Trajectories of Change in Psychotherapy

Jesse Owen,1 Jill Adelson,2 Stephanie Budge,3 Bruce Wampold,4 Mark Kopta,5
T. Minami,6 and Scott Miller7
1
University of Denver
2
University of Louisville
3
3University of Wisconsin, Madison
4
University of Wisconsin, Madison, Modum Bad Clinic, Norway
5
University of Evansville
6
University of Massachusetts, Boston
7
International Center for Clinical Excellence
Objective: The current study used multilevel growth mixture modeling to ascertain groups of
patients who had similar trajectories in their psychological functioning over the course of short-term
treatment. Method: A total of 10,854 clients completed a measure of psychological functioning
before each session. Psychological functioning was measured by the Behavioral Health Measure, which
is an index of well-being, symptoms, and life-functioning. Clients who attended 5 to 25 sessions at 46
different university/college counseling centers and one community mental health center were included
in this study. Client diagnoses and the specific treatment approaches were not known. Results: A
3-class solution was a good fit to the data. Clients in classes 1 and 3 had moderate severity in their initial
psychological functioning scores, and clients in class 2 had more distressed psychological functioning
scores. The trajectory for clients in class 1 was typified by early initial change, followed by a plateau,
and then another gain in psychological functioning later in treatment. The trajectory for clients in class
2 demonstrated an initial decrease in functioning, followed by a rapid increase, and then a plateau.
Last, the clients in class 3 had a steady increase of psychological functioning, in a more linear manner.
Conclusion: The trajectories of change for clients are diverse, and they can ebb and flow more than
traditional dose-effect and good-enough level models may suggest.  C 2015 Wiley Periodicals, Inc. J.

Clin. Psychol. 71:817–827, 2015.

Keywords: therapy outcomes; growth mixture modeling; dose effect; good-enough level

The discussion of trajectories of client change in psychotherapy has a long history in the
literature, which is quite appropriate because these trajectories have implications for theory,
practice, and policies. For example, many counseling centers have session limit policies based
on, in part, the empirical data regarding the necessary dosage needed for the majority of clients
to make meaningful change (Lambert & Ogles, 2004; Owen, Levadas, & Rodolfa, 2007).
Additionally, the trajectories can inform theories on how clients change. If trajectories indi-
cate that clients often get worse before they get better, then theoretical explanations for this are
called for. Moreover, trajectories of change inform expectations for the magnitude and timing of
change over the course of treatment. In particular, there are several client outcome monitoring
systems that track clients’ psychological functioning over the course therapy and provide infor-
mation to therapists when their clients are progressing (or not) compared to norms or average
trajectories of change (e.g., Lambert & Shimokawa, 2011; Miller, Duncan, Sorrell, & Brown,
2005; Pinsof, Goldsmith, & Latta, 2012). When clients are not progressing as expected, thera-
pists are encouraged to address alliance issues, change their approach to treatment, or address
other external factors that might be negatively affecting treatment. Accordingly, these systems
rely on accurate information about how clients change over the course of treatment, yet there
are several gaps in the ways the field has examined trajectories of change.

Please address correspondence to: Jesse Owen, Room 320, Education and Counseling Psychology Depart-
ment College of Education and Human Development, University of Louisville, Louisville KY 40292. E-mail:
[email protected]

JOURNAL OF CLINICAL PSYCHOLOGY, Vol. 71(9), 817–827 (2015) 

C 2015 Wiley Periodicals, Inc.

Published online in Wiley Online Library (wileyonlinelibrary.com/journal/jclp). DOI: 10.1002/jclp.22191

818 Journal of Clinical Psychology, September 2015

Dose Effect and Good-Enough Level Models

Two primary models have been proposed to describe clients’ rate of change in psychotherapy:
the dose-effect and the good-enough level (GEL) models. The dose-effect model is founded on
the notion that each session is a “dose,” akin to taking a medication. Therefore, it assumes that
the benefit clients will receive is primarily dependent on the total number of sessions. Modeling
change in psychotherapy this way, the relationship between the number of sessions and outcome
appear to fit a negatively accelerating curve (often modeled log-linearly), wherein change occurs
more rapidly early in treatment and then plateaus (e.g., Howard, Kopta, Kruause, & Orlinsky,
1986; Lutz, Lowry, Kopta, Einstein, & Howard, 2001; Lutz, Martinovich, & Howard, 1999).
In contrast, proponents of the GEL model posit that the benefit that clients receive from
therapy is less dependent on the number of sessions, resulting in significant variability in how
long a client will stay in therapy. Whereas the dose-effect model supposes that the number of
sessions is crucial in determining the client’s benefit, the GEL model assumes that clients stay in
therapy until they feel good enough. This model also has empirical support and, compared to
the dose effect model, has typically captured clients’ trajectories of change better (e.g., Baldwin,
Berkeljon, Atkins, Olsen, & Nielsen, 2009; Barkham et al., 2006; Reese, Toland, & Hopkins,
2011; Stiles, Barkham, Connell, & Mellor-Clark, 2008). However, note that this is largely a
modeling artifact between the two models; given that the dose-effect model traditionally aims
to create a single trajectory based on the number of sessions, it lacks the flexibility implied by
the GEL model, which allows as many trajectories as it deems appropriate based on the data.
Pertinent to our study, two notable findings from these studies have emerged. First, from the
dose-effect studies, the log-linear model was a good fitting model for clients who attended three
to five sessions, but it was a poor fit for clients who attended more than five sessions (Stulz,
Lutz, Kopta, Minami, & Saunders, 2013). Accordingly, there might be alternative models for
trajectories of change for those who attend more than five sessions. Second, from the GEL
studies, the rates of change significantly varied across clients (e.g., Baldwin et al., 2009; Owen,
Adelson, Budge, Kopta, & Reese, 2014). This suggests that using the number of sessions as a
moderator for clients’ rate of change reflects only a part of the story. There are likely other types
of change patterns that could account for this variability.

Multiple Trajectories
Clinical wisdom would suggest that single trajectories, whether or not they are based on a
number of sessions, are not likely to be representative of all clients. Rather, there is likely to be
more complexity in the trajectories of change, resulting in distinct groups of clients who have
different trajectories. A relatively new approach for modeling growth trajectories, growth mixture
modeling (GMM), generates classes or groups of clients who have similar profiles of change
(Muthén, 2004; Vermunt, Tran, & Magidson, 2008). Several studies have used this approach to
detect patterns in change trajectories. For example, across three randomized clinical trials for the
treatment of major depressive disorders (N = 162; Lutz, Stulz, & Köck, 2009), chronic depression
(N = 504; Stulz, Thase, Klein, Manber, & Crits-Christoph, 2010), and cocaine dependency (N =
346; Stulz, Gallop, Lutz, Wrenn, & Crits-Christoph, 2010), there were three groups who differed
in their trajectories and initial level of symptomology. Two groups demonstrated moderate rates
of change but differed in the level of their initial symptomology; the last group demonstrated
rapid initial rates of change and had moderate to severe initial symptomology.
Consistent with these results, in a naturalistic study of 1,116 clients, Stulz and Lutz (2007)
found three distinct groups when modeling initial severity and rates of change. The first group
demonstrated moderate initial severity and rates of change following a log-linear pattern. The
second likewise reported moderate initial severity but demonstrated more rapid rates of change.
Finally, the last group had high initial severity and also demonstrated rapid rates of change.
Taken together, these studies have set a strong foundation for a new wave of statistical procedures
aimed at enhancing the understanding of client change patterns in psychotherapy.
This analytical approach requires large sample sizes for each class and over the course of
treatment to provide stable estimates (Vermont et al., 2008). Indeed, smaller sample sizes could
 International Center for Clinical Excellence 819

mask the trajectories of some groups of clients. Accordingly, the current study employs a large
naturalistic sample of clients, all of whom attended 5 to 25 sessions. Those who attend fewer ses-
sions have rates of change that are adequately modeled by log-linear and linear models (Baldwin
et al., 2009; Stulz et al., 2013). Moreover, longer term therapy cases (> 25 sessions) would require
larger samples of clients who attended more sessions to adequately model their trajectories.
Last, we test whether the classes differed in their initial psychological functioning scores.

Method
Participants
The sample included a total of 10,854 clients out of an original 38,985 clients. We first excluded
clients who attended only one session because no trajectory could be formed for these clients
(n = 13,850). We also excluded clients who attended four or fewer (n = 13,076) because the
trajectory for up to four sessions is quite simple and best modeled via log-linear models (Stulz
et al., 2013). Given issues with convergence we also excluded those clients who attended over
25 sessions (n = 1205); this group represents 3% of the sample and the upper range of sessions
was 167 sessions, which made the stability of the estimates for longer term therapy uncertain.
Clients were included if they did not have a session for over 90 days; that is, clients who had not
attended therapy for three months were determined to be completed or dropped out of therapy.
This cutoff is consistent with previous studies on trajectories of change in psychotherapy (e.g.,
Baldwin et al., 2009). Unfortunately, we did not have clients’ or therapists’ ratings for how
therapy ended (e.g., dropout, mutual decision).
The majority of the sample participants were female (61.8%, n = 6,706), 31.5% (n = 3,424)
were male, and 6.7% (n = 724) did not report their gender. Only 8,162 (75.2%) clients re-
ported their ethnicity, and of those who did, the following demographics were reported: White
67.7% (n = 5,526); Asian/Pacific Islander 8.0% (n = 651); African American 5.5% (n = 450);
Latina/Latino/Hispanic 5.1% (n = 414); Native American/American Indian 0.1% (n = 67);
multiracial/ethnic 2.3% (n = 186); and “other” 10.6% (n = 868). For the entire sample, the
average number of sessions was 9.41 (standard deviation [SD] = 4.39, median = 8).
A total of 513 therapists were included in the analysis. Although detailed assessments of
the therapists were not conducted, the majority of them were female (71%) and White (72%).
Therapists included psychologists, counselors, psychiatrists, and social workers. The data came
from 47 different treatment centers, with an average of 231 clients per center. Nearly all of the
centers were university/college counseling centers (n = 46 of 47). No diagnostic information
was collected from the clients or therapists.

Measure
Behavioral Health Measure-20 (BHM-20; Kopta & Lowry, 2002). The BHM-20 is
a 21-item self-report clinical outcome measure and comprises three main clusters of items; Well-
Being, Psychological Symptoms, and Life Functioning. The Well-Being cluster includes three
items that assess overall distress, life satisfaction, and motivation (e.g., “How satisfied have you
been with your life?”). The Psychological Symptoms cluster contains 14 items that assess clinical
symptoms such as depression, anxiety, and substance use (e.g., “Feeling sad most of the time”).
The Life Functioning cluster contains four items that assess relationships, life enjoyment, and
work/school functioning.
For purposes of this study, we used the General Mental Health (GMH) index, which is
the average score of all items except for the 21st item that is answered only by clients who
reported more than minimal suicidal tendencies. All items are scored on a 5-point Likert-type
scale ranging from 0 to 4, with higher scores indicating better functioning. The BHM scores
were converted from a 0 to 4 scale to a 0 to 400 scale. Concurrent validity of the BHM-20 has
been supported by significant associations with a variety of psychological functioning measures
(Kopta & Lowry, 2002). In the current study, Cronbach’s alpha was .92.
820 Journal of Clinical Psychology, September 2015

Procedure
Analyses for this study were conducted using data from clients at 46 university counseling centers
and one community mental health center across the United States. Clients completed the BHM
via a computer-based system (CelestHealth System-MH; Bryan, Kopta, & Lowes, 2012; Kopta
et al., in press), which derives electronic input from the client and provides an electronic output
that is available to the psychotherapist. Prior to each session, clients completed the BHM-20.
Clients were informed that they were being administered the measures in order to help inform
future research and provide feedback to the therapists and all data was de-identified.

Data Analysis
To examine whether there were groups of patients who had different trajectories of change, we
conducted multilevel GMMs. Briefly, GMM is analogous to a cluster analysis, in which classes
(i.e., groups) of individuals are identified based on the similarities of their scores or, in this case,
based on their intercept scores and trajectories of change. GMM results in each class having a
unique intercept and slope. Technical explanations of GMM are beyond the scope of this article
and interested readers can find useful descriptions elsewhere (e.g., Muthén, 2004; Vermont et al.,
2008).
In our models, we accounted for the nested nature of our data (i.e., observations nested within
clients who were nested within therapists).1 To develop the trajectories and respective latent
classes, we modeled linear, quadratic, and cubic rates of change. We estimated only the latent
classes at the client level; however, the between-therapist latent class model did not converge,
nor were we able to model random effects at the therapist level due to model nonconvergence.2
We conducted the models using maximum likelihood estimation and Mplus version 7.0 software
(Muthén & Muthén, 1998–2012; further information regarding the models can be found in
footnote 2 and additional information can be requested by contacting the lead author).
For comparisons, we started with two classes and added an additional class until the
likelihood-based test (i.e., Lo-Mendell-Rubin or LMR) was not statistically significant, sug-
gesting that adding more classes is not value added. As Nylund, Asparouhov, and Muthén
(2007) noted, “the first time the p value of the LMR is nonsignificant might be a good indication
to stop increasing the number of classes” (p. 563). The LMR is typically an upper bound estimate
of number of classes, because it typically favors more models with more classes (Nylund et al.,
2007). We also used information criteria estimates, in particular Akaike’s information criteria
(AIC) and the adjusted Bayesian information criteria (BIC), to gauge the value of respective
models. For both the AIC and the adjusted BIC, lower estimates are preferred (McCoach &
Black, 2008). Last, we also examined the models for the practical meaning of the classes.

Results
We initially tested the models with two, three, and four classes.3 The LMR test compar-
ing the two- and three-class solutions was statistically significant, suggesting that the three-
class solution was superior to the two-class model (p < .001). Accordingly, we compared the

1 We were not able to model the third level (counseling center) due to nonconvergence for the models.

2 We used the COMPLEX MIXTURE command in Mplus, which accounts for the clustering of the data and

uses MLR (According to Muthén & Muthén, “maximum likelihood parameter estimates with standard er-
rors and a chi-square test statistic (when applicable) that are robust to non-normality and non-independence
of observations” (Muthén & Muthén, 1998-2012, p. 603)). Residual variances of the outcome variables were
freely estimated at each time and not correlated. The means of the intercept and slope growth factors were
freely estimated for each class as were the residual variances.

3 We also tested a one-class model; however, it was a worse fit to the data as compared to the 2-class solution

(AIC = 1,092,205, BIC = 1,092,468, LMR < .001).

 International Center for Clinical Excellence 821

Table 1
Fit Statistics for Classes in Growth Mixture Models

No. of classes AIC Adjusted BIC 2LL Diff LMR p-value

Two classes 1,088,530 1,088,703 – –

Three classes 1,087,760 1,087,957 781.78 < .001
Four classes 1,087,312 1,087,534 452.04 .65

Note. AIC = Akaike’s information criteria; BIC = Bayesian information criteria; LMR = Lo-Mendell-
Rubin Likelihood test. Lower AIC and BIC estimates favor the preferred model. Significant p-values favor
the model with more classes.

Table 2
Summary of Fixed Effects for Three Class Solution

Class 1 Class 2
Early and Late Worse Before Class 3
Changers Better Slow and Steady
Estimate (SE) Estimate (SE) Estimate (SE)

Intercept 255.81 (1.25) 220.47 (4.16) 253.66 (1.70)

Linear 17.76 (0.58) −19.65 (2.75) 7.29 (0.46)
Quadratic −2.95 (0.13) 8.48 (0.54) −0.60 (0.06)
Cubic 0.16 (0.01) −0.56 (0.04) 0.02 (0.002)

Note. SE - standard error. All estimates were statistically significant (p < .0001).

Table 3
Classification Probabilities for Three-Class Solution

Class 1 Class 2
Early and Late Worse Before Class 3
Changers Better Slow and Steady

Class 1 membership .91 .06 .02

Class 2 membership .25 .74 .01
Class 3 membership .06 .002 .94

Note. The bolded numbers reflect correct classification probabilities. The numbers of the diagonal reflect the
probability of being incorrectly classified.

four- and three-class solutions. The LMR test, however, was not statistically significant (p =
.65), which suggests that the four-class solution was not superior to the three-class solution.
Given this finding, we decided not to add more classes to the model. The fit statistics for the
two-, three-, and four-class solutions are presented in Table 1.
Based on the model fit statistics, the four-class model demonstrated better fit (i.e., lower AIC
and adjusted BIC estimates) compared to the two- and three-class models. The three-class model
had lower AIC and BIC estimates compared to the two-class model. Thus, it appears that the
two-class model was inferior to the three- or four-class solutions based on both model fit indices
and the LMR test. Although the four-class solution demonstrated better model fit, there was
not a statistically significant difference between the three- and four-class solutions according
to the LMR test. Examining the addition of the fourth class for practical value did not reveal
any substantive benefit (i.e., the trajectory essentially mirrored those of the clients in class one).
Accordingly, we opted for the more parsimonious three-class solution.
Table 2 presents the fixed effects for the intercepts, linear, quadratic, and cubic rates of change
for the three classes. All of the effects were statistically significant (p < .001). The classification
probabilities are presented in Table 3. The probability of correctly classifying clients in class 2
822 Journal of Clinical Psychology, September 2015

Figure 1. Actual trajectory scores for the three latent classes.

Note. LC = Latent class. The solid line reflects the clinical cutoff. LC 1 = Early and Late Changers; LC 2 =
Worse before Better; LC 3 = Slow and Steady.

Figure 2. Predicted latent class trajectories for the three latent classes.
Note. LC = Latent class. The solid line reflects the clinical cutoff. LC 1 = Early and Late Changers; LC 2 =
Worse before Better; LC 3 = Slow and Steady.

was .740, which was lower than classes 1 and 3, .912 and .937, respectively. The actual trajectories
and the estimated trajectories for the three classes are presented in Figures 1 and 2.
Class 1 was the largest, with 8,174 (75.3%) clients attending an average of 7.61 sessions (SD =
2.38; range = 5 to 17). As can be seen in the figures, clients in class 1 were typified by moderate
initial psychological functioning, with improvements evidenced early in treatment (i.e., by session
five), followed by a plateau of functioning until session 11 when another notable improvement in
functioning was demonstrated. We opted not to display the trajectories after session 14 because
less than 1% of clients in this class (n = 25) attended 15 to 17 sessions, and these estimates were
not reliable and were notably different than the actual trajectories.4 Given the trajectory pattern,
we refer to this group as “Early and Late Change.”

4 We decided to not illustrate the trajectories at the point where the estimate became less reliable and was
a notable deviation from the anticipated path. Interested readers can contact the lead author for more
information regarding the statistical information and complete figures.
 International Center for Clinical Excellence 823

Table 4
Means and Standard Deviations for BHM Scores for Sessions 1–3 by Class

Class 1 Class 2
Early and Late Worse before Class 3
Changers Better Slow and Steady
Session Number M (SD) M (SD) M (SD)

Session 1 253.97 (61.67) 209.15 (66.40) 251.14 (61.24)

Session 2 271.44 (57.26) 190.41 (58.28) 262.77 (59.61)
Session 3 280.74 (56.80) 181.20 (51.58) 267.79 (60.41)

Note. BHM = Behavioral Health Measure-20; M = mean; SD = standard deviation.

Table 5
Differences Between Classes on Initial Scores and Changes Over the First Three Sessions

Coefficient (SE) t-ratio p-value

Initial BHM Score

Class 3 vs. class 2 40.77 (2.86) 14.25 <.001
Class 1 vs. class 2 45.19 (2.45) 18.48 <.001
Class 1 vs. class 3 4.42 (1.56) 2.84 .005
Slope for BHM
Class 3 vs. class 2 22.75 (1.27) 17.86 <.001
Class 1 vs. class 2 27.69 (0.99) 27.91 <.001
Class 1 vs. class 3 4.93 (0.59) 8.32 <.001

Note. BHM = Behavioral Health Measure-20; SE = standard error; class 1 = Early and Late Changers;
class 2 = Worse before Better; class 3 = Slow and Steady.

Class 2 was the smallest, with 589 (5.4%) clients attending an average of 7.83 sessions
(SD = 2.27, range = 5 to 15). For this class, only 48 clients attended more than 11 sessions.
Similar to class 1, the scores for sessions 12 and beyond were not reliable and subsequently were
not plotted. Clients in this class reported worse initial psychological functioning compared to
the other two classes. Additionally, this class of clients reported a deterioration in functioning in
the early phase of treatment, and then rapid improvement in functioning starting around session
five and lasting until session nine when the trajectory flattened. For this group we refer to them
as “Worse before Better.”
Class 3 comprised 2,091 (19.3%) clients attending an average of 16.88 sessions (SD = 2.73,
range = 12 to 25). Notably, this class represents those clients who tend to stay in therapy
longer compared to the other two classes. Clients in this class had levels of initial psychological
functioning similar to those in class 1, but class 3 clients seemed to make slow but steady progress
over the course of therapy. Accordingly, we refer to this group as “Slow and Steady Change.”
In the figures, we only plotted this class to session 15 (to stay consistent with the other classes),
even though their trajectories were stable up to session 22.
Next, we examined whether the classes would differ in their initial BHM scores and initial
rates of change (over the first three sessions). To do so, we conducted a three-level model in
which clients’ scores over the first three sessions were nested within clients who were nested
within therapists. Table 4 shows the means and standard deviations for the first three sessions by
class. Table 5 displays the differences between classes on initial scores and changes over the first
three sessions. Clients in the Early and Late Change class and Slow and Steady Change class
reported statistically higher initial BHM scores than the clients in the Worse before Better class
(Cohen’s d effect sizes = 0.71, 0.64). The difference in initial BHM scores between clients in the
Early and Late Change class and the Slow and Steady Change class was statistically significant;
however, the magnitude of the effect was small, d = 0.07. The slopes for the three classes also
varied.
824 Journal of Clinical Psychology, September 2015

Clients in the Worse before Better class demonstrated initial deterioration, and this change in
functioning was statistically different from the initial changes in psychological functioning for
clients in the Early and Late Change and Slow and Steady Change classes, who had noticeably
improved over the first three sessions. Both Early and Late Change and Slow and Steady Change
classes demonstrated statistically significant differences in the initial rate of change. Collectively,
it appears that that the distinguishing characteristics between the Early and Late Change and
Slow and Steady Change classes versus clients in the Worse before Better class were relatively
apparent in their initial scores and trajectories. However, it appears there are more similarities
in the initial scores and trajectories between the Early and Late Change and Slow and Steady
Change classes, with the notable differences being that the Slow and Steady Change class was
more distressed initially and had a slower trajectory of change.

Discussion
In a large naturalistic sample of 10,854 clients, we identified three groups of clients who differed
in their initial scores and trajectories of change. Previous studies that have used GMM have also
found three classes (e.g., Lutz et al., 2009; Stulz & Lutz, 2007; Stulz, Gallop, et al., 2010; Stulz,
Thase, et al., 2010); however, the findings for the current study are unique when compared to
previous research. In the current study, the clients of the largest class– Early and Late Change–
comprising over three quarters of the entire sample, were typified by improvement occurring
within the first five sessions and then relatively no change until session 11, at which time there
was an additional period of improvement. Although this second period of improvement was
evidenced, on average, this class attended approximately eight sessions, which suggests that
the majority of the sample did not experience this additional improvement in psychological
functioning. However, those who do remain in therapy after session 11 seem to benefit, which
seems to run counter to the plateau effects seen in the dose-effect models (e.g., Baldwin et al.,
2009; Howard et al., 1986). Consequently, there appears to be a benefit for some clients remaining
in therapy after the initial change; clients appear to be doing work in the middle phase of therapy
that then results in further improvement later (Howard et al., 1986).
It may also be that clients in the Early and Late Change class focused on different aspects
of psychological functioning at different phases of treatment. For example, initial phases of
treatment often aim to reduce distress and increase hope; whereas other more engrained char-
acterological aspects typically take longer to change (Howard, Lueger, Maling, & Martinovich,
1993). Additionally, the psychological functioning of clients in the Early and Late Change class
corresponds with some of the previous phase model research that indicates clients’ psychological
functioning across domains plateaus around sessions 8 to 16 and then sharply improves (Joyce,
Ogrodniczuk, Piper, & McCallum, 2002).
The clients in the Slow and Steady Change class were the second largest class, comprising
approximately 20% of the sample. Notably, this class remained in therapy for an average of
17 sessions, with no clients staying less than 12 sessions. Thus, this class may represent those
relatively longer term clients whose change trajectories are not as steep as clients in the Early
and Late Change class. At the same time, there is not a remarkable plateau effect for clients in
this class either. The findings from this class coupled with previous research (e.g., Baldwin et
al., 2009; Howard et al., 1993), suggest that more therapy is needed for clients whose progress is
more gradual.
The Worse before Better class was the smallest, comprising only approximately 5.5% of the
sample. Notably, they were significantly more distressed than the other two groups. Addition-
ally, and as the name of the class suggests, these clients initially deteriorated before showing
improvement. Similar to the clients in the Early and Late Change class, clients in this class
attended approximately eight sessions, on average, and only very few clients stayed in treatment
longer than 11 sessions. The trajectories could reflect several dynamics. For example, the initial
deterioration may be evidence of clients who are struggling with crises and throughout the
therapy process client–therapist dyads are generally able to find a way to cope and engender
improvements in psychological functioning.
 International Center for Clinical Excellence 825

Previous research has noted that client deterioration initially presents as increased symptoms
(Swift, Callahan, Heath, Herbert, & Levine, 2010), which could reflect the pain experienced by
facing difficult topics in a more overt manner; yet through the process, client–therapist dyads
are able to work through these difficulties. Although there were relatively fewer clients in this
class, it is interesting and important clinically to realize that an initial deterioration may not
necessarily be a sign that treatment is off track.
Additionally, therapists had access to their clients’ scores, which may have informed their
treatment and ultimately assisted in the clients increase in functioning (see Lambert, 2010).
Clearly, more research is needed to understand the process of psychotherapy for clients who
initially deteriorate but then ultimately (and relatively quickly) make substantial improvement.

Limitations
The outcomes from this study should be understood within the scope of the methodological
limitations. First, the data reflect treatment at mainly university counseling center settings. Thus,
it is uncertain whether these results will generalize to other clinical settings. Second, these data
lack the more rigorous controls and monitoring of treatments that are typically associated
with randomized clinical trials. In particular, we do not know whether therapists’ access to
the clients’ scores informed the treatment trajectories observed in this study. Third, we do not
know what therapeutic processes may be associated with the changes and the various trajectories.
Consistently, we did not assess reasons for termination or staying in treatment. More information
regarding these processes could be useful in disentangling what is occurring in the sessions before
and after psychological functioning has improved. Last, we had sparse data for the latter session
in most of the classes. Thus, future studies should seek to replicate these findings and attempt
to increase the sample size per class.

Implications
A number of practical implications can be drawn from the current study. For example, most
manualized treatments require specific steps as clients proceed through the required number of
sessions. It is likely that, regardless of intervention (e.g., cognitive, psychodynamic), clients do
not follow a common trajectory of change, as was seen in the current study as well as previous
studies with manualized approaches (e.g., Lutz et al., 2009; Stulz, Gallop, et al., 2010; Stulz,
Thase, et al., 2010). Though authors have called for “flexible and creative use” of manualized
treatments (Kendall, Chu, Gifford, Hayes, & Nauta, 1998; Owen & Hilsenroth, 2014), it may be
appropriate to adjust treatments based on the trajectory of change for clients.
The results from this study may also have policy implications related to session limits. The
current system (i.e., insurance, session-limited clinics) has used the dose-effect and GEL models
to determine the amount of sessions that clients should have to indicate improvement. Two of
the three trajectories of change differ from widely used models that estimate linear improvement
for clients. Instead, therapists should have more power to assess the specific trajectory for the
client to determine the amount of sessions for maximal client improvement.
The study also has implications for client-feedback systems, in which therapists track their
clients’ progress session by session and clients’ trajectories are compared to norms to indicate
whether a client is “off track” for a positive therapy outcome (e.g., Lambert, 2010; Miller et al.,
2005). Depending on the class, the feedback may or may not be accurate. Initial deterioration
may, for example, indicate that a given case is off track or simply that difficult material is being
processed and necessary scaffolding for progress is being built. Differentiating between the two
possibilities is challenging to be sure. At the same time, knowing that the latter is not outside the
realm of possibility for a small minority of clients is critical for feedback to work as intended.

Conclusions
Therapists should be aware of the varying trajectories for clients early on in therapy. Clients who
are not responding to treatment may be different from clients who are in the Worse before Better
826 Journal of Clinical Psychology, September 2015

class. Treatment nonresponders may need an entirely different treatment plan (Kocsis et al.,
2009), while clients who decline before they improve may need to stay the course. For instance,
therapists may want to provide booster sessions when the decline begins to shift the trajectory for
these particular clients, because booster sessions have shown to have larger effects for treatments
when compared to treatments without booster sessions (Gearing, Schwalbe, Lee, & Hoagwood,
2013). In a similar vein, clients who present in the Early and Late Change class may mislead
therapists into believing that their treatment has hit its peak, when in reality these clients have
the potential for additional improvement. Ultimately, we hope that this study generates new
questions about what we know regarding the trajectories of change for clients and how we
approach treatment given the diversity evidenced.

References
Baldwin, S. A., Berkeljon, A., Atkins, D. C., Olsen, J. A., & Nielsen, S. L. (2009). Rates of change in
naturalistic psychotherapy: Contrasting dose-effect and good-enough level models of change. Journal
of Consulting and Clinical Psychology, 77, 203–211.
Barkham, M., Connell, J., Stiles, W. B., Miles, J. N., Margison, F., Evans, C., & Mellor-Clark J. (2006).
Dose-effect relations and responsive regulation of treatment duration: The good enough level. Journal
of Consulting and Clinical Psychology, 74, 160–167.
Bryan, C. J., Kopta, S. M., & Lowes, B. D. (2012). The CelestHealth System. Integrating Science and
Practice, 2, 7–11.
Gearing, R. E., Schwalbe, C. S., Lee, R., & Hoagwood, K. E. (2013). The effectiveness of booster sessions
in CBT treatment for child and adolescent mood and anxiety disorders. Depression and Anxiety.
Howard, K. I., Kopta, S. M., Krause, M. S., & Orlinsky, D. E. (1986). The dose-effect relationship in
psychotherapy. The American Psychologist, 41, 159–164.
Howard, K. I., Lueger, R. J., Maling, M. S., & Martinovich, Z. (1993). A phase model of psychotherapy
outcome: Causal mediation of change. Journal of consulting and clinical psychology, 61, 678.
Joyce, A. S., Ogrodniczuk, J., Piper, W. E., & McCallum, M. (2002). A test of the phase model of psychother-
apy change. The Canadian Journal of Psychiatry, 47, 759–766.
Kendall, P. C., Chu, B., Gifford, A., Hayes, C., & Nauta, M. (1998). Breathing life into a manual: Flexibility
and creativity with manual-based treatments. Cognitive and Behavioral Practice, 5, 177–198.
Kocsis, J. H., Gelenberg, A. J., Rothbaum, B. O., Klein, D. N., Trivedi, M. H., Manber, R., . . . & Thase, M.
E. (2009). Cognitive behavioral analysis system of psychotherapy and brief supportive psychotherapy
for augmentation of antidepressant nonresponse in chronic depression: The REVAMP Trial. Archives
of General Psychiatry, 66, 1178.
Kopta, S. M., & Lowry, J. L. (2002). Psychometric evaluation of the Behavioral Health Questionnaire-20:
A brief instrument for assessing global mental health and the three phases of psychotherapy outcome.
Psychotherapy Research, 12, 413–426.
Kopta, S. M., Petrik, M., Saunders, S., Mond, M., Hirsch, G., Kadison, R., & Raymond, D. (in press). The
utility of an efficient outcomes assessment system at university counseling centers. Journal of College
Student Psychotherapy.
Lambert, M. J. (2010). “Yes, it is time for clinicians to monitor treatment outcome.” In B. L. Duncan, S.
C. Miller, B. E. Wampold, & M. A. Hubble (Eds.), Heart and soul of change: Delivering what works in
therapy (2nd ed., pp. 239–266). Washington, DC: American Psychological Association.
Lambert, M. J., & Olges, B. M. (2004). The efficacy and effectiveness of psychotherapy. In M. J. Lambert
(Ed.), Bergin and Garfield’s handbook of psychotherapy (pp. 139–193). New York: John Wiley and Sons,
Inc.
Lambert, M. J., & Shimokawa, K. (2011). Collecting client feedback. Psychotherapy, 48, 72–79.
Lutz, W., Lowry, J., Kopta, S. M., Einstein, D. A., & Howard, K. I. (2001). Prediction of dose-response
relations based on patient characteristics. Journal of Clinical Psychology, 57, 889–900.
Lutz, W., Martinovich Z., & Howard, K. I. (1999). Patient profiling: An application of random coeffi-
cient regression models to depicting the response of a patient to outpatient psychotherapy. Journal of
Consulting and Clinical Psychology, 67, 571–577.
Lutz, W., Stulz, N., & Köck, K. (2009). Patterns of early change and their relationship to outcome and
follow-up among patients with major depressive disorders. Journal of Affective Disorders, 1118, 60–68.
 International Center for Clinical Excellence 827

McCoach, D. B., & Black, A. C. (2008). Evaluation of model fit and adequacy. In A. A. O’Connell & D.
B. McCoach (Eds.), Multilevel modeling of educational data (pp. 245–272). Charlotte, NC: Information
Age Publishing.
Miller, S. L., Duncan, B. L, Sorrell, R., & Brown, G. S. (2005). The partners for change outcome management
system. Journal of Clinical Psychology, 61, 199–208.
Muthén, B. (2004). Latent variable analysis. Growth mixture modeling and related techniques for longitudi-
nal data. In D. Kaplan (Ed.), The Sage Handbook of Quantitative Methodology for the Social Sciences
(pp. 345–368). Thousand Oakes, CA: Sage Publications.
Muthén, L. K., & Muthén, B. O. (1998–2012). Mplus User’s Guide. Seventh Edition. Los Angeles, CA:
Muthén & Muthén.
Nylund, K. L., Asparaouhov, T., & Muthén, B. (2007). Deciding on the number of classes in latent class
analysis and growth mixture modeling: A Monte Carlo simulation study. Structural Equation Modeling,
14, 535–569.
Owen, J. J., Adelson, J., Budge, S., Kopta, S. M., & Reese, R. J. (2014). Goodenough level and
dose-effect models: Variation among outcomes and therapists, Psychotherapy Research, DOI:
10.1080/10503307.2014.966346
Owen, J., Devdas, L., & Rodolfa, E. (2007). University counseling center off-campus referrals: An exploratory
investigation. Journal of College Student Psychotherapy, 22, 13–30.
Owen, J. & Hilsenroth, M. (2014). Treatment adherence: The importance of therapist flexibility in relation
to therapy outcomes. Journal of Counseling Psychology, 61, 280–288.
Pinsof, W. M., Goldsmith, J. Z., & Latta, T. A. (2012). Information technology and feedback research
can bridge the scientist-practitioner gap: A couple therapy example. Couple and Family Psychology:
Research and Practice, 1, 253–273.
Reese, R. J., Toland, M. D., & Hopkins, N. B. (2011). Replicating and extending the good-enough level
model of change: Considering session frequency. Psychotherapy Research, 21, 608–619.
Stiles, W. B., Barkham, M., Connell, J., & Mellor-Clark, J. (2008). Responsive regulation of treatment
duration in routine practice in United Kingdom primary care settings: Replication in a larger sample.
Journal of Consulting and Clinical Psychology, 76, 298–305.
Stulz, N., Gallop, R., Lutz, W., Wrenn, G. L., & Crits-Christoph, P. (2010). Examining differential effects
of psychosocial treatments for cocaine dependence: An application of latent trajectories analyses. Drug
and Alcohol Dependence, 106, 164–172.
Stultz, N., & Lutz, W. (2007). Multidimensional patterns of change in outpatient psychotherapy: The phase
model revisited. Journal of Clinical Psychology, 63, 817–833.
Stulz, N., Lutz, W., Kopta, S.M., Minami, T., & Saunders, S. (2013). The shape of the dose-effect relationship
in outpatient psychotherapy: Does treatment duration matter? Journal of Counseling Psychology, 1–8.
Stulz, N., Thase, M. E., Klein, D. N., Manber, R., & Crits-Christoph, P. (2010). Differential effects of
treatments for chronic depression: A latent growth model reanalysis. Journal of Consulting and Clinical
Psychology, 78, 409–419.
Swift, J. K., Callahan, J. L., Heath, C. J., Herbert, G. L., & Levine, J. C. (2010). Applications of the
psychotherapy phase model to clinically significant deterioration. Psychotherapy, 47, 235–248.
Vermunt, J. K., Tran, B., & Magidson, J. (2008). Latent class models in longitudinal research. In. S. Menard
(Ed.), Handbook of longitudinal research: Design, measurement, and analysis (pp. 373–385). Bulington,
MA: Elsevier.