Professional Documents
Culture Documents
2 SEPTEMBER 1980
Psychological Bulletin
The topic of this article is the analysis of topics such as these has resulted in the collec-
reciprocal causal effects. Often, questions about tion and analysis of large amounts of longi-
reciprocal causal effects have been phrased, tudinal panel data.
Does X cause F or does Y cause X? More Data from a longitudinal panel consist of
formally, many have spoken of a determination observations on n cases at T (I = 1, . . . , T)
of causal predominance or of a preponderant time points or waves. At each time point, ob-
causal effect. Examples of such research ques- servations on one or more variables are ob-
tions in developmental psychology include the tained. Much attention is given to the simplest
reciprocal influences in mother-child interac- relevant panel design : the two-wave, two-vari-
tion (Clarke-Stewart, 1973) and relationships able (2W2V) longitudinal panel. For 2W2V
between infant intelligence and infant behavior panels the two variables are labeled X and Y.
(Crano, 1977). Examples from educational re- These variables are subscripted to indicate the
search include the relationship between teacher time of measurement. Thus for each individual
expectation and student achievement (Crano case, measures X\, Y\, X%, and Y^ are available.
& Mellon, 1978; Humphreys & Stubbs, 1977; Cross-lagged correlation (CLC) is currently
West & Anderson, 1976) and the relationship the most popular procedure in many areas of
between self-concept and achievement (Bach- psychological and educational research for
man & O'Malley, 1977; Calsyn & Kenny, identifying causal effects from longitudinal
1977; Purkey, 1970). Empirical research on panel data. Most often CLC is used to deter-
mine a predominant causal influence—the
causal winner.
This research was supported by a seed grant from the
Spencer Foundation. The assistance of David Brandt Users of CLC often make enthusiastic claims.
and Michele Zimowski in this research is gratefully For example, Crano and Mellon (1978) as-
acknowledged. serted,
Requests for reprints should be sent to David Rogosa, i
who is now at the School of Education, Stanford Uni- With the introduction of the cross-lagged panel correla-
versity, Stanford, California 94305. tional method. . ., causal inferences based on correla-
245
246 DAVID ROGOSA
tional data obtained in longitudinal panel studies can purpose is to identify the parameters in models
be made and enjoy the same logical status as those de- for longitudinal panel data that represent re-
rived in the more standard experimental settings,
(p. 41) ciprocal causal effects. To simplify the exposi-
tion, models for 2W2V data are emphasized;
Although technical deficiencies in CLC have extensions to panel data with additional waves
been noted (Bohrnstedt, 1969; Duncan, 1969; or variables are straightforward in most cases.
Goldberger, 1971; Heisc, 1970), CLC is widely Definitions of causal effects obtained from con-
recommended (e.g., Calsyn, 1976; Calsyn & sideration of structural regression models,
Kenny, 1977; Clarke-Stewart, 1973; Crano, continuous-time models, and multiple time-
1974, 1977; Humphreys & Stubbs, 1977; series models are identical.
Kenny, 1975).
I find that CLC does not provide sound in-
Structural Regression Models
formation about casual effects. CLC may in-
dicate the absence of direct causal influence A structural regression model for longitudinal
when important causal influences, balanced panel data is one plausible model from which
or unbalanced, are present. Also, CLC may the longitudinal panel data could have been
indicate a causal predominance when no causal generated. For two variables, X and Y, the
effects are present. Moreover, CLC may indi- causal influences are represented by the regres-
cate a causal predominance opposite to that of sion parameters of the path from a prior X to a
the actual structure of the data; that is, CLC later Y and from a prior Y to a later X. This
ma)r indicate that X causes F when the reverse representation can be formulated for two-wave
is true. or multiwave panel data. (When the regression
A basic deficiency in CLC is the lack of an model is formulated in terms of latent variables
explicit definition of a causal effect. Without a having multiple indicators at each time point,
clearly defined quantity to be estimated, it is the causal effects are represented by the re-
not surprising that CLC fails to provide sound gression parameters for the structural regres-
information about causal effects. The assess- sion equations that relate the latent variables.)
ment of causal effects should be based on a Previous formulations of regression models
model for the data in which causal effects are for panel data with reciprocal causal effects
identified. In the next section definitions of have focused on models for 2 W2V data (Duncan,
causal effects derived from three different 1969, 1972, 1975; Goldberger, 1971; Heise,
models for longitudinal data are presented. 1970). For this simple configuration, the
Also, the emphasis in CLC on the determi- structural regression model is equivalent to a
nation of a causal winner is unwise. Causal pre- path analysis model. Figure 1 is a representa-
dominance is not the only important question. tion of a specific structural regression model
The reciprocal nature of many social and de- for 2W2V data. This configuration can also be
velopmental processes makes determination represented by the structural regression
only of causal predominance an oversimplifi- equations:
cation of the research problem. Measures of
X, = 7,7!
the strength and duration of the reciprocal
relationship and of the specific causal effects
are more informative than the determination
The parameters /3i and 71 represent the influ-
of a causal winner.
ence of a variable on itself over time. The
parameters fo and 72 represent the lagged,
Causal Effects in Longitudinal Panel Data reciprocal causal effects between X and F.
Thus /32 and 72 are key quantities in the in-
To evaluate the usefulness of CLC, it is vestigation of reciprocal causal effects in
necessary to develop operational definitions of 2W2V panels.
causal effects in longitudinal panel data. Defini- Restrictions on the nature of the causal in-
tions of causal effects are based on models for fluences between X and Y are built into Figure
the panel data. No complete treatment of 1 and Equation 1. Most important is the as-
causality is attempted or claimed; the primary sumption that all causal influences are lagged;
CROSS-LAGGED CORRELATION 247
simultaneous causal influences between X2 and of variables over time (Coleman, 1968; Han-
F2 are not included. Also, assumptions of lin- nan & Tuma, 1979). A simple two-variable
earity and additivity of causal influences are model for rates of change that incorporates
built into this model. reciprocal influences between X and F is
The absence of a causal effect between vari-
ables is represented by a zero value of the rele- dX(f)
vant model parameter. In particular, in
Equation 1 the absence of any direct causal dY(t)
effects between X and F is represented by (2)
dl
/32 = Y2 = 0. (Definitions of the absence of any
causal effects between X and F will prove to be Equation 2 is a system of coupled differential
importantin the examination of "spuriousness" equations which stipulates that the rates of
in CLC.) Also, a causal predominance of X over change of X and F at any time depend linearly
F would be represented by a zero (or negligi- on the levels of X and F. The parameters 62
ble) value of 72 and a large value of fa. A causal and ca represent the cross-effects or couplings
predominance of F over X is represented in between X and F. Note that Equation 2 is
the same manner. deterministic; for my purposes this limitation
Standardized versions of the structural is not crucial.
parameters in Equation 1 also may be used to Although rates of change are not directly
define causal effects. Standardized parameters observable, the solution of the system of dif-
are denoted by an asterisk. Explicitly, ferential equations in Equation 2 yields equa-
tions in terms of the observable variables of
Pi a 1 —l I)
ft* = Pi the same form as in Equation 1. The param-
eters fa and 72 are monotone increasing func-
tions of b'2 and c2 and depend on the time be-
= 7i 72 =
72
tween waves and the parameters of Equation 2
Similar representations of reciprocal causal (Kaufman, 1976). Thus the representation of
effects can be made in structural regression reciprocal effects in the structural regression
models for data from longitudinal panel de- model in Equation 1 is consistent with that of
signs that are far more complex than a 2W2V the model for rates of change in Equation 2.
design. For example, multiple indicators of Also, instead of only corresponding to the ex-
latent variables £ and t\ may be available at perimental lag between waves, regression
each time of measurement. In the structural models for panel data can be thought of as
regression model for these data, the role of fa reflecting a process in which causal influences
and 72 is unchanged; these parameters repre- and resulting adjustments are continuous in
sent the lagged reciprocal causal influences be- time (Coleman, 1968; Hannan & Tuma, 1979).
tween the latent variables £ and j;.
Multiple Time-Series Models
Continuous-Time Models
In econometrics the detection of reciprocal
Another approach to modeling panel data is
causal effects from time-series data has at-
to formulate equations for the rate of change
tracted considerable interest. The analysis by
Sims (1972) of the reciprocal causal influences
of mone}' stock and income is the best known
example of this work. Definitions of reciprocal
causal effects in multiple time-series data were
formulated by Granger (1969). Subsequent
work (Pierce, 1977; Pierce & Haugh, 1977)
has interpreted and extended these definitions.
The definitions of causality are based on
Figure 1. Structural regression model for 2W2V panel predictability criteria. Loosely speaking, one
data. time series, for example, X ( l ) , causes another
248 DAVID ROGOSA
terpretation regardless of the magnitude of the is, only the parameter h can be estimated from
difference and regardless of the magnitudes of the 2W2V data. 3
the individual correlations. Also, no distinction The strategy of investigating the tenability
is made between large and equal cross-lagged of a null hypothesis of a spurious association
correlations and small and equal cross-lagged between X and F is not without merit. How-
correlations. ever, Kenny's (1975) model of spurious associ-
Kenny (1975) interpreted the null hypothe- ation is just one of several models in which
sis of equal cross-lagged correlations as a null direct causal influences between X and F are
hypothesis that the relationship of X and F is absent. The term spuriousness is better con-
a result of the common influence of an unmea- sidered as a generic term for the absence of
sured third variable and not a result of direct direct causal influences between X and F.
influences between X and F. Kenny termed Many alternative models for 2W2V data
this a null hypothesis of spuriousness. Rejection can be constructed that are true to this broad-
of this null hypothesis (under certain assump- ened notion of spuriousness. Two examples are
tions) leads to a conclusion of direct causal the common factor model with lagged effects
effects between X and F. Although Kenny em- considered in Duncan (1972, Figure 12) and a
phasized this test for spuriousness over an at- restricted version of the 2W2V model in Figure
tribution of causal predominance, all applica- 1 in which /32 = 0 and 72 = 0. As /32 and y2
tions of CLC seek an attribution of causal represent the magnitudes of the lagged causal
predominance. effects between X and F, it follows from the
Kenny (1975) posited a very specific model definition of causal effects that the absence of
for his null hypothesis of spuriousness, which direct causal effects between X and F be
is depicted in Figure 3. This 2W2V model with represented by the model in Figure 1, with
one evolving common factor (F) is also ana- & = 72 = 0.
lyzed from the standpoint of path analysis by Rarely is it recognized that the objectives of
Duncan (1972, especially pp. 63-74). Duncan CLC are modest and limited. In its most com-
plete form, CLC purports to distinguish only
showed that this model is underidentified; that
between spuriousness and a causal predomi-
nance for one of the variables. Even if CLC were
valid for its stated objectives, it falls short of
providing an adequate description of causal
influence in panel data. The failure of CLC to
achieve even these limited objectives makes the
status of CLC as the primary analysis method
for panel data in education and psychology
very unfortunate.
3
From examination of the models in Figures 1 and 3,
it might appear that the model in Figure 1 contains the
restriction that the autocorrelations among the residuals
in Figure 3 (denoted by the parameters s and t) are zero.
This is not the case, as nonzero values of these param-
eters are incorporated into the parameters /5i and 71.
That the parameters of the model in Figure 1 do reflect
nonzero values of s and / can he seen from the result that
the restriction s = I — 0 requires that /3m = /Sj-n-
Also, the model in Figure 3 does not require that u and
v in Equation 1 be uncorrelated.
The restrictions that /( (a correlation) not be greater
than one in magnitude requires that Pxtf<fY\Xv <• Px\f\'
Px.2Yr Kenny (1975) stated that a violation of this re-
striction in sample data is indicative of a causal effect.
Figure 3. Model for 2W2V panel data used to represent Also, if J and (are both positive, then the model requires
spuriousness in CLC. that PxiV^Cv^x, < Pjrjjfjpy^j-
250 DAVID ROGOSA
•G)
pxpy { 9 1 comparisons of cross-lagged correla-
tions can be made.) The usual practice in CLC
correlations do not support a conclusion of
the absence of direct causal effects, and un-
equal cross-lagged correlations do not support
analyses of such data (Calsyn, 1976; Crano, a conclusion of causal predominance. The ex-
1977) is to tally the statistically significant amples and generalizations based on Equations
differences between the cross-lagged correla- 3, 4, and 5 almost always incorporate the as-
tions in both directions; the causal winner is sumption that the structural parameters and
the variable with the greater number of sig- the correlations are nonnegative. This assump-
nificant differences in the appropriate direction.
tion is not limiting, but in some cases negative
values would require changes in wording.
Results for Two-Wave Panels Equal cross-lagged correlations. In CLC,
equal cross-lagged correlations indicate a con-
In this section I demonstrate that the differ- clusion of a spurious (or nonexistent) pattern
ence between the cross-lagged correlations is of causal influence between X and Y. This
not a sound basis for causal inference. Neither conclusion is unsound because equal cross-
determinations of causal predominance nor lagged correlations are consistent with many
spuriousness are defensible. Complications re- patterns of direct causal influence. Equal cross-
sulting from measurement error, specification
error, and multiple indicators are also discussed.
All results are presented in terms of population 8
The assumption of stationarity can be partially
parameters. represented in Equations 3, 4, and 5 by incorporating
the restriction of equal synchronous correlations, using
the expression given in Footnote 4. Substitution of this
Difference Between the Cross-Lagged Correlations equality has no effect on the interpretations and on con-
clusions that are based on Equations 3, 4, and 5.
The difference between the population cross- Consequently, this complication is not considered in
lagged correlations can be written in terms of the text.
252 DAVID ROGOSA
.70
Y^ = . 4 0
B* - . 3 0 Yj = .55
B* - .28
lagged correlations require only that and F are equal, the cross-lagged correlations
are equal. Panel a in Figure 4 is one of many
~ * ' -
possible illustrations that a zero difference be-
tween the cross-lagged correlations is consistent
with nonnegligible and equal direct causal in-
0 ~ 72 = ftr.K.J - T ) . fluences. Equal cross-lagged correlations are
In particular, equal cross-lagged correlations
are consistent with (a) large and equal causal 7
In the numerical examples displayed in Figures 4
effects, (b) large and unequal causal effects, and S, stationarity is not grossly violated. In each of
and (c) the absence of causal effects between these examples, the correlations were generated under
X and F. These cases are considered below, the assumption that <ruv = 0. A nonzero,value of a-uo
and numerical examples are displayed in could have been specified in each of these examples that
would make the synchronous correlations identical;
Figure 4.7 however, it was felt that the slight discrepancies among
As can be seen from Equation 4, if fi*z and y*z synchronous correlations did not justify the complica-
are large and equal and if the stabilities of X tion of an additional specification.
CROSS-LAGGED CORRELATION 253
also consistent with unequal causal effects, the Similarly, a causal predominance will be in-
difference between the stabilities offsetting the dicated by CLC when $\ = y* ^ 0 and the
difference between fi\ and y%. Panel b in Figure stabilities are unequal. Causal predominance is
4 presents an example in which &\ is twice as attributed to the variable with the lower stabil-
large as -yj anc^ in which the cross-lagged corre- ity because, as in the previously mentioned
lations are equal. Equal cross-lagged correla- expression, pxlYl—PYIX^ = PXIYI(PYIY^—PXIX^-
tions are consistent with the absence of causal Panel c in Figure 4 presents a numerical
effects—ft\ — y'z — 0—only when the stabilities example in which /3g = y%, and the cross-
of X and Y are equal. lagged correlations indicate a causal predomi-
All of the difficulties with the interpretation nance for X, the variable with the lower
of equal cross-lagged correlations persist when stability.
causal effects are defined in terms of unstan- In addition to indicating a causal predomi-
dardized structural parameters. Because the nance when there is none, CLC may indicate a
difference between the cross-lagged correlations causal predominance opposite to that indicated
increases as ox^xjay^v^ increases (all other by the structural parameters. Whenever px,y2
parameters constant), changes in variances — pytxt and /32 — 7 2 (or 0J — -yj) differ in sign,
over time make equal cross-lagged correlations the direction of causal predominance indicated
consistent with an even wider variety of con- by the cross-lagged correlations is antipodal to
figurations of unstandardized causal param- that determined by the structural parameters,
eters than were indicated previously for fn Panels d and e in Figure 4, the stronger
standardized parameters. causal effect, defined by the standardized
The evidence used in CLC to reach a deter- structural parameters, is from Y to X. How-
mination of spuriousness is equal cross-lagged ever, in each example the cross-lagged correla-
correlations (and equal synchronous correla- tions indicate that the predominant causal in-
tions to satisfy stability). Because this evidence fluence is from X to Y. In each example CLC
is consistent with many patterns of direct picks the wrong causal winner,
causal influence, the determination of spurious- When causal effects are expressed in terms
ness in CLC is unsound. of the unstandardized structural parameters,
Unequal cross-lagged correlations. Unequal changes in variances over time further compli-
cross-lagged correlations indicate a causal pre- cate the interpretation of unequal cross-lagged
dominance in CLC. However, a nonzero differ- correlations. As noted previously, the difference
ence between the cross-lagged correlations is no between the cross-lagged correlations increases
more dependable than a zero difference for as ax^xJffY^ffv^ increases. As a result, CLC
determining the pattern of causal influence. (other parameters constant) favors the attribu-
Unequal cross-lagged correlations are consis- tion of causal predominance to the variable
tent with (a) the absence of causal effects be- with increasing variance over time. One way of
tween the variables, (b) equal causal effects, highlighting the role of changing variances over
and (c) unequal causal effects. Moreover, the time is to rewrite Equation 3 as
predominant cause indicated by the unequal
cross-lagged correlations may be opposite to PXfY, — PYtX, = —
Yn
that denned by the structural parameters; for
example, the cross-lagged correlations indicate
that X is causally predominant, when the re-
verse is true.
For example a relatively large <r,v2 tends to
As before, the shortcomings of CLC are seen
from Equations 3, 4, and 5. First, in the absence make PX,YS — Pvixi positive, thus favoring the
attribution of causal predominance to X,
of causal effects (ft = 72 = 0), pXlvt — pvlxt
= PX^^PY^ — PJ^A-J). Unequal cross-lagged
correlations result whenever px^xt ^ pi^y,. Measurement Error
CLC will assign causal predominance to the In the preceding discussion X and Y were
variable with the lower stability, even though assumed to be measured without error. A num-
causal effects are not present, ber of techniques, both standard psychometric
DAVID ROGOSA
methods and special methods peculiar to CLC, arise in the interpretation of the zero-order
have been used in CLC analyses to deal with correlations in the determination of causal
measurement error. It is not vital to consider effects, specifically from the effects of specifica-
these correction techniques explicitly because tion error on the definitions of causal effects
any scheme for disattenuation can at best re- and on the model underlying CLC. When the
produce a situation with perfect measurement. model in Equation 1 is misspecified because of
But CLC is unsound even when measurement the omission of important causal variables, the
is perfect, and better performance cannot be structural parameters in Equation 1 are not
expected when measurement error is present. valid representations of the causal influences.
Equations 3, 4, and 5 are valid both for fallible The cross-lagged correlations are functions of
and for perfectly measured variables. these structural parameters, and thus their
The presence of measurement error is cited interpretation is also affected. For example,
in discussions of CLC as invalidating alterna- when Ay = Pi-i-,
tive regression-based analyses (see Kenny,
197S). If this claim were sound, it would justify
some despair but would not justify the use of
CLC. The argument in support of CLC is that The causal interpretation of the zero-order
partial regression slopes are more severely af- correlations is based on specific assumptions
fected by measurement error than are zero- about causal structures. The model (Figure 3)
order correlations. With only one fallible mea- underlying CLC includes assumptions about
sure of each variable at each time point, it is unmeasured variables that are unlikely to be
not possible in general to obtain acceptable satisfied when important causal variables are
estimates of the parameters of a regression omitted. For example, the requirements that
model for the panel data. However, additional U'2 and F2 be uncorrelated with each other and
measures of each variable can be used to con- that £/i and V\ be similarly uncorrelated are
sistently estimate the regression parameters. susceptible to violations resulting from speci-
The problem of measurement error can be fication error. Figure 15 in Duncan (1972) con-
viewed as a problem of inadequate research tains some forms of specification errors for
design; a research design that incorporates Figure 3.
multiple indicators of the underlying variables
is not as easily invalidated by fallible measure- Multiple Indicators
ment.
A CLC analysis of panel data with multiple
Specification Error indicators of each variable consists of the com-
Sometimes claims are made that CLC is not parison of all possible 2W2V cross-lagged cor-
affected by omitted causal variables or by relations. In a two-wave panel with px mea-
other forms of specification error, whereas re- sures of the latent variable £ ( and with pv
gression analyses are destroyed by specification measures of the latent variable r/, at each time,
error. A fervent argument of this type was pxpv differences between cross-lagged correla-
made by Humphreys and Stubbs (1977), who tions are computed. Multiple indicators are
asserted that considered to be an opportunity for replications
across different operationalizations of the same
the comparison of correlations among the same set of construct (Kenny, 1975, p. 894).
measures obtained at two or more intervals of time
(cross-lagged correlations) either avoids completely or
The pxpv differences between the cross-
minimizes the difficulties involved in the interpretation lagged correlations obtained from a 2W2V
of regression weights, (p. 262) design with multiple indicators can be written
as
The claim is specious. CLC is not immune to
specification error.
The effects of specification error on CLC are
opaque because a zero-order correlation does
not depend on additional variables, whether
they be measured or omitted. The problems i = \,...,px\ j = 1 , . - . , PY. (6)
CROSS-LAGGED CORRELATION 255
The latent variables £ and •>; are related by the of Equation 7, the differences between the
structural equations, cross-lagged correlations are
£•2 = /3o + / + u, - 72 + Pxjfyl
172 = To + + "• ~ PFsx3 = - 7* + PxtYt(yl
The observed indicators are related to the la-
tent variables by the measurement model,
Xu = &nt;t -f- fit,
Yjt = Xj-,7j, + Vjt
* = 1,..., #*; j = l , . - . , ^ ; 1=1,2. . (8)
Equation 6 shows the additional problems CLC analyses for 2W2V data were shown not
with CLC resulting from the use of multiple to be a sound basis for causal inference. The
indicators for replication. The relations in collection of 2W2V analyses produced by CLC
Equations 3, 4, and 5 can be thought of as re- with multiwave data is no more dependable as
lations for the latent variables £ and r\. Even a basis for causal inference than a single 2W2V
if measures of £ and -q were available, the dif- analysis. In fact, the way CLC uses multiwave
ference between their cross-lagged correlations data is likely to increase the difficulties; each
is not a sound basis for causal inference. two-wave snapshot does not yield dependable
Equation 6 shows that these shortcomings are results, and taken together the two-wave analy-
likely to be compounded by the CLC strategy ses will often be contradictory and misleading.
for using multiple indicators. The difference Equation 8 indicates that the differences be-
between cross-lagged correlations of the multi- tween the cross-lagged correlations may be in-
ple indicators depends on the specific relations consistent in causal attribution, even when
(the loadings 5 and X) of each indicator to the causal influences are pronounced and consistent
latent variable it represents. This dependence across waves. Figure 5 displays one of many
is likely to produce inconsistency and con- possible patterns of inconsistency in the cross-
fusion, not replication, lagged correlations. In Figure 5 the stronger
causal effect, as defined by the standardized
Results for Multiwave Panels structural parameters, is from Y to X consis-
tently across the three waves. The cross-lagged
The analysis of CLC for multiwave data is correlations do not indicate this causal pattern;
based on a 3W2Vpanel design. The representa- Waves 1 and 2 indicate a causal predominance
tion of causal effects for the 3W2V design is a for X; Waves 2 and 3 indicate a causal predomi-
simple extension of Equation 1 (intercepts nance for Y, and Waves 1 and 3 indicate
omitted) : spuriousness.
X* = Unfortunately, inconsistencies in the cross-
lagged correlations over waves generate clever
F2 = substantive interpretations. For example,
X, = Clarke-Stewart (1973) conducted a CLC analy-
F3 = j84Xi + 73F5 + ft,Xi + TsFi + »,. (7) sis on the relationship between maternal at-
tention and infant attachment using three
In this model, causal effects between X and F waves of data. Waves 1 and 2 indicated a causal
may span two waves; both between-variables predominance for maternal attention, and
and within-variable effects are so represented Waves 2 and 3 indicated a causal predomi-
in Equation 7. nance for infant attachment (Wave 1-Wave 3
Three separate comparisons of cross-lagged comparison not reported). Clarke-Stewart sug-
correlations are made in CLC for the 3W2V gested that "as mother and child search for
data : comparisons between Waves 1 and 2, be- harmonious, balanced interaction over the
tween Waves 2 and 3, and between Waves 1 course of development, first one then the other
and 3. In terms of the standardized parameters assumes the 'causal role' " (p. 91).
256 DAVID ROGOSA
vr- 55
3* = .28 |3* = .31 8* = .25 0* = .10 8* = .05 g* = .06
.500 .543
.550
.851 .829
.504
.550
.741
p - p = .130 P P Pv Y - PY v = - 0 0 3
X Y Y X
X Y - Y X X Y Y X
1 2 1X2 2Y3 X
2 3 1 3 1 3
Statistical Inference in CLC power test" (p. 887) and that even with moder-
ate sample sizes (defined as 75 to 300), statis-
In the preceding sections relations among tically significant differences are difficult to
population parameters were used to demon- obtain. With large enough samples, trivial de-
strate the failure of CLC. In this section addi- viations from the null hypothesis lead to re-
tional problems with the application and inter- jection. For example, Crano, Kenny, and
pretation of statistical inference procedures in Campbell (1972) found significant differences
CLC are noted. between cross-lagged correlations of .65 and .67
Rejection of the null hypothesis of equal because the sample size was 5,495. Other
cross-lagged correlations (H 0 : px^ = PivO authors ignore sampling variation and with
often is interpreted with little regard for the small sample sizes interpret differences between
power of the statistical test. Users of CLC are the sample cross-lagged correlations as if these
advised to use large samples; Kenny (1975) sample estimates were population values. The
advises that "cross-lagged analysis is a low- use of interval estimates for px^j ~ pr^ would
CROSS-LAGGED CORRELATION 257