concept and class-specific interest measures were the same as those used
in Study 1, whereas the domain-specific interest
measure was based on a new five-item scale (see the
Appendix). Math achievement in Grade 8 was
measured by 158 items that were part of the official
TIMSS item base. The items were distributed over
eight booklets. Each booklet contained between 30
and 40 items, some specific to that booklet and some
anchor items common to all booklets. Students
worked on one booklet each, thus allowing broad
subject matter coverage without student exhaustion.
All items were checked for curricular validity. Six
409
410
Table 5
Factor Solution Relating Academic Self-Concept, Interest, School Grades, and Test Scores at Times 1 and 2 in Study 2: Full Information Maximum
Likelihood Estimation
Time 1 constructs
MASC
Factor loadings
T1MASC1
.57a
T1MASC2
.74
T1MASC3
.82
T1MASC4
.51
T1MASC5
.83
T1MINT1
T1MINT2
T1MINT3
T1MINT4
T1MGrade
T1Mtest
T2MASC1
T2MASC2
T2MASC3
T2MASC4
T2MASC5
T2MINT1
T2MINT2
T2MINT3
T2MINT4
T2MGrade
T2MTst
Path coefficients
T2MASC
.55
T2Mint
.12
T2MGrd
.26
T2MTst
.16
Residual variances/covariances
T1MASC
1.00
T1Mint
.58
T1MGrd
.52
T1MTst
.36
T2MASC
T2Mint
T2MGrd
T2MTst
Correlationsb
T1MASC
1.00
T1Mint
.58
T1MGrd
.52
T1MTst
.36
T2MASC
.64
T2MInt
.43
T2MGrd
.47
T2MTst
.35
MInt
MGrd
Time 2 constructs
MTst
MASC
MInt
MGrd
MTst
.52a
.76
.80
.54
1.00a
1.00a
.62a
.79
.84
.64
.85
.62a
.76
.81
.62
1.00a
1.00a
.07
.51
.01
.01
.08
.04
.37
.14
1.00
.24
.17
1.00
.40
1.00
.24
.17
.42
.58
.25
.18
1.00
.40
.39
.22
.53
.36
.03
.03
.06
.35
1.00
.58
.25
.21
.14
.64
.13
.06
.66
.13
.73
1.00
.27
.12
.31
.46
1.00
.55
.54
.39
1.00
.33
.20
1.00
.38
1.00
Note. All variables were given a label that identifies the Time (T1 or T2), the construct (MASC 5 math self-concept, MINT 5 math interest,
Mgrd 5 math grade, or MTst 5 math test), and, for the multiple indicators of each latent construct, the item number. All parameter estimates are presented in completely standardized form. Not presented are the uniquenesses and correlated uniquenesses. Although the
full information maximum likelihood chi-square of 1106.477 (df 5 176) was highly significant because of the large sample size, the root
mean square error of approximation (RMSEA) of .054 demonstrated that the model was able to fit the data well. For comparison purposes,
the same model was estimated, in which missing values were imputed using the expectation maximization algorithm. The parameter
estimates were nearly identical to those presented here and the goodness-of-fit statistics indicated a good fit to the data (normal theory
weighted least squares w2 5 1181.9, df 5 176, RMSEA 5 0.0563, non-normed fit index 5 0.970, comparative fit index 5 0.977, standardized
root mean square residual 5 0.0378).
a
In the unstandardized model, the first indicator of each construct was fixed to 1.0 to fix the metric of the factor.
b
Factor correlations are based on the equivalent confirmatory factor analysis model in which all constructs are correlated. Time 1 correlations are the same as in the structural equation model (see residual variance/covariance estimates) but differ from Time 2 estimates in
that the effects of Time 1 constructs are not partialed out of the correlations, whereas they are partialed out for the Time 2 residual
variances and covariances.
po.05. po.001.
of the results, we focus on the model with classspecific measures of interest (Table 5) but also briefly
summarize results based on the other models as well
(see Table 2).
Math self-concept and class-specific interest were
both positively correlated with math grades and test
scores (Table 5), but both self-concept and interest
were more highly correlated with school grades than
with standardized test scores. Also, correlations between self-concept and both achievement constructs
(grades and test scores) were more positive than the
corresponding correlations between interest and
achievement. These patterns of results were consistent across T1 and T2 responses.
In evaluating path coefficients relating T1 and T2
constructs, we focused on the overall model that
contained all four constructs (self-concept, interest,
grades, test scores) and was based on the class-specific measure of interest. However, we also juxtaposed the results of models in which interest was
based on responses to the class-specific items, the
domain-specific items, or both sets of interest items.
As in Study 1, we also conducted supplemental
analyses in which we evaluated various pairs of constructs separately (see Figure 2) that were more like
traditional causal models used in previous research.
Math self-concept and achievement. Consistent with
previous self-concept research (and Study 1), there
were reciprocal effects between math self-concept
and achievement in the overall model that included
all constructs (see Table 5 and Figure 1). Of particular
relevance, the effect of T1 math self-concept was
statistically significant for both T2 math grades (.26)
and T2 math test scores (.16), and the effects of T1
math self-concept were greater for T2 school grades
than for T2 test scores. The effects of T1 math grades
on T2 math self-concept were small (.08) but highly
significant (po.001), whereas the effects of T1 math
test scores were even smaller (.03) and not statistically significant. Hence, whereas the effects were
reciprocal, the effects of self-concept on achievement
are stronger than the effects of achievement on selfconcept. In additional models based on domainspecific measures of interest or both class- and domain-specific measures of interest (Table 2), these
path coefficients were nearly identical to those in
Table 5 based on the class-specific measure of interest (none of the path coefficients differed by more
than .05 and the pattern of significant and nonsignificant effects was the same in all three models).
As in Study 1, we constructed separate models to
evaluate the reciprocal effects of math self-concept
with math school grades and with math test scores
(excluding math interest; see Figures 2.1 and 2.2 for
411
Study 2). The pattern of results in these supplemental models was the same as in the overall model,
but the effects tended to be strongerFparticularly
the effect of T1 math self-concept on T2 math grades
(.27; Figure 2.1) and on T2 math test scores (.21;
Figure 2.2). The effects of T1 math grades and T1
math test scores on T2 math self-concept, although
clearly smaller than the effect of T1 math self-concept
on T2 math achievement, were also statistically significant. Hence, these supplemental analyses supported the reciprocal effects model of self-concept
and achievement.
Math interest and achievement. Math interest was
correlated with both measures of math achievement.
However, in the overall model, the effects of T1 math
interest were nonsignificant for both T2 school
grades and T2 test scores, as were the effects of T1
test scores and T1 school grades on T2 math interest
(Table 5). Furthermore, this consistent pattern of
near-zero, nonsignificant effects between interest and
achievement was consistent across different models
based on the class-specific, domain-specific, and
combined (class- and domain-specific) measures of
math interest (see Table 2). However, in supplemental analyses that considered math interest in
combination with either grades or test scores (excluding math self-concept), some effects involving
math interest were statistically significant. In particular, math interest had statistically significant effects
on both math grades (Figure 2.3) and math test scores
(Figure 2.4), whereas the effects of math grades on
math interest (Figure 2.3) were also significant.
Math self-concept and interest. Based on the full
model (Table 5), there was also some support for the
reciprocal effects of math self-concept and math interest; the effects of T1 math self-concept on T2 math
interest (.12) and of T1 math interest on T2 math selfconcept (.07) were both statistically significant. This
same pattern of results was evident in additional
models using class-specific, domain-specific, and
combined (class- and domain-specific) measures of
math interest (see Table 2). However, even in the
supplemental analyses that excluded the math test
scores and math grades (see Figure 2.5), the reciprocal effects of self-concept on interest and interest
on self-concept did not exceed .10.
Gender differences. As in Study 1, we tested the
generalizability of the results across gender. To facilitate comparisons between Studies 1 and 2, we
only considered the subject-specific measure of interest in Study 2 that was the same as the interest
measure in Study 1. We began with two-group
invariance tests in which we constrained various
sets of parameter estimates to be invariant over
412
responses by males and females. Based on comparison of RMSEA values across models MG1 to MG9,
there were almost no differences in fit between any
of the models. In particular, the most restrictive
model (in which all parameters were constrained to
be the same for males and females) had an RMSEA
of .051, equal to that in the least restrictive model in
which there were no invariance constraints. Hence,
consistent with a priori predictions, the results provide strong support for the generalizability of results
over gender.
Next, we evaluated gender differences in the
means of latent factors representing our four constructs at T1 and T2. A comparison of the RMSEAs
for models MG10 and MG11 provided good support
for the invariance of item intercepts, whereas comparison of models MG11 and MG12 (Table 3, Study
2) indicated that the latent means did differ. In general, males had substantially higher latent means for
math self-concept, math interest, and math tests at T1
and T2 (Table 4, Study 2). Whereas males had
slightly higher math grades at T1, there was no significant difference between males and females at T2
so that females had slightly higher school grades at
T2 after controlling for T1 constructs. A comparison
of the results for Studies 1 and 2 (see Figure 1) shows
that the size and pattern of statistically significant
and nonsignificant path coefficients from each study
are similar.
General Discussion
Although developmental and educational psychologists posit interest and self-concept as primary determinants of outcomes such as achievement,
performance, and choice of behavior, as emphasized
by Wigfield and Eccles (2002), there is a need to integrate the developmental and educational psychology research traditions to provide a more complete
picture. Even though there is substantial overlap of
research into academic self-concept and academic
interest, there have been few if any longitudinal
causal-ordering studies of relations among self-concept, interest, achievement, and gender differences.
We began by arguing that a critical question in
self-concept and motivation research is whether
there are causal links from prior measures of academic self-concept, academic interest, and academic
achievement to subsequent measures of these same
constructs. Our results provide clear evidence that
prior academic self-concept does predict subsequent
academic achievement beyond what can be explained in terms of prior measures of academic interest,
school grades, and standardized achievement test
self-concept were substantially related to coursework selection in different school subjects, domainspecific components of self-concept were much better predictors of course selection.
The focus of reciprocal effects models in academic
self-concept research has been primarily on achievement. However, Wigfield and Eccles (1992, 2002) have
suggested that whereas self-concept is more strongly
related to actual achievement, task values may be
more strongly related to choice behavior (e.g.,
coursework selection). There is a need for further research juxtaposing the combined effects of academic
self-concept, interest, and achievement on a more
varied set of academic choice behaviors. Theoretically,
our research helps bridge gaps between the educational and developmental research literatures, and
among large bodies of self-concept research, interest
research, and more general approaches to motivation
such as expectancy-value theory.
Of course, the direction of causality among academic self-concept, academic interest, and achievement has important practical implications for
educators, as well as for educational and developmental psychologists who work in school settings.
If the direction of causality were from academic
self-concept and interest to achievement (a selfenhancement model), teachers should concentrate
more effort on enhancing students self-concepts and
intrinsic interest rather than focusing on achievement. On the other hand, if causality were from
achievement to self-concept and interest (a skill development model), teachers should focus on improving academic skills as the best way to improve
self-concept and interest. In contrast to both these
apparently simplistic (either or) models, the reciprocal effects model implies that academic selfconcept, interest, and academic achievement are
reciprocally related and mutually reinforcing. Improved academic self-concepts and interest will lead
to better achievement, and improved achievement
will lead to better academic self-concepts and interest. Thus, for example, if teachers enhance students
academic self-concepts and interest without
improving achievement, the gains in self-concept
and interest are likely to be short-lived. However,
if teachers improve students academic achievement
without fostering students self-beliefs in their academic capabilities and intrinsic interest, the
achievement gains are also unlikely to be long lasting. If teachers focus on one construct to the exclusion of the other, both are likely to suffer. The
reciprocal effects model suggests that the most effective strategy is to improve academic self-concept,
interest, and achievement simultaneously.
413
References
Allison, P. D. (2001). Missing data [Sage University Papers
Series on Quantitative Applications in the Social Sciences, 07-136]. Thousand Oaks, CA: Sage.
Baumert, J., Lehmann, R. H., Lehrke, M., Schmitz, B.,
Clausen, M., Hosenfeld, I., et al. (1997). TIMSS: Mathematisch-Naturwissenschaftlicher Unterricht im internationalen Vergleich [TIMSS: Mathematics and science
instruction in an international comparison]. Opladen,
Germany: Leske & Budrich.
Baumert, J., Schnabel, K., & Lehrke, M. (1998). Learning math in school: Does interest really matter? In
L. Hoffmann, A. Krapp, K. A. Renninger, & J. Baumert
(Eds.), Interest and learning (pp. 327 336). Kiel,
Germany: Institut fur die Padagogik der Naturwissenschaften IPN.
Beaton, A. E., Mullis, I. V. S., Martin, M. O., Gonzales, E. J.,
Kelly, D. L., & Smith, T. A. (1996). Mathematics achievement in the middle school years: IEAs Third International
Mathematics and Science Study. Chestnut Hills, MA: Boston College.
Bollen, K. A. (1989). Structural equations with latent variables.
New York: Wiley.
Bouffard, T., Marcoux, M., Vezeau, C., & Bordeleau, L.
(2003). Changes in self-perceptions of competence
and intrinsic motivation among elementary schoolchildren. British Journal of Educational Psychology, 73,
171 186.
Byrne, B. M. (1996). Academic self-concept: Its structure,
measurement, and relation to academic achievement. In
B. A. Bracken (Ed.), Handbook of self-concept (pp. 287
316). New York: Wiley.
Byrne, B. M. (1998). Structural equation modeling with LISREL, PRELIS, and SIMLIS: Basic concepts, applications and
programming. Mahwah, NJ: Erlbaum.
Byrne, B. M., & Shavelson, R. J. (1987). Adolescent selfconcept: Testing the assumption of equivalent structure
across gender. American Educational Research Journal, 24,
365 385.
Calsyn, R., & Kenny, D. (1977). Self-concept of ability and
perceived evaluations by others: Cause or effect of academic achievement?. Journal of Educational Psychology,
69, 136 145.
Crain, R. M. (1996). The influence of age, race, and gender
on child and adolescent multidimensional self-concept.
In B. A. Bracken (Ed.), Handbook of self-concept: Developmental, social, and clinical considerations (pp. 395 420).
Oxford, England: Wiley.
Csikszentmihalyi, M., & Schiefele, U. (1993). Die Qualitat
des Erlebens und der Proze des Lernens [The quality of
experience and the process of learning]. Zeitschrift fur
Padagogik, 39, 207 221.
Deci, E. L., & Ryan, R. M. (1985). Intrinsic motivation and
self-determination in human behavior. New York: Plenum.
Eccles, J. S. (1983). Expectancies, values, and academic
choice: Origins and changes. In J. Spence (Ed.),
Achievement and achievement motivation (pp. 87 134). San
Francisco: Freeman.
414
415
416
Appendix
Self-Concept and Interest Items Used in
Studies 1 and 2
Math Self-Concept (Studies 1 and 2)
I would much prefer math if it werent so hard.
(1 5 strongly disagree to 4 5 strongly agree)
Although I make a real effort, math seems to be
harder for me than for my fellow students.
(1 5 strongly disagree to 4 5 strongly agree)
Nobodys perfect, but Im just not good at math.
(1 5 strongly disagree to 4 5 strongly agree)
Some topics in math are just so hard that I know
from the start Ill never understand them.
(1 5 strongly disagree to 4 5 strongly agree)
Math just isnt my thing. (1 5 strongly disagree to
4 5 strongly agree)