INTRODUCTION
Recent estimates indicate that there are approximately 8.9 million cancer
survivors in the United States, representing nearly 3% of the population
(NCI, 1998). Virtually all survivors have been
treated with either surgery, radiation, chemotherapy, biologics, hormonal
therapies, or some combination of these. As treatments for cancer have
improved in efficacy and length of survival has increased for many patients,
extensive research has been done to document the acute and long-term
side effects of these treatments (e.g., Bonadonna et
al., 1985; Brundage, 1997; Fisher et al., 1994; Petros,
2002; Saphner et al., 1991; Videtic, 2001).
While researchers have focused their attention on physical side
effects, another potentially important side effect, impaired
neuropsychological functioning, has received less attention. Possible
neuropsychological side effects of these treatments may include
difficulty concentrating, impaired verbal and visual memory, difficulty
organizing information, decreased motor skills, and language problems
(e.g., word finding difficulty). Health care professionals and
researchers have become increasingly aware of possible cognitive
effects through the anecdotal reports of some cancer survivors. Some
patients have reported that problems in cognitive functioning have
precluded a smooth transition to life after cancer (e.g., problems
resuming complex tasks at work or multi-tasking at home). These
neuropsychological impairments are so well known among cancer survivors
that they have been described in several cancer patient newsletters and
have been dubbed “chemobrain” (Mann,
19991
See
www.mynewspirit.com/chemobrain.htm.
). Thus, clinicians and
researchers concerned with quality of life issues relevant to patients have
begun to turn their attention to this matter.
Scientific studies of the chemobrain phenomenon began in the 1970s
(e.g., Weiss et al., 1974a, 1974b), and have appeared with increasing frequency
in the literature since then. A number of reviews have been published
regarding the possible neuropsychological effects of a variety of
cancer treatments, but interpretations of the findings have not been
definitive (e.g., Ahles & Saykin, 2001;
Bender et al., 2001; Fleishman & Kalash, 1998; Ganz, 1998; Meyers, 2000;
Meyers & Scheibel, 1990; Meyers et al., 1994; Olin,
2001; Peterson & Popkin, 1980;
Redd et al., 1991; Silberfarb, 1983; Silberfarb
& Oxman, 1988; Tope et al., 1993;
Trask et al., 2000; Troy
et al., 2000; Walch et al., 1998).
Earlier reviews had fewer studies to work with and lacked the
methodological rigor that is becoming more common in current research
on possible neuropsychological effects. Ganz (1998) notes that there are some equivocal results
and points to the possibility of a dose effect. Redd et al. (1991) interpreted the existing literature as
suggesting cognitive deficits as a result of treatment, but also
expressed concern about the tools used in the measurement of these
effects. Some recent reviews have focused on effects within a specific
diagnosis, such as Olin's review of breast cancer (2001), while other reviews have focused on a specific
treatment, such as the review by Trask and colleagues of Interferon (2000), or the review by Troy and colleagues of Cisplatin
(2000). The most recently published qualitative
review of the literature, by Ahles and Saykin (2001), concluded that standard-dose chemotherapy is
associated with neuropsychological impairments in a subgroup of adult
cancer survivors. The authors concluded that the impairments include
subtle decrements in memory and concentration and that these subtle
changes can have a significant impact on cancer survivors' quality
of life.
The present review is different because it is a quantitative review
of all available studies. None of the above reviews was quantitative in
nature and therefore none could address the question of the
size of an effect cancer treatments might be having.
Additionally, a meta-analysis has strength in its numbers, in that
patients from across many studies are pooled, allowing for an increase
in power to detect an effect when one exists.
The present meta-analysis attempts to clarify the variety of
conclusions from qualitative reviews by bringing together the results
of quantifiable inquiries into clearly defined cognitive domains
measured with standardized assessment tools. This meta-analysis
examined neuropsychological effects of cancer treatments across
different types of cancers and given different types of treatments.
While cancer treatments are not equal and therefore unlikely producing
a unitary effect on cognitive functioning, the limited availability of
research in this area leads to this synthesis in order to provide a
preliminary and general sense of possible effects. It is anticipated
that additional research will need to be done so that more definitive
conclusions can be drawn about the specific cognitive domains that each
type of therapy affects and the potential mechanism(s) that create any
cognitive deficits.
If any effects do exist in this first analysis of existing
literature, we wanted to know for which domain(s) of cognitive
functioning there exists an effect, and also how big that effect is.
All effects were converted to a common metric (Cohen's
d), so that outcomes could be compared quantitatively across
measures and domains of neuropsychological function. Results of studies
were pooled according to the three types of research designs that have
been employed in the existing literature. Two of the designs used were
between subjects designs, one allowing for comparisons of
post-treatment results with control subjects and the other allowing for
comparisons with normative data. The third design used was a within
subjects design, allowing for comparisons of post-treatment results
with a subjects' own baseline performance. Thus, this latter
design is longitudinal and the former two designs are
cross-sectional.
Synthesizing these results is an important advantage of meta-analysis
since these three methods of comparison might be expected to produce
somewhat different results. Comparisons of post-treatment results with
baseline might be considered the ideal method since using a
subject's own baseline controls many extraneous variables (e.g.,
holds constant prior experience and education). However, significant
practice and learning effects are common with many tests (Dikmen et al., 1999; McCaffrey
et al., 2000; Salinsky et al., 2001;
Temkin et al., 1999). For example, it can be
problematic if a testee is asked to recall the same list of words on a
memory task that had been used during a previous test session, or if a
testee is asked to repeat a particular procedure with which they have
already been familiarized, such as connecting dots in order. A
relatively negligible change from pre- to post-test might not appear to
be a real decline, but it could be a diminution of function
counterbalanced and obscured by practice effects (which normally lead
to a substantial increase from pre- to post-test; Dikmen et al., 1999; McCaffrey
et al., 2000; Salinsky et al., 2001;
Temkin et al., 1999). The design that uses
comparisons with control subject performance also has its benefits
(e.g., eliminates practice and learning effects, while providing a
theoretically comparable contrast group such as persons of similar ages
and education from the same geographic regions). Patients with a
similar diagnosis who are not prescribed a particular treatment could
be said to serve as excellent control subjects; however, they might
differ on some third variable (e.g., no treatment controls might differ
on overall health status, such that their performance status may have
been too low to tolerate chemotherapy and so it was withheld).
Comparisons with healthy controls and normative data are also helpful,
but a drawback is that patients may differ on some important third
variable (e.g., fatigue, distress level). We might expect the greatest
effects to show up in normative comparisons since normative samples
typically contain individuals with a wider range of education and
experience. One strength of the present meta-analysis is that we will
be able to contrast all three methods of comparison and look for
convergence. Ultimately, the meta-analysis allows for quantification of
the magnitude of any effects that are significantly different from
zero.
For this meta-analysis, we selected all identifiable research on the
neuropsychological effects of systemic therapies for adults with
cancer. While there are numerous published studies on the
neuropsychological sequelae of treatments for childhood cancer, we
chose not to include those studies in this analysis. The effects of
cancer treatments on the developing brain are thought to be
significantly different from the effects on the adult brain (for
reviews see Butler, 2002; Copeland et al., 1985; Eiser,
1998; Gotay, 1987). We also chose to
exclude studies that focused exclusively on the effects of treatments
for brain cancers. Treatments for these cancers often involve brain
irradiation and surgical interventions, in addition to direct effects
on brain tissue secondary to the lesions (Roman
& Sperduto, 1995; Weitzner & Meyers,
1997). In such cases, changes in neuropsychological functioning
would be expected and could have skewed the results of this study.
Similarly, we excluded studies in which all patients received some
direct treatment to the brain (e.g., cranial irradiation). While some
studies in our sample include a percentage of patients with gliomas,
central nervous system (CNS) lesions or brain irradiation, these
studies were retained since the samples also included other cancer
diagnoses and treatments.
In addition to examining the available literature for an overall
pattern of results, we also considered the role of possible mediating
and moderating variables. We hypothesized that the severity of
diagnosis and the intensity of treatment would impact any effect found
on neuropsychological functioning. Thus, we conducted analyses that
included only those studies in which diagnoses were relatively less
severe (e.g., non-metastatic) and treatments were relatively less
intense (e.g., excluding any brain irradiation or bone marrow
transplant, BMT). We conducted these additional sub-group analyses to
evaluate whether these two groups of studies produced similar results
and thus, whether or not the “more severe” studies might be
having an untoward effect on overall effect sizes.
METHODS
Literature Search and Criteria for Inclusion
We performed a literature search of the PsychInfo, Medline and
CancerLit databases. Key words included: cancer,
neoplasm, oncology, oncologic,
chemotherapy, systemic treatment, drug
effects, behavior, cognition,
cognitive, tests, measures,
assessment, neuropsychology, and
neuropsychological. Pre-print, unpublished, and file-drawer
papers were requested via the Division 38 (Health Psychology) of the
American Psychological Association list-serve and via email to authors
who had already published on this topic. Research articles retrieved in
this manner were also inspected for relevant references to locate
additional articles.
Nearly 100 articles were located that addressed the issue of cancer
treatment and neuropsychological functioning. To be included in
analyses, the report needed to contain original study data (e.g., not a
review), the sample needed to be adult, diagnoses could not be entirely
metastatic or gliomas, and treatment could not involve brain
irradiation of the whole sample. To qualify for inclusion, it was
necessary that the study had reported quantitative measurement or an
inferential statistic regarding some aspect of neuropsychological
functioning. If not, an attempt was made to retrieve such data (e.g.,
means, standard deviations) directly from the study's author(s).
Thirty of the studies that had been identified met criteria and were
included in this meta-analysis (leading to 29 samples: two samples came
from Ahles et al., 2002, and in two instances
a single sample spanned two articles by the same first author: Kaasa et
al., 1988a, 1988b;
and Walker et al., 1996, 1997). Excluded studies did not contain
quantitative data or statistics that could be analyzed. Table 1 displays information on the studies that
were included in our analyses.
Empirical studies included that examined cognitive effects of
systemic treatments
Recruitment of Participants into Studies
It is important to consider how participants came to be enrolled in
the studies included in this meta-analysis, since a selection bias in
favor of clinically identified participants could artificially inflate
any results. A review of the methods used by each study included here
revealed that, in general, the researchers used exhaustive methods to
identify and recruit all eligible participants at their sites (based on
tumor and/or treatment type, not based on referral by self or other
due to cognitive concerns). One study did not clarify the method of
selection, but also did not specify a referral basis, and one study
specified that only one participant was referred because of concerns
about mental status changes.
Cognitive Domains and Measures
A variety of cognitive domains were assessed across studies and many
different measures were utilized to tap these domains. The following
are the domains that could be compiled based on measures used in the
literature: Attention was measured by six tests: (1), Trails
A; (2) Wechsler Adult Intelligence Scale (WAIS)–Digit Span; (3)
High Sensitivity Cognitive Screen (HSCS)–Attention; (4) Dementia
Rating Scale (DRS)–Attention; (5) Automated Performance Test
System (APTS)–Sternberg Test and Simple Reaction Time; (6)
Stroop–A. Information processing speed was measured by 5
tests: (1) WASI–Digit Symbol; (2) Monroe-Sherman Reading
Comprehension; (3) Zazzo's Attention Test–Speed; (4)
APTS–Code Substitution; (5) Computer Drug Research System
(CDRS)–Number Vigilance and Memory Scanning. Verbal
memory was measured by 8 tests: (1) Wechsler Memory Scale
(WMS)–Logical Memory; (2) HSCS–Memory; (3)
DRS–Memory; (4) Rey Auditory Verbal Learning Test (RAVLT); (5)
Memory Scanning Test; (6) Concept Shifting Test (CST)–A and B;
(7) California Verbal Learning Test (CVLT); (8) Buschke Selective
Reminding Task. Visuospatial memory was measured by three
tests: (1) Benton Visual Retention Task; (2) Rey-Osterrieth Complex
Figure–Recall; (3) WMS–Non-Verbal Memory. Visuospatial
skill were measured by 5 tests: (1) Rey-Osterrieth Complex
Figure–Copy; (2) HSCS–Spatial; (3) Road Map Sense Test; (4)
APTS–Pattern Comparison and Manikin Test; (5) WAIS–Block
Design. Executive function was measured by nine tests: (1)
Controlled Oral Word Association; (2) Trails B; (3) Design Fluency; (4)
HSCS–Planning; (5) Wisconsin Card Sort Test (WCST); (6)
Stroop–C/Interference; (7) MST–C; (8) Paced Auditory
Serial Addition Test (PASAT); (9) Similarities. Psychomotor
skill was measured by four tests: (1) Grooved Pegboard; (2) Finger
Tapping; (3) DRS–Construction; (4) APTS–Tapping.
Methods of Comparison
Three research designs have been used in evaluating possible
neuropsychological effects of cancer treatments: (1) treatment group
outcomes compared with published normative data; (2) treatment
group outcomes compared with control group outcomes; and (3)
treatment group outcomes compared with subjects' own
baseline (pre-treatment) scores. In the second type of
contrast, control groups were comprised of either healthy age-matched
controls or individuals with a similar diagnosis, but who were
receiving some presumably more benign treatment (e.g., radiation only).
These two subgroups were subsequently compared to rule out possible
differences.
Statistical Analysis
Findings from each study were assigned to a cognitive domain category
and converted to Cohen's d using a random effects model
(Shadish & Haddock, 1994). Cohen's
d is interpreted as the number of standard deviations the
average cancer treatment group member differs from the average
individual in the control or normative sample, or from her/his own
baseline score in each of the cognitive domains. When means and
standard deviations were not available, effect sizes were calculated
from other reported statistics (e.g., p values, percentages)
using the methods described by Rosenthal and DiMatteo (2001), and Cooper and Hedges (1994).
Effect sizes were assigned such that negative values indicated poorer
performance in the treatment group (relative to the control group,
relative to the normative sample, or relative to subjects' prior
performance, as appropriate). Allowing for a conservative bias and in
order to protect against Type I errors, we assigned a value of zero to
studies that reported nonsignificant results, but which did not include
sufficient data to specifically estimate effect sizes. Cohen's
(1988) guidelines for interpreting effect
sizes suggest that absolute values less than .2 indicate a negligible
effect, values between .2 and .5 indicate a small effect, values
between .5 and .8 indicate a medium effect, and values of .8 and
greater indicate a large effect.
If more than one measurement was made of a particular cognitive
domain in a single study, an average of the effect sizes was used in
the final analyses. This averaging of multiple within-study effect
sizes yields one effect size per cognitive domain (per method) for each
study and limits the degree of statistical nonindependence of the
results. Some overlap in the reporting of results was inevitable since
many studies measured more than one cognitive domain and used multiple
methods of comparison. Given the variability in sample sizes among the
studies included, effect sizes were weighted by the sample size. The
combining, weighting, confidence interval, and statistical significance
calculations for groups of effect sizes were computed via the
Comprehensive Meta-Analysis computer program (Version 1.25; Borenstein & Rothstein, 2002).
A significance level of .01 is inferred when the 99% confidence limit
does not cross zero (Shadish & Haddock,
1994). Additionally, for each mean effect size, a
“fail-safe N” was calculated (Rosenthal, 1991). This value estimates the number
of unpublished, nonsignificant studies of comparable or larger sample
sizes that would have to exist for the obtained probability value to be
rendered nonsignificant. Cut-offs are computed for each effect based on
the number of studies used in calculating the effect (see below).
RESULTS
Included in analyses were data from 838 patients from 29 samples
(Table 1). The mean age of the patients was
49 years (SD = 10.3) and the average amount of time from
either diagnosis or treatment was 86 weeks (SD = 124.1). See
Table 2 for pooled means and standard
deviations of some of the most commonly reported measures.
Pooled means and standard deviations for the most commonly reported
measures
From the 30 reports with sufficient data, 173 effect sizes
(Cohen's d) were extracted across seven cognitive domains
(attention, information processing, verbal memory, visuospatial memory,
visuospatial skill, executive function, and psychomotor skill; Table 3). As can be seen in Table 3, these effect sizes were categorized by one
of the three methods used to evaluate the patients'
neuropsychological functioning in the study or post hoc by the
first author (e.g., comparing patient's outcomes with a published
normative group's performance, a control group's performance,
or their own baseline prior to treatment).
Unweighted effect sizes (Cohen's d) for individual
studies by type of comparison and cognitive domain
Effect sizes were weighted and combined (Shadish
& Haddock, 1994) to yield a single effect size for each of
the seven cognitive domains across each of the three methods used to
evaluate patients' neuropsychological functioning post-treatment
(norm comparison, control comparison, or baseline comparison; see Table 4 and Figure 1).
Each effect size was evaluated for its significance (p ≤
.05) via the Comprehensive Meta-Analysis computer program (Version
1.25; Borenstein & Rothstein, 2002).
Additionally, to control for escalating Type I error due to multiple
statistical tests, the results are also reported with a Bonferroni
Correction (p ≤ .001). Table 4
and Figure 1 each reveal a number of
significant effect sizes across a variety of cognitive domains and
methods of evaluating change in patient performance post treatment. The
results that remained significant with the conservative Bonferroni
correction will be emphasized henceforth.
Results for cognitive domains by method of comparison
Effect size results: Cognitive Domains × Method of Comparison.
*p ≤ .05.
··········□··········
indicates d and CI adjusted for practice/learning
effects (Dikmen et al., 1999).
The results will be reviewed in several ways: (1) considering
statistical significance, (2) considering absolute magnitude of
effects, (3) considering relative magnitude of effects by examining
patterns of findings within cognitive domains and methods of
comparison, and (d) considering possible moderator variables.
Statistical Significance of Effects
It is noteworthy that 20 of 21 averaged weighted effect sizes across
cognitive domains were in the negative direction, indicating a general
trend toward decrements in functioning. Highly significant effects
(p ≤ .001) were found across five cognitive domains for two
of the three methods of comparison (i.e., normative and control
comparisons), while none of the seven cognitive domains were
significant when the method of comparison was with one's own
baseline. Thus, the findings appear more robust and consistent for the
methods using between-subject comparisons as compared with studies that
involved within-subject comparisons.
Absolute Magnitude of Effects
According to Cohen's (1988) guidelines
(i.e., small effect = .2, medium effect = .5, large effect = .8), of
the five highly significant effects, two effects exceeded the large
effect size, while the remaining three approached or exceeded the
medium effect size. The effect sizes for the normative comparisons of
verbal memory and executive function were exceptionally large (both
approximately .9).
Additionally, the number of studies needed to nullify the significant
findings (“the file drawer statistic” or “fail-safe
N”) was calculated for each of these effect sizes (Rosenthal, 1991; see Table
4). Some of the effect sizes emerged as likely to remain
significant even with a large number of additional studies yielding
non-significant results. Rosenthal (1991)
recommends a tolerance level for the file drawer statistic of
5K + 10 (K = the number of studies used to calculate
the effect size). Fail-safe Ns of 711 and 154 were calculated for the
executive functioning effects for normative and control, respectively.
This indicates the number of additional studies of each comparison type
with null results that would be required for the significance of the
reported finding to be reduced a non-significant level. The fail-safe
Ns for normative and control comparisons exceed Rosenthal's
recommended cut-offs of 95 and 80 suggesting that these are very
resilient findings unlikely to be disproven by a large number of
studies yet unretrieved. Similarly, 168 (exceeds 55 cut-off) studies
would be required to refute the verbal memory results for normative
comparisons, while only 36 (does not exceed 45 cut-off) normative
comparison studies would be necessary for refuting motor functioning
findings.
Relative Magnitude of Effects Patterns Across
Cognitive Domains and Methods of Comparison
Three cognitive domains revealed fairly consistent results across two
methods of comparison. Effect sizes for executive function, verbal
memory, and motor functioning were statistically significant (at either
the .01 or .001 levels) whether the method of comparison was with
normative data or control subjects. The effect sizes for executive
function and verbal memory ranged from medium to large for both methods
of comparison, while the effects for motor function ranged from small
to large.
The baseline comparisons for these three domains were non-significant
and the averaged weighted effect sizes were in the small to negligible
range. It is possible that this is in part an artifact of the method of
comparison since there are substantial learning and practice effects
expected on some neuropsychological tasks when one's own baseline
experience is compared with later retesting (even if alternate forms
are used, which was not often the case in these studies; Dikmen et al., 1999). For a relative comparison of
the baseline data, see the revised point estimates and dashed line
confidence intervals in Figure 1 that
represent an adjustment given the typical effect size found when using
repeated measures (Dikmen et al., 1999; Temkin et al., 1999). A quick comparison reveals
that the adjusted weighted effect sizes start to look more like the
effect sizes of the normative and control comparisons than their
unadjusted counterparts. The relative differences of the unadjusted and
adjusted effect sizes and confidence intervals suggest that perhaps a
different mode of analysis is required when evaluating cognitive domain
outcomes compared with one's own baseline. Rather than an absolute
effect size contrasted with zero, a relative change from the expected
normative practice/learning curve effect could indicate a
“significant” effect. For example, while the baseline
comparison for executive function reveals an effect size of −.28
compared with zero (a small, non-significant effect), the adjusted
effect size (taking into account the average learning/practice
normative effect size on executive function tests) is approximately
−.46, approaching a medium effect size (see adjusted point
estimate and dashed confidence interval; CI, in Figure 1).
Because of the different types of control subjects utilized across
studies (e.g., healthy controls vs. cancer patients with
differing treatments), we re-analyzed these two groups of studies
separately for each of the seven cognitive domains. This was done in an
effort to rule out a possible bias (e.g., if most of the studies
utilized healthy controls who might not have similar levels of fatigue
or distress, this could artificially inflate the difference between
cancer patients and these controls). Analyses revealed that five
studies used healthy controls, while nine studies used cancer patients
as controls. All seven comparisons with cancer controls remained
statistically significant, even when omitting the health controls. In
contrast comparisons with health controls sometimes became
non-significant (with effect sizes shrinking) or sometimes became more
highly significant (with effect sizes enlarging). In some domains there
were only a handful of studies that utilized a health control group and
this could have skewed results as well. It seems clear that even if
studies are excluded that utilized health controls (which theoretically
could artificially increase treatment effects due to the relative
comparisons), the significant findings reported here across all seven
cognitive domains are retained.
It is also possible to look for consistency, not only across
cognitive domains, but also across methods of comparison. In applying
this approach, it appears that normative and control comparisons
yielded significant effects across several cognitive domains. In
contrast, baseline comparisons yielded no significant effects; however,
as noted above there are problems in interpreting this apparent lack of
effect. The significance test determines if the effect size is
significantly different than zero, but for baseline comparisons it may
make more sense to compare any effect to the typically positive effect
size that occurs with repeated measures (see above and dashed
confidence intervals in Figure 1 which
represent practice/learning effects found in normative samples due
to repeated measures, even with alternate forms).
Possible Moderator Variable
Given the possible untoward effects of the heterogeneity of diagnoses
and treatments on effect sizes, we recalculated effect sizes excluding
those studies that included patients with more severe diagnoses or
treatments (e.g., some patients within the study's sample may have
had total brain irradiation, metastatic disease, CNS disease, or
immunologic therapies). Twenty studies were deemed more severe and
thus, nine studies were retained for this re-analysis (see Table 1). Table 4 reveals
that 9 of the 10 statistically significant findings with the entire
sample of studies remain significant even with only the subset of less
severe diagnoses and treatments. Thus, it appears that the studies with
more severe components were not having an untoward effect on the
results.
DISCUSSION
The number of individuals living with or having survived cancer is
increasing as a function of earlier disease detection and the
development of more advanced treatments (National
Cancer Institute, 1998). Given this reality, there is a growing
need to understand the long-term impact of cancer treatments and their
side effects on those who receive them. Recent research has begun to
examine the impact of cancer treatments on short- and long-term quality
of life, and some of these studies have explored the possibility that
there are specific cognitive effects of cancer treatments such as
chemotherapy. Reviews of the literature in this area have thus far been
limited to strictly qualitative discussions (e.g., Ahles & Saykin, 2001; Olin,
2001). The intent of this meta-analysis was to provide a
quantitative review that more clearly delineates the nature and
magnitude of the effects detected in existing studies that have
examined the impact of systemic cancer therapies on cognitive function.
Additionally, it was hoped that the results of this quantitative review
would guide future research in this area.
This meta-analysis encompassed 30 studies (leading to 29 samples) for
which sufficient information was available to calculate an effect size
for mean differences between the cognitive function of individuals who
had received systemic treatments for cancer compared to either
normative data, a control group, or their own pre-treatment baseline
functioning. There are several important findings and implications of
this meta-analysis.
When considering all identified studies, we found that for all
averaged effects calculated across various cognitive domains and study
designs, all significant findings were in the negative direction
indicating that mean cognitive test scores for cancer patients having
received systemic treatment were on average lower than those obtained
from normative samples, study control groups, or pretreatment baseline
assessment of the same cancer patients. The absolute magnitude of these
effects across all studies ranged from negligible to large in size. The
most consistent results were found across two methods of comparison
(normative and control), for executive function, verbal memory and motor
function2
It is interesting to note that a study
that was published just as this manuscript was going to press (Harder et al., 2002) reported results that are consistent
with the results of this meta-analysis. Harder and colleagues similarly
reported that impairments were found in patients years after bone marrow
transplantation when compared with normative data on selective attention,
executive function, information processing speed, verbal learning, and
verbal and visual memory.
. The largest effect sizes were found for
executive function and verbal memory when compared with normative data
(
d = −.89 and −.91, respectively).
Despite these consistencies across the normative and control
comparison, the baseline comparison effects remain somewhat puzzling in
that across the seven cognitive domains, these effect sizes remain
smaller than their normative and control comparison counterparts. In
part, this could be due to the fact that at the time of baseline
testing cancer patients may already been experiencing some cognitive
decline due to the effects of stress, fatigue or the sometimes toxic
byproducts of the cancer process itself (e.g., pro-inflammatory
cytokines). If so, later comparisons will not show much relative
decline, potentially obscuring any real decline from premorbid
functioning. Additionally, the absence of a significant effect in
comparison with baseline functioning could be re-interpreted in light
of normative data (e.g., test–retest), which typically shows
positive effects/improvements in scores, due to practice and
learning effects (Dikmen et al., 1999; McCaffrey et al., 2000; Salinsky
et al., 2001; Temkin et al., 1999). In
contrast, the effects for normative and control comparisons are
interpreted relative to zero, which represents the mean for normative
and control samples. If one instead considers the relative
change, then the baseline comparison effect sizes for each
cognitive domain appear more consistent with the effect sizes computed
via normative and control comparisons (see effect sizes adjusted to
account for this in Figure 1). Thus, the fact
that six of seven cognitive domain effects as estimated by baseline
comparisons are in the negative direction (vs. expected
positive/increase), could be interpreted to mean that indeed
treatments are having some undesirable effects.
In a more conservative analysis of only a subset of studies with the
least intense treatments or least severe diagnoses (omitting any study
with brain irradiation, metastatic disease, etc.), we found fairly
consistent results. Nine of the 10 original statistically significant
findings were retained, with the most consistent findings remaining
(e.g., executive functioning, verbal memory, and motor functioning).
Thus, it appears that even when considering only the least severe
diagnoses and least intense treatments, there are still significant
neuropsychological effects of cancer treatments in adults.
Other possible moderator variables besides intensity of treatment or
severity of diagnosis could be examined in future research (e.g.,
simultaneous/contributing effects of other medicines such as
anti-emetics, role of stress, depression, or fatigue). For example,
newer hormonal treatments (e.g., Tamoxifen and Roloxifene) have led to
preliminary examinations of their effects and Bender's (2001) review of hormonal treatments in breast
cancer suggests detrimental effects. However, research regarding
hormonal treatments in other populations suggests a possible protective
effect (e.g., Paganini-Hill & Clark, 2000;
Yaffe et al., 2001). Similarly, the role of certain
medications given in conjunction with treatments (e.g., anti-fatigue
medications, such as Herceptin), have been thought to cross the blood brain
barrier and could also play a role in affecting cognition. Synergistic effects
of combination treatments have also been noted (e.g., radiation combined
with chemotherapy may intensify effects leading to increased cellular
toxicities). Clearly, more careful research is needed that attempts to
document, control for and parse out the potentially confounding effects
of various components of multimodal treatments.
Comparisons of data from cancer patients after systemic treatment
with normative or control sample data have produced the most sizable
effects, while within-subject comparisons of baseline and
post-treatment assessments have generally resulted in smaller averaged
effects. This is an intriguing pattern of findings given that this
latter study design could provide the most convincing evidence for the
presence or absence of cognitive effects of any cancer treatment due to
the designs' control of possibly confounding variables (e.g.,
diagnosis, education, premorbid cognitive functioning). While the use
of normative and control groups for comparison provides useful
information, it is the longitudinal within-subjects designs that allows
for the control of individual differences in the cognitive performance
of patients prior to treatment—although, as noted above,
longitudinal results can also be clouded if the baseline performance is
affected by tumor metabolites, stress of diagnosis, or other factors.
Given these complexities, it is possible that normative and control
comparisons may overestimate decline, while baseline comparisons may
underestimate decline. Nonetheless, such prospective studies might be
thought to be the best designs to assess the true effects of systemic
treatments on cognitive function and thus one might be tempted to
conclude that cancer treatments have little effect on cognitive
functioning. However, such an interpretation would be premature for
several reasons.
First, there are a number of statistical concerns relevant to the
within subject studies. There were typically only two or three studies
available per cognitive domain, and these studies typically had the
smallest samples and the greatest heterogeneity in terms of diagnoses
and treatment. In contrast, the cross-sectional studies have larger
sample sizes and are more homogenous which could account for the
relatively stronger pattern of results. Furthermore, the standard
deviations used in calculating Cohen's d will tend to be
smaller in the within-subject design and could artificially inflate the
effect size. Thus, the reliability of the effect sizes reported in the
within subject design is subject to further investigation. Second, as
noted above, smaller effects observed in repeated-measures designs
could be due, in part, to practice or learning effects that are
artifacts of the instruments used (possible even with the use of
alternative forms). This could certainly obscure actual detrimental
effects of treatments being investigated. It appears that some
accounting must be made to allow for the impact of practice and
learning effects (Dikmen et al., 1999). Few
investigators have assessed cognitive function across time using
alternate forms of tests. Future research should at least aim to
minimize these effects by using alternate forms, including healthy
controls assessed at similar time intervals so as to measure practice
effects, and statistically controlling for typical practice and
learning effects via methods outlined in the literature (e.g., Temkin et al., 1999).
The effect sizes detected across all studies for executive function,
verbal memory, and motor function indicate that, on average,
individuals having received systemic cancer therapies perform
<->1/3 to nearly 1 standard deviation below normative samples or a
control groups in these domains. This level of disturbance is below
that which would typically warrant a label of impairment by traditional
neuropsychological standards (typical definitions range from −1
to −3 SDs). The deficits found here would not
necessarily be expected to translate into easily observable functional
difficulties in most patients, although given the variability in each
sample there are likely some patients that do exhibit such profound and
obvious problems. However, if these overall effect sizes are indeed
accurate indications of the neuropsychological effects of such
treatments, it is not surprising that they have been difficult to
detect with small samples or clinical observation.
It is possible to better appreciate the level of deficit patients may
be experiencing if we translate our findings into a more easily
interpretable index, that is the percentage of people falling under any
area of the normal curve. Mapping out the results in this way for the
cognitive domain of executive function, we learn that while an average
person who has not received any systemic cancer treatment would perform
at the 50th percentile, the average individual having received such
treatment might be expected to fall somewhere between the 15th and 35th
percentile (i.e., their performance is below that of 65–85% of
the normative comparison sample). Viewed in this way, it becomes more
understandable why some patients may be more disturbed by and convinced
of a deficit in cognitive functioning even when the scientific and
medical communities struggle to verify (or disprove) its existence.
Clearly an individual of average pre-morbid ability who can no longer
perform cognitively as well as most of the individuals s/he once
bested may experience this as distressing. For individuals of high
pre-morbid ability who would perform at the upper end of the normal
curve, the loss of <->1/3 to 1 standard deviation would translate
into a smaller percentile difference. However, even such small
differences may be experienced as detrimental to an individual who is
used to performing in a professional position that demands peak
cognitive skills.
A common theme that may tie these deficits across domains together
could be an underlying phenomenon of slowing in mental and/or
physical processing. Fatigue, which has been well documented as a
frequent side effect of treatment that may not resolve readily with the
end of treatment (Piper, 1998), may account
for these effects. Anecdotally, patients often report that they just
don't feel as “sharp” as they did pre-treatment or
that their mind doesn't seem to “work as quickly” as
it did in the past. Indeed, treatments for fatigue such as central
nervous system stimulants (e.g., Ritalin) have been used to improve
quality of life and apparently cognitive functioning in cancer patients
(e.g., Weitzner et al., 1995). This suggests
a possible underlying mechanism that might lie in subcortical processes
or other pathways that could affect basic speed of neuronal functioning
(as in motor functioning) and/or slow more complex higher order
cognitive functioning seen in executive functioning, perhaps
specifically a working memory component that may overlap with the
verbal memory and executive function deficits found here. Fatigue is
only one of many possible underlying mechanisms that could explain
changes in cognitive functioning. Furthermore, it is possible that
there exists some “third variable” that leads to a syndrome
or cluster of symptoms including both fatigue and cognitive problems,
such as a common biological response to chemotherapy. Other changes in
central nervous system functioning could be influential as well, such
as white matter changes perhaps due to demylination, vascular changes,
or formation of neurotoxic cytokines or free radicals that could cross
the blood-brain barrier. Similarly, impacts of treatments on the
peripheral nervous system (well documented in some cases) could account
for decrements reported here, especially in motor functioning, but also
for other cognitive domains. Clearly more research is needed to
illuminate possible mechanisms.
Limitations
Taking into consideration that published data may be biased toward
positive findings, with other null findings being present in
unpublished data, this meta-analysis is subject to the file drawer
problem described by Rosenthal (1991).
However, calculations suggest largely robust findings for the control
and normative comparisons. While every attempt was made to locate all
available data, it is possible that other studies have been conducted
for which data were not accessible. Such studies could be expected to
support the null hypothesis (no difference between groups or assessment
points) and would contribute to smaller average effect sizes than are
reported here. Additionally, some studies had to be excluded due to a
lack of sufficient statistical information in a report, most likely due
to variability in journal publication requirements (e.g., the use of
cut-offs or categorization of continuous variability vs. the
use of means and variability measures). Future research would do well
to promote more uniformity in reporting of results. Perhaps at a
minimum, means and standard deviations for subgroups need to be
included in all published studies in this area.
Another limitation of this meta-analysis is the non-specificity with
respect to a particular cancer diagnosis or treatment. There was a need
to pool studies including various disease groups and treatment regimens
given the paucity of data on this topic and the fact that studies did
not always specify the diagnostic group(s) or specific systemic
treatment regimen(s) delivered in the cancer patient sample. This led
to the inclusion of a wide variety of diagnoses (e.g., breast,
prostate, lung) and treatments (e.g., cyclophosphomide, interleukin-2,
Tamoxifen) in these analyses. This heterogeneity can be viewed both as
a weakness and a strength of this study. Heterogeneity can make
interpretation of the analyses more difficult, but it can also point to
the generality of the findings. Heterogeneity allows for a more
rigorous test of the hypotheses; when there is a significant finding
despite of heterogeneity, this suggests that the finding is quite
resilient. This meta-analysis is meant to serve as a first, broad
attempt to quantify the possible cognitive effects of cancer treatments
and should not be used as an indicator of any specific treatment's
neuropsychological effects. Only with additional well-designed research
in this area will it be possible to exact conclusions regarding a
specific regimen or drug.
Strengths
In spite of the limitations noted, this meta-analysis provides a
much-needed quantitative overview of the research on neuropsychological
effects of cancer treatments. Rather than describing cognition as a
general function, this meta-analysis isolated various cognitive domains
and calculated separate effect sizes for each domain. In addition,
separate effect-size analyses were conducted for the different study
designs and subsequent methods of comparison used (i.e.,
between-subjects with normative data comparison, between-subjects with
control group comparison, or within-subjects repeated measures
comparison).
Furthermore, the control group comparison was explored in more detail
since one could argue that the comparison of patients with matched
healthy controls could lead to different results than comparison of
patients with cancer patients who were receiving a different (a milder
treatment such as localized radiation), or no treatment at all. It is
possible to argue that the comparison with other cancer patients would
be preferred, in that these patients would presumably be matched on
other variables that could affect cognitive function (e.g., distress,
fatigue). In fact, we found that even when considering cancer patient
controls alone, analyses revealed the same significant results. Thus,
we were reassured that the finding of a significant effect in the
combined control sample is genuine (that is, the healthy controls are
not unduly overpowering an otherwise null effect size).
By examining the effects in specific cognitive domains, the possible
role of study designs and the variety of control subjects, this
meta-analysis provides a more detailed picture of study results as they
relate to both methodology and specific neuropsychological function.
Conclusions and Suggestions for Future Research
Undoubtedly, the strongest statement that can be made based on these
analyses is that there is a great need for additional well-designed
studies investigating this phenomenon. The data are consistent with the
idea that there is an adverse effect, with the magnitude of the effect
being more pronounced in some domains (e.g., executive function, verbal
memory and motor function) than others. The strongest effect found in
these analyses was based on a pooled sample of 511 patients. However,
some other effects were based on fewer than 70 individuals. Clearly it
is necessary to collect data from larger numbers of individuals in
order to draw accurate and reliable inferences. In addition to data
from a larger number of subjects, the studies conducted in this area
must be designed to answer the question at hand: specifically, does a
specific cancer treatment cause cognitive decline in those who receive
it? To date, the data in this area consist primarily of a single
assessment collected post-treatment and compared to a normative or
control sample. Such cross-sectional between-subject designs are not
fully adequate to determine whether any detected difference is truly
due to the cancer treatment received rather than to some other
individual difference such as pre-morbid cognitive ability. Through
research using a longitudinal repeated-measures design, including a
pre-treatment assessment, we will gain a better understanding of the
causal nature of any effects seen. Future studies must incorporate such
designs or the data will add little to what we already know.
Ultimately, it would be best to see a clear convergence of results from
all three types of study designs (methods of comparison). The needs for
data from a larger pool of subjects as well as more within-subjects
designs are both underscored by the fact that the effect sizes from the
within-subject findings in this meta-analysis are the most tenuous.
Given that neuropsychological functioning is a multi-faceted
phenomenon, future studies should continue to use test batteries that
assess various cognitive domains. In this way, differential effects on
specific cognitive abilities may be detected. The results of this
meta-analysis suggest that executive function, verbal memory and motor
function may be areas of particular interest to future investigators.
However, other cognitive domains discussed above should not be ignored
as these phenomena are still in the very early stages of investigation.
Use of alternate test forms should also be part of the standard
repeated-measures study design whenever reliable alternate forms of a
test are available for a given cognitive domain. This will help
minimize the washing out of any true differences due to practice or
learning effects. Additionally, comparisons of results relative to
typical practice or learning effects should be considered as well.
Future studies must also assess additional factors that may play a
role in both cancer treatment and neuropsychological function.
Specifically, depression, anxiety, sleep disturbance and fatigue are
likely candidates that may result from a cancer diagnosis and
treatment, and which have documented impacts on cognitive performance
(Daly et al., 2001; Tiersky et al., 1997). However, assessment of these
factors has not historically been standard practice in the studies
reviewed. In the handful of studies that did measure and control for
one or more of these factors, the conclusion was generally that results
retained their significance despite consideration of such potential
mediators or moderators. Twelve of the studies included in our analyses
attempted to determine the impact of potentially relevant mediator or
moderator variables (e.g., age, type of treatment, fatigue, mood; Brezden et al., 2000; Capuron et
al., 2001; Denicoff et al., 1987;
Meyers et al., 1995; Oxman
& Silberfarb, 1980; Peace et al.,
2002; Riggs et al., 2001; Schagen et al., 1999; Silberfarb et
al., 1980; van Dam et al., 1998; Walker et al., 1996; Wieneke &
Dienst, 1995). Eleven of these twelve studies found that the variables
they measured did not remove or significantly account for the cognitive
declines identified in their samples; only Oxman and Silberfarb (1980) found some significant impact of such
variables. This suggests that the effects observed may not be easily
explained away by other variables.
Nonetheless, future research should still strive to measure potential
mediator or moderator variables since a larger sample of studies might
yield sufficient power to detect potentially small effects. Ganz (1998) argues that treatment characteristics (e.g.,
a dose effect) may explain some equivocal results. Several reviews have
pointed to a number of other possible contributing factors in the
apparent cognitive decline (e.g., side effects of other medications,
stress, etc.), as well as possible physiological explanations for the
presence of some data showing cognitive declines (e.g., frontal lobe
dysfunction seen on EEG with some treatments, a link between Tumor
Necrosis Factor and white matter disease in animal models; Meyers, 2001; Silberfarb,
1983; Silberfarb & Oxman, 1988;
Tope et al., 1993). Even one's
subjective experience could be important to assess, perhaps implicating
a type of self-fulfilling prophecy (Cull et al.,
1996; Devlen et al., 1987a, 1987b). Future research should incorporate these
types of variables into sophisticated analyses that might detect
possible relationships to individual neuropsychological performance.
Investigators must be mindful to assess additional variables, such as
those suggested above, that may impact the effects being explored and
use appropriate analytical techniques to examine those variables as
mediators, moderators, or confounds in the relationships identified.
In summary, given the data available to date, it appears that cancer
treatments in adults can impact neuropsychological functioning in those
who receive them. These effects fall primarily in the mild to moderate
range and warrant further investigation. The most consistent effects
across most methods of comparison are seen in executive function,
verbal memory and motor skills. The implications of these findings for
the clinician are not yet clear. As the first quantitative review,
caution should be used in interpreting this data and future research
should lead to updated meta-analyses that might clarify the finer
points of interest to clinicians and their patients. Future research
may clarify the possible need for a more detailed informed consent
process for cancer treatment and more focused quality of life
discussions between clinicians and their patients. Future research may
provide some validation of some patients' experiences, it might
increase anxiety for others anticipating treatment, or relieve those
who can anticipate smaller or more reversible declines than they might
imagine. Additionally, interventions might be designed to assist
patients in coping with any anticipated cognitive changes (e.g., Ferguson & Ahles, 2002). Clearly, it would be
immensely helpful if we could clarify which specific treatments might
be expected to lead to a detrimental effect and how long any specific
effect might last so that patients can make informed choices and
prepare for any impact on work or relationships. This meta-analysis
serves to highlight some aspects of a complex jigsaw puzzle that still
has many pieces missing.
