International Journal of Epidemiology
©International Epldemlological Association 1389
Vol.
18,
No.
4
Printed in Great Britain
A Meta-Analysis of Acupuncture for
Chronic Pain
MAHESH PATEL,* FEUX GUTZW1LLER," FRED PACCAUD" AND ALRO MARA2ZI**
Patel M (Route de Saint-Loup 4,1290 Versoix, Switzerland), Gutzwiller
F,
Paccaud
F and
Marazzi A. A meta-analysis of
acupuncture for chronic pain.
Internationa/Journal
of
Epidemiology
1989,18: 900-906.
Results
of 14 randomized controlled
trials
of acupuncture for chronic
pain
were pooled
in a
meta-analysis
and
analysed
in three subgroups according to site of
pain;
and in two subgroups each according to type to
trial,
type of treatment,
type of
control,
'blindness' of participating
agents,
trial
size,
and
type
of journal in
which
results
were
published.
While
few individual trials had statistically significant results, pooled results of many subgroups attained statistical signif-
icance in favour of acupuncture. Various potential sources of bias, including problems with blindness, precluded a
conclusive finding although most results apparently favoured acupuncture.
While acupuncture is increasingly used by the general
public and treatment costs are often reimbursed by
health insurance companies, its clinical efficacy
remains scientifically unproven.
u
This study was
undertaken to investigate the hypothesis that the indi-
vidually inconclusive trials performed to date might,
when their results were pooled in a meta-analysis
(MA),
yield a more definitive result.
MATERIAL
This MA
is
based on results of
all
trials of acupuncture
for treatment of chronic pain, published in English,
listed in Index Medicus from 1970 onwards that were
randomized controlled trials (RCTs) of chronic pain
that measured outcome in terms of number of patients
whose condition improved. The World Health Organ-
ization's collection of (English language) Chinese, tra-
ditional, and alternative medical journals and
literature yielded additional trials. References were
obtained from previous reviews
34
and from Catherine
Hill's,
as yet unpublished, excellent and comprehen-
sive bibliography.
Trials were discarded if they were uncontrolled
j
5
"
10
not randomized""
13
or did not measure results in terms
of numbers of patients improved and provide the
number of patients randomized.
14
~
16
A complete list of
excluded trials is not presented.
Published study plans of
the
selected RCTs are sum-
marized in Table 1. Few published baseline data after
•Route de Saint-Loup 4, 1290 Versoix, Switzerland.
"Institut Universitaire de Medecine Sociale et Preventive, 17 Rue du
Bugnon, 1005 Lausanne, Switzerland.
randomization.
17
Additional technical,
18
methodolog-
ical" and paradigmatic
20
problems in evaluation of
alternative medicine have been reviewed elsewhere.
Formula acupuncture (FA) uses a set of fixed points
repeatedly. Classical acupuncturists (CA) traditionally
vary points used from patient to patient, and from
treatment to treatment. Most trials achieved 'Teh Chi'
or a 'needling' feeling, ie numbness in the area of the
needle, proof that a point has been located correctly.
The 'control' was sometimes a continuation of
medical treatment. Although these RCTs focused on
chronic rather than acute pain, thus reducing the prob-
ability of remission, use of continued conventional
treatment controls is still unsatisfactory. Transcutane-
ous neural stimulation (TNS) was often used, some-
times on acupuncture points. Treatment of chronic
pain with medical placebo (sugar pills) was not per-
formed. Placebos most frequently used included
placebo acupuncture and mock TNS. Clearly, blind-
ness of patients was possible only when placebo acu-
puncture was used. Relative costs were not mentioned
in any trial.
METHOD
'Meta-analysis'
2
'"
22
is a set of methodological tech-
niques used to define accumulated knowledge by pool-
ing results of studies.
23
Methods used differ according
to homogeneity of study
outcome.
RCTs
selected were
tested for homogeneity using the 'Q' statistic.
2<
The
overall pooled risk difference and its 95% confidence
interval (CI) between acupuncture and control groups
was
evaluated on the
basis
of
a
'random effects' model,
necessary as a result of the lack of underlying homog-
900
TABLE
1 Acupuncture studies
Study First author
number and year
Low back pain
1 Coan 1980
2 Laitincn 1976
3 Edelist 1976
4 Fox 1976
5 Mendelson 1983
Head and neck
6 Hansen 1983
7 Coan 1982
8 Loh 1984
9 Dowson 1986
10 Petrie 1983
11 Loy 1983
Other
12 Berry 1980
13 Ghia 1976
14 Godfrey 1978
Site of pain
Low back pain
Low back pain
Low back pain
Low back pain
Low back pain
Chronic facial pain
Neck pain
Headache
Headache
Cervical pain
Cervical spondylosis
Shoulder-cuff lesions
Pain below the waist
Musculoskeletal pain
Treatment group
Experimental Control
CA, EA
FA
EFA
FA
CA
CA
CA,ECA
CA, EA
FA
FA
ECA
FA
EFA
FA
Delayed acupuncture
TNS on acupuncture pts
Placebo acupuncture
TNS on acupuncture pts
LJdocaine injection
+ placebo acupuncture
Placebo acupuncture
Medical (continued)
Medical (continued)
Placebo TNS
Mock TNS
Physiotherapy
Placebo
Tender spots on neck
Inappropriate acupuncture
Number in groups
Exper
25
50
15
12
95
20
15
55
25
7
26
12
19
88
Control
25
50
15
cross over
cross over
cross over
15
cross over
23
6
27
12
19
84
Number pats, better
Exper.
19
29
6
8
26
12
12
24
8
7
6
5
8
53
Control
5
23
5
6
22
9
2
9
6
2
3
9
9
45
Test
of sig-
nificance
No
NS
No
NS
NS
No
No
NS
No
NS
NS
NS
Retest
on
rates
NS
NS
NS
NS
NS
NS
*
NS
NS
NS
NS
Blindness
None
Patients
Patients + evaluators
None
Patients + evaluators
Patients
None
None
Evaluators
None
None
Assessor
Patients
Triple blind
1ETA-ANALYSIS
O
n
JPUNCTURE
FOR
C
I
pa
O
MIC
PAIN
1
Note: Studies numbered as in Figure 1. Crossover studies summed over treatment order.
Legend: CA ~ acupuncture point selection varied according to the needs of the patient.
FA = formula acupuncture, a standard set or sets of points applied to all patients.
ECA or EFA
a
acupuncture or formula acupuncture in conjunction with electrical stimulation of the needles.
TNS
—
transcutaneous neural stimulation.
= p<0.05. *• - p<0.01. NS = no significant difference.
902
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
eneity of the studies. Cochran's semi-weighted estima-
tor for the risk difference wasused.
M
The
results of
two
additional indicators, the logarithm of the odds ratio,
and the logarithm of one minus
(1—)
the relative differ-
ence,
were also examined."
Complete information on crossover studies was
unavailable in published trial reports, and these were
treated as two independent samples summed over
treatment order. If there is agreement between treat-
ments, this procedure simply loses power.
25
In order to
be certain that confusion was not generated by pro-
cedural differences between crossover and standard
randomized trials, these were also analysed as two sep-
arate subgroups for each pain site.
For detailed evaluation, trials were classified (Table
2) into one of three subgroups according to the general
anatomical site of pain (lower back, head and neck,
and other sites) and into one of two subgroups accord-
ing the nature of the control (placebo or treatment).
Classical acupuncture was distinguished from formula
acupuncture. Trials in which any agents (patients, ther-
apists,
or evaluators) were blind were identified. Large
trials,
50 patients or more, were identified. Trials pub-
lished
in
journals with the words 'Chinese' or 'acupunc-
ture'
in their title were distinguished from those in
'traditional western' medical journals.
Due to the large number of these classification cri-
teria and the small total number of trials, subgroups
based on combined classification criteria (eg. partially
blind trials using classical acupuncture) were not
exhaustively analysed. There are 192 ways in which
TABLE
2 Composition of subgroups
Subgroup (Homogeneity)
Low back pain**
Head and neck pain**
Other site of pain**
Crossover design**
Standard design*
*
Placebo control
**
Conventional treatment**
Classical acupuncture**
Formula acupuncture*
Some blindness*
No blindness**
Large trial (50 or over)**
Small trial (less than
50)*
*
Chinese medical journal*
*
Mainstream medical journal**
Study numbers'
1,2,3,4,5
6,7,8,9, 10, 11
12,
13, 14
4,
5, 6, 8
1,2,3,7,9, 10, 11, 12, 13, 14
3,6,9, 10, 12, 14
1,2,4,5,7,8, 11, 13
1,5,6,7,8
2,3,4,9, 10, 12, 13, 14
2,3,5,6,9, 12, 13, 14
1,4,7,8,
10, 11
1,2,5,8,
11, 14
3,4,6,7,9, 10, 12, 13
1,2,7
3,4,5,6,8,9,10,11,12,13,14
1
Studies numbered as in Table 1.
* Studies not homogeneous at p<0.05.
** Studies not homogeneous at p<0.01.
two criteria could be combined. Hence residual, and
potentially relevant, heterogeneity within initial
groups could not be systematically avoided.
Editors may have been biased against publishing
inconclusive studies. Pooling of published studies
would then obtain a biased result. The influence of
publication bias was evaluated by estimating the
number of unpublished randomized controlled trials
of
acupuncture for the treatment of chronic pain with
inconclusive results that would need to exist in order to
negate the findings obtained
26
(the 'file drawer'
problem).
In order to ensure that MA methodology was com-
prehensively applied, the analysis was measured
against
a
list of qualities and
a
scoring system proposed
for medical MAs.
27
DATA
Five trials dealing with low back pain (trials 1 to 5 in
Table 1) met the selection criteria. Coan did not report
statistical analysis of
results.
28
Results were statistically
significant though long-term follow-up showed regres-
sion of beneficial effects. Laitinen
29
and Edelist
30
did
not attain significance. Fox
31
inserted only three
needles, unilaterally, for one minute at each point
sequentially. These third and fourth trials offered less
treatment than is conventional. In Mendelson,
32
both
patients and the final evaluator of pain were blind,
potentially confusing effects of treatment order were
noted, and no statistically significant results were
obtained.
The second group treated headache, neck pain, cer-
vical pain, cervical spondylosis and chronic facial pain
and included six trials that met the selection criteria
(trials 6 to 11 in Table 1). In Hansen, placebo treat-
ment involved superficial insertion of acupuncture
needles at non-acupuncture points. A pain index
yielded a Wilcoxson test with 0.05>p>0.025, and a
sign test of subjective preferences of patients yielded
p = 0.035, one-tailed, both in favour of acupuncture.
33
Coan
34
and Loh
35
both attained statistical significance.
These latter results in favour of acupuncture may have
been pure placebo effects, as control groups merely
continued medical treatment. Dowson
36
used a true
placebo (mock
TNS),
and did not obtain statistical sig-
nificance. Cervical pain responded significantly
(p<0.01) to acupuncture in one small study
37
but not in
another.
38
Three trials (trials
12
to
14
in Table
1)
treated varied
diagnoses. A single blind trial of multiple therapies for
shoulder cuff lesion is one of two trials in which treat-
ment did worse than control.
39
The second evaluated
acupuncture for chronic pain below the waist against
A META-ANALYSIS OF ACUPUNCTURE FOR CHRONIC PAIN
903
'tender area needling'. Control did not explicitly
exclude classical acupuncture points.
40
In Godfrey
'most appropriate' acupuncture points, were com-
pared to 'least appropriate' points. Triple-blindness
may have been achieved. Directions were given by an
acupuncturist who had evaluated the patient to
another
who
could not see the patient (whose head was
hidden by a screen). No statistically significant results
were obtained.*
1
RESULTS
Individual
Trials
A Chi-squared test on proportions of patients
improved in studies that had not published statistical
tests of
results,
2834
-
33
yielded values of
p<0.01.
The risk
difference is illustrated in Figure 1 for the 14 trials.
Only two out of 14 trials obtained the result that
patients treated with acupuncture did worse, on aver-
age,
than the control group. The 95% CIs for four of
these
14
trials did not include the 'zero risk difference'
result. All four favoured acupuncture.
Meta-analysis
of
Cumulated Trials
Studies were not homogeneous according to the
Q-statistic (p<0.01). The overall risk difference (indi-
cator 'A') between acupuncture and control groups
was 0.184 (SE = 0.062), in favour of acupuncture
(p<0.01). Acupuncture was also superior overall
according to the logarithm of the odds ratio (indicator
'B') and the logarithm of '1—the relative difference'
(indicator 'C').
1.
Coan -
2.
Laltlnen -
3. Ede)l8t
4.
Fox -
6. Mendel8on -
6. Hansen
7. Coan
8. Loh
8. Dowson
10.
Petrle
11.
Loy
12.
Berry
13.
Ghla
14.
Godfrey
Risk Difference
How, +
Central,
H High
None of the subgroups (Table
2)
within which results
were cumulated were homogeneous according to the
'Q'
test (p<0.05 for all subgroups).
Ninety-five per cent confidence intervals for the risk
difference estimator (indicator 'A') are presented as
Figure 2. The results for all three indicators are pre-
sented as Table 3.
There was quite good agreement between results
obtained from the three summary statistics used. In 14
of the 19 subgroups analysed, indicators 'A', 'B', and
'C,
provided the same result in terms of significance
(or insignificance) of the difference between treatment
and control groups.
Considering pooled results by site of pain, only the
subgroup of trials for head and neck pain attained sig-
nificance for all three indicators. Low back pain
attained significance in favour of acupuncture accord-
ing to indicators 'B' and 'C if crossover trials were
included. Results for other sites of pain showed an
insignificant result in favour of the control group.
Acupuncture compared to conventional treatment
was more favourable to acupuncture than trials against
placebo. Patients receiving classical acupuncture at
sites that varied from treatment
to
treatment did better
than patients receiving formula acupuncture at fixed
sites.
Trials with at least one blind agent were less favour-
able to acupuncture than trials without blind agents.
Trials with some blindness did not attain significance
for any indicator.
All trials
Standard trials
Crossover trials
Low baok palru All
Low baok: Standard
Head/neok
pain:
All
Head/neck: Standard
Other pain site: All
Other site: Standard
Placebo oontrol
Treatment control
Classic acupuncture
Formula aoupuncture
Partially blind
No blindness
Large trial
(->50)
Small trial (<60)
Chinese Journal
Western journal
-1
Risk Difference
I-Low,
+
Central,
-\ High
FIGURE
1. 14 RCTs of acupuncture; 95% Cl for the risk difference.
FIGURE
2. Pooled trial subgroups; 95% Cl for the risk difference.
904
INTERNATIONAL JOURNAL OF EPIDEMIOLOGY
'A'
Weighted average
of risk difference
0.184*
0.196*
0.147*
0.191
0.258
0.303*
0.361*
-0.056
-0.056
0.105
0.235*
0.329*
0.092
0.048
0.398*
0.183*
0.177
0.441*
0.109*
Indicator
'B
1
Logarithm of the
odds ratio
0.645*
0.633
0.680*
0.753*
1.069
1.124*
1.328
-0.173
-0.173
0.202
1.000*
1.411*
0.260
0.228
1.662*
0.810*
0.422
1.947*
0.383*
•c
Logarithm of '1—the
relative difference'
-0.219*
-0.223
-0.221
-0.299*
-0.461
-0.293*
-0.304
0.054
0.054
-0.102
-0.309*
-0 513*
-0.119
-0.081
-0.555*
-0.250*
-0.163
-0.886*
-0.135*
TABLE
3 Results obtained with meta-analysis of results of all trials and subgroups of trials
defined,
using the risk difference, the logarithm of the odds
ration, and the logarithm of
7—the
relative risk difference', with tests of significance at 95% presented (')
Trial group or subgroup
All trials
All standard trials
All crossover trials
All trials for low back pain
Standard trials for low back pain
All trials for head and neck pain
Standard trials for head and neck
All trials for other pain sites
Standard trials for other sites
Placebo control group
Treatment control group
Classical acupuncture
Formula acupuncture
At least partial blindness
No blindness in the trial
50 subjects of more (large trial)
Less than 50 subjects (small)
Chinese or acupuncture journal
Mainstream medical journal
Large trials were more favourable to acupuncture
than small
trials.
Results published
in
journals that had
the word 'acupuncture' or 'Chinese' in their titles were
significantly superior to those reported in 'traditional
western' medical journals, but both these groups
showed results favourable to acupuncture according to
all three indicators.
The methodology of MA was comprehensively
applied. Issues were covered in all six major areas of
quality control of medical MAs. Twelve out of
23
rele-
vant items were addressed. The analysis would be
placed in the top decile of the 86 MAs reviewed by
Sacks.
That review obtained a mean of 7.7 items add-
ressed with standard deviation of 2.7.
While some weaknesses of MA may not yet be fully
appreciated,'
12
known problems certainly include 'pub-
lication bias' and 'author self-selection bias'. The
necessary number of unpublished acupuncture RCTs
for chronic pain that had, on average, inconclusive
results (zero risk difference) to negate the statistical
significance of the pooled result in favour of acupunc-
ture was 26 trials at p =
0.01,
and 67 trials at p = 0.05.
DISCUSSION
Blindness
is
the only
means
of avoiding
bias
due to pre-
conceived notions of the superiority (or inferiority) of
a treatment. The 95% CI for the subgroup of 'trials
with some blindness' includes the zero risk difference
possibility for all indicators. If the four trials with
statistically significant results are considered, only one
common characteristic emerges. None had any degree
of
blindness.
According to Godfrey,
41
full triple blind-
ness is technically feasible. If feasible, it should be
regarded as essential.
There are two potential explanations for the appar-
ent superiority of CA over
FA.
Either efficacy of
CA
is
superior, or the protocols of CA trials were inferior.
Two out of
six
CA trials, and six out of eight FA trials,
displayed some degree of blindness.
The result that trials against conventional treatment
are more favourable to acupuncture than are trials
against placebo may similarly be explained. Five out
of
eight of the former had no blindness, and only one out
of
six
of the latter.
The superior results obtained by trials published in
journals oriented towards Chinese medicine or acu-
puncture may indicate selective publication bias. It is
possible that acupuncture treatment described in
specialist journals was superior, or that study methods
differed between publication subgroups. All RCTs
published
in
journals with the
words
China or acupunc-
ture in their titles were trials against conventional
treatment.