HHS Public Access

Author manuscript
Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
Author Manuscript

Published in final edited form as:

Clin Psychol Rev. 2020 March ; 76: 101824. doi:10.1016/j.cpr.2020.101824.

Beyond Linear Mediation: Toward A Dynamic Network Approach

to Study Treatment Processes
Stefan G. Hofmann1,*, Joshua E. Curtiss1, Steven C. Hayes2
1Boston University
2University of Nevada, Reno
Author Manuscript

Abstract
Few clinical scientists would disagree that more research is needed on the underlying mechanisms
and processes of change in psychological therapies. In the dominant current approach, processes
of change are studied through mediation. The study of mediation has been largely structured
around a distinction between moderation and mediation first popularized by Baron and Kenny’s
(1986) seminal article, which is based on a nomothetic and cross-sectional framework. In this
article, we argue that this approach is unable to adequately address change processes in
psychological therapies, because it falsely assumes that treatment change is a linear,
unidirectional, pauci-variate process and that the statistical assumptions are met to study processes
of change in an individual using a nomothetic approach. In contrast, we propose that treatment is a
dynamic process involving numerous variables that may form bi-directional and complex
relationships that differ between individuals. Such relationships can best be studied using an
Author Manuscript

individual dynamic network approach connected to nomothetic generalization methods that are
based on a firm idiographic foundation. We argue that our proposal is available, viable, and can
readily be integrated into existing research strategies. We further argue that adopting an individual
dynamic network approach combined with experimental analyses will accelerate the study of
treatment change processes, which is necessary as the field of evidence-based care moves toward a
process-based model. We encourage future research to gather empirical evidence to examine this
approach.

Keywords
Therapy; process; Treatment; Mediation
Author Manuscript

The field of psychological intervention has long included processes of change on the list of
key questions to be addressed by treatment research. Gordon Paul’s famous description of
the core question for treatment researchers pointed to the study of change processes in its
last six words: “What treatment, by whom, is most effective for this individual with that

*
Corresponding author. [email protected].
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our
customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of
the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered
which could affect the content, and all legal disclaimers that apply to the journal pertain.
 Hofmann et al. Page 2

specific problem, under which set of circumstances, and how does it come about?” (Paul,
Author Manuscript

1969, p. 44).

In the nearly 50 years since, the reliable production of improved outcomes has been
emphasized far more than processes of change. Although the American Psychological
Association (APA) created Task Forces for both, evidence-based therapies and treatment
processes, considerably more attention was paid to the former. The scientific articles
summarizing the criteria and list of evidence-based treatments became some of the most
widely cited papers in psychology (Chambless & Hollon, 1998; Chambless & Ollendick,
2001) and stimulated entire book series designated to the various treatment manuals of this
list (e.g., Oxford’s Treatments That Work series). In contrast, the Task Force on therapeutic
change processes remained largely unnoticed (Castonguay & Beutler, 2005).

There was a certain historical logic connected to this emphasis on outcome. At the time of
Author Manuscript

Paul’s question, psychological treatment as a field most needed to show it could reliably
produce important results. Furthermore, process evidence unlinked to outcome evidence is
aimless and of unknown importance. Today, however, evidence-based psychological
interventions exist in virtually every area of life that can reliably alleviate human suffering
and promote psychological growth, generally with better long-term outcomes and fewer side
effects than medical or other alternative approaches.

Further progress on outcome demands that processes of change become more front and
center. The mountain of outcome evidence that has been produced regarding hundreds of
different packages and protocols cannot be assimilated without the ability to draw methods
into functional groupings. In addition, after decades of effort, it seems increasingly clear that
the latent disease model that underlies psychiatric syndromes is simply not a progressive
Author Manuscript

pathway to a deeper understanding of mental health difficulties. Meta-analyses of the

“protocols for syndromes” era of psychological intervention in specific areas have at times
struggled to demonstrate increases in effect sizes, leading to calls for greater attention to
more clinically valid approaches to therapy research (Goldfried & Wolfe, 1998) and
identifying the processes of change as a possible route forward (Goldfried, 1980; Greenberg,
1986; Kazdin, 2007; Kazdin & Nock, 2003; Weisz et al., 2018). Similarly, funders are
shifting away from testing new protocols focused on signs and symptoms of syndromes in
favor of attention on underlying mechanisms, such as through the Research Domain Criteria
(RDoC) initiative (Insel et al., 2010), and calls for a renewed focus on mechanisms of
change

The clear trend toward a process-based era of evidence-based therapy (e.g., S. C. Hayes &
Hofmann, 2018; S. C. Hayes et al., in press; Hofmann & Hayes, 2018) requires
Author Manuscript

methodologies that fit the new analytic agenda. In a process-based therapy model the key
empirical question is: “What core biopsychosocial processes should be targeted with this
client given this goal in this situation, and how can they most efficiently and effectively be
changed?” (Hofmann & Hayes, 2018, p. 38). Without well-crafted methods to answer this
question, progress will be hindered.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 3

The primary experimental approach to the study of processes of change has been methods of
Author Manuscript

linear mediation of randomized controlled trials (RCTs). Although a variety of mediation

frameworks exist, most current methods can be traced back to a single classic paper. In one
of the top 100 most cited articles of all time in all of science, in 1986 Baron and Kenny
described a straightforward “causal steps” approach to study mediation and moderation.

The popularity of Baron and Kenny’s approach can be traced to several immediate benefits:
the simplicity, clarity, and applicability of the analytic approach; the ability to go beyond
mere correlational evidence on processes of change; the reassuring idea that causal
mechanisms could be readily uncovered; the claimed ability to distinguish partial and full
mediation; the ability to integrate moderation and mediation into a single model; and the
ability to distinguish functionally important pathways of change from mere socialization to
therapeutic concepts.
Author Manuscript

Although the field of mediation has undergone many upgrades and revisions, the basic
structure of Baron and Kenny’s (1986) approach is still the underlying model in the study of
mediation and moderation in general and for psychological treatment processes in particular.
That persistence is noteworthy in part because classic methods of mediational analysis have
not led to a robust science of processes of change.

Often, RCTs testing specific approaches have yielded disappointing results regarding the
processes of change that may have given rise to outcomes, and consistent replication of
positive findings is unusual. While the clinical field has produced a dizzying number of
treatment models and treatment protocols for virtually every psychiatric and psychological
problem imaginable, increases in understanding of the processes of change in psychotherapy
has been slow to arrive.
Author Manuscript

It is time to ask “why”? If a process-based approach to therapy is key to progress of the field
of psychological intervention it may be time to rethink mediation itself, and the methods we
are using to conceptualize, define, and examine it.

Mediation: The Classic Approach to Test for Treatment Processes

Baron and Kenny (1986) originally defined a mediator as “the generative mechanism
through which the focal independent variable is able to influence the dependent variable of
interest” (Baron & Kenny, 1986, p. 1173) and defined a moderator as a “variable that affects
the direction and/or strength of the relation between an independent or predictor variable and
a dependent or criterion variable” (Baron & Kenny, 1986, p. 1174). Mediators and
moderators of change have been primarily studied by using linear regression models (Baron
Author Manuscript

& Kenny, 1986; Holmbeck, 1997; Judd & Kenny, 1981; Sobel, 1982). Other techniques
include structural equation models, including latent growth models and multilevel models,
and other regression-based approaches, such as conditional process analysis (A. F. Hayes,
2009). These techniques share many of the same problems we will describe below.

Figure 1 depicts the relationship between a mediator, the independent variable, and the
outcome. According to Baron and Kenny’s causal steps approach (1986), the conditions of
mediation are met if (1) treatment x has a significant impact of outcome y (the c path), (2)

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 4

variance of the independent variable accounts for variance in the presumed mediator (i.e.,
Author Manuscript

path a is statistically significant); (3) variance of the presumed mediator, controlling for
treatment, accounts for variance in the dependent variable (i.e., path b is significant), and (4)
the previously significant relation c path is no longer significant when paths a and b are
controlled (yielding what is called the c’ path). Full mediation is assumed to exist when the
“c” relationship disappears, whereas partial mediation exists if this relationship is
significantly reduced (James & Brett, 1984). In cross-sectional designs, the variables that
constitute the mediation model can be arbitrarily assigned. For this reason, treatment process
researchers often insist on temporal precedence of the mediator as compared to the outcome
variable (and the independent variable as compared to the mediator).

Since the publication of Baron and Kenny’s (1986) seminal paper, a number of important
revisions and clarifications and modifications for the testing of moderation and mediation
(and moderated mediation) have been provided (Holmbeck, 1997; A. F. Hayes & Preacher,
Author Manuscript

2014; Judd & Kenny, 1981; Kraemer, Wilson, Fairburn, & Agras, 2002; MacKinnon,
Lockwood, Hoffman, West, & Sheets, 2002). Baron and Kenny’s causal steps approach
never actually tested the significance of the difference between the c and c’ paths, for
example, and the categorical language of “full” and “partial” mediation was linked to alpha
levels, not effect sizes. Sobel’s (1982) test promised a test of the c versus c’ comparison by
assessing the cross-products of the a path and b path coefficients (since in most contexts c –
c’ = a * b), but it improperly assumed a normal distribution when in fact the product of two
normal distributions can be non-normal. That was addressed by non-parametric tests of the
significance of the cross-product of the coefficients, which also allow mediation to be tested
on smaller data sets provided the data sets were internally consistent (A. F. Hayes &
Preacher, 2014; MacKinnon et al., 2002). Other revised approaches to mediation including
structural mean models (Emsley, Dunn, & White, 2010), and principle stratification (Emsley
Author Manuscript

et al., 2010; Gallop et al., 2009), also have been developed to resolve distributional
assumptions in classic tests of mediation, such as sequential ignorability. The problem of
effect size estimation has yet to be solved since no currently available method is agreed to
meet statistical requirements (Kelley & Preacher, 2012).

Some researchers have tightened the logic needed to suggest that mediators were causally
related to outcomes. Adopted to RCTs, Kraemer et al. (2002) suggested that a mediational
relationship exists if: (1) the proposed mediator correlates with treatment choice; (2) the
mediator has either a main or interactive effect on outcome; and (3) changes in the mediator
variable precede changes in the dependent variable. Stice, Presnell, Gau, and Shaw (2007)
similarly emphasized the centrality of temporal precedence of the mediator in relation to
changes in the outcome variable.
Author Manuscript

Some of these modifications and clarifications have been explicit in their desire to go
beyond Baron and Kenny (e.g., A. F. Hayes, 2009). Most, however, retain and share several
key limitations of classical mediational analysis, and indeed some of the solutions make the
applicability of this approach even more removed from the key questions faced by process-
oriented intervention.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 5

Limitations of the Traditional Approach to Test for Mediation and

Author Manuscript

Moderation
The four most significant problems that cannot be addressed using a classical mediational
approach are: (1) the violation of the key statistical assumptions necessary to apply
mediational results tested at the level of the collective to individuals; (2) change processes
may involve multiple variables extended over time; (3) mediator, outcome, and independent
variables typically are not commonly in a strict unidirectional and stable relationship, but
instead form bidirectional relationships that often change over time; and (4) change
processes are often nonlinear. As a result of these issues, we argue here that treatment
processes cannot be properly examined using linear, unidirectional, and pauci-variate models
(i.e., models containing only a few variables) that are based on the collective as a unit of
analysis. Instead, we argue that treatment processes require idiographic methods that can
Author Manuscript

examine relationships between many variables, some of which form bi-directional

relationships that change in a non-linear, dynamic way in response to treatment. Firmly
established findings from a large number of individuals can then be gathered into nomothetic
statements about processes of change.

Methods now exist to study processes of change in this way, and the field can integrate them
into existing outcome research designs. Dynamic network approaches are now available that
seem adequate to this task, and adopting these methods might significantly advance the field
of psychotherapy research. The following will provide some details about the limitations of
contemporary mediation tests.

A Nomothetic Approach is Unable to Capture the Complexity of Therapy in a Given

Individual
Author Manuscript

Clinical mediation approaches use group data and mean values to understand the process of
change. It is well established that group averages can easily obscure clinically relevant
individual differences (Barlow et al., 2009), but it is not as widely known that
mathematicians have long agreed that what is true for measures of central tendency for a
group cannot be assumed to reflect individuals within the group. In the physical sciences, it
has been clear for nearly a century that analyses of collections of events (e.g., a volume of
gas) can be assumed mathematically to apply to the behavior of individual events (e.g.,
individual gas molecules) only under a highly constrained set of circumstance that define the
events as “ergodic” (the original mathematical proof was offered by Birkhoff, 1931).
Examples of such circumstances in the case of Gaussian processes are that the events are
identical and do not change over time, but while these conditions sometimes exist the
physical domain (some ideal gases for example: Volkovysskii & Sinai, 1971), it is unheard
Author Manuscript

of in biobehavioral areas.

The concern over meeting necessary “ergodic assumptions” is just beginning to get attention
in the biobehavioral sciences (e.g., Fisher, 2015). The implications are stark and only
reinforce long held concerns that psychological processes are not well studied using
variable-centric analyses of collectives, instead of person-centered longitudinal and
contextual analyses of action (e.g., Barlow, & Hersen, 1984; Sidman, 1960; von Eye &

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 6

Bergman, 2003). The single biggest implication is this: standard statistical analysis
Author Manuscript

techniques based on inter-individual variation cannot properly model a developmental or

change processes (Molenaar, 2008a).

Although it is not usually explicitly stated as such, the concern over ergodic assumptions is
reflected in the familiar homogeneity assumptions of group outcome research (Molenaar,
2004). The polite fiction of a latent disease promised such homogeneity if syndromal criteria
are carefully applied, which may have deflected proper attention within the field toward this
key statistical issue. The ergodic assumption rarely holds true in psychopathology
(Borsboom & Cramer, 2013), but it is an impossible assumption when processes of change
are the focus. Even if all persons with a specific psychiatric diagnosis are deeply the same,
no one would assume that the many events that can influence syndromal symptoms or other
outcomes unfold in the same sequence and pattern for all. Which change processes are most
relevant, how they interact over time, and the trajectories of outcome changes that result, are
Author Manuscript

obviously not homogeneous across people. To the extent that is the case, there is no
legitimate way to apply findings of classical mediational studies to individual recipients of
care without making entirely implausible statistical assumptions.

It helps to have a common-sense example of the difference between relationships seen as a

collective level and those seen as an individual level so that the level of analysis problem can
be separated from the analytic need to study many people in order to make reliable
nomothetic generalizations. Consider the relationship between the speed of typing and the
number of errors made while typing. Virtually any large group of people will contain both
beginning and expert typists, and experts will type faster with fewer errors. Thus, almost
every group of notable size will show in a statistical test that typing speed is negatively
related to errors. However, for practically every single individual within all of these
Author Manuscript

collectives, beginner and expert alike, efforts to type faster over a series of tests will produce
more errors. This example shows that analyses at the level of a collective can yield findings
that are orthogonal or even consistently opposite to what occurs at the level of individuals
over time. In this common sense example, corrections could be made post hoc at the group
level to account for the effects of typing expertise but in real world applications to processes
of change, there is no way to know a priori why, whether, or how a specific group level
analysis would or would not apply to individuals within the group.

Once the violations of statistical assumptions are admitted, it becomes clearer that if
information about processes of change is to be applied to individuals, that information needs
to be based on statistically legitimate, idiographic, intra-individual analyses of processes of
change and only then drawn into nomothetic generalizations. In order to study the treatment
Author Manuscript

processes that underlie effective interventions, it seems necessary to move away from
traditional nomothetic approaches that emphasize average group differences derived from
large samples of individuals and move toward idiographic approaches (Kazdin, 2016;
Barlow, Nock, & Hersen, 2009).

Although there have been calls to study mediation idiographically (e.g., Faldowski, 2009), to
date little research has been done to study evidence-based processes of change using
idiographic methods. Historically that may have been both of the weak focus on processes of

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 7

change and because of the limited range of statistical tools available. Recent statistical and
Author Manuscript

methodological developments appear to be overcoming those barriers, however, as we will

discuss later.

Treatment Processes Are Temporally Extended Multivariate Events

The response to one event often serves as the occasion for another action and any one action
can be influenced by a multitude of other variables. This is common sense and is evident in
the careful examination of any complex human action (reading a scholarly article, for
example), never mind such complex treatment matters as emerging from a struggle with
anxiety, or recovering from an addiction. Classical mediation and moderation models are too
cumbersome analytically to take the multivariate and temporally extended nature of
processes of change seriously. Some studies do exist that test multiple mediators but not at
multiple times, and the vast majority of studies applying classical mediation only consider a
Author Manuscript

very few variables at a single point in time to mediate and moderate treatment.

In the original formulation, Baron and Kenny (1986) acknowledged that full mediation is
unlikely to be observed given that multiple mediators are likely to be involved for any given
phenomenon. Specifically, they wrote: “If the residual Path c is not zero, this indicates the
operation of multiple mediating factors. Because most areas of psychology, including social,
treat phenomena that have multiple causes, a more realistic goal may be to seek mediators
that significantly decreases Path c rather than eliminating the relation between independent
and dependent variables altogether” (p. 1176). When traditional studies on mediation and
moderation consider multiple variables, however, they are unable to characterize how these
variables are interrelated as processes of change extended over time.

When traditional multiple mediator analyses are used to try to explain multivariate events
Author Manuscript

their analytic weakness also become evident in other ways. For example, researchers may
encounter suppressor variables that at times are virtually impossible to interpret. An example
is provided in a study by Papandonatos et al. (2012), which applied a multiple mediator
model to the 12-month physical activity outcomes of a motivation intervention. When well-
known measures unexpectedly acted as suppressor variables in a multiple mediator model,
the authors could only blame “possible collinearity” with other measures, even though no
such collinearity was actually documented.

Elements of Treatment Processes Can Form Bi-directional Relationships and Feedback

Loops
Classical mediational analysis is entirely unidirectional – the a, b, c, and c’ paths are all one
headed arrows. This is not a plausible assumption even in static analyses, but it becomes
Author Manuscript

incoherent in temporally extended relationships.

The therapeutic process is usually happening between two or more individuals during which
a skillful therapist establishes a supportive environment for a patient to explore and
overcome certain problems and move toward personal goals. Therapeutic change as it is
actually produced is not a simple summation of the patient’s responses to the therapist’s
input. The client is responding to the therapist, but the therapist is also responding to the
client (Stiles, Honos-Webb, & Surko, 2006). Earlier responses set the occasion for later ones

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 8

– for example, as the patient gains insight about her or his behavior, events are responded to
Author Manuscript

in new and self-sustaining ways. Positive therapeutic events often lead to even more of the
same, and as positive affect becomes stronger, negative affect weakens via positive and
negative feedback loops.

This phenomenon has been described in the broaden-and-build model (e.g., Fredrickson,
2000); positive affect loosens the influence of negative affect on the person and at the same
time broadens the behavioral repertoire by enhancing physical, social, and intellectual
resources. A positive therapeutic relationship can provide the fertile ground to develop such
a process. As the patient become more emotionally open, therapists may free up as well, and
in complex ways. For example, the therapist may push into even more difficult areas, trying
to take advantage of the building momentum.

Common therapeutic events of this kind cannot be properly characterized by the simple,
Author Manuscript

unidirectional relationships imagined in classical mediational analysis. Analytic models are

needed that can characterize bi-directional relationships and complex feedback loops. This
point is especially obvious in the context of the therapeutic relationship, but it is not limited
to it. The efficacy data reported by many internet-based therapies contradict the notion that
therapeutic success is entirely due to therapeutic alliance and the so-called common factors
(Ahn & Wampold, 2001). Self-amplifying and bi-directional tendencies such as “broaden
and build,” or carrying gains into previously avoided areas, apply whether therapy is
delivered by a live person or through a computer. Classic mediational analysis is unable to
characterize and evaluate such common bidirectional and recursive relationships.

Treatment Processes Are Typically Dynamic Non-Linear Events

Treatment progress is rarely linear (A. M. Hayes, Laurenceau, Feldman, Strauss,
Author Manuscript

Cardaciotto, 2007). When high frequency measures are added to RCTs, a wide variety of
shapes and patterns appear that will simply be missed by classical mediational analysis (e.g.,
A. M. Hayes et al., 2007). Individual trajectories are typically marked by local peaks and
valleys, even if the general trend might be upward as is the case of treatment responders.
Sometimes, there are rapid changes from one session to the next that predict positive long-
term outcomes such as sudden gains (Tang & DeRubeis, 1999) and rapid early response
(Ilardi & Craighead (1994). These phenomena have been replicated in numerous studies
(e.g., Aderka et al., 2012; A. M. Hayes, Feldman, Beevers, Laurenceau, Cardaciotto, &
Lewis-Smith, 2007). Interestingly, the depression spike pattern observed by A. M. Hayes et
al. (2007) appears to be due to the activation of a depressive network, similar to activation of
a fear or trauma network in exposure-based therapies, predicting later treatment success (A.
M. Hayes, Feldman et al., 2007).
Author Manuscript

Network theorists have examined such events as tipping points (e.g., A. M. Hayes, Feldman
et al., 2007; Hofmann, Curtiss, McNally, 2016; Nelson et al., 2017; Scheffer et al, 2009,
2012; Schiepek, & Strunk, 2010; van de Leemput et al., 2014; Wichers et al., 2016) that
appear to be predictable and serve as precursors of dramatic changes in the patient’s mental
health state. Because traditional tests of mediation and moderators assume linear
relationships between independent variables, mediators, and outcome and the influence of

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 9

moderators is considered to be constant, however, these kinds of patterns of change are

Author Manuscript

impossible to detect.

Consider how the dynamic and non-linear nature of therapeutic change might complicate the
timing of mediation. In a classic RCT with assessment at pre, mid, post, and follow up, a
mediator that may have been collected at the mid-point is tested in its ability to explain
outcome changes at follow up. That mid-session might be a sudden gain for one client, and a
depression spike for another. Both patterns could be positive indicators but they may cancel
each other out, leading to a false conclusion about the relevance of given change processes.

The Conflation of Function and Cause

Traditional mediational analysis is often spoken of in mechanistic causal terms, but no
amount of information about change processes can eliminate the possibility that other
unmeasured variables are responsible for the relationships observed. Biobehavioral
Author Manuscript

relationships are contextual and historical. Simple causal language may at times have
practical utility, but it is important not to allow a focus on causes to diminish the role of
context. For example, it might be useful to speak of a spark that “caused” an explosion in a
grain elevator, in part because combustible dust and oxygen can be assumed as context;
when welding combustible metal in a vacuum it might be useful to speak of a breach of the
vacuum chamber that “caused” an explosion, because a spark and combustible material can
be assumed as contextual events. In either case, explosions are the coming together of
multiple variables, none of which have their functional roles except in the context of the
other participating events: they are not causes in a simple ontological sense. Causality of that
kind has long since been waived away in philosophy of science (Russell, 1912) but it
continues to create problems methodologically. For example, if a researcher conducting the
typical “pre-mid-post-follow-up” RCT is unlucky enough to have seen significant
Author Manuscript

improvement in the outcome variable at the mid-point, it may be impossible to publish any
resulting data on change processes because temporality will be argued to have been violated.
Rather than modeling the degree of strength of functional relationships between two
variables over time, the conflation of cause and function will mean that it is considered
“conservative” for the field to receive no guidance whatever regarding processes of change
from any study that shows unexpectedly rapid outcome changes. That is not conservative:
we think it is nonsensical.

In sum, psychotherapy is a process that cannot be reduced to one or even a few mediators
and moderators, and it cannot be assumed that these variables are independent from each
other or that they form simple, unidirectional, and linear relationships. Moreover, a
nomothetic approach is inadequate to study treatment processes in a way that is known to
Author Manuscript

apply to individuals. As a result, changes processes in intervention are better characterized

as a pattern of multiple interrelated variables that form a complex and dynamic network over
time, person by person. A nomothetic, pauci-variate, unidirectional, and linear approach is to
mediation is simply unable to capture and study the nature and complexity of processes of
change. It is thus not surprising that despite the large number of efficacious psychological
treatment, relatively little is known about the underlying therapeutic change processes. In the
coming era of process-based therapy, these methodological limitations must be overcome.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 10

From Mediators and Moderators to Treatment Processes

Author Manuscript

Understanding and intervening at the level of treatment process is a particularly promising

path forward for intervention science (S. C. Hayes & Hofmann, 2017, 2018; Hofmann &
Hayes, 2019). Consistent with the overall trend toward personalized and precision medicine,
focusing on change processes provides a way for evidence-based interventions to be tailored
to fit the person or persons being treated. We have defined a therapeutic process as a set of
theory-based, dynamic, progressive, contextually bound, and multilevel changes that occur
in predictable empirically established sequences oriented toward the desirable outcomes.
These processes are theory–based and associated with falsifiable and testable predictions;
they are dynamic, because processes may involve feedback loops and non-linear changes;
they are progressive in the long-term in order to be able to reach the treatment goal; they are
contextually bound and modifiable, so as to focus on practical targets of change; they form a
multilevel system, because some processes supersede others; and they are oriented toward
Author Manuscript

both immediate and long-term goals (Hofmann & Hayes, 2019, p. 38; Hayes, Hofmann, &
Ciarrochi, 2019).

This definition is consistent with the notion that therapy includes learning processes that,
together with therapeutic procedures, form the mechanisms of change of a certain treatment
(Bruijniks et al., 2019). In the case of cognitive therapy for depression, for example,
cognitive change processes, such as changes in negatively-biased thinking, have been
proposed in cognitive therapy that involves cognitive procedures. However, other procedures
may target the same processes, leading also to improvements in depression (Lorenzo-
Luaces, German, & DeRubeis, 2015). Therefore, it is important to distinguish the theoretical
process from the practical therapeutic strategy leading to change. Moreover, traditional use
of the term therapeutic process is sometimes used to refer to the patient-therapist
Author Manuscript

relationship, such as the therapeutic alliance, other so-called common factors of the
therapeutic relationship. Our definition of the term therapeutic process is considerably
broader, although it may include this more traditional use of the term as long as it refers to
change processes that are based on a clearly defined and testable theory and assessed in a
characteristic way, which is not yet often the case. To help with the distinction, we will use
the term change processes or its equivalent as a synonym for therapeutic processes.

In our definition, a therapeutic process is thus a theoretically integrated set of changes that
are empirically shown to help achieve desirable treatment goals – the defined and
measurable outcomes that the therapist and client have agreed upon. While such goals need
to be clearly specified, they should not necessarily be seen as static goal posts. Often, they
are moving targets that change as treatment progresses. Typically, therapy is directed toward
Author Manuscript

multiple proximal and longer-term goals, which may be arranged in a hierarchy depending
on such factors as priority, difficulty, generalizability, and immediacy. Therapy usually
comprises a number of different treatment processes that are designed to lead to the
attainment of a variety of desirable treatment goals within the case.

The term treatment process so defined is not synonymous with mediation. Some mediators
may be found to be treatment processes and vice versa, but the change processes used in
psychological intervention is a much larger set. Change processes are less restrictive and

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 11

allow for a more complex interplay between the elements underlying a successful course of
Author Manuscript

treatment, including bidirectional, non-linear, and recursive relationships and that may be
arranged into feedback loops. Baron and Kenny (1986) used the term mediator to describe
the element of a mechanism or the mechanism itself. Tryon (in press), however, argued that
the term mediator should not be used synonymously with generative mechanism because
other variables besides the mediators can also influence the change process. Similarly,
Kazdin (2007) cautioned that even when mediators are known, it is still necessary to develop
a pragmatically useful and theoretically consistent explanation of how a treatment can lead
to the outcome.

We agree with these points, and they are nested within our approach to understanding
processes of change. Classical methods of identifying mediators and moderators can only be
considered as a possibly useful first step in identifying and considering change processes,
which may include what has been thought of as clinician factors, client factors, and actual
Author Manuscript

mechanisms of change (Kazdin, 2007; Kazdin & Nock, 2003; Nock, 2007). Although
classical mediation may be useful as an initial “rough sort” or a guide for places to look, it is
neither sufficient nor necessary to the identification of change processes. Mediators and
moderators identified with classic methods still need to be re-examined in ways that are
statistically defensible in terms of their underlying assumptions, which suggests that it is
best for knowledge to be developed at the level at which it will be applied.

Traditional Non-Mediational Approaches for Studying Change Processes

There are a variety of traditional methods that can be useful for identifying change processes
beyond mediational analysis, but each need modification to be more fully useful to process-
based therapy models. These methods are described next.
Author Manuscript

Traditional Functional Analysis

Functional analysis is one of the original approaches developed for identifying change
processes in a single individual. The primary objective of functional analysis involves
identifying the functional relationship between variables that contribute to behavior by
assessing both a target behavior and the context in which this behavior occurs (e.g., Haynes
& O’Brien, 1990). A functional analysis isolates specific variables that cause or maintain a
behavior of interest by identifying the antecedents and consequences of the behavior
(Yoman, 2008). This approach provides a unique framework for designing individually-
tailored treatments. As the outcome of the intervention varies over the course of treatment,
the intervention strategies can be adapted in such a way as to allow the therapist to discover
which processes contribute to symptom improvement.
Author Manuscript

There is a strong empirical basis underlying functional analysis as a component of

efficacious interventions for some conditions (Ghaderi, 2006; Hurl, Wightman, Haynes, &
Virues-Ortega, 2016; Miller & Lee, 2013). However, traditional functional analysis has
several limitations. Often functional analysis also relies on a powerful but excessively simple
model in which relationships abide between antecedents, behaviors, and consequences.
Furthermore, functional analytic models often only consider a relatively small number of
contingency variables that are thought to be responsible for a target behavior, while

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 12

cognitive or other important variables have been deemphasized (S. C. Hayes, Long, Levin,
Author Manuscript

& Follette, 2013). This has had the effect of narrowing the available empirical literature on
applied functional analysis largely to developmental disorders, in which direct active
contingency principles can be sufficient to understand the target behavior in children with
developmental problems (Friman, 2010). Furthermore, the time-intensive and complicated
nature of existing methods of functional analysis have prevented it from being easily
replicable among practitioners (Follette, Naugle, & Linneroth, 1999). In principle, functional
analysis might be a theoretically strong approach to create individually tailored treatments,
but it needs to be recast in a form that is capable of identifying a sufficiently large range of
change processes so as to be broadly applicable. This new form of functional analysis we
believe may yet emerge from the steps we are suggesting here.

Single Case Design and Graphical Time Series Analysis

Author Manuscript

Characterization of a change process in a single individual (or a single couple; or a single

classroom and so on) has historically been done using single case experimental designs and
graphical time series analysis (S. C. Hayes, Barlow, & Nelson, 1999). While these analyses
are logical and reside at the right level of analysis, they need to be rethought somewhat to fit
the need to characterize dynamical non-linear change processes. For example, when a
baseline is interrupted it is commonly advised to focus on the immediate impact of change.
Superficially this is logical but it is not wise. Well-established processes exist that predict a
period of seemingly negative change as a part of a positive change process. A telling
example arises if the treatment involves extinction. In that case, an extinction burst will be
predicted and will be taken as a positive sign of change. That correct interpretation depends
on past idiographic research helps the psychologist correctly identify a positive non-linear
change process, but in the absence of such guidance a common sense methodological rule of
Author Manuscript

thumb (“focus on the immediate impact of change”) might have severely interfered with the
identification of key processes of change.

Instead of placing this distinction in the hands of the psychologist, it seems less likely to
lead to error if it is placed in the hands of well-crafted statistical methods that can
accommodate intensive time series data. Such methods now do exist and they begin to
suggest a generally useful way forward.

Statistically Characterized Time Series Analyses

A number of studies have employed sophisticated analyses using time series data to examine
a variety of clinical phenomena (e.g., Schiepek et al., 2016). In a recent proposal to embrace
a dynamic assessment model to identify individual expressions of psychopathology, Fisher
(2015) examined the underlying structure of generalized anxiety disorder (GAD)
Author Manuscript

symptomatology in 10 individuals receiving daily assessments over a 60-day period.

Consistent with the dynamic assessment model, considerable heterogeneity characterized
each individual’s dynamic structure of symptoms. For instance, some individuals exhibited a
pattern in which greater levels of avoidance led to subsequent reductions in worry and
anxiety, whereas others demonstrated the opposite pattern. An advantage of the approach
proposed by Fisher (2015) is that it considers the time series characteristics of clinical
constructs at the individual level.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 13

Other studies have also adopted a similarly intensive time series approach; often however, it
Author Manuscript

has been conducted at the collective level. For example, some studies have used time series
analysis to estimate patterns of change in constructs targeted by transdiagnostic treatments
(e.g., reappraisal, mindfulness) in individuals with depression and anxiety (Boswell,
Anderson, & Barlow, 2014; Boswell & Bugatti, 2016). The results provided some evidence
that specific intervention strategies led to greater improvement in expected change
constructs (e.g., a present-moment nonjudgmental mindfulness module was associated with
increased mindfulness ratings). Although such studies provide invaluable information, in the
area of change processes in particular it is important to move beyond approaches to
longitudinal data that relying entirely on a collective level, and instead to capture important
change processes that occur at the level at which knowledge will be applied.

Nomothetic approaches have long been used to test for mediation even in longitudinal
datasets. Common statistical techniques include the use of growth curve analyses using
Author Manuscript

multilevel models (e.g., Moscovitch et al., 2005) and latent variables (e.g., Gao, Curtiss, Liu,
& Hofmann et al., in press) to examine longitudinal mediation. A focus on individual change
patterns does not change the level of analysis, however, if intensive longitudinal data at the
level of individuals are used to generate error terms for statistical tests at a collective level.

In sum, to further our understanding of change processes, more sophisticated statistical tools
are needed that test and lock down findings at the desired level of analysis first (generally
the individual, although in some clinical or applied contexts it may be a couple or small
group) before seeking nomothetic generalizations. To that topic, we now turn.

Adopting a Dynamic Network Approach to Study Treatment Processes

Recently, the network approach has been gaining ground as a methodology for studying
Author Manuscript

psychological events. For example, a number of studies have employed network analyses to
address empirical questions pertaining to classification of psychopathology and its treatment
(e.g., Hofmann, Curtiss, & McNally, 2016; Borsboom & Cramer, 2013). Recently, advances
have been made in constructing networks for intensive time series data even at the level of
the individual that then can lead to nomothetic findings (e.g., Fisher, Newman, & Molenaar,
2011). Such dynamic networks might be able to be used to (1) facilitate case
conceptualization, to (2) select the most appropriate treatment strategies, and to (3) predict
the course of treatment.

A Case Example
The dynamic network approach provides the methodological framework to systematically
Author Manuscript

examine individual change processes while considering the functionality of multiple

variables involved in this process (Epskamp, van Borkulo, et al., 2019; Epskamp, Waldorp,
Mottus, & Boorsboom, 2018; Hofmann, Curtiss, & McNally, 2016; Nelson et al., 2017;
Scheffer et al., 2009, 2012; Wichers & Groot, 2016; van de Leemput et al., 2014). Dynamic
networks can inform how processes unfold over time at the individual level. We begin this
discussion on an individual level. The goal, however, is to develop theory-based models and
aggregate across individuals. As we will describe further below, the Group Iterative Multiple

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 14

Model Estimation method (GIMME; Gates & Molenaar, 2012) is an example of such an
Author Manuscript

approach.

A network can represent variables of interest as nodes and relationships between nodes as
edges. In dynamic networks, temporal information can be encoded in the edges, which can
convey important insight into the time-series relationships abiding between nodes. Several
types of networks can be estimated using time series data. The temporal network can
provide information about the relationship between nodes across different measurement
windows, which might reveal whether certain nodes prospectively predict other nodes.
Directed edges can be specified to represent partial regression coefficients connecting
different nodes. The contemporaneous network provides information about associates
occurring between nodes within the same time window, identifying larger patterns of
relationships. In the temporal network, both autoregressive and cross-lagged effects are
possible because each node is regressed onto both itself and other time lagged nodes (t-1).
Author Manuscript

There are a number of advantages a network approach confers beyond the traditional Baron
and Kenny mediation framework for understanding treatment processes. To illustrate this
approach in a specific client, we can explore how a network approach to functional analysis
and case conceptualization enriches our knowledge of the processes underlying a clinical
problem. Figure 2 shows an example of a typical Baron and Kenny mediation model.

In our hypothetical example, the client, Sam, pursued treatment after undergoing a recent
break-up from his girlfriend for the purpose of addressing his low mood. The mediation
model depicted in Figure 1 suggests that his break-up prompted increased levels of
rumination that partially mediated his low mood. Also, the break-up contributed to low-
mood independently from rumination.
Author Manuscript

This type of mediation model is quite typical in contemporary clinical science and consistent
with Figure 1 of Baron and Kenny’s (1986) mediation approach. However, we argue, such
an approach is quite likely to be insufficient to capture the complexity of Sam’s
psychological problems because rarely does change in psychopathology involve only one or
two variables that are uni-directionally, independently, linearly, and causally linked to a
problem. The limitations of such an approach can become very evident once more
information is collected from Sam throughout the course of treatment. We provide such an
example for illustrative purposes.

Suppose, for example, that the break-up and rumination are not the only problems leading to
Sam’s depression – perhaps low self-esteem and loneliness also contribute to Sam’s
problems. These variables might not be connected in a simple unidirectional manner, as
Author Manuscript

would be assumed by the traditional mediation model. Furthermore, some of the problems
might be more central than others to Sam’s suffering. An alternative model of Sam’s
problems is presented in Figure 3, which offers an example of a simple network perspective
as applied to an individual case.

As depicted in Figure 3, low self-esteem and loneliness are the two central problems for
Sam’s overall problem space. Both variables have edges that are stronger in magnitude, as
reflected by their thickness, and are more influential than the other variables as indicated by

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 15

the thickness of the node borders. This network model permits bi-directional relationships,
Author Manuscript

as shown by the connections between loneliness and low self-esteem and between low self-
esteem and low mood. Thus, Sam’s break-up led to more loneliness, which has a recursive
and mutually reinforcing relationship with low self-esteem and low mood.

The Figure illustrates that the simple mediation model of Figure 2 is not necessarily wrong,
but it is unlikely to represent the complexity of Sam’s suffering. Figure 3 simply adds a
different level of analysis, which introduces additional variables and their connections. In
fact, the traditional mediation model is still represented in Figure 3, but it only constitutes a
limited subset of Sam’s underlying problem space. The individual network approach allows
for a more complete conceptualization by allowing a nuanced functional analysis of Sam,
which can lead to important treatment decisions.

For example, the therapist may decide that it is important to intervene on his low self-esteem
Author Manuscript

and loneliness, given that these are the two most influential nodes in the network. Perhaps,
the therapist might introduce self-compassion mediation as a treatment strategy, and teach
the client when he feels lonely apply this skill, perhaps while remembering how he felt as a
lonely child so as to be more kind to himself. Toward the termination of treatment, Sam’s
network conceptualization now reflects the presence of this more adaptive node, which
might change the network significantly (Figure 4).

The new network model shows how self-compassion might change Sam’s entire problem
space. By fostering self-compassion, the influence and dominance of more maladaptive
nodes become weaker, as suggested by thinner borders on all the maladaptive variables.
Furthermore, a number of edges no longer exist (i.e., low self-esteem leading to rumination,
loneliness leading to low self-esteem, and low mood leading to low self-esteem), while
Author Manuscript

others became weaker in strength. In this model, self-compassion functions as moderator on

the relationship between rumination and low-mood, because self-compassion dampens the
influence of rumination on low-mood. Self-compassion also now completely mediates the
relation of loneliness to low self-esteem.

Sam’s example illustrates how the network approach could be used as a tool for case
conceptualization and treatment planning. In contrast to the uni-directional and pauci-
variable assumptions underlying traditional mediation tests (Baron & Kenny, 1986) the
network approach allows clinicians to more adequately represent the complexity of a case.
Sam’s pathological network has been perturbed by introducing self-compassion mediation.
The two features that were originally most influential in Figure 3 (i.e., loneliness and low
self-esteem) lost their dominance in Sam’s problem space. Both Figures 3 and 4 substantiate
the fact that Sam’s functional analysis required a more sophisticated framework than that
Author Manuscript

provided by the original mediation model presented in Figure 2. The next case might
likewise include Figure 2 as part of a quite different set of variables (that person’s own
“Figure 3”) that lead to different treatment suggestions (that person’s own “Figure 4”).

It is important to feed complex network analyses the right kind of information, assessed with
high fidelity and frequency. Thus, both adequate theory and assessment technology are
needed to mount the use of dynamical systems in case conceptualization. To date, many of

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 16

the network analyses have been based on assessments focused on self-reports of syndromal
Author Manuscript

features (e.g., signs and symptoms), as distinct from contextual factors, biological measures,
overt behavioral measures, or measures focused specifically on change processes such as
cognitive flexibility or emotional openness.

High temporal density measures of known adequacy are needed for change processes to be
modeled as nodes in complex networks. Traditional psychometrics is likely not an adequate
filter since it too is based on implausible ergodic assumptions (Molenaar, 2008b). Any
weakness in assessment necessarily limits network-based case conceptualization and its
treatment utility.

This problem can only be solved over time as a network approach becomes more dominant
and EMA studies that identify possible variables of importance to case conceptualization
(e.g., Vilardaga, Hayes, Atkins, Bresee, & Kambiz, 2013) lead to assessment protocols that
Author Manuscript

will allow these processes to be examined in a dynamical system fashion. In considering

how to begin to assemble a more adequate empirical base for network analysis, theoretically
coherent processes that are sensitive to context, history, and the individual need to be
emphasized. Technological advances in assessment make this approach more possible, not
just the use of EMA but also the use of automatic recording of such things as activity,
location, biological reactions (e.g., heart rate variability), or social context. Whether this
would lead to better overall treatment outcomes is an empirical matter, but testing that
possibility requires first that the characteristics of such an approach be appreciated and
pursued. It is important to create high temporal density measures tied to functionally and
theoretically relevant variables. Signs and symptoms can be included if they prove helpful,
but ironically a better starter set may be variables that have been regularly identified as
mediators using traditional methodology, since a significant “a path” in traditional mediation
Author Manuscript

ensures that variables are responsive to treatment for at least some people. The existing
literature contains scores if not hundreds of successful mediators, some of which have been
shown to be functionally important to outcomes in several studies. This “initial rough sort”
can immediately provide the field with a range of theoretically meaningful variables to draw
from in creating and using measures that model change processes ideographically.

Prediction of Changes
Dynamical systems framework can be applied to treatment research to identify plausible
change processes. This seems an especially useful use of dynamical systems which have
historically been used to identify interconnected features that transition from one state to
another. Mental health can usefully be thought of as a complex system that alternates
between psychopathology and health (A. M. Hayes et al., 2015; Lutz et al., 2018; van de
Author Manuscript

Leemput et al., 2014; Schiepek & Strunk, 2010; Wichers et al., 2016). As we noted earlier,
such transitions have been referred to as tipping points. An example of an important sign of
an impending transition or tipping point is critical slowing down (Scheffer et al., 2009). If
someone exhibits critical slowing down with regard to mental health, this would mean that it
would take the individual more time to recover from a life stressor. Critical slowing down
can be modeled statistically by using metrics such as auto-correlation, with increases in
auto-correlation of a certain symptom signifying increases in critical slowing down (van de

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 17

Leemput et al., 2014; Wichers et al., 2016). We will illustrate this in an empirical case
Author Manuscript

described further below.

By collecting EMA data in a single individual throughout treatment, it might be possible to

identify whether treatment response can be predicted using critical slowing down and other
early warning signs common to complex network analysis. Such early warning signs of a
tipping point can be used to reveal possible change processes and mechanisms underlying
psychotherapy. If a certain variable or feature of a disorder demonstrates critical slowing
down or characteristics of an early warning signal, then that might reveal insights about its
ability to function as a change process in the context of an intervention. Expanding
mechanism research to dynamic processes, such as those measured by these metrics, will
foster a more idiographic approach to change processes in psychotherapy. A goal of such
research would be to identify early warning signs that are of clinical significance.
Considerable individual differences are likely to emerge in identifying which variables lead
Author Manuscript

to these early indicators of change, but by building this process knowledge on a firm
conceptual and empirical foundation these differences can enter into nomothetic
generalizations rather than being dismissed as “error.”

Selection of Treatment Strategies

When a temporal network structure provides clues about the impact a node has on the whole
network and how this might differ across separate individuals, and may predict changes
within temporal series, individualized networks have the potential to inform treatment
strategies. For example, strategies targeting influential nodes of the network are more likely
to perturbate the network than strategies that target nodes that are not influential or nodes
that are not even part of the network. Similarly, identifying sections of a larger network that
are self-amplifying in a negative direction, might direct treatment toward key pathological
Author Manuscript

change processes that need modification. Notice for example in Figure 4, that a recursive
loop is available between loneliness, low mood, and low self-esteem. Once the break up
kicked off this part of the network it could take on a life of its own, leading to a depressive
downturn. If the therapist was aware of this interlocking set of relationships, self-
compassion might not only be used to undermine these linkages, it might be done with
greater precision. For example, the attempt to specifically encourage the use of self-
compassion in reaction to loneliness (establishing the mental guide of applying compassion
to oneself as a lonely child) could anchor deployment of new methods of adjustment to
ecologically sensible events targeted to maximize the disruption of a pathological and self-
amplifying set of relations.

Preliminary attempts to validate personalized interventions based on symptom dynamics

Author Manuscript

have been explored by Fisher et al. (2019). In this open trial, individuals with low mood and
anxiety were assessed using EMA during a baseline phase for 30 days. Afterwards, person-
specific factor analyses and dynamic factor models were constructed to identify the
symptom domains and unique order of treatment modules for each person. These procedures
were in accordance with Dynamic Assessment Treatment Algorithm (DATA) developed by
Fernandez, Fisher, and Chi (2017). The results of the open trial indicated that personalized
treatment using the DATA algorithm resulted in a Hedges’ g of 2.33 for depression, which

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 18

was larger than the average Hedges’ g of 1.72 the authors found after analyzing 31 clinical
Author Manuscript

trials for depression (Fisher et al., 2019).

Despite these early examples, more research is needed to continue to validate the use of
dynamic networks in clinical practice. The point of this article is that methodological
changes are needed to test that potential empirically. One way to build a more adequate
empirical base that can assess the relationship between intervention and temporal network
structures, is to include high density longitudinal EMA data into large RCTs themselves. A
startling number of such studies include no process assessment, or merely examine
correlational evidence with a small number of such processes collected in a temporally lean
fashion. Part of that situation appears to be due to the fact that mediation analysis is
presented as the main way of determining the functional role of process changes. That very
decision then leads the researcher to a pauci-variate approach, often including only
theoretically preferred possible mediators rather than processes drawn from competing
Author Manuscript

models, and to assess them infrequently, all the while worrying that early outcome changes
will make the effort moot. By building low burden but high density process data into RCTs,
large data sets suitable for network analysis emerge.

One strength of traditional mediational analysis is the ability to control for treatment
assignment in the “b path”, which prevents mere markers of treatment assignment from
being significant mediator (e.g., using terms assessed by self-report that only those in the
treatment group are likely to recognize). This is difficult to do entirely within the individual
but if the intervention arm includes features such as baseline periods of different durations,
the linkage between treatment and specific patterns of system perturbation can be examined
person by person. If positive change processes are identified idiographically at the point at
which treatment occurs, at least occasionally these should also be seen in control conditions
Author Manuscript

and during baseline periods. For example, if self-compassion was found to be a key change
process for Sam and others like him (Figure 4), in a large study at least some people who
develop greater self-compassion toward themselves when lonely without explicit treatment
intervention should also show positive changes. If nomothetic generalizations based on
idiographic networks overlap across conditions or baseline and treatment phases, a
functionally important pathway can be inferred. Network analysis could thus represent a
promising method to identify the processes of change that can best be targeted in treatment.

Recent research has embraced randomized single-case trials using intensive longitudinal
data to examine when the introduction of an intervention leads to improvement (Hallford,
Sharma, & Austin, 2019). For instance, Hallford et al. (2019) introduced an intervention to
enhance episodic future thinking for depressed individuals using a randomized start point
Author Manuscript

after a baseline control phase. The results suggested that introduction of the intervention was
associated with large increases in anticipatory pleasure in seven of the eight participants.
Although examination of the network dynamics was not pursued in this study, this type of
single-case design using intensive time-series data represents a meaningful methodological
pathway to apply network analyses for studying processes of change in treatment.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 19

Ongoing Functional Analysis

Author Manuscript

When information on case conceptualization, prediction of changes, and selection of

treatment strategies care combined into theoretically coherent models, a new form of
ongoing functional analysis has arrived (Hofmann & Hayes, 2019). There are experimental
methods that purport to yield such knowledge, such as SMART designs (Almirall, Compton,
Gunlicks-Stoessel, Duan, & Murphy, 2012), but they can be very expensive to conduct and
only a very small number of treatment choice points can be examined. Recent attempts have
been made to apply network analyses to intervention research to examine how treatment
impacts different symptom nodes across each sequential week of treatment (Blanken et al.,
2019); however, this research still uses a cross-sectional approach. A more longitudinally
intensive approach could afford idiographic knowledge and foster a better functional
analysis. Once high-density process-relevant data are more available within existing
outcome studies, dynamic network methods seem far better suited to our current state of
Author Manuscript

knowledge development to develop valid knowledge about change process. With such data,
change processes will more often be evident both at the level of specific nodes, and via
edges and interdependent subsections of networks as researchers and theoreticians state clear
rules for the identification of change processes at these levels.

Fortunately, there are existing analytic methods that can be applied. For example, Configural
Frequency Analysis (CFA), allows mediation hypotheses to be tested at the level of
individual configurations instead of variables, and do so person by person (e.g. von Eye,
Mun, & Mair, 2009). Person-specific factor analysis of baseline data (Fisher, 2015) has been
experimentally linked to modular treatment with observed improved outcomes (Fernandez,
Fisher, & Chi, 2017).

The Group Iterative Multiple Model Estimation method (GIMME; Gates & Molenaar, 2012)
Author Manuscript

applies network analysis at the individual level, and then statistically generates nomothetic
population parameters for sub-types or entire populations if and only if the generalization
does not significantly distort findings at the level of the individual. This essentially turns
classic statistical analysis on its head, in which instead of viewing between individual
variability as error for central tendency estimates at the level of the group, findings are based
on statistical models accounting for intra-individual consistency in the context of intra-
individual variability and then are generalized nomothetically provided that step does not
add excessive variability. GIMME implements unified structural equation models (uSEMs;
Gates et al., 2010) to estimate vector autoregressive (VAR) models, which contain both
contemporaneous and lagged parameters. The particular networks GIMME creates are
directional and individual to each person. Although GIMME has largely been circumscribed
to fMRI data, recent efforts have been made to validate these procedures for behavioral
Author Manuscript

EMA data (Lane et al., 2019).

All of these methods can be harnessed to the task of generating methods of ongoing
functional analysis, and it is becoming more common to see such methods included in the
design of RCTs (e.g., Bell et al., 2018). If researchers understand the importance of process-
based research, the single most important step is to begin to include high-density
longitudinal assessment of processes of change in research and to do so in the theoretically
catholic manner.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 20

Because there is a vast number of specific terms and concepts using in intervention science,
Author Manuscript

a process-based approach methods for integrating evidence across schools, models, and
systems of therapy are essential to long-term progress. One suggestion for how to assure
broader coverage of treatment processes has been provided by Hayes, Hofmann et al (2019),
which is to apply multi-dimensional, multi-level evolutionary science concepts to the
assessment of possible change processes. In this approach researchers should consider the
relevance of the assessment of maladaptive and adaptive change processes in each of six
dimensions (cognition, affect, attention, self, motivation, and overt behavior), across three
levels (physiological/genetic; psychological; social/cultural) through the filter of an extended
evolutionary synthesis.

Evolutionary accounts are essentially a type of complex network dealing with variation and
selective retention in particular contexts (Wilson, S. C. Hayes, Biglan, & Embry, 2014).
Thus, in the S. C. Hayes, Hofmann et al (2019) approach, idiographic assessments should
Author Manuscript

consider variation and selective retention in context, for each of the domains and levels. The
word “consider” does not mean “include” – rather the point is to place process assessment
decisions in a broad matrix-like context that encourages breadth and theoretical
heterogeneity in the context of the consilience that evolutionary theory alone provides in the
other life sciences.

It should be noted that such an evolutionary approach toward psychotherapy is in no way

restricted to any one form of treatment orientation. An evolutionary model is broad enough
to encompass any psychotherapy orientation, as long as the underlying treatment processes
are theory-based, naturalistic, and empirically testable. The variety of existing mediators that
have been identified by different theoretical and clinical approaches can readily be thought
of in evolutionary terms. For example, the importance of healthy cognitive variation or
Author Manuscript

flexibility helps bring traditional cognitive behavioral mediators such as cognitive

reappraisal (Goldin, Morrison, Jazaieri, Brozovich, Heimberg, & Gross, 2016) and more
recent mediating processes emerging from the acceptance and mindfulness wing of the
cognitive and behavior tradition such as cognitive defusion (Arch, Wolitzky-Taylor, Eifert, &
Craske, 2012) under a single broad umbrella. Similarly, a common interest in healthy
variation in sense of self and perspective taking might draw together psychodynamically
grounded mediators such as mentalization (Rossouw & Fonagy, 2012) with cognitive
behavioral mediators such as decentering (Hoge et al., 2014). An extended evolutionary
approach might thus provide the consilience needed for a process-based approach to apply
equally and fairly across theoretical traditions, and to provide the grounds for a genuine
conversation across the walls of these schools and traditions.
Author Manuscript

Empirical Example
A case example with empirical data was presented by Wichers and Groot (2016). The
authors report the case of a 57-year old male with a history of multiple depressive episodes.
The patient was using antidepressants for the last 8.5 years. The researchers collected
ecological momentary assessment data over the course of his gradual discontinuation from
his antidepressant, venlafaxine.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 21

The relatively simple network consisted of negative affect, positive affect, mental unrest, the
Author Manuscript

psychotic experience suspicious, and worrying. All momentary states were rated on a 7-
point Likert scale. The patient reported his momentary state up to 10 times a day over a
period of 239 days. Figure 5 shows the changes in the network during the tapering up until a
depressive episode on day 128. As can be seen, the network structure changed such that
more and stronger connections formed leading up to the depressive episode.

Figure 6 further shows that this information can be used as an early warning signal.
Specifically, critical slowing down of the system has been shown to precede critical
transitions (also known as tipping points). Figure 6 shows that immediately before the
transition to a depressive state at day 128 (the grey zone) the autocorrelation of mental states
increased (middle panel) and variance of these states decreased.

The chosen nodes in this example were simple affect and mental state ratings that may not
Author Manuscript

even be central to the client’s pathology. Moreover, this case only allows for the depiction of
the structure of the covariation of the subjective ratings, rather than directional changes over
time. Therefore, no inference can be made about causation. Testing for causality in a
dynamic network is complicated by the fact that it is limited to a vector autoregressive
model (VAR) over the same time sample, VAR (1). As noted by one of our reviewers, even if
a VAR (1) uncovers relationships from Node A to Node B and Node B to Node C, it is
important to remember that these are concurrent events, because the effects are both from t
to t+1. Therefore, it is important to include the functional analytic case formulation when
interpreting dynamic network models.

It should be noted, however, that Molenaar and others have been examining hybrid VARs
(Molenaar, 2018) and phase resampling method to test for Granger causality in the
Author Manuscript

frequency domain (Liu & Molenaar, 2016). Additionally, novel methods have been adapted
to examine the duration of one node’s effect on another node. Impulse response functions
(IRF) can investigate how one node influences other nodes in the network by simulating a
sudden increase (or ‘shock’) in that particular node (Bos et al., 2018). Of the many aspects
of a network that can be examined using IRFs, questions can be addressed regarding how
many time windows one node is influenced by another node after a simulated shock (Blaauw
et al., 2017). Along with the other aforementioned approaches, IRF may represent a
promising direction to address issues regarding time lags that extend beyond VAR (1).
Future dynamic systems algorithms, combined with changes in sampling methods might
resolve these issues in the near future.

For the time being, changes in the dynamic network structure can only be understood in
combination with a functional analytic case formulation to formulate hypotheses about the
Author Manuscript

causal links between specific problem areas. These can be depicted as uni- or bidirectional
edges of a network. This functional analytic network can then serve as a framework to gather
empirical data to test some of the clinical hypotheses derived from the functional analytic
case conceptualization with the ultimate goal to effectively perturbate the maladaptive
network. Just as covariation does not necessarily imply causation, causation also does not
necessarily imply covariation. Nevertheless, the changes in dynamic network structure can

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 22

provide clinicians with valuable clues about possible causal links between nodes that can be
Author Manuscript

subject to treatment target.

Conclusion
Traditional tests of mediation and moderation have used simple regression models to test the
influence of a small set of variables on a specified outcome. Baron and Kenny’s (1986)
seminal article established the framework for this approach. Although this framework was
not originally developed to understand change processes, it has dominated the search for
processes of change for too long. When applied to treatment process, the traditional
mediational approach rests on a set of implausible assumptions about processes of change:
1) they are ergodic and thus are the same for all individuals, at the same level and sequence,
2) they are few in number and can be assessed a small number of times, 3) they are related
linearly to outcome and in a way that does not change over time, and 4) they do not enter
Author Manuscript

into dynamic, recursive, or self-organizing systems of relationships with context, or other

biopsychosocial variables.

All of these assumptions are bizarre when applied to therapeutic change. A number of
different variables influence the treatment process, in a person by person fashion, and these
variables are rarely uni-directionally related to outcome. They enter into context sensitive
idiographic dynamic relationships that change over time, and are at times recursive, self-
amplifying, and self-organizing. Traditional methodological approaches are simply inert in
the face of such assumptions, which may help explain the failure of these methods yet to
lead to a robust science of processes of change.

It is becoming clearer by the day that change processes need to be a focus of intervention
science. We have argued that new more idiographic methodological approaches will be
Author Manuscript

needed to do so in a way has plausible assumptions, and that provides the kind of knowledge
researchers and practitioners must have to implement a process-based approach. The needed
analytic and strategic changes are unlikely to occur overnight, but the simple step of placing
high-density, functional, and theoretically meaningful process assessment inside our existing
research studies provides an important beginning.

The present article offers a proposal to rethink our approach to studying treatment processes.
The main limitation of our argument is its very objective – this is a strategic proposal, not a
final answer to a complicated question. We believe that a dynamic network approach offers
the potential to study processes of change at the level of analysis at which such processes
legitimately exist, moving to nomothetic summaries of the role of multiple variables in
subgroups and populations only when we can do so without doing violence to analytic
Author Manuscript

assumptions. We would like to emphasize that many details still need to be worked out
before we can fully transition to this approach, such as how to hone in on the most clinically
relevant measures assessed with the right time-interval in gathering this data. But with such
data in hand it will be more possible to explore idiographic analytic strategies in a
theoretically catholic manner, and to link successful results to increasingly refined
theoretical models. There are a vast number of possible change processes to consider but
careful examination of existing mediational studies should be able to shrink the available

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 23

range of robust treatment process to more manageable set from which to generate the high
Author Manuscript

density longitudinal information needed.

It should be noted that other related factors might contribute to the slow progress in
intervention science, aside from those we have focused on in this article. For example, the
construct under investigation is often assessed very infrequently (typically only pre-post
assessments), few measures are transtheoretical, there is an overemphasis on single
processes and diagnostic categories, and there is a similar overemphasis on outcome
research.

Intervention science needs to become better able conceptually and methodologically to

address the challenge and opportunity of a process-based therapy era within evidence-based
care. Both the progress of our science and the ability to improve our patients’ lives
increasingly depends upon it.
Author Manuscript

Acknowledgments
Dr. Hofmann receives financial support from the Alexander von Humboldt Foundation (as part of the Humboldt
Prize), NIH/NCCIH (R01AT007257), NIH/NIMH (R01MH099021, U01MH108168), and the James S. McDonnell
Foundation 21st Century Science Initiative in Understanding Human Cognition - Special Initiative. He receives
compensation for his work as an advisor from the Palo Alto Health Sciences and for his work as a Subject Matter
Expert from John Wiley & Sons, Inc. and SilverCloud Health, Inc. He also receives royalties and payments for his
editorial work from various publishers.

Dr. Hayes receives financial support from NIH/NCCIH (R44AT006952). He also receives compensation for his
work as a content expert from New Harbinger Publications. He also receives royalties and payments for his editorial
work from various publishers.

Biography
Author Manuscript

Stefan G. Hofmann, Ph.D. is Professor of Psychology at Boston University, where he is the

Director of the Psychotherapy and Emotion Research Laboratory. He was born and raised in
Germany and has been living in Boston, USA, since 1996. He has published widely as an
author of more than 300 articles and 15 books. He is a Highly Cited Researcher by Thomson
Reuters and Clarivate Analytics. His research focuses on the mechanism of treatment
change, translating discoveries from neuroscience into clinical applications, emotion
regulation, and cultural expressions of psychopathology, especially anxiety disorders.
Because of this expertise, he served as an advisor to the DSM-5 Development Process and
was a member of the DSM-5 Anxiety Disorder Sub-Work Group. He is the recipient of
numerous awards, including the ABCT’s 2010 Outstanding Service Award, the Aaron T.
Beck Award for Excellence in Contributions to CBT by Assumption College, and the Aaron
T. Beck Award for Significant and Enduring Contributions to the Field of Cognitive Therapy
Author Manuscript

by the Academy of Cognitive Therapy. He is a fellow of the American Psychological

Association and the Association for Psychological Science. He has also been president of
numerous national and international professional societies, including the Association for
Behavioral and Cognitive Therapies (ABCT) and the International Association for Cognitive
Therapy (IACP). He is presently editor-in-chief of Cognitive Therapy and Research and
associate editor of Clinical Psychological Science. His research has been funded through
generous research grants from the National Institutes of Health private foundations.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 24

Steven C. Hayes is Nevada Foundation Professor in the Behavior Analysis program at the
Author Manuscript

Department of Psychology at the University of Nevada. An author of 44 books and nearly

600 scientific articles, his career has focused on an analysis of the nature of human language
and cognition and the application of this to the understanding and alleviation of human
suffering. He is the developer of Relational Frame Theory, an account of human higher
cognition, and has guided its extension to Acceptance and Commitment Therapy (ACT), a
popular evidence-based form of psychotherapy that uses mindfulness, acceptance, and
values-based methods. Dr. Hayes has been President of Division 25 of the APA, of the
American Association of Applied and Preventive Psychology, the Association for Behavioral
and Cognitive Therapy, and the Association for Contextual Behavioral Science. He was the
first Secretary-Treasurer of the Association for Psychological Science, which he helped form
and has served a 5 year term on the National Advisory Council for Drug Abuse in the
National Institutes of Health. In 1992 he was listed by the Institute for Scientific Information
Author Manuscript

as the 30th “highest impact” psychologist in the world and Google Scholar data ranks him
among the top ~1,500 most cited scholars in all areas of study, living and dead (http://
www.webometrics.info/en/node/58). His work has been recognized by several awards
including the Exemplary Contributions to Basic Behavioral Research and Its Applications
from Division 25 of APA, the Impact of Science on Application award from the Society for
the Advancement of Behavior Analysis, and the Lifetime Achievement Award from the
Association for Behavioral and Cognitive Therapy

References
Aderka IM, Nickerson A, Bøe HJ, & Hofmann SG (2012). Sudden Gains during psychological
treatments of anxiety and depression: A meta-analysis. Journal of Consulting and Clinical
Psychology, 80, 93–101. doi: 10.1037/a0026455. [PubMed: 22122290]
Ahn H-N, & Wampold BE (2001). Where oh where are the specific ingredients? A meta-analysis of
Author Manuscript

component studies in counseling and psychotherapy. Journal of Counseling Psychology, 48, 251–
257. doi: 10.1037/0022-0167.48.3.251
American Psychiatric Association (2013). Diagnostic and statistical manual of mental disorders (5th
ed) (DSM-5). Washington, DC: Author.
Almirall D, Compton SN, Gunlicks-Stoessel M, Duan N, & Murphy SA (2012). Designing a pilot
sequential multiple assignment randomized trial for developing an adaptive treatment strategy.
Statistics in Medicine, 31, 1887–1902. doi: 10.1002/sim.4512 [PubMed: 22438190]
Arch JJ, Wolitzky-Taylor KB, Eifert GH, & Craske MG (2012). Longitudinal treatment mediation of
traditional cognitive behavioral therapy and acceptance and commitment therapy for anxiety
disorders. Behaviour Research and Therapy, 50, 469–478. doi: 10.1016/j.brat.2012.04.007
[PubMed: 22659156]
Barlow DH & Hersen M (1984) Single-case experimental designs: Strategies for studying behavior
change (2nd ed.). New York: Pergamon.
Barlow DH, Nock M, & Hersen M (2009). Single case experimental designs: Strategies for studying
Author Manuscript

behavior change (3rd ed.). Boston, MA: Allyn & Bacon ISBN-13: 978–0205474554
Baron RM, & Kenny DA (1986). The moderator-mediator variable distinction in social psychological
research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social
Psychology, 51, 1173–1182. 10.1037/0022-3514.51.6.1173 [PubMed: 3806354]
Bell IH, Fielding-Smith SF, … Thomas N (2018) Smartphone-based ecological momentary assessment
and intervention in a coping-focused intervention for hearing voices (SAVVy): Study protocol for a
pilot randomised controlled trial. Trials, 19, 262. doi: 10.1186/s13063-018-2607-6 [PubMed:
29720208]

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 25

Blaauw FJ, van der Krieke L, Emerencia AC, Aiello M, & de Jonge P (2017). Personalized advice for
enhancing well-being using automated impulse response analysis---AIRA. arXiv preprint
Author Manuscript

arXiv:1706.09268
Blanken TF, Van Der Zweerde T, Van Straten A, Van Someren EJ, Borsboom D, & Lancee J (2019).
Introducing network intervention analysis to onvestigate sequential, symptom-specific treatment
effects: A demonstration in co-occurring insomnia and depression. Psychotherapy and
Psychosomatics, 88, 52–54. doi: 10.1159/000495045 [PubMed: 30625483]
Birkhoff GD (1931). Proof of the ergodic theorem. Proceedings of the National Academy of Sciences
of the United States of America, 17, 656–660. doi: 10.1073/pnas.17.2.656 [PubMed: 16577406]
Borsboom D, & Cramer AO (2013). Network analysis: an integrative approach to the structure of
psychopathology. Annual Review of Clinical Psychology, 9, 91–121. doi: 10.1146
Bos FM, Blaauw FJ, Snippe E, Van der Krieke L, de Jonge P, & Wichers M (2018). Exploring the
emotional dynamics of subclinically depressed individuals with and without anhedonia: an
experience sampling study. Journal of Affective Disorders, 228, 186–193. doi: 10.1016/
j.jad.2017.12.017 [PubMed: 29253685]
Boswell JF, Anderson LM, & Barlow DH (2014). An idiographic analysis of change processes in the
Author Manuscript

unified transdiagnostic treatment of depression. Journal of Consulting and Clinical Psychology, 82,
1060–1071. doi: 10.1037/a0037403 [PubMed: 25045911]
Boswell JF, & Bugatti M (2016). An exploratory analysis of the impact of specific interventions: Some
clients reveal more than others. Journal of Counseling Psychology, 63, 710–720. doi: 10.1037/
cou0000174 [PubMed: 27709971]
Bruijniks SJE, DeRubeis RJ, Hollon SD, & Huibers MJH (2019). The potential role of learning
capacity in cognitive behavior therapy for depression: A systematic review of the evidence and
future directions for improving therapeutic learning. Clinical Psychological Science, 7, 668–692.
doi: 10.1177/2167702619830391
Castonguay LG, & Beutler LE (2005). Principles of therapeutic change that work. Oxford University
Press.
Chambless DL, & Hollon SD (1998). Defining empirically supported therapies. Journal of Consulting
and Clinical Psychology, 66, 7–18. doi: 10.1037/0022-006X.66.1.7 [PubMed: 9489259]
Chambless DL, & Ollendick TH (2001). Empirically supported psychological interventions:
Controversies and evidence. Annual Review of Psychology, 52, 685–716. 0.1146/
Author Manuscript

annurev.psych.52.1.685
Epskamp S, van Borkulo CD, van der Veen DC, Servaas MN, Isvoranu AM, Riese H, & Cramer AO
(2019). Personalized network modeling in psychopathology: The importance of contemporaneous
and temporal connections. Clinical Psychological Science, 6, 416–427. doi:
10.1177%2F2167702617744325
Epskamp S, Waldorp LJ, Mõttus R, & Borsboom D (2018). The gaussian graphical model in cross-
sectional and time-series data. Multivariate Behavioral Research. doi:
10.1080/00273171.2018.1454823
Faldowski RA (2009). Rediscovering the individual in mediation analyses. Integrative Psychological
and Behavioral Science, 43, 253. doi:10.1007/s12124-009-9099-6 [PubMed: 19543948]
Fernandez KC, Fisher AJ, & Chi C (2017). Development and initial implementation of the Dynamic
Assessment Treatment Algorithm (DATA). PLOS ONE, 12(6): e0178806. doi:10.1371/
journal.pone.0178806 [PubMed: 28654667]
Fisher AJ (2015) Toward a dynamic model of psychological assessment: Implications for personalized
Author Manuscript

care. Journal of Consulting and Clinical Psychology, 83, 825–836. doi: 10.1037/ccp0000026.
[PubMed: 26009781]
Fisher AJ, Bosley HG, Fernandez KC, Reeves JW, Soyster PD, Diamond AE, & Barkin J (2019). Open
trial of a personalized modular treatment for mood and anxiety. Behaviour Research and Therapy,
116, 69–79. doi: 10.1016/j.brat.2019.01.010 [PubMed: 30831478]
Fisher AJ, Newman MG, & Molenaar PCM (2011). A quantitative method for the analysis of
nomothetic relationships between idiographic structures: Dynamic patterns create attractor states
for sustained post-treatment change. Journal of Consulting and Clinical Psychology, 79(4), 552–
563. doi:10.1037/a0024069 [PubMed: 21707138]

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 26

Follette WC, Naugle AE, & Linneroth PJN (1999). Functional alternatives to traditional assessment
and diagnosis In Dougher MJ (Ed.), Clinical behavior analysis. Reno, NV: Context Press.
Author Manuscript

Fredrickson BL (2000). What good are positive emotions? Review of General Psychology, 2, 300–319.
doi: 10.1037/1089-2680.2.3.300
Friman PC (2010). Cooper, Heron, and Heward’s Applied Behavior Analysis: Checkered flag for
students and professors, yellow flag for the field. Journal of Applied Behavior Analysis, 43, 161–
174. doi:10.1901/jaba.2010.43-161
Gao L, Curtiss J, Liu X, & Hofmann SG (in press). Differential treatment mechanisms in mindfulness
meditation and progressive muscle relaxation. Mindfulness.
Gates KM, Molenaar PC, Hillary FG, Ram N, & Rovine MJ (2010). Automatic search for fMRI
connectivity mapping: an alternative to Granger causality testing using formal equivalences among
SEM path modeling, VAR, and unified SEM. NeuroImage, 50, 1118–1125. doi: 0.1016/
j.neuroimage.2009.12.117 [PubMed: 20060050]
Gates KM, & Molenaar PCM (2012). Group search algorithm recovers effective connectivity maps for
individuals in homogeneous and heterogeneous samples. NeuroImage, 63, 310–319. doi: 10.1016/
j.neuroimage.2012.06.026 [PubMed: 22732562]
Author Manuscript

Goldfried MR (1980). Toward the delineation of therapeutic change principles American Psychologist,
35, 991–999. 10.1037/0003-066X.35.11.991. [PubMed: 7436119]
Goldin PR, Morrison A, Jazaieri H, Brozovich F, Heimberg R, & Gross JJ (2016). Group CBT versus
MBSR for social anxiety disorder: A randomized controlled trial. Journal of Consulting and
Clinical Psychology, 84, 427–437. doi: 10.1037/ccp0000092 [PubMed: 26950097]
Ghaderi A (2006). Does individualization matter? A randomized trial of standardized (focused) versus
individualized (broad) cognitive behavior therapy for bulimia nervosa. Behaviour Research and
Therapy, 44, 273–288. doi:10.1016/j.brat.2005.02.004 [PubMed: 16389065]
Greenberg LS (1986). Change process research. Journal of Consulting and Clinical Psychology, 54, 4–
9. doi: 10.1037/0022-006X.54.1.4 [PubMed: 3958300]
Hallford DJ, Sharma MK, & Austin DW (2019). Increasing anticipatory pleasure in major depression
through enhancing episodic future thinking: A randomized single-case series trial. 10.31234/osf.io/
9uy42
Hayes AF (2009). Beyond Baron and Kenny: Statistical mediation analysis in the new millennium.
Author Manuscript

Communication Monographs, 76, 408 420. doi: 10.1080/03637750903310360

Hayes AF, & Preacher KJ (2014). Statistical mediation analysis with a multicategorical independent
variable. British Journal of Mathematical and Statistical Psychology, 67, 451–470. doi: 10.1111/
bmsp.12028 [PubMed: 24188158]
Hayes AM, Feldman GC, Beevers CG, Laurenceau J-P, Cardaciotto L, & Lewis-Smith J (2007).
Discontinuities and cognitive changes in an exposure-based cognitive therapy for depression.
Journal of Consulting and Clinical Psychology, 75(3), 409–421. Doi: 10.1037/0022-006X.75.3.409
[PubMed: 17563158]
Hayes AM, Laurenceau J-P, Feldman G, Strauss JL, Cardaciotto LA (2007). Change is not always
linear: The study of nonlinear and discontinuous patterns of change in psychotherapy. Clinical
Psychology Review, 27, 715–723. doi: 10.1016/j.cpr.2007.01.008 [PubMed: 17316941]
Hayes AM, Yasinski C (2015). Pattern destabilization and emotional processing in cognitive therapy
for personality disorders. Frontiers in Psychology, 6, 107. doi: 10.3389/fpsyg.2015.00107
[PubMed: 25755647]
Hayes AM, Yasinski C, Ben Barnes J & Bocking CL (2015). Network destabilization and transition in
Author Manuscript

depression: New methods for studying the dynamics of therapeutic change. Clinical Psychology
Review, 41, 27–39. doi: 10.1016/j.cpr.2015.06.007. [PubMed: 26197726]
Hayes SC, Long DM, Levin ME, & Follette WC (2013). Treatment development: Can we find a better
way? Clinical Psychology Review, 33, 870–882. doi:10.1016/j.cpr.2012.09.009 [PubMed:
23647855]
Hayes SC & Hofmann SG (Eds.) (2018). Process-based CBT: The science and core clinical
competencies of cognitive behavioral therapy. Oakland, CA: New Harbinger Publications
ISBN-13: 978–1626255968.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 27

Hayes SC & Hofmann SG (2017). The third wave of CBT and the rise of process-based care. World
Psychiatry, 16, 245–246. doi: 10.102/wps.20442 [PubMed: 28941087]
Author Manuscript

Hayes SC, Hofmann SG, & Ciarrochi J (in press). Creating an alternative to syndromal diagnosis:
Needed features of processes of change and the models that organize them. Chapter to appear in
Hofmann SG & Hayes SC (Eds.), Beyond the DSM: Toward a process-based alternative for
diagnosis and mental health treatment. Oakland, CA: Context Press / New Harbinger Publications.
Hayes SC, Hofmann SG, Stanton CE, Carpenter JK, Sanford BT, Curtiss JE, & Ciarrochi J (2019). The
role of the individual in the coming era of process-based therapy. Behaviour Research and
Therapy, 117, 40–53. doi: 10.1016/j.brat.2018.10.005 [PubMed: 30348451]
Haynes SN & O’Brien WH (1990). Functional analysis in behavior therapy. Clinical Psychology
Review, 10, 649–668. 10.1016/0272-7358(90)90074-K
Hofmann SG, Curtiss J, & McNally RJ (2016). A complex network perspective on clinical science.
Perspectives on Psychological Science, 11, 597–605. doi: 10.1177/1745691616639283 [PubMed:
27694457]
Hofmann SG, & Hayes SC (2019). The future of intervention science: Process-based therapy. Clinical
Psychological Science, 7, 37–50. doi: 10.1177/2167702618772296 [PubMed: 30713811]
Author Manuscript

Hofmann SG & Hayes SC (2019). Functional analysis is dead: Long live functional analysis. Clinical
Psychological Science, 7, 63–67. doi: 10.1177/2167702618805513 [PubMed: 30713812]
Hoge EA, Bui E, … & Simon NM (2014). Change in decentering mediates improvement in anxiety in
mindfulness-based stress reduction for generalized anxiety disorder. Cognitive Therapy and
Research, 39, 228–235. doi:10.1007/s10608-014-9646-4 [PubMed: 28316355]
Holmbeck GN (1997). Toward terminological, conceptual, and statistical clarity in the study of
mediators and moderators: Examples from the child-clinical and pediatric psychology literatures.
Journal of Consulting and Clinical Psychology, 4, 599–610. 10.1037/0022-006X.65.4.599
Hurl K, Wightman J, Haynes SN, & Virues-Ortega J (2016). Does a pre-intervention functional
assessment increase intervention effectiveness? A meta-analysis of within-subject interrupted time-
series studies. Clinical Psychology Review, 47, 71–84. doi:10.1016/j.cpr.2016.05.003 [PubMed:
27372279]
Ilardi SS, & Craighead WE (1994). The role of nonspecific factors in cognitive- behavior therapy for
depression. Clinical Psychology: Science and Practice, 1, 138–156. do: 10.1111/
j.1468-2850.1994.tb00016.x
Author Manuscript

James LR, & Brett JM (1984). Mediators, moderators and tests for mediation. Journal of Applied
Psychology, 69, 307–321. doi:10.1037/0021-9010.69.2.307
Judd CM, & Kenny DA (1981). Process analysis: Estimating mediation in treatment evaluations.
Evaluation Review, 5, 602–619. doi:10.1177/0193841X8100500502
Kazdin AE (2007). Mediators and mechanisms of change in psychotherapy research. Annual Review
of Clinical Psychology, 3, 1–27. doi:10.1146/annurev.clinpsy.3.022806.091432
Kazdin AE (2016). Single-case experimental research designs In Kazdin AE (Ed.), Methodological
issues and strategies in clinical research (pp. 459–483). Washington, DC, US: American
Psychological Association ISBN-13: 978–1433820915
Kazdin AE, & Nock MK (2003). Delineating mechanisms of change in child and adolescent therapy:
methodological issues and research recommendations. Journal of Child Psychology and
Psychiatry, 44, 1116–1129. doi:10.1111/1469-7610.00195 [PubMed: 14626454]
Kelley K, & Preacher KJ (2012). On effect size. Psychological Methods, 17, 137–152. doi: 10.1037/
a0028086 [PubMed: 22545595]
Author Manuscript

Kraemer HC, Wilson T, Fairburn CG, & Agras WS (2002). Mediators and moderators of treatment
effects in randomized clinical trials. Archives of General Psychiatry, 59, 877–883. doi:10.1001/
archpsyc.59.10.877 [PubMed: 12365874]
Lane S, Gates K, Pike H, Beltz A, & Wright AG (2019). Uncovering general, shared, and unique
temporal patterns in ambulatory assessment data. Psychological Methods, 24, 54–69. doi: 10.1037/
met0000192 [PubMed: 30124300]
Liu S & Molenaar P (2016). Testing for Granger causality in the frequency domain: A phase
resampling method. Multivariate Behavioral Research, 51, 53–55. doi:
10.1080/00273171.2015.1100528. [PubMed: 26881957]

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 28

Lorenzo-Luaces L, German RE, & DeRubeis RJ (2015). It’s complicated: The relation between change
procedures, cognitive change, and symptom change in cognitive therapy for depression. Clinical
Author Manuscript

Psychology Review, 41, 3–15. doi: 10.1016/j.cpr.2014.12.003 [PubMed: 25595660]

Lutz W, Schwartz B, Hofmann SG, Fisher AJ, Husen K, & Rubel JA (2018). Using network analysis
for the prediction of treatment dropout in patients with mood and anxiety disorders: A
methodological proof-of-concept study. Scientific Reports. 8:7819. doi: 10.1038/
s41598-018-25953-0 [PubMed: 29777110]
MacKinnon DP, Lockwood CM, Hoffman JM, West SG, & Sheets V (2002). A comparison of methods
to test mediation and other intervening variable effects. Psychological Methods, 7, 83–104.
doi:10.1037/1082-989X.7.1.83 [PubMed: 11928892]
Miller FG, & Lee DL (2013). Do functional behavioral assessments improve intervention effectiveness
for students diagnosed with ADHD? A single-subject meta-analysis. Journal of Behavioral
Education, 22, 253–282. doi:10.1007/s10864-013-9174-4
Molenaar PCM (2008a). On the implications of the classical ergodic theorems: Analysis of
developmental processes has to focus on intra-individual variation. Developmental Psychobiology,
50, 60–69. doi: 10.1002/dev.20262 [PubMed: 18085558]
Author Manuscript

Molenaar P (2008b). Consequences of the ergodic theorems for classical test theory, factor analysis,
and the analysis of developmental processes In Hofer SM & Alwin DF (Eds.), Handbook of
cognitive aging: Interdisciplinary perspectives (pp. 90–104). Thousand Oaks, CA: SAGE. doi:
10.4135/9781412976589.n5
Molenaar P (2018). Granger causality testing with intensive longitudinal data. Prevention Science, 20,
442–451. doi: 10.1007/s11121-018-0919-0.
Moscovitch DA, Hofmann SG, Suvak MK, & In-Albon T (2005). Mediation of changes in anxiety and
depression during treatment of social phobia. Journal of Consulting and Clinical Psychology, 73,
945–952. doi: 10.1037/0022-006X.73.5.945 [PubMed: 16287394]
Nelson B, McGorry PD, Wichers M, Wigman JT, & Hartmann JA (2017). Moving from static to
dynamic models of the onset of mental disorder: a review. JAMA Psychiatry, 74, 528–534. doi:
10.1001/jamapsychiatry.2017.0001 [PubMed: 28355471]
Nock MK (2007). Conceptual and design essentials for evaluating mechanisms of change. Alcoholism:
Clinical and Experimental Research, 31 (Suppl. 3), 4S–12S. doi:10.1111/
j.1530-0277.2007.00488.x
Author Manuscript

Papandonatos GD, Williams DM … & Marcus BH (2012). Mediators of physical activity behavior
change: Findings from a 12-month randomized controlled trial. Health Psychology, 31, 512–520.
doi: 10.1037/a0026667 [PubMed: 22149130]
Paul GL (1967). Strategy of outcome research in psychotherapy. Journal of Consulting Psychology, 31,
109–118. doi: 10.1037/h0024436. [PubMed: 5342732]
Paul GL (1969). Behavior modification research: design and tactics In Franks CM (Ed.), Behavior
therapy: Appraisal and status (pp. 29–62). New York: McGraw-Hill.
Rossouw TI & Fonagy P (2012). Mentalization-based treatment for self-harm in adolescents: A
randomized controlled trial. Journal of the American Academy of Child and Adolescent
Psychiatry, 51, 1304–1313.e3. doi: 10.1016/j.jaac.2012.09.018. [PubMed: 23200287]
Russell B (1912). On the notion of cause. Proceedings of the Aristotelian Society, 13, 1–26.
Scheffer M, Bascompte J, Brock WA, Brovkin V, Carpenter SR, Dakos V, … & Sugihara G (2009).
Early-warning signals for critical transitions. Nature, 461, 53–59. doi: 10.1038/nature08227
[PubMed: 19727193]
Author Manuscript

Scheffer M, Carpenter SR, Lenton TM, Bascompte J, Brock W, Dakos V, … & Pascual M (2012).
Anticipating critical transitions. Science, 338, 344–348. doi: 10.1126/science.1225244 [PubMed:
23087241]
Schiepek G, Aichhorn W, Gruber M, Strunk G, Bachler E & Aas B (2016) Real-Time Monitoring of
Psychotherapeutic Processes: Concept and Compliance. Frontiers in Psychology, 7, 604. doi:
10.3389/fpsyg.2016.00604 [PubMed: 27199837]
Schiepek G, & Strunk G (2010). The identification of critical fluctuations in phase transitions in short
term and coarse-grained time series-a method for the real-time monitoring of human change

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 29

processes. Biological Cybernetics, 102, 197–207. doi: 10.1007/s0042200903621 [PubMed:

20084517]
Author Manuscript

Sidman M (1960). Tactics of scientific research. Oxford, England: Basic Books.doi: 10.1126/
science.1225244
Sobel ME (1982). Asymptotic confidence intervals for the indirect effects in structural equations
models In Leinhart S (Ed.), Sociological methodology (pp. 290–321). San Francisco: Jossey-Bass.
Stiles WB, Honos-Webb L & Surko M (2006). Responsiveness in psychotherapy. Clinical Psychology:
Science and Practice, 5, 439–458. doi: 10.1111/j.1468-2850.1998.tb00166.x
Tang TZ, & DeRubeis RJ (1999). Sudden gains and critical sessions in cognitive-behavioral therapy
for depression, Journal of Consulting and Clinical Psychology, 67, 894–904. doi:
10.1037//0022-006x.67.6.894 [PubMed: 10596511]
Tryon WW (in press). Mediators and mechanisms. Clinical Psychological Science. doi:
10.1177/212167702618765791
van de Leemput IA, Wichers M, Cramer AO, Borsboom D, Tuerlinckx F, Kuppens P, … & Derom C
(2014). Critical slowing down as early warning for the onset and termination of depression.
Proceedings of the National Academy of Sciences, 111(1), 87–92. doi: 10.1073/pnas.1312114110
Author Manuscript

Vilardaga R, Hayes SC, Atkins DC, Bresee C & Kambiz A (2013). Comparing experiential acceptance
and cognitive reappraisal as predictors of functional outcome among individuals with psychotic
symptoms. Behaviour Research and Therapy, 51, 425–433. 10.1016/j.brat.2013.04.003 [PubMed:
23747581]
Volkovysskii RL & Sinai YG (1971). Ergodic properties of an ideal gas with an infinite number of
degrees of freedom. Functional Analysis and Its Applications (Funktsional’nyi Analiz i Ego
Prilozheniya), 5(3), 185–185.
von Eye A, & Bergman L (2003). Research strategies in developmental psychopathology: Dimensional
identity and the person-oriented approach. Development and Psychopathology, 15, 553–580.
doi:10.1017/S0954579403000294 [PubMed: 14582932]
von Eye A, Mun EY, & Mair P (2009). What caries a mediation process? Configural analysis of
mediation. Integrative Psychological & Behavioral Science, 43, doi: 10.1007/s12124-009-9088-9.
Weisz JR, Kuppens S … Corteselli KA (2018). Are psychotherapies for young people growing
stronger? Tracking trends over time for youth anxiety, depression, attention-deficit/hyperactivity
Author Manuscript

disorder, and conduct problems. Perspectives on Psychological Science. doi:

10.1177/1745691618805436.
Wichers M, & Groot PC (2016). Critical slowing down as a personalized early warning signal for
depression. Psychotherapy and Psychosomatics, 85, 114–116.doi: 10.1159/000441458 [PubMed:
26821231]
Wilson DS, Hayes SC, Biglan T, & Embry D (2014). Evolving the future: Toward a science of
intentional change. Behavioural and Brain Sciences, 34, 395–416. doi:10.1017/
S0140525X13001593
Woolfolk RL (2015). Clinical Trials in Psychiatry and Clinical Psychology: Science or Product
Testing? Acta Psychopathologica, 1, 1–12. doi: 10.4172/2469-6676.100012
Yoman J (2008). A primer on functional analysis. Cognitive and Behavioral Practice, 15, 325–340.
doi:10.1016/j.cbpra.2008.01.002
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 30

Highlights
Author Manuscript

Linear Mediation is inadequate to study treatment processes

Treatment processes are dynamic

Dynamic network processes are more appropriate to study mediation

Author Manuscript
Author Manuscript
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 31
Author Manuscript

Figure 1.
Traditional Mediation Model
Reprinted with permission from Baron, R. M., & Kenny, D. A. (1986). The moderator-
mediator variable distinction in social psychological research: Conceptual, strategic, and
statistical considerations. Journal of Personality and Social Psychology, 51, 1173–1182.
doi.org/10.1037/0022-3514.51.6.1173, p. 1176
Author Manuscript
Author Manuscript
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 32
Author Manuscript
Author Manuscript

Figure 2.
Simple Mediation Model of Sam
Note: Example of a traditional mediation model based on Baron & Kenny (1986). The
circles (nodes) represent different problems Sam experiences. The arrows (edges)
connecting the nodes are directed, indicating how they are temporally and “causally”
connected.
Author Manuscript
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 33
Author Manuscript
Author Manuscript
Author Manuscript

Figure 3.
Network Model of Sam
Note: Example of an individual network model. Here, the traditional mediation model only
occupies one part of the entire network. In addition to rumination, low self-esteem and
loneliness are considered. These two nodes might be even more central to Sam’s problem
space than rumination. The thickness of the borders around the nodes indicate how much
they contribute to the entire problem space, and the thickness of the edges indicate the
relative strength of the influence that one node has on the other.
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 34
Author Manuscript
Author Manuscript
Author Manuscript

Figure 4.
Network Model of Sam During Treatment
Note: Example of a network that includes an adaptive element (i.e., self-compassion). Black
borders and edges are maladaptive, whereas the gray border and edges are adaptive. The
thickness of the border around self-compassion indicates that it is a relatively important
adaptive node with a strong influence on self-esteem, as shown by the thick edge and a
moderating influence on the relationship between rumination and low mood.
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 35
Author Manuscript

Figure 5.
Change in network structure over trime.
Author Manuscript

Note: Change in network structure over the course of the antidepressant discontinuation
period and prior to the day of a transition to a depressive episode on day 128. The dashed
arrows indicate negative, the solid arrows positive time-lagged effects. AD: antidepressant;
PA: positive affect; na: negative affect; MU: mental unrest; WO: worry; SU: suspicious.
Author Manuscript
Author Manuscript

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.
 Hofmann et al. Page 36
Author Manuscript
Author Manuscript
Author Manuscript

Figure 6.
Experimental design (upper panel), raw weekly measurements of depressive symptoms
using the SCL-90-R depression subscale (second panel), the moving window (width = 30
days) in autocorrelation (third panel), and variance results (fourth panel) based on the
detrended overall mental state variable (total mental state score based on a moving window
Author Manuscript

over time). Reprinted with permission from Wichers and Groot (2016).
Note: Changes in depressive symptoms, autocorrelation, and variance of the mental states
ratings. The grey zone shows the period immediately before the transition showing the
typical pattern of critical slowing of the system.

Clin Psychol Rev. Author manuscript; available in PMC 2021 March 01.