Pain, 5(1 (1992) 281-285 281

© 1992 Elsevier Science Publishers B.V. All rights reserved 0304-3959/92/$05.00

PAIN 02095

Regression to the mean in treated versus untreated chronic pain *

Coralyn W. Whitney a and Michael Von Korff b

"Departments of Dental Public Health Sciences and Oral Medicine, SM-35, Unicersity of Washington, Seattle, WA 98195 (USA)
and b Center for Health Studies, Group Health Cooperatice of Puget Sound, Seattle, WA (USA)
(Received 7 June 1991, revision received 12 February 1992, accepted 25 February 19921

Summary The course of pain associated with temporomandibular disorders (TMD) and other chronic pain
conditions is typically episodic. Its expression may influence when a person seeks treatment, for example, when the
level of pain flares up or exceeds its characteristic severity. Improvement in pain status subsequent to entering
treatment may be due to: (1) specific effects of treatment; (2) non-specific effects of treatment ('placebo effects'); or
(3) regression to the mean. Due to regression to the mean, uncontrolled evaluation of treatment in persons
self-selected by a pain flare-up may lead to erroneous conclusions concerning effects of treatment by patients,
providers, and/or researchers. For this report, the magnitude of regression to the mean due to self-selection for
treatment is estimated by comparing subjects who sought treatment for TMD pain (n = 1471 to a random sample of
subjects with TMD pain not seeking treatment (n = 95). Among subjects seeking treatment, a significant 14.7-point
reduction in VAS pain intensity was observed at 1-year follow-up. A control group of TMD subjects not seeking
treatment showed no mean reduction in pain intensity but reported lower pain intensity at baseline than the group
seeking care. When both groups of subjects were stratified on baseline VAS pain values, the reduction in pain
increased as the baseline pain level increased, but no differences between comparable treated and untreated cases
in the extent of improvement were observed. The before-after differences in both groups may be attributed to
regression to the mean. We conclude that before-after differences in pain intensity can be large and that such
improvement may be largely due to regression to the mean. This sugg~:sts the need for research which differentiates
change due to regression to the mean (due to homeostatic processes, random within-subject variation, or
measurement error) from change due to specific and non-specific effects of treatment. In clinical practice, the
personal experience of patients and clinicians who observe improvement after initiation of treatment should be
regarded as an unreliable guide to treatment efficacy due to regression to the mean. This phenomenon may
contribute to the proliferation and continued use of treatments of unproven efficacy for pain management and
suggests caution in the use of costly or risky pain treatments the efficacy of which is unknown.

Key words: Regression to the mean

Introduction the dominant characteristic of most long-lived condi-

"One characteristic of almost everything that per-
sists for a long time is that it changes with time. It is
changeability and variation, not stability, that is in fact
* Supported by NIDR Grant No. P01 DE08773 and AHCPR Grant
No. R01 HS06168.
Correspondence to: Dr. Coralyn W. Whitney, Dept. of Dental
Public Health Sciences, SM-35, University of Washington, Seattle,
WA 98195, USA. Tel.: (206) 543-2034.
tions" (Sartwell and Merrell 19521.
Like many other pain conditions, the course of
temporomandibular disorder (TMD) pain is episodic
(Von Korff et al. 1988). Thus, the level of pain at any
single point in time does not indicate where individuals
are in relation to their characteristic level of pain (or
across-time central tendency). The present status of a
person in their cycle of pain, along with the variability
and central tendency of their cycle, determines the
potential magnitude of change in the level of pain
between the present and some future point in time.
282
The likelihood of a change of a particular magnitude is
determined by the proximity of the initial measurement
to its characteristic level. The closer the level of pain is
to an extreme of an individual's across-time distribu-
tion of pain, the greater the likelihood of the next
measurement being less extreme. In statistical terms,
this phenomenon is known as 'regression to the mean'
(Galton 1886; Davis 1986). That is, a variable that is
extreme when it is measured will tend, by chance
alone to be closer to its central tendency on a subse-
quent measure. Studies which base subject selection on
the outcome variable of interest, in this case pain, are
at risk of inducing this phenomenon (Davis 1976). As a
consequence, if someone seeks treatment when their
level of pain is at its peak, then pain can be expected to
decrease towards its characteristic level whether or not
treatments are given. Similarly, if someone is experi-
encing the lowest level of pain in their cycle, then an
increase in pain would be expected. Pain cycles with
little within-cycle variability will yield small changes
over time and cycles with greater variability will yield
greater changes.
The phenomenon of cyclic variation is significant in
clinical settings. Subjects seeking treatment for pain do
so when there is a flare-up in the level of pain or when
the level of pain is no longer tolerable (Von Korff et al.
1991). If the level of pain and treatment-seeking behav-
ior are related, the efficacy of pain treatment as it
appears to the patient and to the clinician may exceed
the actual reduction in pain due to treabner, t. In such
situations treatment may appear efficacious due to the
initiation of treatment at a point in the pain cycle when
pain would naturally decrease (Turk and Rudy 1990).
Without additional data to adjust for episodic varia-
tion, it is not possible to determine what part of pain
reduction is attributable to treatment and what is due
to cyclic variation.
While regression to the mean is a well-established
and fundamental statistical principle, pain researchers
have paid little attention to establishing its empirical
validity in pain research or to demonstrating its poten-
tial for influencing interpretations of pain treatment
outcome studies.
When a patient seeks treatment for a pain condi-
tion, subsequent improvement may be due to 3 types of
effects: (1) specific effects of treatment, such as the
biologic effects of a medicine or a procedure; (2)
non-specific effects of treatment, sometime referred to
as 'placebo effects'. Miller (1989) identified non-specific
treatment effects as including: the providers attitude
toward the treatment; the provider's attitude toward
the patient (warmth, interest, empathy); the faith of
the patient in the treatment; the reputation, expense,
or impressiveness of the procedure; suggestibility; and
other psychological mechanisms. And, (3) regression to
the mean due to measurement error, to random
within-subject variation across time, or to homeostatic
processes tending to bring an individual's status back
to its characteristic level or set point. McDonald and
Mazzuca (1983), based on a literature review compar-
ing randomized controlled trials with and without mul-
tiple baseline assessments, have argued that regression
to the mean has often been incorrectly attributed to
non-specific, or placebo, effects of treatments. They
suggest that the potency of the placebo effect may have
been overstated due to its confusion with regression to
the mean and advise exercising "caution in interpreting
patient improvements as causal effects of our actions
... [we] should avoid the conceit of assuming that our
personal presence has strong healing powers".
While regression to the mean has been studied in
the context of epidemiologic and clinical research on
medical conditions (Ederer 1972; Gardner and Heady
1973; Davis 1976; Shepherd 1981) it has not been
extensively studied among pain patients. The present
report uses empirical data from an on-going longitudi-
nal study of pain associated with TMD to elucidate the
extent to which statistical regression to the mean might
contribute to the magnitude of apparent effect size in
pain treatment outcome.
The method for calculating the amount of reduction
in pain due to regression to the mean requires the
assumption that the before- and after-treatment levels
of pain follow a bivariate normal distribution and are
correlated (James 1973; McDonald and Mazzuca 1983).
Measurements of pain over time on the same individ-
ual are typically correlated but do not follow a bivari-
ate normal distribution and are generally too skewed to
be normalized. This situation arises, for example, when
post-treatment measures reflect an appreciable num-
ber of subjects who report zero pain on a visual analog
scale (VAS), The result is a non-normal, positively
skewed distribution of post-treatment pain scores. Al-
though the impact of regression to the mean cannot be
calculated by established methods under these condi-
tions, this does not mean the effect cannot be evalu-
ated.
Estimation of the magnitude of regression to the
mean due to cyclic variation is possible through the use
of a control group. The purpose of control groups is to
adjust for sources of variation other than the specific
effects of treatment, yielding a less biased estimate of
treatment effect. The sources of variation of interest
here are the variability of pain over time and the
self-selection for treatment when pain is higher than
usual. However, most controlled studies cannot distin-
guish l~on-specific effects of treatment from regression
to the mean. In addition, studies which classify treated
subjects into responders and non-responders may con-
fuse factors related to elevated pain at baseline with
those predicting treatment response. The primary pur-
pose of this paper is to estimate empirically the possi-
 283

ble m a g n i t u d e o f r e g r e s s i o n to t h e m e a n by c o m p a r i n g magnitude of regression to the mean is shown graphically by select-

p e r s o n s s e e k i n g p a i n t r e a t m e n t to a c o n t r o l g r o u p w i t h ing subjects who, based on their level of pain at baseline, exceeded a
series of VAS threshold values: pain greater than 0, 5, 10, 20, 30, 40,
t h e s a m e p a i n c o m p l a i n t i d e n t i f i e d by a p o p u l a t i o n 50.
s u r v e y r a t h e r t h a n by s e e k i n g care. T h i s o b s e r v a t i o n a l
comparison may provide some insight into the magni-
t u d e of r e g r e s s i o n to t h e m e a n r e l a t i v e to t h e com-
Results
b i n e d specific a n d n o n - s p e c i f i c e f f e c t s o f t r e a t m e n t .

Method
Tile data used are from a longitudinal study evaluating the
clinical signs and symptoms associated with TMD (Dworkin et al.
1990). The treated cases and untreated controls were identified from
the enrollees in the Group Health Cooperative of Puget Sound, a
Health Maintenance Organization with over 320,000 enrollees at the
time of this study. The case group consisted of a consecutive series of
patients who sought treatment and were referred for evaluation of
TMD pain. This group of enrollees will be referred to as clinic cases
(CLCA, n = 147). The control group consists of all individuals identi-
fied in a random sample survey of Group Health Cooperative en-
rollees who reported TMD pain in the prior 6 months but who had
not sought referral for treatment. This group is referred to as the
community cases (COCA, n = 95). TMD pain was defined as "facial
ache or pain in the jaw muscles, the joint in front of the ear or inside
the ear (excluding infection)" in the 6 months prior to interview.
The intensity of pain was evaluated by three 100-mm VAS
ratings: 'worst pain', 'average pain', and 'pain right now' at the time
of interview. 'Pain right now', a measure commonly used to assess
the level of pain for pre-treatment and post-treatment evaluations, is
used in this paper. These pain ratings were obtained at entry into the
study (baseline) and at l-year follow-up. COCAs receiving treatment
between baseline and I-year follow-up were excluded. Some COCAs
had a zero score on the 'pain right now' measure at baseline.
The overall comparison of pain intensity across time between
CLCAs and COCAs, with the COCAs serving as a control group,
was done by analysis of cowu'iance adjusting for pain at baseline. The
E a c h c a s e w a s c l a s s i f i e d into a series o f g r o u p s
b a s e d o n t h e t h r e s h o l d s s t a t e d above. T h e m e a n level
o f p a i n w a s e s t i m a t e d for b o t h C O C A s a n d C L C A s at
b a s e l i n e a n d 1-year f o l l o w - u p . F o r this c o m p a r i s o n , the
C O C A s p r o v i d e a n e s t i m a t e o f t h e a v e r a g e c h a n g e in
p a i n in subjects n o t t r e a t e d in t h e i n t e r i m . T h e result-
ing m e a n V A S v a l u e s a n d p e r c e n t c h a n g e s a r e s h o w n
in T a b l e I.
T h e c h a n g e in m e a n p a i n i n t e n s i t y b e t w e e n b a s e l i n e
a n d 1-year f o l l o w - u p is i l l u s t r a t e d for C L C A s (Fig. l a )
a n d C O C A s (Fig. l b ) for t h e set of t h r e s h o l d values.
O v e r a l l , t h e C L C A s (n = 147) s h o w e d a 14.7-point
r e d u c t i o n in t h e i r p a i n i n t e n s i ~ . T h i s p r e - p o s t differ-
e n c e w a s statistically s i g n i f i c a n t (t = 5 . 4 8 , df= 146,
P < 0.01). In c o n t r a s t , t h e e n t i r e s a m p l e o f C O C A s
s h o w e d n o r e d u c t i o n in p a i n i n t e n s i t y r a t i n g f r o m
b a s e l i n e to follow-up. T h e i r m e a n i n t e n s i t y w a s 12.6 at
b a s e l i n e a n d 14.7 at 1-year follow-up. H o w e v e r , t h e
C O C A s s h o w e d t h e s a m e p a t t e r n in p a i n r e d u c t i o n as
t h e C L C A s w h e n r e s t r i c t e d to t h e s a m e b a s e l i n e
t h r e s h o l d c a t e g o r i e s as t h e C L C A s . F o r t h e t h r e s h o l d
v a l u e o f V A S p a i n g r e a t e r t h a n 20, a 5 3 % r e d u c t i o n in
m e a n V A S p a i n f r o m b a s e l i n e was o b s e r v e d in the
C L C A s , c o m p a r e d to t h e 5 9 % r e d u c t i o n a m o n g C O -
C A s (see Fig. 1). A s i m i l a r p a t t e r n held for t h e h i g h e r

TABLE !
MEAN LEVEL OF PAIN AT BASELINE AND l-YEAR FOLLOW-UP WITH MEAN AND PERCENT CHANGE ACROSS TIME FOR
CASES EXCEEDING BASELINE PAIN THRESHOLD

Pain n Baseline 1 Year Change

threshold Mean S.D. * Mean S.D. Mean Percent
(I CLCA 147 33.70 28.10 19.00 25.20 - 14.70 -44
COCA 59 20. I 1 22.66 15.42 23.64 - 4.80 - 24
All ÷ COCAs 95 12.56 2(I.35 14.65 21.73 2.09 17
5 CLCA 116 42.10 25.74 22.29 26.65 - 19.81 - 47
COCA 36 32.06 21.91 22.56 26.71 - 9.50 - 30
10 CLCA 103 46.44 24.03 24.83 27.20 - 21.61 - 47
COCA 30 36.63 21.20 20.10 24.69 - 16.53 - 45
20 CLCA 83 53.94 20.57 25.40 26.78 - 28.54 - 53
COCA 22 44.73 18.96 18.23 24.04 - 26.50 - 59
30 CLCA 70 59.21 17.49 26.09 27.90 -33.12 -56
COCA 16 52.56 16.12 "~2.06 26.75 - 30.50 - 58
40 CLCA 58 64.60 14.69 27.7; 28.60 - 36.86 - 57
COCA 12 58.00 14.94 27.17 28.66 - 30.83 - 53
50 CLCA 46 69.35 12.66 31.93 29.93 - 37.42 - 54
COCA 7 65.43 15.94 26.43 32.30 - 39.00 - 60

÷ Includes COCAs with no pain at baseline.

* S.D. = standard deviation.
 284

(a) (b)
Pain I n t e n s i t y
Pain Intensity 80
80

• 50 70
70 ~ ' 50
40
60-t 30 60 40
20 30
50 50
10 20
5 40
40 , 10
0 5
3O 30

20 0

>= 0
t 10

t t t 0 t I
Baseline One Year Baseline O n e Year
Clinic Cases Community Cases
Fig. i. Mean pain intensity at baseline and l-year fl~llow-up of cases exceeding the specified VAS baseline threshold: (a) clinic cases (CLCA)
(b) community cases (COCA).

thresholds. Thus, an appreciable reduction in pain
intensity occurred in patients not seeking treatment
when they were sampled on the same basis as patients
entering treatment.
Finally, the baseline distributions of pain for the
CLCAs and COCAs were different. Fig. 2 shows the
relative frequency (%) distributions of VAS pain for
both groups. In the CLCAs 56% of the subjects had a
VAS rating greater than 20 and 31% had a VAS rating
greater than 50. This compared to 22% and 7%, re-
spectively, for the COCAs. For those subjects not seek-
ing treatment the level of pain at baseline was skewed
towards zcro with far more cases at the lower levels of
pain compared to the CLCAs. in this situation, an
analysis of covariance might bc used to adjust h)l"
baseline differences in pain intensity. Comparing tl.e
two groups using their respective mean changes in
pre-post VAS scores ( - 14.7 tbr the CLCAs compared
to 2.09 for the COCAs)is not valid unless it can be
assumed that the CLCAs, with their higher baseline
40 7
Relative Frequency (%)
I In
J COCA
3° i
!
i
f
!
i possible to interpret changes in pain intensity over time
! as due to treatment efficacy, given subjects self-selected
0 -- -- _ _
~' 1- 11- 21- 31- 41- 51- 61- 71-
Baseline Pain
Fig. 2. Relati~,,: frequency (C,~) distributions of baseline pain for clinic
(CLCA) and com,,unity (COCA) cases. Corresponding frequencies
(no.) for CLCAs are: O, 44, 2(I, 13, 12, 12, 15, 12, Ill, 6. 3: for COCAs:
36,29.8,6,3,6,4, 1,0, I, 1.
pain levels, would have had the same change in pain
intensity as the COCAs over time in the absence of
treatment (Anderson et al. 1980). Given self-selection
and the cyclic nature of pain intensity, such an assump-
tion is not tenable. The results of an analysis of covari-
ance showed that the two groups of subjects (CLCAs
and COCAs) did not differ in mean changes of pre-post
VAS scores (F = 0.80, df = 1,239, P = 0.37) after con-
trolling for pain level at baseline.
Discussion
As previously shown, subjects seeking treatment have
higher levels of pain at baseline compared to persons
with pain not seeking treatment. Under such circum-
stances, there is potential for pain levels to show re-
gression to the mean. After stratifying on dtreshold
levels for baseline pain intensity, it was apparent that
reduction in pain at follow-up was greater as the initial
level of pain increased for both the treatment and
contro! groups. Since regression to the mean is ob-
served for both groups, the greater reductions observed
CLCA '
in the CLCAs are in all likelihood an artifact of self-
selection. Apparently, effects of self-selection am5 re-
gression to the mean can be large, in our sample we
estimate, overall, that a 50% reduction in mean pain
intensity from baseline to l-year follow-up among cases
entering treatment may be attributed to this effect.
Thus, in the absence of any comparison group, it is not
for treatment based on heightened pain experience.
81- 91-
Proper evaluation of treatment efficacy requires a con-
trol group. Thus, the trend toward evaluatiott of pain
treatments using randomized, controlled designs, with
the decreased frequency of uncontrolled designs in the
literature, is appropriate.

Moreover, control groaps are also susceptible to
regression to the mean so it is important that the
distribution of pain levels is comparable to the distri-
bution in the treated group. Thus, the cyclic variability
of pain and self-selection for treatment of pain patients
requires that the assumption of baseline comparability
of pain level be explicitly tested. When it is not possi-
ble to construct a completely comparable control group,
an analysis of covariance can be used to adjust for
imbalances. Controlled trials of pain treatments might
benefit from measuring pain on multiple occasions
prior to treatment intervention to reduce regression to
the mean as a source of within-subject change, as is
increasingly done in clinical trials in other areas of
medicine (McDonald and Mazzuca 1983).
The phenomenon of regression to the mean among
patients self-selecting treatment during a pain flare-up
may be important in shaping clinicians' beliefs regard-
ing treatment efficacy. The clinician who routinely
observes improvement in patients following initiation
of pain treatment may attribute the improvement to
the treatment rather than to the natural history of the
condition. Such faulty reasoning may lead clinicians to
regard expensive or risky treatments of limited efficacy
as being valuable in the management of the patients
they see. The antidote for such faulty reasoning is
insistence on evaluation of treatments by controlled
trials.
The phenomenon of regression to the mean also has
implications for patients' beliefs regarding treatment
efficacy. Patient populations may embrace ineffective
pain treatments that are expensive, risky, or both when
their use is followed by a reduction in pain. In a study
currently underway, 164 TMD pain patients were iden-
tified who said they had improved somewhat or a great
deal 3-6 weeks after seeking care. Among these pa-
tients, 85% attributed some of their improvement to
the recommended treatments, suggesting that this phe-
nomenon is common. It has been observed (Shapiro
1960; Miller 1989) that useless or dangerous treatments
have been prescribed by physicians throughout the
history of medicine, seemingly without undermining
patient beliefs in the efficacy. Regression to the mean
may play a role, along with non-specific effects of
treatment, in maintaining patient beliefs in the potency
of medical management of chronic pain. If the phe-
nomenon of regression to the mean plays a significant
role in shaping treatment choices of patients, profes-
sionals who treat patients in pain need to consider the
ethics of use of costly or risky treatments the worth of
which has not been evaluated by controlled clinical
trials.
Finally, the mechanisms of regression to the mean
deserve increased research attention along side re-
search evaluating specific and non-specific treatment
285
effects. Possible explanations of regression to the mean
may include measurement error, within-subject ran-
dom variation in pain levels and homeostatic processes
that tend to return pain to a characteristic level after a
flare-up. Among patient groups in whom significant
regression to the mean occurs, the ~:atural process of
improvement may provide opportunities to reinforce
patient self-care behaviors and enhance patient beliefs
in their own abilities to control pain, rather than rein-
forcing patient beliefs in the efficacy of medical care
for chronic pain. If so, the phenomenon of regression
to the mean may have the potential to enhance patient
autonomy in managing chronic pain just as it may now
contribute to the dependence of pain patients on health
care providers.
References
Anderson, S., Auquier, A., Hauck, W.W., Oakes, D., Vandaele, W.
and Weisberg, H.I., Statistical Methods for Comparative Studies,
Wiley, New York, 1980.
Davis, C.E., The effect of regression to the mean in epidemiologic
and clinical studies, Am. J. Epidemiol., 104 (1976)493-498.
Davis, C.E., Regression to the mean. in: N.L. Johnson and S. Kotz
(Eds.), Encyclopedia of Statistical Sciences, Vol. 7, Wiley, New
York, 1986, pp. 706-708.
Dworkin, S.F., Huggins, K.H., LeResche, L., Von Korff, M., Howard,
J., Truelove, E. and Sommers, E., Epidemiology of signs and
symptoms in temporomandibular disorders. Clinical signs in cases
and controls, J. Am. Dent. Assoc., 120 (1990) 273-281.
Ederer, F., Serum cholesterol changes: effects of diet and regression
toward the mean, J. Chronic Dis., 25 (1972) 277-289.
Galton, F., Regression towards mediocrity in hereditary stature, J.
Anthro. Inst., 15 (1886) 246-263.
Gardner, M..I. and Heady, J.A., Some effects of within-person vari-
ability in epidemiologic studies, J. Chronic Dis., 26 (1973) 781-
795.
James, K,E., Regression toward the mean in uncontrolled clinic~d
studies, Biometrics, 29 ( ! 973) 121 - 13[].
McDonald, C.J. and Mazzuca, S.A., How much of the placebo
~effect' is really statistical regression?, Star. Med., 2 (1983) 417-
427.
Miller, N.E., Placebo factors in lreatment: views of a psychologist.
In: M. Shepherd and N. Sartorius (Eds.), Non-Specific Aspects of
Treatment, Hans Huber, Toronto, 1989, pp. 39-55.
Sartwell, P.E. and Merrell, M., Influence of the dynamic character of
chronic disease on the interpretation of morbidity rates, Am. J.
Public Hlth, 42 (1952) 579-584.
Shapiro, A.K., A contribution to a history of the placebo effect.
Behav. Sci., 5 (1960) 109-135.
Shepherd, D.S., Reliability of blood pressure measurements: implica-
tions for designing and evaluating programs to control hyperten-
sion, J. Chronic Dis., 34 (1981) 191-209.
Turk, D.C. and Rudy, R.E., Neglected factors in chronic pain treat-
ment outcome studies: referral patterns, failure to enter treat-
ment, and attrition, Pain, 43 (1990) 7-25.
Von Korff, M., Dworkin, S.F., LeResche, L. and Kruger, A.. An
epidemiologic comparison of pain complaints, Pain, 32 (1988)
173-183.
Von Korff, M., Wagner, E.H., Dworkin, S.F. and Saunders, K.W..
Chronic pain and use of ambulatory health care. Psychosom.
Med., 53 (1991) 61-79.