current study examined the available data
(i.e., completer outcomes) to test whether an
association between dropout rate and
outcome might be present. Is dropout an indi-
cator of treatment effectiveness for those who
complete the therapy? In other words, when
comparing two treatments, is the rate of
dropout from one treatment relative to
another predictive of the relative effectiveness
of those treatments for the completers? If
dropout is a sign of treatment failure, then
one would predict that higher rates of
dropout herald a less effective treatment. If
dropout is, as many have suggested, a more
complex phenomenon, then one would
expect for the relationship between dropout
and outcome to vary in its strength or direc-
tion depending on the type of dropout. For
example, higher rates of dropout due to dis-
content with the treatment might be expected
to be more highly related to treatment ineffec-
tiveness than dropout due to logistical con-
straints. Furthermore, dropout due to
improvement might be expected to be associ-
ated with better outcomes. Therefore, the
third set of analyses examined comparative
treatment trials in which the rates and
reasons given for dropout as well as the thera-
peutic outcomes of these treatments were
reported for each treatment group.
It may also be possible that dropout is only relevant
for treatment outcome under certain conditions.
Therefore, analyses also examined whether the rel-
evance of dropout for treatment outcome varied as
a function of treatment, therapist, client, and study
factors.
Overall Method
Studies
Search methods. The 669 articles of the most
recent meta-analysis of dropout rates (Swift & Green-
berg, 2012) were first screened for eligible articles.
Studies published after the Swift and Greenberg
(2012) meta-analysis search were identified through
PsycINFO and PsycARTICLES using the terms
dropout, attrition, or termination combined with
either psychotherapy or treatment from the period
of 2010 to 2017. Finally, issues of psychotherapy
journals were hand searched including the American
Journal of Psychiatry, Behavior Therapy, Behaviour
Research and Therapy, British Journal of Psychiatry,
Journal of the American Medical Association Psychiatry,
Journal of Clinical Psychology, Journal of Consulting and
Clinical Psychology, Psychotherapy, and Psychotherapy
Research also from the period of 2010 to 2017. To
be selected, the published articles had to be available
in English.
Selection criteria. A full explanation of exclusion
criteria can be found in Figure 1. Notably, studies
with treatments that took place in inpatient units or
in prison as well as studies of individuals with psycho-
tic, child, or family/couple populations were excluded
because in each case there may be external pressures
beyond the individual’s preference guiding the
decision to continue treatment.
Procedure
Coding. All screening and coding followed
PRISMA guidelines for meta-analyses and was con-
ducted by the first author. Ten percent of the
studies were randomly selected and rated by a
second coder. The reliability between the first and
second coder was excellent for both effect size data
(ICC = .91) as well as all variables pertaining to
dropout and timing such as treatment length (ICC
= .92). Dropout rates were computed as proportions
of the total dropped over the total sample at the start
of the treatment. These proportions were corrected
for abnormal distribution using the arcsine trans-
formation for all analyses (Keppel & Wickens,
2004, Chapter 7, p. 155).
Cohen’s(1977) d was calculated as a measure of
effect size to indicate the strength of the difference
being assessed (Borenstein, Hedges, Higgins, &
Rothstein, 2009 ). When no information was provided
in an article except that the comparison was “not stat-
istically significant,” the effect size was assumed to be
zero. This method could be randomly biased as some
studies may have had a nonsignificant but slightly
positive or slightly negative finding. Excluding such
cases could also bias the findings toward inclusion
of studies that reported significant findings. There-
fore, the results below are reported both including
and excluding effect sizes assumed to be zero when
the results were meaningfully different. In all cases,
Hedges’ g (Hedges & Olkin,
1985,
p. 81) was used
to correct for sampling bias.
Additional variables were coded as available for
descriptive purposes and to explore potential moder-
ating effects. Moderators were selected based on
those variables noted to moderate dropout rates in
the Swift and Greenberg (2012 ) meta-analysis as
well as availability of data in articles. Additionally,
two variables related to dropout were coded: the
study definition of dropout and the average number
of sessions before drop out occurred. When the
Psychotherapy Research 3