A State Space Modeling Approach

to Mediation Analysis

Fei Gu

McGill University

Kristopher J. Preacher

Vanderbilt University

Emilio Ferrer

University of California, Davis

Mediation is a causal process that evolves over time. Thus, a study of

mediation requires data collected throughout the process. However, most

applications of mediation analysis use cross-sectional rather than longitudinal data. Another implicit assumption commonly made in longitudinal designs

for mediation analysis is that the same mediation process universally applies

to all members of the population under investigation. This assumption ignores

the important issue of ergodicity before aggregating the data across subjects.

We first argue that there exists a discrepancy between the concept of mediation and the research designs that are typically used to investigate it. Second,

based on the concept of ergodicity, we argue that a given mediation process

probably is not equally valid for all individuals in a population. Therefore, the

purpose of this article is to propose a two-faceted solution. The first facet of

the solution is that we advocate a single-subject time-series design that aligns

data collection with researchersâ€™ conceptual understanding of mediation. The

second facet is to introduce a flexible statistical methodâ€”the state space

modelâ€”as an ideal technique to analyze single-subject time series data in

mediation studies. We provide an overview of the state space method and illustrative applications using both simulated and real time series data. Finally, we

discuss additional issues related to research design and modeling.

Keywords: mediation, state space model

1. Introduction

Mediation is a causal process that evolves over time. In the simplest case, the

causal variable (X) exerts an effect on the outcome variable (Y) partially or

Journal of Educational and Behavioral Statistics

2014, Vol. 39, No. 2, pp. 117â€“143

DOI: 10.3102/1076998614524823

# 2014 AERA. http://jebs.aera.net

117

completely through a mediator variable (M) over time. Clearly, time plays an

explicit role in the mediation process. In one of the earliest articles devoted specifically to mediation, Judd and Kenny (1981) emphasized the role of time, even

placing the words â€˜â€˜process analysisâ€™â€™ in the title of their article. More recently,

Schmitz (2006) provided an overview of the necessity of process analyses in the

context of learning and instruction. Ideally, the empirical study of mediation

requires (1) data collected throughout the process and (2) pertinent statistical

methods that can capture the dynamic mechanism underlying the causal process.

However, process-oriented methods that explicitly consider the role of time are

not common in educational and psychological research (Schmitz, 2006). In the

context of mediation, Cole and Maxwell (2003) discussed that most applications

of mediation analysis use cross-sectional rather than longitudinal data, and not all

methodological treatments acknowledge the necessity of considering time

(see Maxwell & Cole, 2007; Maxwell, Cole, & Mitchell, 2011).

Another implicit assumption commonly made in mediation analysis is that the

same mediation process universally applies to all members of the population

under investigation. This assumption is made in a handful of longitudinal models

recently developed within the structural equation modeling (SEM) framework

proposed to study mediation processes. Cheong, MacKinnon, and Khoo (2003)

used a parallel process latent growth curve model to investigate the effect of a

causal variable on the change in an outcome variable through change in a mediator variable. Gollob and Reichardt (1991) and Cole and Maxwell (2003)

implemented a cross-lagged panel model emphasizing longitudinal relations

between the absolute level of the causal variable and the outcome variable

through the mediator variable. Another variant of a longitudinal SEM approach

is the latent difference score model, in which differences between adjacent observations are treated as latent variables (e.g., Hamagami & McArdle, 2007;

MacKinnon, 2008; McArdle, 2001; McArdle & Nesselroade, 1994; Selig &

Preacher, 2009). In the traditional regression framework, Judd, Kenny, and

McClelland (2001) advocated the use of within-subject designs (in contrast to

between-subjects designs) to assess mediation and moderation, where individuals are put into both treatments. In all models mentioned previously, parameters

are estimated by pooling information across subjects. Although the longitudinal

designs in which these models are applied take the role of time into consideration, they are still limited in that none of them considers the important issue

of ergodicity before aggregating the data across subjects (ergodicity is described

later). As we will show shortly, the conclusions drawn from pooling across subjects may not be informative about how single subjects behave (e.g., Ferrer &

Widaman, 2008; Molenaar, 2004).

In this article, we contend that (a) there is a discrepancy between the concept

of mediation and the research designs that are typically undertaken to investigate

it in practice and (b) based on the concept of ergodicity, there is not necessarily a

universal mediation process that is equally valid for everyone in the population.

A State Space Modeling Approach to Mediation Analysis

118

Therefore, the purpose of this article is to propose a two-faceted solution to the

persistent problem of using cross-sectional designs in mediation analysis. The

first facet of the solution is to encourage researchers to rethink how they

approach the research design for mediation studies; that is, we advocate moving

from the traditional cross-sectional design to a single-subject time-series design

that aligns data collection with researchersâ€™ conceptual understanding of mediation. The second facet is to introduce a flexible statistical methodâ€”the state

space model (SSM)â€”as an ideal technique to analyze single-subject1 timeseries data in mediation studies. In Section 2, we elaborate on the theoretical

foundation of the single-subject time-series design for mediation analysis. In

Section 3, we provide an overview of state space methods and two illustrative

applications using simulated and empirical data sets. In Section 4, we discuss

additional issues related to research design and modeling.

2. Foundation of the Single-Subject Time-Series Design

In a cross-sectional design, the focus is on the analysis of interindividual

variationâ€”differences among different units at a single point in time. According

to a survey in 2005 of the five American Psychological Association journals publishing the most articles studying mediation, Maxwell and Cole (2007) reported

that more than half of the mediation studies were based on cross-sectional data.

However, according to the concept of mediation as well as the basic requirements

for causal inference, mediation must involve at least two relations that unfold

over time. Specifically, the effect of a causal variable (X) is first exerted on the

mediator variable (M, i.e., X ! M), and then, this effect is carried over to the outcome variable (Y, i.e., M ! Y). Thus, it is immediately clear that a certain amount

of time must elapse for the effect of X to reach Y. Therefore, a requirement for

any mediation analysis is the consideration of the role of time, that is, the necessity of the analysis of intraindividual variationâ€”changes in the same unit over

time. By comparing the concept of mediation and how mediation analysis was

conducted in the literature, it is evident that there exists a large discrepancy

between how mediation is theoretically conceptualized and how it has actually

been modeled in the past. This discrepancy raises an important validity issue concerning the equivalence between the analysis of interindividual variation and the

analogous analysis of intraindividual variation.

Cole and Maxwell (2003) argued and demonstrated that very restrictive conditions are required to ensure accurate results from mediation analysis based on

cross-sectional data. In reality, such restrictive conditions almost never occur,

and bias in cross-sectional analyses of longitudinal mediation has been amply

demonstrated (Maxwell & Cole, 2007; Maxwell et al., 2011). The divide between

interindividual variation and intraindividual variation not only exists in mediation analysis but also appears in areas such as test theory, factor analysis, and

developmental psychology (Molenaar, 2004, 2008a, 2008b).

Gu et al.

119

Given the existence of this divide in various areas, an important question

becomes whether this division between orientations is justified. Unfortunately,

as a direct consequence of the classical ergodic theorems, the answer to the question is â€˜â€˜no.â€™â€™ In fact, equivalence between the analysis of inter- and intraindividual variation is established only for ergodic processes (Molenaar, 2004). It is

worth noting that, outside the context of mediation and in the broader sense of

studying behavior, the inter-/intraindividual debate can be traced back to an older

distinction between nomothetic lawfulness, emphasizing generality in the population, and idiographic characterization, emphasizing the uniqueness of the individual (e.g., Allport, 1937; Lamiell, 1981, 1988; Molenaar, 2004; Rosenzweig,

1958; van Kampen, 2000; Zevon & Tellegen, 1982). When individuals differ

qualitatively rather than quantitatively, â€˜â€˜Qualitative differences mistaken for

quantitative differences can seriously distort relationships and are a prescription

for diluted nomothetic relationshipsâ€™â€™ (Nesselroade, Gerstorf, Hardy, & Ram,

2007, p. 219). The following subsection gives a brief, heuristic description of the

concept of ergodicity as the foundation of the single-subject time-series design.

2.1. Ergodicity

From the perspective of dynamical systems, a process is said to be ergodic if

the average of a single trajectory over time (structure of intraindividual variation)

is equal to the average of the ensemble of trajectories at a single point in time

(structure of interindividual variation by pooling across subjects). In order to

understand the consequences of ergodicity in psychology, the development of

human behavior over time is conceived of as a unique high-dimensional space

that contains dynamic processes and all the relevant information about the subject (cf. Molenaar, 1994, 2004, 2008a, 2008b; Molenaar & Ram, 2009, 2010;

Sinclair & Molenaar, 2008). For a particular individual, a finite sample of the

dynamic process over consecutive time points (usually evenly spaced) constitutes a trajectory in his or her behavior space. This trajectory carries information

about intraindividual variation. Correspondingly, a finite sample of the same

behavior space from a group of individuals at a single point in time represents

an ensemble of trajectories, carrying information about interindividual variation

from this group of individuals. Hence, the question of whether the divide

between orientations is justified becomes a question of whether the developmental trajectory of human behavior is ergodic. This, in turn, reduces to the question

of whether stationarity and homogeneity hold for a Gaussian process. Therefore,

examining the stationarity and homogeneity criteria for a given process are

essential empirical steps.

For a Gaussian process, two criteria are required to be met simultaneously for

a process to be considered ergodic: stationarity and homogeneity. In terms of the

first criterion, a stationary process refers to a stochastic process whose joint probability distribution is time-invariant. For a Gaussian process, stationarity2

A State Space Modeling Approach to Mediation Analysis

120

requires that the first two moments of the process are time-invariant. That is, the

mean function of the time series is a constant, and the covariance function of the

time series depends only on relative time differences (i.e., â€˜â€˜lagâ€™â€™). Nonstationary

processes, however, are the norm in psychology, for example, learning and

developmental trajectories.

Regarding the second criterion, homogeneity refers to the situation in which each

member of a population obeys the same dynamic law and follows the same statistical

model, constituting exchangeable replications of each other, much as molecules of a

homogeneous gas. The reality, however, is quite the opposite. In fact, heterogeneity

is a general characteristic of human populations. Furthermore, besides the widely

recognized genetic and environmental effects that cause heterogeneity, Molenaar,

Boomsma, and Dolan (1993) arguedthatthere exists a third source of developmental

differences: self-organization of nonlinear epigenetic processes.

Since nonstationarity, heterogeneity, or both are thought to be the rule rather

than the exception in most psychological processes, such processes are then

nonergodic, which means that the structure of interindividual variation is not

equivalent to the structure of intraindividual variation. Therefore, we conclude

that there is no equivalence between measurement orientations in the majority

of cases. This implies that there are not necessary lawful relationships between

the analysis of inter- and intraindividual variation. Thus, in cases concerning processes that unfold over time, statistical analyses should focus on intraindividual

variation, with greater emphasis on single-subject time-series designs. This conclusion, therefore, also applies to mediation analysis.

Discussions of the implications and consequences of ergodicity started to

appear in many areas of psychological research about two decades ago (e.g.,

Molenaar, 1994; Molenaar, 2004, 2008a, 2008b; Molenaar & Campbell, 2009;

Molenaar & Ram, 2009, 2010; Molenaar, Sinclair, Rovine, Ram, & Corneal,

2009; Nesselroade & Molenaar, 1999; Sinclair & Molenaar, 2008, and most

recently Hamaker, 2012). However, there is virtually no mention of ergodicity

in the mediation literature (but see Roe, 2012).

3. Overview of State Space Methods

After establishing the theoretical foundation of the time-series design, processoriented methods are required to analyze the time series data. In this section, we

discuss the second facet of our proposed solution, that is, an overview of SSM. This

is, we believe, the first effort to utilize SSM to investigate mediation. As the first

application of SSM in the context of mediation, we provide some essential basics

of the most straightforward and frequently discussed variant of SSM in the timeseries and econometrics literature, that is, the linear Gaussian SSM. Due to space

limitations, our introduction is brief. For more comprehensive treatments of the

state space methodology in general, we refer readers to Commandeur and

Koopman (2007), Harvey (1989), and Durbin and Koopman (2001).

Gu et al.

121

3.1 History of State Space Modeling in Psychology and Its Application to

Mediation Analysis

State space methods have their origin in control theory, beginning with the

groundbreaking article by Kalman (1960). Applications in astronautics were initially developed (and are still used) for accurately tracking the position and velocity of moving objects such as aircraft, missiles, and rockets. Shortly after its

application in engineering, SSM also found application in time series analysis and

econometrics (e.g., Aoki, 1987; Harvey, 1989). More recently, quantitative social

and behavioral scientists have begun applying SSM because of its statistical flexibility for evaluating both the measurement properties and the leadâ€“lag relationships among latent variables in psychological processes. Analytic similarities

and differences between the currently dominant SEM and the relatively newly

emerging SSM in the psychology literature are discussed by several authors. Specifically, MacCallum and Ashby (1986) noted that SSM is a special case of SEM,

while Otter (1986) showed the reverse. Chow, Ho, Hamaker, and Dolan (2010)

reconciled the two approaches and provided a more detailed discussion of the relative strengths and weaknesses of both approaches vis-a`-vis their use in representing

intraindividual dynamics and interindividual differences.

Another line of research involving SSM in psychology is built upon the state

space representation of the dynamic factor model (DFM; Molenaar, 1985).

Dynamic factor analysis was proposed to combine P-technique factor analysis

(Cattell, 1963; Cattell, Cattell, & Rhymer, 1947) and time series analysis (Browne

& Nesselroade, 2005; Molenaar, 1985; Molenaar, de Gooijer, & Schmitz, 1992;

Nesselroade, McArdle, Aggen, & Meyers, 2002). Some recent work devoted to

methodological discussions and substantive applications of DFM can be found

in the psychology literature (e.g., Chow, Nesselroade, Shifren, & McArdle,

2004; Ferrer & Nesselroade, 2003; Hershberger, Corneal, & Molenaar, 1994;

Nesselroade & Molenaar, 1999; Sbarra & Ferrer, 2006; Shifren, Hooker, Wood,

& Nesselroade, 1997; Wood & Brown, 1994; Zhang & Browne, 2006). Given the

strong similarity between SSM and DFM, the two terms frequently are used interchangeably, and DFMs often are expressed in state space form to exploit better

parameter estimation properties (Hamaker, Dolan, & Molenaar, 2005; Ho, Ombao,

& Shumway, 2005; Ho, Shumway, & Ombao, 2006; Molenaar, 1994; Song & Ferrer, 2009, 2012; Zhang, Hamaker, & Nesselroade, 2008).

Although SSM models have been used in many areas of research, their application in questions about mediation is not available in the literature. Statistically,

probably all existing longitudinal models for studying mediation can be represented in their state space form, and identical results can be obtained. However,

we do not pursue this direction because of the consequences of the ergodic theorems stated before. Our proposed approach, instead, is based on the specification of an SSM for analyzing time series data from a single subject (especially

when the number of measurement occasions is large).

A State Space Modeling Approach to Mediation Analysis

122

In practice, modeling the mediation process as an SSM has important benefits.

First, it allows the researcher to investigate time-related sequences among variables (i.e., predictor ! mediator ! outcome), as the process represented by

these variables unfolds over time. Although some longitudinal structural equation models also allow such investigation, the state space approach provides a

better and more thorough depiction because some mediation processes need a

longer time to unfold. Second, SSM can accommodate complex specifications

such as measurement structures and second-order factors.3 Third, state space

analyses at the individual level provide a theoretically sound, bottom-up

approach to create homogeneous subpopulations. If a hypothetical model fits

separately the time series data from several subjects, a homogeneous subpopulation can be created from the analyses at the individual level, and generalized conclusions can be drawn for this subpopulation. The bottom-up approach, however,

can be labor-intensive. More discussion of multiple-subject time series is given

in subsequent sections.

3.2. The Linear Gaussian SSM

Currently, there is no standard notation in the literature for SSM, and different

authors have different preferences. Based on the similarity between SEM and

SSM, we choose to use LISCOMP notation to present the formulation of SSM.

The benefit of using LISCOMP notation is that each matrix has gained a standard

interpretation in the literature, thus providing a convenient and familiar notation.

Let yt be a p-variate vector representing p manifest variables, Zt a q-variate vector

representing q-latent variables (p q 1), and t Â¼ 1, 2, … , T denotes the time

point for the corresponding vector. The general linear Gaussian SSM contains a

measurement equation

yt Â¼ tt Ã¾ LtZt Ã¾ et; et MVNÃ°0; tÃž;

and a transition equation,

Zt Â¼ at Ã¾ BtZt1 Ã¾ zt; zt MVNÃ°0; CtÃž;

where tt is a p 1 vector for intercepts, Lt is a p q loading matrix, et is a p 1

vector for measurement errors (also known as innovations in the time series literature), t is a p p diagonal covariance matrix, at is a q 1 vector for means,

Bt is a q q transition matrix, zt is a q 1 vector for residuals, and Ct is a q q

covariance matrix. The measurement errors and residuals are assumed to be serially independent and independent of each other at all time points. We denote as yt

the vertical vector that collects all parameters in tt, Lt, t, at, Bt, and Ct, and the

subscript t means that yt is time varying, thus resulting in the time-varying SSM.

Applications of the time-varying SSM can be found in recent articles by Molenaar, Sinclair, Rovine, Ram, and Corneal (2009), Sinclair and Molenaar

(2008), and Chow, Zu, Shifren, and Zhang (2011).

Gu et al.

123

Here we consider only the time-invariant SSM, with the understanding that

the model could be extended to include time-varying parameters. The measurement and transition equations are thus simplified to

yt Â¼ t Ã¾ LZt Ã¾ et; et MVNÃ°0; Ãž

Zt Â¼ a Ã¾ BZt1 Ã¾ zt; zt MVNÃ°0; CÃž;

in which the parameter vector, y, is time-invariant.

3.3. Parameter Estimation and the Kalman Filter

Unknown parameters of the linear Gaussian SSM are estimated via a recursive

algorithm, called the Kalman filter (KF). The KF4 algorithm is initialized with the

latent variable, Z0|0, and the associated covariance matrix, P0|0, and proceeds with

the prediction and filtering steps iteratively at each time point. When t Â¼ 1, the prediction step gives the predicted latent variable and its covariance matrix, that is,

Z1j0 Â¼ a Ã¾ BZ0j0

P1j0 Â¼ BP0j0B0 Ã¾ C:

As a byproduct, the one-step-ahead prediction error and its associated covariance

matrix are obtained, that is,

e1 Â¼ y1 y1j0 Â¼ y1 Ã°t Ã¾ LZ1j0Ãž

D1 Â¼ LP1j0L0 Ã¾ :

Then, the filtering step uses the observed value at t Â¼ 1 to update the predicted

values, giving

K1 Â¼ P1j0L0

D1

1 Ã°Kalman gain matrixÃž

Z1j1 Â¼ Z1j0 Ã¾ K1e1 Â¼ Z1j0 Ã¾ P1j0L0

D1

1 e1

P1j1 Â¼ P1j0 K1D1K0

1 Â¼ P1j0 P1j0L0

D1

1 LP1j0:

When t Â¼ 2, Z1|1 and P1|1 are used in the prediction step, followed by the filtering

step to calculate Z2|2 and P2|2, and so on. In sum, for t Â¼ 1, 2, … , T, the recursive

KF algorithm can be written as

Ztjt1 Â¼ a Ã¾ BZt1jt1

Ptjt1 Â¼ BPt1jt1B0 Ã¾ C

et Â¼ yt ytjt1 Â¼ yt Ã°t Ã¾ LZtjt1Ãž

Dt Â¼ LPtjt1L0 Ã¾

Kt Â¼ Ptjt1L0

D1

t

Ztjt Â¼ Ztjt1 Ã¾ Ktet Â¼ Ztjt1 Ã¾ Ptjt1L0

D1

t et

Ptjt Â¼ Ptjt1 KtDtK0

t Â¼ Ptjt1 Ptjt1L0

D1

t LPtjt1:

Inserting et and Dt at each time point into the log-density function of the multivariate normal distribution and summing all log-density functions, the prediction

error decomposition (PED; Schweppe, 1965) function is obtained:

A State Space Modeling Approach to Mediation Analysis

124

PED Â¼ 1

2

XT

tÂ¼1

p logÃ°2Ãž log Dt j j e0

t

D1

t et

:

Giving certain starting values in y and maximizing the PED function with respect

to y provides the parameter estimates.

3.4. Testing Mediation: Bootstrapping the Time Series Data

As described in the previous sections, early definitions of direct and indirect

effects in mediation analysis were based on cross-sectional designs (Baron &

Kenny, 1986). These definitions were theoretically inaccurate because of the lack

of consideration of time. Modern definitions of the concepts advocate that the

causal variable, the mediator variable, and the outcome variable of a mediation

process should be obtained at different occasions (Collins, Graham, & Flaherty,

1998; Gollob & Reichardt, 1991). Figure 1 illustrates an example of the simplest

three-variate model in which time is taken into account. This figure illustrates the

concepts of direct and indirect effects in mediation analysis. In this model, a is

defined as the direct effect of Xt1 on Mt, b is the direct effect of Mt on YtÃ¾1, and

c is the direct effect of Xt1 on YtÃ¾1. The indirect effect of Xt1 on YtÃ¾1 is defined

as the product of a and b, that is, ab. For more complicated models, the indirect

effect can be a product of more parameters linking several different occasions.

In order to evaluate the direct effects, several statistical tests are available, for

example, the Wald test and the likelihood ratio test. Testing the significance of

the indirect effects is also of great importance but more difficult, and the common

direct methods just mentioned are not appropriate because of the nonnormality of

the sampling distribution of the indirect effect. As an alternative, the use of bootstrap confidence intervals (CIs) is recommended (Bollen & Stine, 1990; Hayes,

2009; MacKinnon, Lockwood, & Williams, 2004; Preacher & Hayes, 2004,

2008a, 2008b; Shrout & Bolger, 2002). The standard nonparametric bootstrap

involves two steps. In the first step, a resample of size N is drawn with replacement from the original sample. In the second step, model parameters are estimated from this resample. The two steps are replicated B times (where B is

large), so that the sampling distribution of the statistic of interest can be obtained.

FIGURE 1. Direct and indirect effects in the simplest mediation model.

Gu et al.

125

At the .95 level, the 2.5th and 97.5th percentiles are chosen to construct the CI to

conduct the significance test of a single parameter, or of a product of parameters,

by examining whether the CI excludes 0 (indicating a significant effect). Bootstrapping the time series data, however, poses yet another difficulty because of

the temporal dependence of the time series data (Zhang & Browne, 2006). In general, the standard nonparametric bootstrap is not appropriate for time series data

because it destroys the inherent time dependency in the data. In this section, we

introduce two bootstrap methods appropriate for SSM, namely the parametric

bootstrap and the residual-based bootstrap. Besides these two methods, there are

other approaches appropriate for SSM (e.g., Zhang & Chow, 2010).

The parametric bootstrap is essentially a Monte Carlo simulation, in which

repeated bootstrap samples are simulated from a specified model, where the estimates from the original sample are treated as parameters. The underlying

assumption is that the specified model is correct in the population. To generate

each random sample, a number of steps are followed:

1. Generate Z0 from MVN (0, 100 Iq).

2. Set the iteration number t Â¼ 1.

3. Generate zt from MVN (0, C^ ).

4. Calculate Zt using Zt Â¼ a^ Ã¾ B^Zt1 Ã¾ zt.

5. Generate et from MVN (0, ^ ).

6. Calculate yt using yt Â¼ ^t Ã¾ L^Zt Ã¾ et.

7. Set t Â¼ t Ã¾ 1 and return to Step 3.

8. Repeat Steps 3 to 6 until t > T Ã¾ 1,000.

9. Save the data from 1,001 to T Ã¾ 1,000.

Although not always necessary, the first 1,000 observations are typically

discarded as the burn-in period.

The residual-based bootstrap was first applied to linear Gaussian SSM by

Stoffer and Wall (1991) in assessing the precision of maximum likelihood estimates, and it is considered a semiparametric approach. Similar to the parametric

bootstrap, population parameters are taken to be sample estimates in the

residual-based bootstrap, and the underlying assumption of a correctly specified

model is also made. On the other hand, random samples are drawn, with replacement, from the standardized residuals as in the standard nonparametric bootstrap.

Specifically, the residual-based bootstrap procedure is based on the innovations

form of the KF:

et Â¼ yt t LZtjt1

Dt Â¼ LPtjt1L0 Ã¾

Kt Â¼ Ptjt1L0

D1

t

ZtÃ¾1jt Â¼ BZtjt1 Ã¾ BKtet

yt Â¼ t Ã¾ LZtjt1 Ã¾ et:

A State Space Modeling Approach to Mediation Analysis

126

Then, the algorithm proceeds as follows:

1. Calculate standardized innovations using D^ 1=2

t ^et, denoted ~et.

2. Draw, with replacement, a random sample from ~et to obtain ~e

t .

3. Construct a bootstrap sample by fixing the initial conditions of the KF and

iteratively using the following two equations:

ZtÃ¾1jt Â¼ B^Ztjt1 Ã¾ B^K^tD^1=2

t ~e

t

yt Â¼ ^t Ã¾ L^Ztjt1 Ã¾ D^1=2

t ~e

t :

The idea behind the residual-based bootstrap is that the standardized residuals

are independent and identically distributed, and therefore exchangeable, after all

the dynamic and measurement relationships have been accounted for by the

model. This procedure, however, is not robust against model misspecification

(Stoffer & Wall, 1991, 2004; Zhang & Chow, 2010).

For the examples considered in this article, we set B Â¼ 2,000 for both bootstrap procedures to approximate the sampling distribution of the product. As a

general rule, large numbers are required to allow enough simulated cases in both

tails of the sampling distribution of the indirect effect so that the percentiles can

be accurately estimated for constructing a CI (Yung & Chan, 1999).

3.5. Illustration 1: A Simulated Lag-2 Example

In order to illustrate the parameter estimation in SSM and the two bootstrap

procedures just described, a three-variate, single-subject time series data set is

generated from a lag-2 model (depicted in Figure 2). Note that the figure represents a temporal slice of three time points from the entire process. Three equations are involved in this model:

FIGURE 2. A temporal slice of the lag-2 model.

Gu et al.

127

Yt Â¼ aYt1 Ã¾ bMt1 Ã¾ cXt1 Ã¾ gXt2 Ã¾ eYt

Mt Â¼ dMt1 Ã¾ eXt1 Ã¾ eMt

Xt Â¼ fXt1 Ã¾ eXt;

where a, d, and f are autoregressive parameters of the outcome variable, the mediator variable, and the causal variable, separately; b, c, and e are the lag-1 crossregressive parameters; g is the lag-2 cross-regressive parameter; and eYt, eMt, and

eXt are residuals in each equation. If g is equal to 0, it reduces to the lag-1 model,

which is a particular case of the lag-2 model. In addition, if c is also equal to 0, it

corresponds to one of the models in Cole and Maxwell (2003, model 5). As we

will show immediately, the lag-2 model can be expressed in state space form. In

principle, as long as we can write the model in state space form, the parameter

estimation and bootstrap procedures can be readily applied.

By defining the following measurement equation and the transition equation,

the state space form of the lag-2 model is obtained:

Yt

Mt

Xt

0

[email protected]

1

CA Â¼

100000

010000

001000

0

[email protected]

1

CA

Yt

Mt

Xt

Yt1

Mt1

Xt1

0

[email protected]

1

CCCCCCCCA

; and Â¼

000

000

000

0

[email protected]

1

CA

Yt

Mt

Xt

Yt1

Mt1

Xt1

0

[email protected]

1

CCCCCCCCA

Â¼

abc 0 0 g

0 d e 000

0 0 f 000

100000

010000

001000

0

[email protected]

1

CCCCCCCCA

Yt1

Mt1

Xt1

Yt2

Mt2

Xt2

0

[email protected]

1

CCCCCCCCA

Ã¾

eYt

eMt

eXt

0

0

0

0

[email protected]

1

CCCCCCCCA

;

and C Â¼

cY 0 0 000

0 cM 0 000

0 0 cX 000

0 0 0 000

0 0 0 000

0 0 0 000

0

[email protected]

1

CCCCCCCCA

:

We simulated four time series data sets with lengths equal to 50, 100, 150, and

200. Compared with the typical lengths in the time series literature, the lengths

considered here are relatively short. Although most social scientists typically collect at best only a handful of repeated measures, intensive longitudinal data are

highly desirable for SSM analyses. Some precedent examples have already

emerged in emotion studies using daily diary data (Chow, Hamaker, Fujita, &

A State Space Modeling Approach to Mediation Analysis

128

Boker, 2009; Song & Ferrer, 2009, 2012), a cognitive study using daily cognitive

assessment data (Chow, Hamaker, & Allaire, 2009), and a neuropsychology

study using functional magnetic resonance imaging data (Ho et al., 2005).

We estimated the parameters using the KF algorithm with results shown in

Table 1. The estimated parameters in four different length conditions roughly

show that length of the time series is inversely related to the magnitude of parameter bias. A challenge, however, is that when the sampling frequency (i.e., the

time elapsed between consecutive measurement occasions) is fixed, longer

time series data are more expensive and difficult to collect. One strategy to

obtain longer time series data is to increase the sampling frequency. The issue

of length and sampling frequency will be discussed briefly in the last section.

The indirect effect from X(t2) to Y(t) is the product of e and b. The two

bootstrap methods are used to construct 95% CIs for this product. For the four

simulated examples, the 95% CIs from both the parametric bootstrap and the

residual-based bootstrap are similar, and none of the CIs contains zero, indicating a significant lag-2 indirect effect from X(t2) to Y(t). In addition, the CIs from

the parametric bootstrap are a bit larger than those from the residual-based

bootstrap when T Â¼ 50 and 100; this difference almost disappears when

T Â¼ 150 and 200.

TABLE 1

Parameter Estimates From a Lag-2 Model Fitted to Simulated Data

Parameter True Value T Â¼ 50 T Â¼ 100 T Â¼ 150 T Â¼ 200

a: autoreg of Y .5 .440 .460 .469 .432

b: Yt on Mt1 .4 .295 .356 .351 .338

c: Yt on Xt1 .4 .373 .342 .360 .386

d: autoreg of M .5 .414 .444 .463 .476

e: Mt on Xt1 .4 .458 .462 .443 .400

f: autoreg of X .8 .820 .741 .778 .792

g: Yt on Xt2 .3 .298 .327 .281 .309

cY .1 .110 .082 .091 .096

cM .4 .282 .318 .332 .349

cX .9 .753 .758 .815 .996

eb .16 .135 .164 .155 .135

resid parm resid parm resid parm resid parm

Number converged 2,000 2,000 2,000 2,000 2,000 2,000 2,000 2,000

Lower bound of eb .067 .055 .119 .110 .117 .115 .104 .105

Upper bound of eb .220 .245 .228 .237 .209 .208 .176 .178

Range of the 95% CI .153 .190 .109 .127 .092 .093 .072 .073

Note. CI Â¼ confidence interval; resid Â¼ residual-based bootstrap; parm Â¼ parametric bootstrap; eb is

the indirect effect from Xt2 to Yt.

Gu et al.

129

3.6. Illustration 2: An Empirical Study

In this section, we fit the lag-2 model displayed in Figure 2 to the time series of

two manâ€“woman dyads. Each series contains self-reported daily stress and affect

for 91 days. These data are part of a larger study designed to examine dyadic interactions (for more details see, e.g., Ferrer, Steele, & Hsieh, 2012; Ferrer &

Widaman, 2008). The variables used in these analyses are female perceived stress

(X), male positive affect specific to his relationship (M), and female negative affect

specific to her relationship (Y). Relationship-specific affect (RSA) was measured

using the RSA scale (Ferrer et al., 2012), 18 items intended to tap into the participantsâ€™ positive and negative emotional experiences specific to their relationships.

Examples of the positive items include â€˜â€˜emotionally intimate,â€™â€™ â€˜â€˜trusted,â€™â€™ and

â€˜â€˜loved.â€™â€™ Examples of the negative items include â€˜â€˜sad,â€™â€™ â€˜â€˜trapped,â€™â€™ and â€˜â€˜discouraged.â€™â€™ The stress construct was measured using 5 items from the Positive and Negative Affect Schedule (Watson, Clark, & Tellegen, 1988) including â€˜â€˜distress,â€™â€™

â€˜â€˜upset,â€™â€™ â€˜â€˜scared,â€™â€™ â€˜â€˜nervous,â€™â€™ and â€˜â€˜afraid.â€™â€™ For all analyses, we created unitweighted composites for each person, using all the items in each of the scales.

The results of fitting the lag-2 models to these data are reported in Table 2. For

Dyad 1, the estimated values for parameters b and e are statistically significant,

indicating reliable evidence to support positive relations between the male

partnerâ€™s positive affect on a given day and his female partnerâ€™s negative affect

the following day, as well as between the female partnerâ€™s perceived stress on a

given day and the male partnerâ€™s positive affect the following day. However, the

estimated c and g parameters are not statistically significant. The nonsignificant

estimate of g may suggest a more parsimonious model that does not include the

lag-2 structure between the female partnerâ€™s stress and her subsequent negative

affect.

For Dyad 2, the estimated b parameter is not statistically significant (.011/.031

Â¼ .355), whereas the estimated e and g are both significant. It is not unreasonable

to expect that the nonsignificant estimate of b may result in a model with a

nonsignificant indirect effect linking the causal variable at t 2 to the outcome

variable at t. For researchers interested in emotion, these preliminary results

provide motivation for further investigation of each dyadâ€™s data separately. Particularly, the counterintuitive sign of the estimated b for both dyads may flag

some problems or uncertainties in the model, the data, or both. In reality, the

affective processes underlying different dyads may be qualitatively different

(e.g., including a feedback process).

For illustrative purposes, we also present the 95% CIs from the two bootstrap

methods for both dyads, but we acknowledge that the lag-2 model may not be

plausible for either dyad. The caveat is that these CIs are not readily interpretable

because neither bootstrap method is robust to model misspecification. Further

efforts are necessary to determine the best final model for each dyad and, in order

to make proper statistical inferences, the two bootstrap methods need to be

A State Space Modeling Approach to Mediation Analysis

130

applied to each final model separately. This is beyond the scope of the current

illustration, but is deserving of separate study. In summary, the message from this

illustration for mediation researchers is clear; that is, mediation analyses should

be conducted for each subject (i.e., individual, dyad, or other unit of analysis)

separately to accommodate heterogeneity in the units.

4. Additional Issues

4.1. Causal Inference

Establishing causality is an important component of longitudinal research.

There are several recent treatments of causal inference in mediation analysis,

most inspired by Rubinâ€™s causal model (or the potential outcomes framework;

see Albert, 2008; Ten Have & Joffe, 2010). In this framework, causality is

defined with reference to potential outcomes that might have been obtained

under different counterfactual conditions. Because it is not possible to observe

outcomes under all possible conditions, certain assumptions are commonly

invoked to permit causal inference, for example, the assumption that key paths

TABLE 2

Parameter Estimates From a Lag-2 Model Fitted to Two Dyadsâ€™ Time Series Over 91

Days

Parameter Dyad 1 Dyad 2

a: autoreg of Y (female negative affect relationship

specific)

.420 (.137) .267 (.121)

b: Yt on Mt1 .220 (.043) .011 (.031)

c: Yt on Xt1 .053 (.132) .473 (.214)

d: autoreg of M (male positive affect relationship specific) .707 (.049) .736 (.066)

e: Mt on Xt1 .669 (.110) .792 (.205)

f: autoreg of X (stress) .927 (.040) .973 (.020)

g: Yt on Xt2 .042 (.099) .274 (.174)

cY: Var(pagf) .184 (.027) .113 (.017)

cM: Var(nasf) .385 (.057) .588 (.088)

cX: Var(pasf) .309 (.046) .049 (.007)

eb .147 .009

resid parm resid parm

Number converged 2,000 2,000 2,000 2,000

Lower bound of eb .100 .095 .029 .028

Upper bound of eb .225 .246 .076 .049

Range of the 95% CI .125 .151 .105 .077

Note. Standard errors are in parentheses. CI Â¼ confidence interval; resid Â¼ residual-based bootstrap;

parm Â¼ parametric bootstrap; eb is the indirect effect from Xt2 to Yt.

Gu et al.

131

composing an indirect effect are not confounded by omitted variables, and the

assumption that X does not moderate the effect of M on Y, among others (Imai,

Keele, & Tingley, 2010; Pearl, 2010, 2012; VanderWeele & Vansteelandt, 2009).

The single-subject design is not an attempt to establish causal relationships

per se. Instead, it provides a temporally plausible way to model and test a

hypothetical mediation process. That is, the parameters associated with the

mediator/mediators can be tested against a null hypothesis using a sampling

distribution (perhaps obtained by bootstrapping), so that the researcher can gain

more insights into the data and determine to what degree the data are consistent

with the hypothesized underlying process. Because it is not explicitly couched

in a potential outcomes framework, researchers should be cautious in making

causal inferences using the state space modeling approach to mediation analysis. However, it should be noted that SSM has in its favor that key effects are

within-subject rather than between-subject, and measurements on key variables

are necessarily separated in time. These features provide a stronger basis for

causal inference than other methods that are designed for use with betweensubject and/or cross-sectional data. Extending the logic of the potential outcomes framework to single-subject longitudinal designs is an interesting avenue for future research.

4.2. Extension to the Multiple-Subject Time Series

Gathering time series data simultaneously on multiple persons (as in our

empirical study) is a common practice. In terms of modeling strategy, it can

be considered a straightforward extension to the model presented here. We call

this extension the multiple-subject time-series design. Since the introduction of

DFM (Molenaar, 1985), the single subject is always emphasized as the unit of

analysis. This emphasis may give researchers the impression that DFM is

restricted to analyzing the time series data of a single subject. This impression

is not incorrect for the standard application of DFM. As is demonstrated in the

modern literature, DFM and SSM can and should be extended to the multiplesubject time-series design (e.g., Chow, Hamaker, & Allaire, 2009; Chow,

Hamaker, Fujita, & Boker, 2009; Chow, Zu et al., 2011; Hamaker et al., 2005;

Molenaar, 2010b; Nesselroade, 2010; Song & Ferrer, 2012).

The term panel model is often used to refer to models used to analyze timeseries data (of any length from short to long) collected from a group of subjects

(of any sample size). In mediation analysis, panel models under the SEM framework usually take very short lengths (e.g., a handful of repeated measurements)

and large sample sizes (e.g., Cole & Maxwell, 2003).

Methodologically, multiple-subject modeling can be implemented easily by

fitting a qualitatively and quantitatively identical model to multiple subjects,

treating each individual as a group. By qualitatively identical, we mean that each

subject can be characterized by the same dynamic process implied by the

A State Space Modeling Approach to Mediation Analysis

132

specified model; whereas by quantitatively identical, we mean that the parameters of the specified model are equated across different subjects. Then, it

is possible to test parameter invariance across subjects by means of likelihood

ratio tests. This procedure is akin to the standard multiple-group SEM analysis

of interindividual variation in searching for commonality across subjects

(nomothetic lawfulness). Following the new definition of parameter invariance

proposed by Nesselroade, Gerstorf, Hardy, and Ram (2007), we suggest that

parameter invariance tests can be better examined at an appropriate level of

abstraction. Specifically, parameters in the measurement equation can differ

to some arbitrary degree to recognize and isolate idiosyncrasy, while those in

the transition equation can be equated and tested for similarities that reflect

nomothetic lawfulness.

However, as described previously, some empirical examples suggest that

heterogeneity is the rule rather than the exception. In addition to our empirical

example, for instance, economists studying workersâ€™ levels of satisfaction

encountered the problem that each individual anchors his or her scale at a different level (Winkelmann & Winkelmann, 1998). This renders interindividual

comparisons of responses meaningless in a cross-sectional study. Another

example, as discussed by Nesselroade (2010), concerns participantsâ€™ idiosyncratic use of language. Specifically, one debriefed participant from Mitteness

and Nesselroade (1987) reported that she interpreted the term â€˜â€˜anxiousâ€™â€™ to

mean â€˜â€˜eager.â€™â€™

This empirical evidence further supports the general conclusion of heterogeneity across people. On the other hand, Kelderman and Molenaar (2007) provided counterintuitive evidence of the insensitivity of the standard factor

analysis of interindividual variation to the presence of extreme qualitative heterogeneity of the factor loadings in the population of subjects. This proven insensitivity can have serious practical and ethical consequences, yielding individual

assessments and decisions that are biased to unknown degrees (Molenaar,

2008a, 2008b). Given these considerations, great caution should be used if homogeneity is to be assumed.

Another popular modeling strategy for multiple-subject time series data

that allows some degree of heterogeneity is the multilevel modeling framework. Song and Ferrer (2012) recently proposed a random coefficient DFM

(which can be conceptualized as a multilevel SSM) to investigate both intraand interindividual variation. Specifically, they assumed that the parameters

in the transition matrix are drawn from some distribution, so that differences

in dynamics at the appropriate level of abstraction can be accommodated. A

further extension of their model is to allow the parameters in the loading

matrix to be random to capture heterogeneity in the between-subject factor

structure. Such an extension, however, is complex and difficult in terms of

model specification, and estimation may render a model infeasible and/or

indefensible in practice.

Gu et al.

133

4.3. Length and Sampling Frequency

Longer time series are desired for state space analysis and other time-series

techniques in general. Whereas increasing sampling frequency is a way to

obtain more data points for a given time span, the optimal balance of length and

sampling frequency is almost always context-dependent (Brose & Ram, 2012;

Sliwinski & Mogle, 2008). Researchers should consider carefully the relative

benefits of extending the length of a study versus making more observations

within a fixed length. A thorough discussion of the issue is beyond the scope

of this article; see Collins and Graham (2002), Nesselroade (1991), Nesselroade and Boker (1994), Nesselroade and Jones (1991), and Windle and Davies

(1999).

4.4. Use of Latent Variables in Mediation Analysis

According to Cole and Maxwell (2003, question 6: What are the effects of random measurement error?), unmodeled measurement error variance can cause

both under- and overestimation of other model parameters. A natural extension

of the simple mediation model that we have illustrated involves the use of latent

variables, each of which can be indicated by multiple manifest variables. By

explicitly modeling error variance () in the measurement equation, the bias

in parameter estimation can be almost completely resolved, provided that the

model is otherwise correctly specified. Further, psychometric properties of the

measurement instruments (e.g., tau-equivalence) can be evaluated by equating

some loading parameters in the L matrix. As for studying mediation, dynamics

of the process can be examined at the latent level while taking into consideration

the factorial structure of the data.

4.5. Use of Exogenous Variables

Exogenous variables, or fixed external inputs, may enter into the observation equation, the transition equation, or both (e.g., LuÂ¨tkepohl, 2005, p. 613;

Shumway & Stoffer, 2011, p. 320). In this case, the linear Gaussian SSM is

extended to

yt Â¼ tt Ã¾ LtZt Ã¾ txt Ã¾ et; et MVNÃ°0; tÃž

Zt Â¼ at Ã¾ BtZt1 Ã¾ txt Ã¾ zt; zt MVNÃ°0; CtÃž:

Perhaps the most common purpose for including exogenous variables in statistical models is to account for the variance in some random component (e.g.,

Snijders & Bosker, 2011). Molenaar, de Gooijer, and Schmitz (1992) used a discrete time variable in the transition equation to accommodate a linear time trend.

Beyond parameter estimation in time-varying SSM, Molenaar (1994, 2010a,

2010b) presented the theory, methods, and application of optimal control in psychopathological processes. Basically, a feedback function is derived after the

A State Space Modeling Approach to Mediation Analysis

134

SSM parameters are estimated to determine the optimal level of the external

inputs such that it is possible to manipulate the external inputs (e.g., the insulin

dose) by the controller (e.g., the therapist) to guarantee that the outcome variable

(e.g., the blood glucose level of a patient) will be as close as possible to the

desired level. This type of application presents a grand opportunity to develop

more effective personalized treatment, as opposed to the general dosage of the

clinical medication which is not optimal, less effective, and often brings some

undesired side effects.

Moreover, it is possible to collect data with exogenous variables in the

multiple-subject time-series design. If the exogenous variable is time varying

within a person, the multiple-group analysis discussed before might be applied

(e.g., Molenaar, 2010b). If the exogenous variable is time-invariant within a person but varying between persons (i.e., a Level-2 variable), the multilevel (or random coefficient) state space framework can be applied (Song & Ferrer, 2012). In

both scenarios, moderated mediation (or other complex interactions) can be

examined (Card, 2012; Preacher, Rucker, & Hayes, 2007). Future research on

this topic is warranted.

4.6. Missing Data

An attractive feature of SSM is that missing values in the time series data are

easily handled by the KF algorithm. The full-information maximum likelihood

(FIML) procedure varies the dimension of the data vector for observations that

contain one or more missing values. A computationally easier variant of the

FIML procedure is to zero out the missing values and retain the same dimension

of the data vector throughout the observations. No additional effort is needed to

preprocess the missing values as in multiple imputation procedures. Shumway

and Stoffer (2011, chapter 6, subsection 6.4) outline the details of the FIML procedure and its easier variant, and they also describe the necessary modifications

for the expectationâ€“maximization optimization algorithm.

4.7. Software Implementation

Estimating time-varying SSMs is often a difficult task in that extensive programming skills are required to write oneâ€™s own software. EKFIS, a Fortran program developed by Peter Molenaar to implement the Extended KF with Iteration

and Smoothing algorithm has been used to illustrate the examples in several articles and chapters (e.g., Molenaar et al., 2009; Molenaar & Ram, 2009, 2010; Sinclair & Molenaar, 2008). However, the flexibility of time-varying SSM can

become a liability as well, as EKFIS requires the ability to write and compile Fortran code (Molenaar & Ram, 2010). Other authors have implemented the

(extended) KF algorithm using MATLAB and Ox/Ssfpack to estimate timevarying SSMs (e.g., Chow, Hamaker, & Allaire, 2009; Chow, Hamaker, Fujita,

& Boker, 2009; Chow, Zu et al., 2011; Zu, 2008). The programming efforts

Gu et al.

135

involved are still, unfortunately and inevitably, demanding, which is one of the

reasons for the scarcity of modeling work along these lines.

The programming task is relatively easier for time-invariant SSM than for its

time-varying counterpart. In 2011, several articles were published to illustrate different software packages (e.g., EViews, MATLAB, R, SAS, Stata, and several others) in a special volume (Vol. 41) of the Journal of Statistical Software. However,

each software package has certain limitations. Readers are urged to consult this

special volume to get a flavor of the package with which they are the most familiar.

Itis worth notingthat MKFM6, a Fortran program provided by Dolan (2005), and

a SAS/IML program, provided by Gu and Yung (2013), are available to estimate

time-invariant linear Gaussian SSMs. MKFM6 is free and has been used by several

authors (e.g., Chow, Ho, Hamaker, & Dolan, 2010; Hamaker et al., 2005; Zhang

et al., 2008). The programming tasks in this article are implemented in SAS/IML,

which was developed by modifying and extending the SAS/IML code provided

by Gu and Yung. All the programs can be obtained by request from the first author.

Finally, a Bayesian approach to parameter estimation for DFM and SSM is

emerging (e.g., Bhattacharya, Ho, & Purkayastha, 2006; Bhattacharya & Maitra,

2011; Chow, Tang, Yuan, Song, & Zhu, 2011; Song & Ferrer, 2012; Zhang &

Nesselroade, 2007). This newer approach requires estimation methods that are

computationally heavy (e.g., Markov chain Monte Carlo). The rapid development of specialized software programs (e.g., WinBUGS, Mplus, OpenBUGS),

however, makes use of Bayesian methods both manageable and appealing.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research,

authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication

of this article.

Notes

1. Application ofthe state space model (SSM)is not restrictedto single-subjecttime

series data. For example, Chow, Ho, Hamaker, and Dolan (2010) provide an

example in which SSM is applied to cross-sectional data. Moreover, we will

discuss the extension of multiple-subject time series data in Subsection 4.1.

2. Note that our definition of stationarity is consistent with that used in the timeseries literature, but not strictly parallel to the same term described in the mediation literature (Cole & Maxwell, 2003; Kenny, 1979).

3. More discussion is provided in Subsection 4.4.

4. Proof of the Kalman filter (KF) algorithm can be found, for instance, in LuÂ¨tkepohl

(2005, pp. 630â€“631) or Shumway and Stoffer (2011, pp. 326â€“327).

A State Space Modeling Approach to Mediation Analysis

136

References

Albert, J. M. (2008). Mediation analysis via potential outcomes models. Statistics in Medicine, 27, 1282â€“1304.

Allport, G. W. (1937). Personality: A psychological interpretation. New York, NY: Holt,

Rinehart, and Winston.

Aoki, M. (1987). State space modeling of time series. Berlin, Germany: Springer.

Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in

social psychological research: Conceptual, strategic, and statistical considerations.

Journal of Personality & Social Psychology, 51, 1173â€“1182.

Bhattacharya, S., Ho, M.-H. R., & Purkayastha, S. (2006). A Bayesian approach to

modeling dynamic effective connectivity with fMRI data. NeuroImage, 30,

794â€“812.

Bhattacharya, S., & Maitra, R. (2011). A nonstationary nonparametric Bayesian approach

to dynamically modeling effective connectivity in functional magnetic resonance imaging experiments. The Annals of Applied Statistics, 5, 1183â€“1206.

Bollen, K. A., & Stine, R. (1990). Direct and indirect effects: Classical and bootstrap estimates of variability. Sociological Methodology, 20, 115â€“140.

Brose, A., & Ram, N. (2012). Within-person factor analysis: Modeling how the individual fluctuates and changes across time. In M. R. Mehl & T. S. Conner (Eds.),

Handbook of research methods for studying daily life (pp. 43â€“61). New York, NY:

The Guilford Press.

Browne, M. W., & Nesselroade, J. R. (2005). Representing psychological processes with

dynamic factor models: Some promising uses and extensions of autoregressive moving

average time series models. In A. Maydeu-Olivares & J. J. McArdle (Eds.), Contemporary psychometrics (pp. 415â€“452). New York, NY: Routledge.

Card, N. A. (2012). Multilevel mediational analysis in the study of daily lives. In M. R.

Mehl & T. S. Conner (Eds.), Handbook of research methods for studying daily life

(pp. 479â€“494). New York, NY: The Guilford Press.

Cattell, R. B. (1963). The structuring of change by P-technique and incremental R-technique. In C. W. Harris (Ed.), Problems in measuring change (pp. 167â€“198). Madison:

University of Wisconsin Press.

Cattell, R. B., Cattell, A. K. S., & Rhymer, R. M. (1947). P-technique demonstrated in

determining psychophysical source traits in a normal individual. Psychometrika, 12,

267â€“288.

Cheong, J., MacKinnon, D. P., & Khoo, S. T. (2003). Investigation of mediational processes using parallel process latent growth curve modeling. Structural Equation

Modeling, 10, 238â€“262.

Chow, S.-M., Hamaker, E. L., & Allaire, J. C. (2009). Using innovative outliers to detect

discrete shifts in dynamics in group-based state-space models. Multivariate Behavioral

Research, 44, 465â€“496.

Chow, S.-M., Hamaker, E. L., Fujita, F., & Boker, S. M. (2009). Representing timevarying cyclic dynamics using multiple-subject state-space models. British Journal

of Mathematical and Statistical Psychology, 62, 683â€“716.

Chow, S.-M., Ho, M.-H. R., Hamaker, E. L., & Dolan, C. V. (2010). Equivalence and differences between structural equation modeling and state-space modeling techniques.

Structural Equation Modeling, 17, 303â€“332.

Gu et al.

137

Chow, S.-M., Nesselroade, J. R., Shifren, K., & McArdle, J. J. (2004). Dynamic structure

of emotions among individuals with Parkinsonâ€™s disease. Structural Equation Modeling,

11, 560â€“582.

Chow, S.-M., Tang, N., Yuan, Y., Song, X., & Zhu, H. (2011). Bayesian estimation of

semiparametric nonlinear dynamic factor analysis models using the Dirichlet process

prior. British Journal of Mathematical and Statistical Psychology, 64, 69â€“106.

Chow, S.-M., Zu, J., Shifren, K., & Zhang, G. (2011). Dynamic factor analysis models

with time-varying parameters. Multivariate Behavioral Research, 46, 303â€“339.

Cole, D. A., & Maxwell, S. E. (2003). Testing meditational models with longitudinal data:

Questions and tips in the use of structural equation modeling. Journal of Abnormal

Psychology, 112, 558â€“577.

Collins, M. L., & Graham, J. W. (2002). The effect of timing and spacing of observations

in longitudinal studies of tobacco and other drug use: Temporal design considerations.

Drug and Alcohol Dependence, 68, 85â€“96.

Collins, M. L., Graham, J. W., & Flaherty, B. P. (1998). An alternative framework for

defining mediation. Multivariate Behavioral Research, 33, 295â€“312.

Commandeur, J. J. F., & Koopman, S. J. (2007). An introduction to state space time series

analysis. New York, NY: Oxford University Press.

Dolan, C. V. (2005). MKFM6: Multi-group, multi-subject stationary time series modeling based

on the Kalman filter. Retrieved December 1, 2004, from http://users.fmg.uva.nl/cdoman/

Durbin, J., & Koopman, S. J. (2001). Time series analysis by state space methods.

New York, NY: Oxford University Press.

Ferrer, E., & Nesselroade, J. R. (2003). Modeling affective processes in dyadic relations

via dynamic factor analysis. Emotion, 3, 344â€“360.

Ferrer, E., Steele, J., & Hsieh, F. (2012). Analyzing dynamics of affective dyadic interactions using patterns of intra- and inter-individual variability. Multivariate Behavioral

Research, 47, 136â€“171.

Ferrer, E., & Widaman, K. F. (2008). Dynamic factor analysis of dyadic affective processes with inter-group differences. In N. A. Card, J. P. Selig, & T. D. Little (Eds.),

Modeling dyadic and interdependent data in the developmental and behavioral

sciences (pp. 107â€“137). Hillsdale, NJ: Psychology Press.

Gollob, H. F., & Reichardt, C. S. (1991). Interpreting and estimating indirect effects

assuming time lags really matter. In L. M. Collins & J. L. Horn (Eds.), Best methods

for the analysis of change: Recent advances, unanswered questions, future directions

(pp. 243â€“259). Washington, DC: American Psychological Association.

Gu, F., & Yung, Y.-F. (2013). A SAS/IML program using the Kalman filter for estimating

state space model. Behavior Research Methods, 45, 38â€“53.

Hamagami, F., & McArdle, J. J. (2007). Dynamic extensions of latent difference score

models. In S. M. Boker & M. L. Wegner (Eds.), Data analytic techniques for dynamical systems (pp. 47â€“85). Mahwah, NJ: Lawrence Erlbaum.

Hamaker, E. L. (2012). Why researchers should think â€˜â€˜within-personâ€™â€™: A paradigmatic

rationale. In M. R. Mehl & T. S. Conner (Eds.), Handbook of research methods for

studying daily life (pp. 43â€“61). New York, NY: The Guilford Press.

Hamaker, E. L., Dolan, C. V., & Molenaar, P. C. M. (2005). Statistical modeling of the

individual: Rational and application of multivariate stationary time series analysis.

Multivariate Behavioral Research, 40, 207â€“233.

A State Space Modeling Approach to Mediation Analysis

138

Harvey, A. C. (1989). Forecasting, structural time series models and the Kalman filter.

Princeton, NJ: Princeton University Press.

Hayes, A. F. (2009). Beyond Baron and Kenny: Statistical mediation analysis in the new

millennium. Communication Monographs, 76, 408â€“420.

Hershberger, S. L., Corneal, S. E., & Molenaar, P. C. M. (1994). Dynamic factor analysis:

An application to emotional response patterns underlying daughter/father and stepdaughter/stepfather relationships. Structural Equation Modeling, 2, 31â€“52.

Ho, M.-H. R., Ombao, H., & Shumway, R. (2005). A state-space approach to modelling

brain dynamics. Statistica Sinica, 15, 407â€“425.

Ho, M.-H. R., Shumway, R., & Ombao, H. (2006). State-space approach to modeling

dynamic processes: Applications in biological and social sciences. In T. A. Walls &

J. L. Schafer (Eds.), Models for intensive longitudinal data (pp. 148â€“170). New York,

NY: Oxford University Press.

Imai, K., Keele, L., & Tingley, D. (2010). A general approach to causal mediation analysis.

Psychological Methods, 15, 309â€“334.

Judd, C. M., & Kenny, D. A. (1981). Process analysis: Estimating mediation in treatment

evaluations. Evaluation Review, 5, 602â€“619.

Judd, C. M., Kenny, D. A., & McClelland, G. H. (2001). Estimating and testing mediation

and moderation in within-participant designs. Psychological Methods, 6, 115â€“134.

Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Transactions of the ASME-Journal of Basic Engineering (Series D), 82, 35â€“45.

Kelderman, H., & Molenaar, P. C. M. (2007). The effect of individual differences in factor

loadings on the standard factor model. Multivariate Behavioral Research, 42, 435â€“456.

Kenny, D. A. (1979). Correlation and causality. New York, NY: John Wiley.

Lamiell, J. T. (1981). Toward an idiothetic psychology of personality. American Psychologist, 36, 276â€“289.

Lamiell, J. T. (1988). Once more into the breach: Why individual differences research

cannot advance personality theory. Paper presented at the annual meeting of the American Psychological Association, August, Atlanta, GA.

LuÂ¨tkepohl, H. (2005). New introduction to multiple time series analysis. Berlin, Germany:

Springer.

MacCallum, R. C., & Ashby, G. F. (1986). Relationships between linear systems theory

and covariance structure modeling. Journal of Mathematical Psychology, 30, 1â€“27.

MacKinnon, D. P. (2008). Introduction to statistical mediation analysis. New York, NY:

Lawrence Erlbaum.

MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confidence limits for the

indirect effect: Distribution of the product and resampling methods. Multivariate

Behavioral Research, 39, 99â€“128.

Maxwell, S. E., & Cole, D. A. (2007). Bias in cross-sectional analyses of longitudinal

mediation. Psychological Methods, 12, 23â€“44.

Maxwell, S. E., Cole, D. A., & Mitchell, M. A. (2011). Bias in cross-sectional analyses of

longitudinal mediation: Partial and complete mediation under an autoregressive

model. Multivariate Behavioral Research, 46, 816â€“841.

McArdle, J. J. (2001). A latent difference score approach to longitudinal dynamic structural

analyses. In R. Cudeck, S. du Toit, & D. Sorbom (Eds.), Structural equation modeling:

Present and future (pp. 341â€“380). Lincolnwood, IL: Scientific Software International.

Gu et al.

139

McArdle, J. J., & Nesselroade, J. R. (1994). Using multivariate data to structure developmental change. In S. H. Cohen & H. W. Reese (Eds.), Life-span developmental psychology: Methodological contributions (pp. 223â€“267). Hillsdale, NJ: Lawrence Erlbaum.

Mitteness, L. S., & Nesselroade, J. R. (1987). Attachment in adulthood: Longitudinal

investigation of mother-daughter affective interdependencies by p-technique factor

analysis. The Southern Psychologist, 3, 37â€“44.

Molenaar, P. C. M. (1985). A dynamic factor model for the analysis of multivariate time

series. Psychometrika, 50, 181â€“202.

Molenaar, P. C. M. (1994). Dynamic latent variable models in developmental psychology.

In A. van Eye & C. C. Clogg (Eds.), Latent variable analysis (pp. 155â€“180). Thousand

Oaks, CA: Sage.

Molenaar, P. C. M. (2004). A manifesto on psychology as idiographic science: Bring the

person back into scientific psychology, this time forever. Measurement, 2, 201â€“218.

Molenaar, P. C. M. (2008a). On the implications of the classical ergodic theorems: Analysis of developmental processes has to focus on intra-individual variation. Developmental Psychobiology, 50, 60â€“69.

Molenaar, P. C. M. (2008b). Consequences of the ergodic theorems for classical test

theory, factor analysis and the analysis of developmental processes. In S. M. Hofer

& D. F. Alwin (Eds.), Handbook of cognitive aging: Interdisciplinary perspectives. (pp. 90â€“104). Thousand Oaks, CA: Sage.

Molenaar, P. C. M. (2010a). Note on optimization of individual psychotherapeutic

processes. Journal of Mathematical Psychology, 54, 208â€“213.

Molenaar, P. C. M. (2010b). Testing all six person-oriented principles in dynamic factor

analysis. Development and Psychopathology, 22, 255â€“259.

Molenaar, P. C. M., Boomsma, D. I., & Dolan, C. V. (1993). A third source of developmental differences. Behavior Genetics, 23, 519â€“524.

Molenaar, P. C. M., & Campbell, C. G. (2009). The new person-specific paradigm in

psychology. Current Directions in Psychological Science, 18, 112â€“117.

Molenaar, P. C. M., de Gooijer, J. G., & Schmitz, B. (1992). Dynamic factor analysis of

nonstationary multivariate time series. Psychometrika, 57, 333â€“349.

Molenaar, P. C. M., & Ram, N. (2009). Advances in dynamic factor analysis of psychological processes. In J. Valsiner, P. C. M. Molenaar, M. C. D. P. Lyra, & N. Chaudhary

(Eds.), Dynamic process methodology in the social and developmental sciences (pp.

255â€“268). Dordrecht, the Netherlands: Springer Science.

Molenaar, P. C. M., & Ram, N. (2010). Dynamic modeling and optimal control of intraindividual variation: A computational paradigm for nonergodic psychological processes.

In S.-M. Chow, E. Ferrer, & F. Hsieh (Eds.), Statistical methods for modeling human

dynamics: An interdisciplinary dialogue (pp. 13â€“37). New York, NY: Routledge.

Molenaar, P. C. M., Sinclair, K. O., Rovine, M. J., Ram, N., & Corneal, S. E. (2009). Analyzing developmental processes on an individual level using nonstationary time series

modeling. Developmental Psychology, 45, 260â€“271.

Nesselroade, J. R. (1991). Interindividual differences in intraindividual change. In L. M.

Collins & J. L. Horn (Eds.), Best methods for the analysis of change (pp. 92â€“105).

Washington, DC: American Psychological Association.

Nesselroade, J. R. (2010). On an emerging third discipline of scientific psychology. In P.

C. M. Molenaar & K. M. Newell (Eds.), Individual pathway of change: Statistical

A State Space Modeling Approach to Mediation Analysis

140

models for analyzing learning and development (pp. 209â€“218). Washington, DC:

American Psychological Association.

Nesselroade, J. R., & Boker, S. M. (1994). Assessing constancy and change. In T. F.

Heatherton & J. L. Weinberger (Eds.), Can personality change? (pp. 121â€“147).

Washington, DC: American Psychological Association.

Nesselroade, J. R., Gerstorf, D., Hardy, S. A., & Ram, N. (2007). Idiographic filters for

psychological constructs. Measurement, 5, 217â€“235.

Nesselroade, J. R., & Jones, C. J. (1991). Multi-model selection effects in the study of

adult development: A perspective on multivariate, replicated, single-subject, repeated

measures designs. Experimental Aging Research, 17, 21â€“27.

Nesselroade, J. R., McArdle, J. J., Aggen, S. H., & Meyers, J. M. (2002). Dynamic factor

analysis models for representing process in multivariate time-series. In D. M. Moskowitz

& S. L. Hershberger (Eds.), Modeling intraindividual variability with repeated measures

data: Advances and techniques (pp. 235â€“265). Mahwah, NJ: Lawrence Erlbaum.

Nesselroade, J. R., & Molenaar, P. C. M. (1999). Pooling lagged covariance structures based

on short, multivariate time series for dynamic factor analysis. In R. H. Hoyle (Ed.), Statistical strategies for small sample research (pp. 223â€“250). Thousand Oaks, CA: Sage.

Otter, P. (1986). Dynamic structure systems under indirect observation: Identifiability and

estimation aspects from a system theoretic perspective. Psychometrika, 51, 415â€“428.

Pearl, J. (2010). The foundations of causal inference. Sociological Methodology, 40, 75â€“149.

Pearl, J. (2012). The causal mediation formulaâ€”A guide to the assessment of pathways

and mechanisms. Prevention Science, 13, 426â€“436.

Preacher, K. J., & Hayes, A. F. (2004). SPSS and SAS procedures for estimating indirect

effects in simple mediation models. Behavior Research Methods, 36, 717â€“731.

Preacher, K. J., & Hayes, A. F. (2008a). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research

Methods, 40, 879â€“891.

Preacher, K. J., & Hayes, A. F. (2008b). Contemporary approaches to assessing mediation

in communication research. In A. F. Hayes, M. D. Slater, & L. B. Snyder (Eds.), The

Sage sourcebook of advanced data analysis methods for communication research (pp.

13â€“54). Thousand Oaks, CA: Sage.

Preacher, K. J., Rucker, D. D., & Hayes, A. F. (2007). Addressing moderated mediation

hypothesis: Theory, methods, and prescriptions. Multivariate Behavioral Research, 42,

185â€“227.

Roe, R. (2012). What is wrong with mediators and moderators? The European Health

Psychologist, 14, 4â€“10.

Rosenzweig, S. (1958). The place of the individual and of idiodynamics in psychology:

A dialogue. Journal of Individual Psychology, 14, 3â€“20.

Sbarra, D. A., & Ferrer, E. (2006). The structure and process of emotional experience following nonmarital relationship dissolution: Dynamic factor analyses of love, anger,

and sadness. Emotion, 2, 224â€“238.

Schmitz, B. (2006). Advantages of studying processes in educational research. Learning

and Instruction, 16, 433â€“449.

Schweppe, F. (1965). Evaluation of likelihood functions for Gaussian signals. IEEE

Transactions on Information Theory, 11, 61â€“70.

Gu et al.

141

Selig, J. P., & Preacher, K. J. (2009). Mediation models for longitudinal data in developmental research. Research in Human Development, 6, 144â€“164.

Shifren, K., Hooker, K., Wood, P., & Nesselroade, J. R. (1997). Structure and variation of

mood in individuals with Parkinsonâ€™s disease: A dynamic factor analysis. Psychology

and Aging, 12, 328â€“339.

Shrout, P. E., & Bolger, N. (2002). Mediation in experimental and nonexperimental studies: New procedures and recommendations. Psychological Methods, 7, 422â€“445.

Shumway, R. H., & Stoffer, D. S. (2011). Time series analysis and its applications: With R

examples. New York, NY: Springer.

Sinclair, K. O., & Molenaar, P. C. M. (2008). Optimal control of psychological processes: A

new computational paradigm.Bulletin de la Societe des Sciences Medicales Luxembourg,

13â€“33.

Sliwinski, M., & Mogle, J. (2008). Time-based and process-based approaches to analysis

of longitudinal data. In S. M. Hofer & D. F. Alwin (Eds.), Handbook of cognitive

aging: Interdisciplinary perspectives (pp. 477â€“491). Thousand Oaks, CA: Sage.

Snijders, T. A. B., & Bosker, R. J. (2011). Multilevel analysis: An introduction to basic

and advanced multilevel modeling. Thousand Oaks, CA: Sage.

Song, H., & Ferrer, E. (2009). State-space modeling of dynamic psychological processes

via the Kalman smoother algorithm: Rationale, finite sample properties, and applications. Structural Equation Modeling, 16, 338â€“363.

Song, H., & Ferrer, E. (2012). Bayesian estimation of random coefficient dynamic factor

models. Multivariate Behavioral Research, 47, 26â€“60.

Stoffer, D. S., & Wall, K. D. (1991). Bootstrapping state-space models: Gaussian maximum likelihood estimation and the Kalman filter. Journal of the American Statistical

Association, 86, 1024â€“1033.

Stoffer, D. S., & Wall, K. D. (2004). Resampling in state space models. In A. Harvey, S. J.

Koopman, & N. Shephar (Eds.), State space and unobserved component models:

Theory and applications (pp. 171â€“202). New York, NY: Cambridge University Press.

Ten Have, T. R., & Joffe, M. M. (2010). A review of causal estimation of effects in

mediation analyses. Statistical Methods in Medical Research, 21, 77â€“107.

VanderWeele, T. J., & Vansteelandt, S. (2009). Conceptual issues concerning mediation,

interventions and composition. Statistics and Its Interface, 2, 457â€“468.

van Kampen, V. (2000). Idiographic complexity and the common personality dimensions

of insensitivity, extraversion, neuroticism, and orderliness. European Journal of

Personality, 14, 217â€“243.

Watson, D., Clark, L. A., & Tellegen, A. (1988). Development and validation of brief

measures of positive and negative affect: The PANAS scales. Journal of Personality

and Social Psychology, 54, 1063â€“1070.

Windle, M., & Davies, P. T. (1999). Developmental research and theory. In K. E. Leonard

& H. T. Blane (Eds.), Psychological theories of drinking and alcoholism (2nd ed., pp.

164â€“202). New York, NY: Guilford.

Winkelmann, L., & Winkelmann, R. (1998). Why are the unemployed so unhappy? Evidence from panel data. Econometrica, 65, 1â€“15.

Wood, P., & Brown, D. (1994). The study of intraindividual differences by means of

dynamic factor models: Rationale, implementation, and interpretation. Psychological

Bulletin, 116, 166â€“186.

A State Space Modeling Approach to Mediation Analysis

142

Yung, Y.-F., & Chan, W. (1999). Statistical analyses using bootstrapping: Concepts and

implementation. In R. H. Hoyle (Ed.), Statistical strategies for small sample research

(pp. 81â€“105). Thousand Oaks, CA: Sage.

Zevon, M., & Tellegen, A. (1982). The structure of mood change: Idiographic/nomothetic

analysis. Journal of Personality and Social Psychology, 43, 111â€“122.

Zhang, G., & Browne, M. W. (2006). Bootstrap standard error estimates in dynamic factor

analysis. Multivariate Behavioral Research, 45, 453â€“482.

Zhang, G., & Chow, S.-M. (2010). Standard error estimation in stationary multivariate time

series models using residual-based bootstrap procedures. In P. C. M. Molenaar & K. M.

Newell (Eds.), Individual pathway of change: Statistical models for analyzing learning

and development (pp. 169â€“182). Washington, DC: American Psychological Association.

Zhang, Z., Hamaker, E. L., & Nesselroade, J. R. (2008). Comparisons of four methods for

estimating a dynamic factor model. Structural Equation Modeling, 15, 377â€“402.

Zhang, Z., & Nesselroade, J. R. (2007). Bayesian estimation of categorical dynamic factor

models. Multivariate Behavioral Research, 42, 729â€“756.

Zu, J. (2008). Examining nonlinear changes of coefficients in time-varying dynamic factor

models (Unpublished masterâ€™s thesis). University of Notre Dame, South Bend, IN.

Authors

FEI GU is an assistant professor at the Department of Psychology, McGill University,

1205 Doctor Penfield Avenue, Montreal, QC H3A 1B1, Canada; e-mail: [email protected]

gill.ca. He is interested in innovative applications of state space methods. Specifically,

his work extends the state space model to analyze cross-sectional, multilevel, and

(intensive) longitudinal data from multiple subjects and unifies many existing statistical

models into the state space approach.

KRISTOPHER J. PREACHER is an assistant professor at the Psychology & Human

Development,Vanderbilt University, PMB 552, 230 Appleton Place, Nashville, TN

37203-5721; e-mail: [email protected] His research focuses on the use

of latent variable analysis and multilevel modeling to analyze longitudinal and

correlational data. Other interests include developing techniques to address mediation,

moderation, and model evaluation and selection.

EMILIO FERRER is a professor at the Department of Psychology University of

California, Davis, One Shields Avenue, Davis, CA 95616-8686; e-mail: [email protected]

vis.edu. He is interested in methods to analyze change and intraindividual variability, in

particular latent growth analysis and linear and nonlinear dynamical systems. His

current research in this area involves techniques to model dyadic interactions and developmental processes underlying fluid reasoning.

Manuscript received July 10, 2013

Revision received November 8, 2013

Accepted January 21, 2014

Gu et al.

143

Any academic writer who wishes to join our team of professional writers must possess all the following qualities:

- They must have a degree in a particular field of science.
- They must show a high level of academic writing skills.
- They must have a high level of English proficiency.
- They must be highly punctual and responsible.
- They must have great interpersonal skills.

To write an exemplary academic paper, you must have good critical thinking skills, possess the proper knowledge of the discipline, and be knowledgeable about applying an academic writing style.

As such we have a rigorous recruitment process. We only collaborate with professional academic writers. We believe in offering the highest quality academic writing services. Our writers pass various grammar and academic writing tests. They have to provide documents about their personal information and credentials to prove their level of expertise.

As a result, our clients receive papers that are thoroughly-researched, properly cited, and written within academic standards. We are proud that any academic writer from our writer's team can complete the paper at a high standard.

We work with the student’s budget because we know that students are usually on a budget majority of the time. We do not compromise on quality because of low prices. On the contrary, we love to foster a good relationship with our clients. That is why we charge our clients reasonable prices, and we are willing to negotiate and work with their budget.

It may be quite disturbing to decide whether to hire an academic writing company. We care about our client’s privacy and confidentiality. We never disclose your information to third parties. We never publish your paper online. You can use our academic writing service without any fear or anxiety.

Many students struggle with writing academic papers. Some barely have time to do their assignments because of their job and family responsibilities. Others have difficulty applying critical thinking skills or meeting time or assignment requirements. Whatever the reason is, you can always have time to do the things you love and other important things. All you need is a reliable and quality academic writing service. Unfortunately, even if you strongly desire to write the paper yourself, you sometimes face unexpected challenges. As we all know, life is unpredictable! Your teacher may be unconcerned about helping students and may not answer your questions. The good news is that there is a way out! You can hire an online academic writer to help you with your assignments. All you need to do is stipulate your paper requirements in the order form, and you can spend your time as you like.

Our company commits towards delivering high-quality custom papers to our clients. We seek to offer reliable essay writing services to our customers in various subject areas. Our customers are very valuable to us. As such, we commit to ensuring that they derive the utmost satisfaction from the essays we deliver. We have a mission to promote our clients' educational and professional lives by providing high-quality essays for their use. We also have a mission to offer a convenient essay writing system where our customers can easily order and pay for the services. We value quality and professionalism in our company.

We write papers on any subject area, and we also write various types of papers for various purposes. We have a team of able writers who are eager to help our customers with writing services of exceptional quality. We offer custom writing services for customers across the globe and charge affordable prices for our services. We are the best essay writing company offering personalized services to all our customers. We ensure our customers receive maximum satisfaction from the essays we deliver. Our website is the place to be if you are seeking high-quality essays.

By using our academic writing service, we match your paper subject with a writer with a degree in the subject. The writer is able to apply their skills, knowledge, and expertise to the paper. You receive an original, unique, well-research paper, properly cited and formatted. As such, you are assured of a high-quality paper.

Truth be told, writing assignments can be stressful and difficult for any student. There is no shame in getting academic writing help. When you search the term “get academic writing help” there are numerous sites that pop up on the results and our website is among them. So, why is it a great idea to choose us?

During your course, your instructor will assign various types of homework. Our academic writers can prepare essays, presentations, speeches, case studies, research papers, dissertations, thesis papers, and more. Our writer’s department is capable of tackling any assignment of any complexity easily. All you need is to give us detailed instructions to help our experts understand the task.After doing so, you can rest assured that everything is in control, and we will deliver a paper of unmatchable quality.

Are you a college or university student pursuing your Bachelor’s , Masters, PhD, or Doctorate degree? Sometimes juggling schoolwork with work, family and hobbies can seem like a daunting task. You have to sacrifice one or the other. The sole purpose of our website is to alleviate your academic burdens. We ensure that you do not fail in your classes and you get good grades consistently. We understand that there is a need for academic help.

We acknowledge that our clients are not dumb or lazy but only need academic life need help in order to live a balanced life and make ends meet. We make it our core priority to ensure that all assignments are done and submitted before the stipulated deadlines. All our writers are graduates. They are competent in handling the clients’ assignments. We step in to help you with any and all of your assignments. Our assignment help service ensures that you never miss a grade or deadline.

Students barely have time to read. We got you! Have your literature essay or book review written without having the hassle of reading the book. You can get your literature paper custom-written for you by our literature specialists.

Do you struggle with finance? No need to torture yourself if finance is not your cup of tea. You can order your finance paper from our academic writing service and get 100% original work from competent finance experts.

Computer science is a tough subject. Fortunately, our computer science experts are up to the match. No need to stress and have sleepless nights. Our academic writers will tackle all your computer science assignments and deliver them on time. Let us handle all your python, java, ruby, JavaScript, php , C+ assignments!

While psychology may be an interesting subject, you may lack sufficient time to handle your assignments. Don’t despair; by using our academic writing service, you can be assured of perfect grades. Moreover, your grades will be consistent.

Engineering is quite a demanding subject. Students face a lot of pressure and barely have enough time to do what they love to do. Our academic writing service got you covered! Our engineering specialists follow the paper instructions and ensure timely delivery of the paper.

In the nursing course, you may have difficulties with literature reviews, annotated bibliographies, critical essays, and other assignments. Our nursing assignment writers will offer you professional nursing paper help at low prices.

Truth be told, sociology papers can be quite exhausting. Our academic writing service relieves you of fatigue, pressure, and stress. You can relax and have peace of mind as our academic writers handle your sociology assignment.

We take pride in having some of the best business writers in the industry. Our business writers have a lot of experience in the field. They are reliable, and you can be assured of a high-grade paper. They are able to handle business papers of any subject, length, deadline, and difficulty!

We boast of having some of the most experienced statistics experts in the industry. Our statistics experts have diverse skills, expertise, and knowledge to handle any kind of assignment. They have access to all kinds of software to get your assignment done.

Writing a law essay may prove to be an insurmountable obstacle especially when you need to know the peculiarities of the legislative framework. Take advantage of our top-notch law specialists and get superb grades and 100% satisfaction.

HIGH SCHOOL

from
$10 page

COLLEGE

from
$13 page

UNIVERSITY

from
$14 page

Our prices depend on the urgency of your assignment, your academic level, the course subject, and the length of the assignment. Basically, more complex assignments will cost more than simpler ones. The level of expertise is also a major determinant of the price of your assignment.

If you need professional help with completing any kind of homework, is the right place to get it. Whether you are looking for essay, coursework, research, or term paper help, or with any other assignments, it is no problem for us. At our cheap essay writing service, you can be sure to get credible academic aid for a reasonable price, as the name of our website suggests. For years, we have been providing online custom writing assistance to students from countries all over the world, including the US, the UK, Australia, Canada, Italy, New Zealand, China, and Japan.

Our cheap essay writing service has already gained a positive reputation in this business field. Understandably so, since all custom papers produced by our academic writers are individually crafted from scratch and written according to all your instructions and requirements. We offer APA, MLA, or a Chicago style paper in almost 70 disciplines. Here, you can get quality custom essays, as well as a dissertation, a research paper, or term papers for sale. Any paper will be written on time for a cheap price.

It might seem impossible to you that all custom-written essays, research papers, speeches, book reviews, and other custom task completed by our writers are both of high quality and cheap. It is surprising, but we do have some tricks to lower prices without hindering quality.

To start using our services, it’s enough to place a request like “I need a writer to do my assignment” or “Please, write an essay for me.” We have a convenient order form, which you can complete within minutes and pay for the order via a secure payment system. The support team will view it after the order form and payment is complete and then they will find an academic writer who matches your order description perfectly. Once you submit your instructions, while your order is in progress and even after its completion, our support team will monitor it to provide you with timely assistance.

Hiring good writers is one of the key points in providing high-quality services. That’s why we have entry tests for all applicants who want to work for us. We try to make sure all writers working for us are professionals, so when you purchase custom-written papers, they are of high quality and non-plagiarized.

Our cheap essay writing service employs only writers who have outstanding writing skills. The quality of all custom papers written by our team is important to us; that is why we are so attentive to the application process and employ only those writers who can produce great essays and other kinds of written assignments. All our writers are graduates. They are competent in handling the clients’ assignments. We step in to help you with any and all of your assignments. Our assignment help service ensures that you never miss a grade or deadline.

All our cheap essays are customized to meet your requirements and written from scratch. Our writers have a lot of experience with academic papers and know how to write them without plagiarism. Moreover, at our academic service, we have our own plagiarism-detection software which is designed to find similarities between completed papers and online sources. You can be sure that our custom-written papers are original and properly cited.

Our essay writing service has a 0% plagiarism tolerance. We are well aware of the dangers of plagiarism. Plagiarism is academic suicide. Our essay writing service ensures that all papers are original. We do not sell pre-written papers. All papers are written from scratch as per the instructions. We pass our papers through powerful anti-plagiarism software such as SafeAssign and TurnItIn.

Our cheap essay writing service tries to always be at its best performance level, so each customer who pays money for paper writing can be sure that he or she will get what is wanted. On the off chance that you don’t like your order, you can request a refund and we will return the money according to our money-back guarantee.

There can be a number of reasons why you might not like your order. If we honestly don’t meet your expectations, we will issue a refund. You can also request a free revision, if there are only slight inconsistencies in your order. Your writer will make the necessary amendments free of charge. You can find out more information by visiting our revision policy and money-back guarantee pages, or by contacting our support team via online chat or phone.

We know how important any deadline is to you; that’s why everyone in our company has their tasks and perform them promptly to provide you with the required assistance on time. We even have an urgent delivery option for short essays, term papers, or research papers needed within 8 to 24 hours.

We appreciate that you have chosen our cheap essay service, and will provide you with high-quality and low-cost custom essays, research papers, term papers, speeches, book reports, and other academic assignments for sale. We beat all deadlines. We can also handle urgent orders with deadlines as short as 1 hour. Our urgent paper writing service does not compromise on quality due to the short deadline. On the contrary, our essay writers have a lot of experience which comes in handy in such situations.

We provide affordable writing services for students around the world. That’s why we work without a break to help you at any time, wherever you are located. Contact us for cheap writing assistance. Our impeccable customer support team will answer all your questions and help you out with any issues.

Proceed to order page
The price is based on these factors:

Academic level

Number of pages

Urgency

Basic features

- Free title page and bibliography
- Unlimited revisions
- Plagiarism-free guarantee
- Money-back guarantee
- 24/7 support

On-demand options

- Writer’s samples
- Part-by-part delivery
- Overnight delivery
- Copies of used sources
- Expert Proofreading

Paper format

- 275 words per page
- 12 pt Arial/Times New Roman
- Double line spacing
- Any citation style (APA, MLA, Chicago/Turabian, Harvard)

Delivering a high-quality product at a reasonable price is not enough anymore.

That’s why we have developed 5 beneficial guarantees that will make your experience with our service enjoyable, easy, and safe.

You have to be 100% sure of the quality of your product to give a money-back guarantee. This describes us perfectly. Make sure that this guarantee is totally transparent.

Read moreEach paper is composed from scratch, according to your instructions. It is then checked by our plagiarism-detection software. There is no gap where plagiarism could squeeze in.

Read moreThanks to our free revisions, there is no way for you to be unsatisfied. We will work on your paper until you are completely happy with the result.

Read moreYour email is safe, as we store it according to international data protection rules. Your bank details are secure, as we use only reliable payment systems.

Read moreBy sending us your money, you buy the service we provide. Check out our terms and conditions if you prefer business talks to be laid out in official language.

Read more