A closer look at random and fixed effects panel regression in

structural equation modeling using lavaan

Henrik Kenneth Andersen

Chemnitz University of Technology

Institute of Sociology
[email protected]

30 July, 2021

Abstract
This article provides an in-depth look at random and fixed effects panel regression in the structural
equation modeling (SEM) framework, as well as their application in the lavaan package for R. It is meant
as a applied guide for researchers, covering the underlying model specification, syntax, and summary
output.

This is an Accepted Manuscript of an article published by Taylor & Francis in Structural Equation
Modeling: A Multidisciplinary Journal, available online: https://doi.org/10.1080/10705511.2021.1963255.

1
1 Introduction
Several years ago, Curran and Bauer (2011) reflected positively on the growing use of panel studies in
empirical social research. Some of the strengths of panel data are well-known, e.g., the ability to establish
temporal precedence, increased statistical power and the reduction of potential alternative models. However,
the greatest strength of panel data is that they allow for a more rigorous testing of substantive theories.
Panel data, i.e., repeated measures of the same observed units (people, schools, firms, countries, etc.), allow
researchers to decompose the error term into a part that stays constant within units and the part that
changes over time. The part that does not change over time can be seen as the combined effect of all time-
invariant influences (e.g., sex, nationality, personality traits, intelligence) on the dependent variable (Bollen
and Brand 2010). Random effects (RE) and fixed effects (FE) regression involve accounting for these often
unobserved time-invariant influences via a number of methods.
In the case of RE models, it is assumed that the stable characteristics are unrelated to the other model
covariates. So, the stable characteristics, also known as ‘individual effects’ (Hsiao 2014, the term I prefer),
‘unobserved effects’ (Wooldridge 2002) or ‘unit effects’ (Skrondal and Rabe-Hesketh 2004; Zyphur et al.
2020a, 2020b), will not affect point estimates but their existence violates important regression assumptions.
FE models are useful when the individual effects are assumed to be related to one or more of the model
covariates. In that case, these unobserved stable characteristics will act as confounders and lead to biased
and inconsistent point estimates. In controlling for these individual effects, FE regression thus accounts for
a likely and common source of bias.
Structural equation modeling (SEM) is a popular regression framework. One of its main strengths is its
flexibility (Hoyle 2015b, 142). Not only can complex causal structures with multiple dependent variables be
tested simultaneously, but in longitudinal (and, more generally, hierarchical) studies, both time-varying and
invariant predictors can be included, and effects can easily be allowed to vary over time. Thus researchers can
allow for and study effects that increase or fade over time, or that appear only in specific periods. Beyond
that, with the use of latent variables, SEM provides a way to deal with measurement error and get closer to
the theoretical constructs of interest.
The article focuses on the lavaan (Rosseel 2012) package for R (R Core Team 2017). While Mplus (L.
Muthén and Muthén 1998--2017) is arguably the most robust SEM software currently available (in terms of
features like alignment and latent variable interactions, for example), the lavaan package has many benefits.
First, like R it is open source and completely free. For researchers new to SEM, there is no financial barrier to
entry. Second, the implementation of lavaan in the larger R environment is an enormous advantage. Instead
of poring over reams of plain text, copying out coefficients by hand, every part of the lavaan output is
available as an object. This means that all aspects of the model, from fit indices, to coefficients and standard
errors, to the model matrices, can be accessed and easily integrated into tables and plots. Furthermore, R
can be used for a great deal of applications. It can be used to manage and manipulate as well as simulate
data, perform symbolic algebra, run more traditional analyses (e.g., multiple regression, logistic regression,
principal component analysis), etc. Once one is comfortable using R, there is (at the time of writing) little
need to switch between different software for data preparation and analysis.
There are a number of articles describing the basic concept of panel regression, including RE and FE
regression in SEM (e.g., P. D. Allison 2009; P. Allison 2011; Bollen and Brand 2010; Teachman et al. 2001).
This article is intended as a practical guide for researchers looking for help specifying panel regression models
in SEM. It assumes some basic knowledge of SEM (e.g., identification, the basic logic of estimation, model
fit, etc.).1
The article proceeds by giving a short review of panel regression models (Sections 2 to 2.2 and 2.3) while
touching on the implementation in SEM (Section 2.4). Then the lavaan package is discussed in Section 3
and an annotated syntax for specifying RE and FE models in SEM is shown in detail in Section 3.2. An
example using simulated data is given in Section 4 to show and discuss the model output and results. Section
5 provides a brief summary and points to some extended panel SEMs that build on the basic RE and FE
1 For an introduction to SEM, Bollen (1989) is a classic for good reason, and Ferron and Hess (2007) lay out exactly how

SEM works in one easy to follow article.

2
framework (including latent growth curves and dynamic models). Finally, supplementary materials touch
on further topics, including some drawbacks to panel SEM and their potential remedies, comparability with
non-SEM methods, and extensions to the basic models (relaxing assumptions, dealing with measurement
error).
For experienced SEM-users with little knowledge of static panel regression, the entire article may be of some
use. For SEM-users already familiar with panel regression in SEM, the supplementary materials may be of
the most interest. For those new to SEM but familiar with panel regression, Sections 3 and 4 (as well as the
supplementary materials) are likely most relevant.

2 Random and Fixed Effects Panel models

It is typically the case that the values for a given unit on a variable at one point in time will tend to
tell us something about that unit’s values on the same variable at another point in time. There are two
main explanations for this ‘empirical regularity’ (Heckman 1981; Hsiao 2014, 261; Brüderl and Ludwig
2015; Bianconcini and Bollen 2018). First, it could be that an experience at one point in time has an
effect on future experiences. In other words, experiencing an event could change the probability of the
same or a similar event taking place in the future. For example, if employment increases wages, then the
incentive to continue working should increase over time, thereby increasing the likelihood that someone who
was employed at one point in time will continue to be employed in the future (Heckman 1981). When
past experiences impact future events, the empirical regularity is referred to as ‘state dependence.’ So-
called ‘dynamic’ panel models that include the lagged dependent variable in the equation for the current
dependent variable, like autoregressive and autoregressive cross-lagged models, are examples of panel models
that account for state dependence (Zyphur et al. 2020a, 2020b). The second explanation is that correlations
over time are due to stable characteristics, like sex, place of birth, motivation, ability or personality.2 In other
words, stable unit-specific characteristics might predispose individuals to experience events with a certain
likelihood over time. For example, stable characteristics like sex or motivation could be part of the reason why
some individuals tend to be continuously employed, while others experience spells of unemployment, or are
habitually unemployed. This second source of empirical regularity is referred to as individual heterogeneity
or, more generally, unobserved heterogeneity (Wooldridge 2012). So-called ‘static’ panel models (that do
not include the lagged dependent variable in the equation for the current one) focus on this second source
of empirical regularity (Zyphur et al. 2020a, 2020b). The random and fixed effects models that will be
discussed here are examples of models that attempt to account for unobserved heterogeneity as a source of
empirical regularity.

2.1 A review of static panel models and their implementation in SEM

Let us begin with a simplified cross-sectional model with a single covariate

yi = β0 + β1 xi + β2 zi + εi , i = 1, . . . , N

where yi is the dependent variable for unit i, β0 is the intercept, xi and zi are scalar covariates that are
considered random variables in most social science applications with observational data, β1 and β2 are the
associated regression coefficients and εi is the normal idiosyncratic error. We can think of the intercept, β0 ,
as ‘absorbing’ the expectation of the error, E(εi ), so that it is safe to assume that E(εi ) = 0.3
2 This article will tend to refer to individuals as the unit of analysis, but others, such as families, schools and countries, are

of course possible.
3 Take a simplified equation without an intercept: y = β x + ε . We can add and subtract the expectation of ε without
i 1 i i i
changing the equation: yi = β1 xi + εi + E(εi ) − E(εi ). Now, we can define β0 = E(εi ) and ε∗i = εi − E(εi ) for yi = β0 + β1 xi + ε∗i
where E(ε∗i ) = 0, see Wooldridge (2009), p. 25, 61; or https://stats.stackexchange.com/questions/231549/expected-value-of-
error-term-equals-zero-formal-proof. Throughout, this is taken for granted and the notation with the star is dropped.

3
We typically assume the following (Wooldridge 2002; B. O. Muthén, Muthén, and Asparouhov 2016; Bollen
and Brand 2010):

1. Linearity: the model can be written as yi = β0 + β1 xi + β2 zi + εi

2. Variation in covariates, no perfect multicollinearity

3. Zero conditional mean: E(εi |xi , zi ) = E(εi ) = 0, which implies Cov(εi , xi ) = Cov(εi , zi ) = 0 and
E(yi |xi , zi ) = β0 + β1 xi + β2 zi
4. Var(εi |xi , zi ) = Var(εi ) = σε2 and Cov(εi , εj |xi , zi , xj , zj ) = Cov(εi , εj ) = 0, i ̸= j which implies
homoskedasticity (constant variance) and uncorrelated errors

5. Normally distributed errors: εi |xi , zi ∼ N (0, σε2 ), for significance testing.

In a cross-sectional setting, we could estimate the parameters of interest, β0 , β1 , β2 by ordinary least squares
(OLS) or maximum likelihood (ML), where it is widely known that if the assumption of normally distributed
errors holds, the OLS estimator is the ML estimator.4
Now, assume we are dealing with repeated measures of the same units over time so we can write

yit = β0 + β1 xit + β2 zi + εit , i = 1, . . . , N, t = 1, . . . , T, (1)

where the subscript it stands for unit i at time t (P. D. Allison 2009). Now, xit is referred to as a ‘time-
varying’ covariate because it changes across i and t, while zi is a ‘time-invariant’ covariate because it changes
over i but not t. Time-varying covariates encompass characteristics that can change over time, like mood,
stress level, income, number of children, etc. Time-invariant covariates do not change over time, or at least
not usually over the period of observation. These can be things like sex, nationality, upbringing, etc. β0
is the overall intercept, absorbing the expected value of εit . Sometimes we will want to allow the intercept
to vary over time. These are sometimes referred to as ‘occasion effects’ (Zyphur et al. 2020b) and can be
modeled by including time dummies. We will ignore these for now, and allow the overall intercept to absorb
the expectation of the error over i and t.
Now, there is nothing practical stopping us from stacking all N T observations on top of one another and
estimating the model using OLS or ML. This is called the ‘pooled’ OLS (POLS) estimator with an analogous
ML estimator. The problem with this approach is that it ignores the panel nature of the data. As such, the
assumption of uncorrelated errors is likely violated within units. In other words, there are likely unit-specific
characteristics that do not change over time affecting the dependent variable, many of which will remain
unobserved. We can call the sum of all these effects on the dependent variable individual effects or unobserved
heterogeneity and write them as αi with the composite error: εit = αi + νit . Then Equation (1) becomes

yit = β0 + β1 xit + β2 zi + αi + νit , i = 1, . . . , N, t = 1, . . . , T. (2)

The individual effects αi are treated as a random variable, even if the use of the Greek letter makes it look like
another parameter to be estimated. This is the common notation used in the literature on panel regression
in SEM, see for example P. Allison (2011); Bollen and Brand (2010). When the number of cross-sectional
units is large, Wooldridge (2002) (p. 247 and p. 252) suggests always treating the unobserved heterogeneity
as a random variable rather than a fixed unknown parameter since little is lost in doing so. Notice this
equation still contains an overall intercept, β0 . This will absorb the expected value of αi as well.
Now, because
4 In this section, we will focus on the typical least squares-based treatment of RE and FE models because the link between

the ML-based multilevel or hierarchical models and multilevel and panel SEM is already arguably fairly strong, see for example
Hox (2010), in which SEM is made reference to constantly, and the final two chapters are dedicated to SEM-based approaches
to multilevel modeling.

4
 E(εit εjs ) = E[(αi + νit )(αj + νjs )]

the assumption about uncorrelated errors becomes


2
σε , i = j, t = s

E(εit εjs ) = σα2 , i = j, t ̸= s

0, otherwise


where σε2 = σα2 + σν2 , which is the sum of the variances of the individual effects and idiosyncratic errors. This
follows if we assume the individual effects are unrelated to the idiosyncratic errors at all points in time while
retaining the independent errors assumption from above as it pertains to the idiosyncratic errors (Hsiao
2014, 40). This means that the variance of the errors can no longer be described “up to a multiplicative
constant” (Wooldridge 2002, 153), i.e., σε2 I (where I is the identity matrix with ones down the diagonal and
zeros everywhere else), and the pooled model is no longer appropriate.
Let us write Equation (2) in matrix notation

yi = Xi β + εi ,
= Xi β + ιT αi + νi , i = 1, . . . , N (3)

where yi = (yi1 , yi2 , . . . , yiT )⊺ , Xi = (ιT , xi , ιT zi ), β = (β0 , β1 , β2 )⊺ , xi = (xi1 , xi2 , . . . , xiT )⊺ , ιT is a vector
of T ones (sometimes written as 1), and write the covariance matrix of the errors as

σε2 σα2 . . . σα2

 
σα2 σε2 . . . σα2 
⊺
(ε ε ) = Ω = ..  . (4)
 
E i i i  .. .. ..
 . . . . 
σα2 σα2 . . . σε2

The full covariance matrix over all i takes a block-diagonal form with Ωi down the diagonal and 0 everywhere
else (Wooldridge 2002, 259; Schmidheiny 2019). Notice the homoskedasticity assumption here now applies
across time, as well.

2.2 Random effects

In a typical random effects (RE) model, a transformation to Equation (3) is carried out such that the
covariance matrix of the errors becomes the identity matrix, i.e.,

−1/2 −1/2 −1/2

Ωi yi = (Ωi Xi )β + Ωi εi ,
yi∗ = Xi∗ β + ε∗i (5)

(Wooldridge 2002, 155). This is akin to dividing a random variable by its standard deviation in order to
standardize it with a variance of one. In this case, we have

∗ ∗⊺ −1/2 −1/2
E(εi εi ) = E[(Ωi εi )(Ωi ε i )⊺ ]
−1/2 −1/2⊺
= E(Ωi εi ε⊺i Ωi )
= IT

5
(Wooldridge 2002, 155). The variances of σε2 and σα2 are not known, but they can be estimated by the sample
variances, see Wooldridge (2002) p. 260 f.; Schmidheiny (2019). With the transformation, Equation (5) can
be estimated by OLS which is referred to as the (feasible) generalized least squares (GLS) estimator Brüderl
and Ludwig (2015).
Notice the model puts the unobserved characteristics in αi in the composite error term. Consistency and
unbiasedness thus hinge on the assumption that the model covariates are unrelated to the composite error,
i.e., E(εit |Xi ) = E(εit ) = 0 or E(νit |Xi ) = E(νit ) = 0 and E(αi |Xi ) = E(αi ) = 0:5

XN XN
β̂GLS = ( Xi⊺ Ω−1
i X i )−1
( Xi⊺ Ω−1
i yi )
i=1 i=1
N
X N
X
=β+( Xi⊺ Ω−1
i Xi )
−1
( Xi⊺ Ω−1
i εi )
i=1 i=1
N
X N
X
=β+( Xi⊺ Ω−1
i Xi )
−1
( Xi⊺ Ω−1
i (ιT αi + νi ))
i=1 i=1
N
X XN N
X
=β+( Xi⊺ Ω−1
i X i )−1
( X ⊺ −1
i Ω i ι α
T i + Xi⊺ Ω−1
i νi ). (6)
i=1 i=1 i=1

Equation (6) shows that consistency and unbiasedness depend on the relationship between the model co-
variates and the individual effects and idiosyncratic errors (take the probability limit and expectation to see
that consistency requires the covariances to be zero, while unbiasedness requires conditional independence).
Notice, as well, that the RE model requires the assumption that the composite errors at each point in time
are unrelated to the model covariates at all points in time.6 This is the strict exogeneity assumption, and it
comes from the transformation in Equation (5), where

⊺ −1
E(Xi Ωi εi ) = 0

involves multiplying Xi (which holds xit at all points in time) by the errors at each point in time. For this
to be zero, it is therefore not enough to assume contemporaneous exogeneity, which is sufficient for pooled
OLS (Wooldridge 2002, 146).

2.3 Fixed effects

The critical assumption in the RE model is E(εit |Xi ) = 0, which implies E(νit |Xi ) = 0 and E(αi |Xi ) = 0.
This means that any of the stable characteristics contained in αi are assumed mean independent of the model
covariate(s), though typically, it is enough to look at the covariance between αi and the model covariate(s),
which accounts for any linear relationship, at least (see, for example, Bollen and Brand 2010, 6; Wooldridge
2002, 252). As an example, if we were interested in the effect of social isolation (xt ) on well-being (yt ), this
would mean that any and all (potentially unobserved) personality characteristics (α), like extroversion or
neuroticism, affecting someone’s usual well-being level cannot be correlated with social isolation (Seifert and
Andersen 2020). This is obviously implausible because extroverted people tend to engage more socially and
neurotic people less so. In this case, the assumption of unrelated effects is likely violated and the RE model
will return biased and inconsistent coefficient estimates for the effect of social isolation on well-being.
5 In applied settings, it is normally enough to establish that the correlation between the individual effects and the model

covariates are zero (Bollen and Brand 2010; Wooldridge 2002, 254). Also note that assuming the unconditional expectations
are zero is unproblematic as long as an intercept is included (Wooldridge 2002, 257).
6 Contemporaneous exogeneity would be (ε |x ) = 0, where, in this case, x = (1, x , z ) (Wooldridge 2002).
E it it it it i

6
The conventional FE model relaxes this assumption by transforming Equation (2) such that the individual
effects are eliminated. This usually entails demeaning the equation by subtracting the respective person- or
unit-means from each of the variables

yit − ȳi = (xit − x̄i )β + αi − ᾱi + νit − ν̄it

ÿit = ẍit β + ν̈it ,
ÿi = Ẍi β + ν̈i

PT
where the variables with the dots above them represent the demeaned versions and ȳi = T −1 t=1 yit and
the person-means of the other variables are calculated analogously. Notice the average of something that does
not change is that thing itself, so αi , the individual effects, as well as any time-invariant variables in xit are
eliminated by the transformation (Brüderl and Ludwig 2015; Wooldridge 2002; Angrist and Pischke 2009).
Now, the strict exogeneity assumption applies only to the idiosyncratic errors ν̈it . That is, the idiosyncratic
error at each point in time is assumed to be uncorrelated with the model covariate(s) at all points in time.

2.4 Random and fixed effects in SEM

In this section, I explain the implementation of RE and FE models in SEM in a non-technical way. For a
more detailed explanation, see the supplementary materials.
We can use path models like the one in Figure 1 to help us convey the model and its assumptions quickly
and intuitively. Figure 1 shows a four-wave RE-SEM like the one discussed above. The observed variables
are conventionally shown as squares or rectangles. Latent variables, like αi , are shown as circles or ellipses.
Effects and factor loadings are shown as directional arrows, coming out of the independent variable and
pointing towards the dependent variable. Covariances or correlations are shown as bidirectional arrows
between variables. The errors, here νit , are not enclosed in anything. They point to the time-varying
dependent variable, yit , with an implicit coefficient of one. The circular arrows around the errors indicate
an endogenous error variance to be estimated. The fact that each is labelled σν2 signifies that the variance
has been constrained to be equal over time. Sometimes, intercepts are shown as a triangle labelled 1, with
arrows pointing to each of the variables for which an intercept or mean should be estimated. For the sake
of legibility, this is not shown here, although we allow an intercept to be estimated for each yit , as well as
unconditional means for each xit and zi . The mean of αi is again set to zero for identification purposes (we
can only estimate as many intercepts and means as there are observed variables). The unconditional mean
of αi will be absorbed into the time-varying intercepts κyt .

[Figure 1 about here.]

For the RE and FE models, we switch from the long format setup (where each model variable is N T × 1) to
wide format, i.e., T individual N × 1 columns for each variable. Then, we write T individual equations:

yi1 = κy1 + βxi1 + γzi + αi + νi1 , i = 1, . . . , N

yi2 = κy2 + βxi2 + γzi + αi + νi2 , i = 1, . . . , N
..
.
yiT = κyT + βxiT + γzi + αi + νiT , i = 1, . . . , N. (7)

where we have a single time-varying covariate, xit , and a time-invariant covariate, zi , along with the unob-
served individual effect, αi , that captures the sum of all other time-invariant unobserved variables (Bollen
and Brand 2010; Wooldridge 2002). Note that the change of notation compared to the base text (from β1
to β, β2 to γ) is inconsequential.

7
In this setup, with a separate equation per timepoint, there is nothing stopping us from allowing each of
the coefficients (even the implicit coefficient in front of αi ) to be estimated freely. In fact, while it is the
standard practice in the long-format least squares models discussed above, when we write the model like this,
it may even seem arbitrary to constrain the coefficients to be time-invariant across equations. I retain these
constraints for now, in order to establish a baseline equivalence with the least squares-based models. In the
supplementary section, models will be discussed that do allow for time-varying coefficients. Something that
does separate the SEMs from the conventional models are the intercepts. With T individual equations, it is
straightforward to have a freely estimated intercept, κyt , per timepoint. Note as well that we could easily
include other time-varying and invariant coefficients and turn xit and zi into vectors.
Now, the assumption of constant effects over time is conveyed by the lack of a subscript t on the coefficients, β
and γ. The latent individual effects are also assumed constant and constrained to 1 at each timepoint. Again,
these can easily be relaxed but are retained for now for the sake of establishing baseline equivalence with
the least squares-based models. Strict exogeneity in the RE model means the idiosyncratic errors, as well
as the individual effects are uncorrelated with the model covariates at each point in time. This is conveyed
by the lack of any two-headed arrows between them. Notice the idiosyncratic errors, νit , are specific to a
given timepoint, while the individual effects, αi , is common to all timepoints (P. D. Allison 2009). This is
how we translate the structure seen in Ωi in Equation (4). Once we have converted the data to wide-format,
the conditional covariance between any two columns of the dependent variable, Cov(yit , yis |xit , xis , zi ), is
just the conditional variance of the individual effects, Var(αi |xit , xis , zi ). In the RE case, it is assumed that
αi is independent of all the other model covariates, so Var(αi |xit , xis , zi ) = Var(αi ). In other words, by
regressing yit onto the latent variable αi at each point in time, we capture the unobserved heterogeneity; the
common variance in yit due to unobserved time-invariant characteristics. Finally, homoskedasticity of the
idiosyncratic errors over time is conveyed by labelling each of the idiosyncratic error variances (the circular
arrows around νit ) the same over time. It can be difficult to create path models that effectively convey all
the important information of a model, but they at least have the potential to express such information in a
fairly intuitive and non-technical way.
For the FE model, Equation (7) from the RE setup pertains here as well. We relax the assumption of
unrelated effects by allowing the individual effects, αi , to correlate with the time-varying model covariates
at each point in time, rather than fixing them to zero. The assumption of Cov(αi , zi ) = 0 must be retained
for identification. Again, we can represent the FE-SEM in a non-technical way by altering the path diagram
to include the covariances between each xit and αi , see the new two-headed arrows in Figure 2.

[Figure 2 about here.]

3 Random and fixed effects in lavaan

The basic panel models discussed in this tutorial consist of three to four main parts. First, we want to
create a latent variable representing the individual effects. Second, we regress the time-varying dependent
variable on the time-varying and time-invariant covariates. Third, we specify the correlations depending
on our assumptions. If we believe the individual effects are unrelated to the time-varying covariates (we
must assume they are unrelated to the time-invariant ones, otherwise the model is not identifiable), then we
apply the RE assumptions and constrain the covariances between the individual effects and the time-varying
covariates to zero. If we believe the individual effects are indeed related to the model covariates (or more
realistic assumption in many circumstances), then we must specify these correlations between the individual
effects and the time-varying covariates. Finally, in an optional step, we can constrain the residual variances
to be equal over time, if we want or need to, potentially in order to save degrees of freedom. The following
section will explain these steps in lavaan in detail.

8
3.1 The lavaan package in R

The package lavaan needs to be installed once with install.packages("lavaan"). This can be entered
directly into the console or at the top of the script. To be able to use the package, we need to load it for
every new R session:

library(lavaan)

For users unfamiliar with R, SEM analyses can be carried out with almost no knowledge of the language.
Typically, someone unfamiliar with R would prepare their data using some other statistical software, and
then save the intended dataset as a .csv, .xlsx, .dta, .sav, etc. file. The user must then import the data,
preferably as a dataframe, and the rest occurs using the lavaan syntax.7
To use lavaan, we create an R object using the assignment operator <-, see the model syntax example below.
Here, the object has been called fe_sem because in the following example we will be assuming an FE model
is appropriate. The object can be named anything that complies with naming conventions in R (e.g., the
object name must start with a letter or dot, underscores and dots can be used to separate words, etc.). The
model syntax is enclosed in quotes, either single '' or double "". This means that the model syntax is
essentially a string that the lavaan package interprets in a second step. Once the model has been specified,
we use the sem() function to ‘fit’ the model. Notice a second object is made out of the fitted lavaan object.
Here the fitted lavaan object has been named fe_sem.fit.
To reiterate, we specify the SEM by writing the model syntax as a string and saving it as an object. Then, in
a second step, we run the sem() function on that object. The sem() function requires at least two arguments:
model, i.e., the model object (here: fe_sem), and data, i.e., the dataframe or covariance matrix (along with
the mean vector, if desired). That is, at a bare minimum, we must tell lavaan how the model is specified
and where the data is. There are a number of other optional arguments that can be included. If they are
not, the defaults of the sem() wrapper are used.8 For this example, even though maximum likelihood is
the default estimator, estimator = "ML" has been included as an optional argument to emphasize the fact.
The supplementary materials provide some guidance on dealing with nonnormal endogenous variables and
missing values, both of which can be addressed with optional arguments in the sem() function call. Finally,
we add the argument meanstructure = TRUE. If we leave this out, the model is essentially fit to centered
data so that each variable has a mean of zero. When we turn the mean structure on, the default behaviour
is to estimate an intercept or mean (depending on whether the variable is exogenous or endogenous) per
equation.

3.2 Model syntax

Again, specifying the most basic random or fixed effects model, like the one shown in Bollen and Brand
(2010) and described in Equation (2) involves three to four components. First, we define the latent individual
effects variable using the =~ ‘measured by’ or ‘manifested by’ (Rosseel 2012) operator while at the same time
constraining the factor loadings at each timepoint to one with 1* (see line 3 in the model code below). I will
call the latent variable alpha. Constraining all of the factor loadings to one reflects our implicit assumption
that the combined effect of the unit-specific unobserved factors is constant over time (Zyphur et al. 2020a,
660). This is the default behaviour of traditional least squares-based approaches to RE and FE that use the
stacked long-format data.
7 There are many online tutorials for importing data in various formats, see for example one of the many posts on stackoverflow

(https://stackoverflow.com/search?q=r+import+data).
8 The main defaults of the sem() wrapper are: intercepts of the latent variables set to zero, the first factor loading in factor

models is set to one, the residual variances and variances of exogenous latent variables are freely estimated, exogenous latent
variables are set to covary. Further details can be found at https://rdrr.io/cran/lavaan/man/sem.html and https://cran.r-
project.org/web/packages/lavaan/lavaan.pdf, or by entering lavOptions() into the console to get a full list of defaults. An
explanation of the optional arguments can be found by entering ?lavOptions in the console. There are other ‘wrappers’ with
slightly different default options, like cfa() for example, see the lavaan tutorial website at https://lavaan.ugent.be/tutorial/
cfa.html.

9
 Second, we regress the dependent variable on the independent variable using the ~ regression operator (see
lines 5–8). With stacked, long-format data, only one regression coefficient per covariate is estimated over all
observed timepoints. To have our model mimicking this behaviour, we need to constrain the the estimated
coefficients to equal over time. We do so by adding the same label to the regression coefficient at every time
point. We will use the labels beta and gamma (though we could use any letter or string of characters for
labels) and have them act as equality constraints for the regression coefficient of interest, β and γ.
The key to a FE model, as opposed to an RE model are our assumptions about the relatedness of our
time-varying covariates and the individual effects. For an FE model, we want to partial out any potential
covariance between the independent variable and the individual effects. This accounts for any linear relation-
ship between xit and the unit-specific characteristics influencing the dependent variable. Further, allowing
unrestricted covariances between the independent variable itself over time will not affect how the coefficients
are estimated, but will improve model fit (see lines 10–14).9 To mimic the behaviour of a conventional FE
model, we allow the independent variable to be correlated with the individual effects and itself over time.
Covariances (including covariances between a variable and itself, i.e., variances) are specified using the ~~
operator. For the RE model, we would constrain the covariances between alpha and the time-varying and
invariant covariates with alpha ~~ 0*x1 + 0*x2 + 0*x3 + 0*x4 + 0*z.
The last component of our code involves the variances of the residuals (lines 16–19). This component is
optional, but we can constrain the residual variances to be equal over time to again mimic the behaviour of
a conventional RE/FE model using least squares-based methods on stacked data. Here, again, we use labels
to make equality constraints. Because yit is endogenous, the ~~ operator specifies the variances of residuals,
i.e., νit .

1 fe_sem <- '

2 # Define individual effects variable
3 alpha =~ 1*y1 + 1*y2 + 1*y3 + 1*y4
4 # Regressions, constrain coefficient to be equal over time
5 y1 ~ beta*x1 + gamma*z
6 y2 ~ beta*x2 + gamma*z
7 y3 ~ beta*x3 + gamma*z
8 y4 ~ beta*x4 + gamma*z
9 # Correlations, individual effects covary with the covariate
10 alpha ~~ x1 + x2 + x3 + x4 + 0*z
11 x1 ~~ x2 + x3 + x4 + z
12 x2 ~~ x3 + x4 + z
13 x3 ~~ x4 + z
14 x4 ~~ z
15 # Constrain residual variances to be equal over time
16 y1 ~~ nu*y1
17 y2 ~~ nu*y2
18 y3 ~~ nu*y3
19 y4 ~~ nu*y4
20 '
21 fe_sem.fit <- sem(model = fe_sem,
22 meanstructure = TRUE,
23 data = dfw,
24 estimator = "ML")
9 If we do not specify the correlations between the covariate over time, we are telling the model we believe them to be zero.

This is unrealistic in most cases and discrepancy between the model-implied covariances of zero and the likely non-zero observed
covariances will be a source of misfit. I.e., the χ2 statistic will be larger than it needs to be (there is not usually a good reason
to fix covariate correlations to zero) potentially leading to otherwise well-fitting models being rejected by the test statistic.

10
4 A simulated example
To demonstrate the application of panel regression in SEM, a dataset can be simulated that embodies the FE
assumptions. The code for the simulation can be found in the supplementary materials. The simulated data
was constructed such that the time-varying covariate of yit is correlated with two separate time-invariant
variables. One of these time-invariant variables will be considered observed; the other will be treated as
unobserved. This means that approaches that fail to account for this confounding influence, such as POLS
or RE, will be biased.
The equations for the data generating process (DGP) can be described as:

xit = βxt ,z zi + βxt ,α αi + δit

yit = βyt ,xt xit + βyt ,z zi + βyt ,α αi + νit

where αi ∼ N (2, 12 ), zi ∼ N (1.5, 32 ), δit ∼ N (3, 22 ), νit ∼ N (3, 12 ) and the means and variances were chosen
arbitrarily. Correlations between xit and both zi and αi are induced by making xit dependent on each of
them. Note that the mean of νit is not zero, but that it will be absorbed into the intercept for yit . The
coefficients were also chosen arbitrarily and set to βxt ,z = 0.5, βxt ,α = 0.85, βyt ,xt = 0.30, βyt ,z = 0.45 and
βyt ,α = 0.75. For the following example, a sample size of 1,000, observed over four waves, was chosen.
We can get a summary of the model with summary(). Optional arguments for summary() are, for example,
standardized = TRUE for standardized coefficients, fit.measures = TRUE for further (e.g., comparative)
fit indices. The first portion of the summary output gives an overview of some basic information and fit
statistics. The maximum likelihood estimator is the default, so it did not have to be explicitly selected in
the fitting function call. Other estimators are available, including generalized and unweighted least squares
(GLS and ULS, respectively), robust standard errors maximum likelihood (MLM) and several others (see the
lavaan online tutorial at https://lavaan.ugent.be/tutorial/est.html as well as the supplementary materials).
This part of the summary output also tells us that the analysis is based on 1,000 observations (missings would
be shown here as well if there were any), and that the χ2 statistic is 18.21 based on 22 degrees of freedom (45
observed (co)variances minus 1 error variance, 2 coefficients, 1 latent variable variance, 5 exogenous variable
variances and 14 covariances for 45 − 23 = 22 df). The p-value on the χ2 statistic is not significant with p =
0.693 which suggests we should retain the model (given how the data was generated, it would be surprising
if this were not the case).

summary(fe_sem.fit)

## lavaan 0.6-8 ended normally after 92 iterations

##
## Estimator ML
## Optimization method NLMINB
## Number of model parameters 41
## Number of equality constraints 9
##
## Number of observations 1000
##
## Model Test User Model:
##
## Test statistic 18.210
## Degrees of freedom 22
## P-value (Chi-square) 0.693
##
## Parameter Estimates:
##

11
## Standard errors Standard
## Information Expected
## Information saturated (h1) model Structured
...

Next, the summary output shows the measurement models for the latent variables, if any. In this case the
latent variable, alpha, is measured by each of the four observed dependent variables with factor loadings
fixed to 1.0.

...
## Latent Variables:
## Estimate Std.Err z-value P(>|z|)
## alpha =~
## y1 1.000
## y2 1.000
## y3 1.000
## y4 1.000
...

The regressions are shown next. Here, because we have constrained the regression coefficients to be equal
over time (the equality constraint label (beta) and (gamm) is listed to the left of the estimates), the estimates
of β = 0.298 (0.01) and γ = 0.456 (0.01) are both repeated four times. The corresponding z- and p-values
show that the coefficients are, unsurprisingly based on the simulated DGP, significant.

...
## Regressions:
## Estimate Std.Err z-value P(>|z|)
## y1 ~
## x1 (beta) 0.298 0.010 31.112 0.000
## z (gamm) 0.456 0.010 43.476 0.000
## y2 ~
## x2 (beta) 0.298 0.010 31.112 0.000
## z (gamm) 0.456 0.010 43.476 0.000
## y3 ~
## x3 (beta) 0.298 0.010 31.112 0.000
## z (gamm) 0.456 0.010 43.476 0.000
## y4 ~
## x4 (beta) 0.298 0.010 31.112 0.000
## z (gamm) 0.456 0.010 43.476 0.000
...

Next, the covariance estimates are listed. First, the covariances between the latent individual effects and the
independent variables are shown, followed by the covariances between the independent variables themselves.
One should always take care to double-check that there are no unintended covariances listed here. Like
Mplus, the lavaan package estimates some covariances per default, without the user explicitly having to
add them to the model syntax. For example, covariances between latent variables are estimated per default.
If one does not wish for them to covary, it must be explicitly stated, e.g., with f1 ~~ 0*f2, assuming the
latent variables are called f1 and f2, or by overriding the default behaviour for the entire model by adding
orthogonal = TRUE (which sets the correlation between all latent variables to zero) to the fitting call.10
10 This is at least the current behaviour of both the cfa and sem wrappers. In fact, both wrappers seem to be identical in

terms of the default settings, see Rosseel et al. (2020).

12
...
## Covariances:
## Estimate Std.Err z-value P(>|z|)
## alpha ~~
## x1 0.680 0.071 9.640 0.000
## x2 0.617 0.070 8.797 0.000
## x3 0.660 0.070 9.470 0.000
## x4 0.700 0.066 10.547 0.000
## z 0.000
## x1 ~~
## x2 3.472 0.260 13.337 0.000
## x3 3.528 0.256 13.791 0.000
## x4 3.118 0.237 13.153 0.000
## z 4.919 0.311 15.814 0.000
## x2 ~~
## x3 3.536 0.263 13.459 0.000
## x4 3.534 0.249 14.186 0.000
## z ~~
## x2 5.363 0.326 16.459 0.000
## x3 ~~
## x4 3.241 0.240 13.498 0.000
## z ~~
## x3 5.115 0.316 16.192 0.000
## x4 4.739 0.295 16.065 0.000
...

Next the intercepts and means are shown. The intercepts of the endogenous variables are marked with a
dot (.) beside the variable name. The variables without the dot are exogenous and so the estimates refer to
unconditional means. xit is treated as an exogenous variable in the model, but its unconditional mean can
be derived from its equation,

E(xit ) = βxt ,z E(zi ) + βxt ,α αi + E(δit )

= 0.5(1.5) + 0.85(2.0) + 3.0 = 5.45,

where we simply plug in the values from the simulation description above. For yit , since it is an endogenous
variable, we are interested in the intercept. We do not have enough empirical information to estimate means
for the individual effects and the idiosyncratic errors, so we let them get ‘absorbed’ into the intercepts. The
expected value of yit is

E(yit ) = κyt + βyt ,xt E(xit ) + βyt ,z E(zi )

= (0.75(2.0) + 3.0) + 0.30(5.45) + 0.45(1.5) = 6.81

where, again, κyt = E(αi ) + E(νit ) ‘absorbs’ the expectations of αi and νit . The equation can be rearranged
in terms of κyt ,

κyt = E(yit ) − βyt ,xt E(xit ) + βyt ,z E(zi )

= 6.81 − 0.30(5.45) + 0.45(1.5) = 4.5.

which is roughly the intercept shown below in the output.11

11 The values in the output will differ slightly from the theoretical values due to sampling error.

13
 ...
## Intercepts:
## Estimate Std.Err z-value P(>|z|)
## .y1 4.469 0.062 71.781 0.000
## .y2 4.520 0.062 72.818 0.000
## .y3 4.496 0.061 73.230 0.000
## .y4 4.546 0.062 72.789 0.000
## x1 5.438 0.085 63.970 0.000
## z 1.431 0.100 14.286 0.000
## x2 5.412 0.088 61.642 0.000
## x3 5.315 0.086 62.067 0.000
## x4 5.467 0.080 68.186 0.000
## alpha 0.000
...

Finally, the variance estimates are listed. Here, we see that in order to mimic the behaviour of a traditional
FE model, the error variances over time were specified to be equal using the equality constraint (nu). Notice
again the . beside y1, y2, etc.: this indicates that the listed variance refers to an endogenous variable, and
that it is thus an error variance. In this case, these refer to the variances of νt . After that, the variances of
the exogenous variables, both observed and unobserved are listed.

...
## Variances:
## Estimate Std.Err z-value P(>|z|)
## .y1 (nu) 1.039 0.027 38.730 0.000
## .y2 (nu) 1.039 0.027 38.730 0.000
## .y3 (nu) 1.039 0.027 38.730 0.000
## .y4 (nu) 1.039 0.027 38.730 0.000
## x1 7.227 0.323 22.361 0.000
## z 10.040 0.449 22.361 0.000
## x2 7.709 0.345 22.361 0.000
## x3 7.332 0.328 22.361 0.000
## x4 6.429 0.288 22.361 0.000
## alpha 0.611 0.042 14.711 0.000

4.1 Testing against a random effects model

The RE model differs from the FE model in that it assumes there is no correlation between the model
covariates and the individual effects. We can test this assumption by running an RE model and then
comparing it with the FE one. If the model fit for the RE model is not substantially worse, it would be an
indication that the could retain the assumption Cov(xit , αi ) = 0 and choose a RE specification. Because the
RE model is nested within the FE model, i.e., we can get the RE model by fixing some of the parameters
in the FE model to zero Bollen and Brand (2010), we can perform a likelihood ratio test and see if the
difference in model fit is significant, or not.
The RE model is achieved by simply fixing the covariances in line 10 in the code above to zero to represent
the assumption that Cov(xit , αi ) = 0.

1 re_sem <- '

2 # Define individual effects variable
3 alpha =~ 1*y1 + 1*y2 + 1*y3 + 1*y4
4 # Regressions, constrain coefficient to be equal over time
5 y1 ~ beta*x1 + gamma*z

14
 6 y2 ~ beta*x2 + gamma*z
7 y3 ~ beta*x3 + gamma*z
8 y4 ~ beta*x4 + gamma*z
9 # Correlations, individual effects do not covary with the covariate
10 alpha ~~ 0*x1 + 0*x2 + 0*x3 + 0*x4 + 0*z
11 x1 ~~ x2 + x3 + x4 + z
12 x2 ~~ x3 + x4 + z
13 x3 ~~ x4 + z
14 x4 ~~ z
15 # Constrain residual variances to be equal over time
16 y1 ~~ nu*y1
17 y2 ~~ nu*y2
18 y3 ~~ nu*y3
19 y4 ~~ nu*y4
20 '
21 re_sem.fit <- sem(model = re_sem,
22 meanstructure = TRUE,
23 data = dfw,
24 estimator = "ML")

The RE model shows a χ2 statistic of 299.997 based on 26 degrees of freedom. Contrast this with the χ2
statistic of the FE model of only 18.210 (22). The difference in degrees of freedom comes from the four
covariances that are constrained to zero in the RE model and estimated freely in the FE one.
The estimated coefficient for xit is β̂RE = 0.363∗∗∗ (0.009), which is substantially larger than the effect of
the FE model, which was β̂F E = 0.298∗∗∗ (0.010). This is because we have a kind of omitted variable bias
stemming from the incorrect assumption that Cov(xit , αi ) = 0.
As mentioned, because the RE model is nested within the FE one, we can perform a likelihood ratio test
to determine if the difference in model fit is significant, or not. In R, we can use the anova() function and
supply it with the fitted models from above to perform a likelihood ratio test:

anova(fe_sem.fit, re_sem.fit)

## Chi-Squared Difference Test

##
## Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)
## fe_sem.fit 22 34895 35052 18.209
## re_sem.fit 26 35169 35307 299.997 281.79 4 < 2.2e-16 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

The table shows a comparison of the nested models, in descending order according to degrees of freedom.
The RE model does not estimate the correlations between the individual effects and the covariates, so it
is more parsimonious and thus listed at the bottom. The Chisq column shows the χ2 statistic for both
models and the Chisq diff column calculates the difference between the two. Obviously, according to the
simulated data generating process (DGP), the correlation between the individual effects and xit is not zero,
so fixing these to zero leads to a substantial amount of misfit. The last column puts the χ2 difference in
relation to the difference in degrees of freedom and gives a p-value that we can use to decide on whether to
discard the null hypothesis that the models fit equally well. Here, the change in χ2 is highly significant, so
the more general FE model should be retained.

15
5 Conclusion
Panel regression in SEM has been outlined in well-known articles by P. Allison (2011); Bollen and Brand
(2010); Teachman et al. (2001). This article provides a focused look at the implementation of the basic
model using the lavaan package in R. The supplementary materials further discuss common extensions and
some tools for evaluating and loosening model assumptions.
The benefits of the SEM-based approach as opposed to traditional least squares-based FE-models are largely
the same ones that apply to the SEM framework in general: for one, SEM allows for a great deal of flexibility.
For example, it is easy to loosen model constraints as necessary. Measurement error in both the dependent
and independent variables can be dealt with using latent variables to address attenuation bias in variables
measured with error. Researchers interested in time-invariant predictors can integrate them into a hybrid
FE/RE model with ease. Further extensions, like measurement invariance testing (Schoot, Lugtig, and
Hox 2012; Millsap 2011; Steenkamp and Baumgartner 1998) as well as lagged dependent variables (Bollen
and Brand 2010; P. Allison, Williams, and Moral-Benito 2017) for example, can also be implemented in a
straightforward fashion.
There are drawbacks or at least important caveats associated with the SEM-based approach to panel regres-
sion. For one, the models can become very cumbersome to specify. Even fairly simple models can turn in
into dozens of lines of code, depending on the number of observed waves of data. By introducing multiple
indicator measurement models, the code can quickly expand to over 100 lines. The normality assumption
is often a point of contention, and while there are numerous ways to estimate such models with nonnormal
and categorical outcomes in SEM (discussed briefly in the supplementary materials), there remains some
uncertainty regarding which alternative to use in a given situation, and under what circumstances one can
expect desirable results. Simulation studies can help guide one’s choice, but there is always a chance that
some neglected boundary conditions could thwart one’s analysis.
The most basic static panel SEM regression model is the basis for a variety of currently popular extended
models, such as Latent Curve Models in general (Curran and Bollen 2001; Bollen and Curran 2004), as well
as special implementations like the Dynamic Panel Model (P. Allison, Williams, and Moral-Benito 2017),
the Random-Intercept Cross-Lagged Panel Model (Hamaker, Kuiper, and Grasman 2015) and the Latent
Curve Model with Structured Residuals (Curran et al. 2014). For this reason, it is all the more important
for researchers to have a good grasp on the method of applying panel regression in SEM, and understanding
the intuition of controlling for time-invariant confounders. This article is meant to serve as a consolidated
resource for researchers looking for concrete advise on specifying RE, FE and more general panel models in
SEM.

References
Allison, Paul. 2011. Fixed Effects Regression Models. Thousand Oaks: Sage Publications.
Allison, Paul D. 2009. Fixed Effects Regression Models. Thousand Oaks: Sage Publications. https://doi.
org/10.4135/9781412993869.
Allison, Paul, Richard Williams, and Enrique Moral-Benito. 2017. “Maximum Likelihood for Cross-Lagged
Panel Models with Fixed Effects.” Socius 3: 1–17. https://doi.org/10.1177/2378023117710578.
Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An Empiricist’s Com-
panion. Princeton, Oxford: Princeton University Press.
Bianconcini, Silvia, and Kenneth A. Bollen. 2018. “The Latent Variable-Autoregressive Latent Trajec-
tory Model: A General Framework for Longitudinal Data Analysis.” Structural Equation Modeling: A
Multidisciplinary Journal 25 (5): 791–808.
Bollen, Kenneth. 1989. Structural Equations with Latent Variables. New York, Chichester: Wiley.

16
Bollen, Kenneth, and Jennie Brand. 2010. “A General Panel Model with Random and Fixed Effects: A
Structural Equations Approach.” Social Forces 89(1): 1–34.
Bollen, Kenneth, and Patrick Curran. 2004. “Autoregressive Latent Trajectory (ALT) Models: A Syn-
thesis of Two Traditions.” Sociological Methods and Research 32(3): 336–83. https://doi.org/10.1177/
0049124103260222.
Brüderl, Josef, and Volker Ludwig. 2015. “Fixed-Effects Panel Regression.” In The Sage Handbook of
Regression Analysis and Causal Inference, edited by Henning Best and Christof Wolf, 327–57. London,
Thousand Oaks: Sage Publications.
Curran, Patrick, and Daniel Bauer. 2011. “The Disaggregation of Within-Person and Between-Person Effects
in Longitudinal Models of Change.” Annual Review of Psychology 62: 583–619.
Curran, Patrick, and Kenneth Bollen. 2001. “The Best of Both Worlds: Combining Autoregressive and
Latent Curve Models.” In New Methods for the Analysis of Change, edited by L. Collins and A. Sayer,
107–35. Washington, DC: American Psychological Press. https://doi.org/%20doi:10.1037/10409-004.
Curran, Patrick, Andrea Howard, Sierra Bainter, Stephanie Lane, and James McGinley. 2014. “The Sep-
aration of Between-Person and Within-Person Components of Individual Change over Time: A Latent
Growth Curve Model with Structured Residuals.” Journal of Consulting and Clinical Psychology 82(5):
879–94. https://doi.org/10.1037/a0035297.
Ferron, John M., and Melinda R. Hess. 2007. “Estimation in SEM: A Concrete Example.” Journal of
Educational and Behavioral Statistics 32 (1): 110–20.
Hamaker, Ellen, Rebecca Kuiper, and Raoul Grasman. 2015. “A Critique of the Cross-Lagged Panel Model.”
Psychological Methods 20(1): 102–16. https://doi.org/10.1037/a0038889.
Heckman, James J. 1981. “Heterogeneity and State Dependence.” In Studies in Labor Markets, edited by
Sherwin Rosen, 91–140. Chicago: University of Chicago Press.
Hox, Joop J. 2010. Multilevel Analysis: Techniques and Applications. Second Edition. New York, Hove:
Routledge.
Hoyle, Rick H. 2015a. “Introduction and Overview.” In Handbook of Structural Equation Modeling, edited
by Rick H. Hoyle, 3–16. New York, London: The Guilford Press.
———. 2015b. “Model Specification in Structural Equation Modeling.” In Handbook of Structural Equation
Modeling, edited by Rick H. Hoyle, 126–44. New York, London: The Guilford Press.
Hsiao, Cheng. 2014. Analysis of Panel Data. Third Edition. New York: Cambridge University Press.
Millsap, Roger. 2011. Statistical Approaches to Measurement Invariance. New York, London: Routledge.
Muthén, Bengt O., Linda K. Muthén, and Tihomir Asparouhov. 2016. Regression and Mediation Analysis
Using Mplus. Los Angeles, CA: Muthén & Muthén.
Muthén, Linda, and Bengt Muthén. 1998--2017. Mplus User’s Guide. Eighth Edition. Los Angeles, Califor-
nia: Muthén & Muthén.
R Core Team. 2017. R: A Language and Environment for Statistical Computing. Vienna, Austria: R
Foundation for Statistical Computing. https://www.R-project.org/.
Rosseel, Yves. 2012. “lavaan: An R Package for Structural Equation Modeling.” Journal of Statistical
Software 48 (2): 1–36. http://www.jstatsoft.org/v48/i02/.
Rosseel, Yves, Terrence D. Jorgensen, Daniel Oberski, Jarrett Byrnes, Leonard Vanbrabant, Victoria Savalei,
Ed Merkle, et al. 2020. lavaan: Latent Variable Analysis. https://CRAN.R-project.org/package=lavaan.
Schmidheiny, Kurt. 2019. “Panel Data: Fixed and Random Effects.” Unversität Basel.
Schoot, Rens van de, Peter Lugtig, and Joop Hox. 2012. “A Checklist for Testing Measurement Invariance.”
European Journal of Developmental Psychology 9(4): 486–92.

17
Seifert, Nico, and Henrik Andersen. 2020. “The Impact of Social Isolation on Subjective Well-Being.”
December 16, 2020. https://www.researchgate.net/publication/348548953_The_impact_of_social_
isolation_on_subjective_well-being.
Skrondal, Anders, and Sophia Rabe-Hesketh. 2004. Generalized Latent Variable Modeling. Boca Raton,
London, New York, Washington, D.C.: Chapman & Hall/CRC.
Steenkamp, Jan-Benedict, and Hans Baumgartner. 1998. “Assessing Measurement Invariance in Cross-
National Consumer Research.” Journal of Consumer Research 25(1): 78–90.
Teachman, Jay, Greg Duncan, Jean Yeung, and Dan Levy. 2001. “Covariance Structure Models for Fixed
and Random Effects.” Sociological Methods and Research 30(2): 271–88.
Wooldridge, Jeffery. 2002. Econometric Analysis of Cross Sectional and Panel Data. Cambridge, Mas-
sachusetts: The MIT Press.
———. 2009. Introductory Econometrics: A Modern Approach, 4th Edition. Mason, Ohio: South-Western
Cengage Learning.
———. 2012. Introductory Econometrics: A Modern Approach, 5th Edition. Mason, Ohio: Thomson
South-Western.
Zyphur, Michael J., Paul D. Allison, Louis Tay, Manuel C. Voelkle, Kristopher J. Preacher, Zhen Zhang,
Ellen L. Hamaker, et al. 2020a. “From Data to Causes i: Building a General Cross-Lagged Panel Model
(GCLM).” Organizational Research Methods 23 (4): 651–87. https://doi.org/10.1177/1094428119847278.
———, et al. 2020b. “From Data to Causes II: Comparing Approaches to Panel Data Analysis.” Organiza-
tional Research Methods 23 (4): 688–716. https://doi.org/10.1177/1094428119847280.

18
 Figure 1: Four-wave random effects model
σν2 σν2 σν2 σν2

ν1 ν2 ν3 ν4

y1 y2 y3 y4

1 1 1 1

β β γ γ γ γ β β

x1 x2 z α x3 x4

19
 Figure 2: Four-wave fixed effects model
σν2 σν2 σν2 σν2

ν1 ν2 ν3 ν4

y1 y2 y3 y4

1 1 1 1

β β γ γ γ γ β β

x1 x2 z α x3 x4

20