Making Sense of Odds and Odds Ratios

David A. Grimes, MD, and Kenneth F. Schulz, PhD, MBA

Downloaded from http://journals.lww.com/greenjournal by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0h

D espite their growing use in the Odds

CywCX1AWnYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC1y0abggQZXdgGj2MwlZLeI= on 09/19/2024

Odds and odds ratios are hard for

medical literature,1,2 odds ra- In contrast to probability, odds are
many clinicians to understand. Odds
are the probability of an event occur-
tios remain poorly understood by not intuitive— except perhaps to
ring divided by the probability of the clinicians (and by some research- gamblers. Odds are simply a differ-
event not occurring. An odds ratio is ers, as well).2 Because of confusion ent expression of the probability:
the odds of the event in one group, about odds and odds ratios, we will the probability of an event divided
for example, those exposed to a drug, provide a brief overview of these by the probability of the event not
divided by the odds in another group terms. Our primary goal is to help happening. Thus, the odds of an
not exposed. Odds ratios always ex- busy clinicians interpret research event would be probability/(1–
aggerate the true relative risk to some reports that use these unfamiliar probability). Because this is a ratio,
degree. When the probability of the terms. Relying on simple examples its values range from zero to infinity.
disease is low (for example, less than from weather forecasts, tubal steril- In the weather example above,
10%), the odds ratio approximates
the probability of rain today is 0.10.
the true relative risk. As the event
Hence, the odds of rain today
becomes more common, the exag-
geration grows, and the odds ratio no For most clinicians, odds ratios would be 0.10/(1– 0.10) or 0.10/
longer is a useful proxy for the rela- 0.90⫽0.11. Here, the odds of rain
will remain . . . well, odd. are 1/9 (one chance “yes” to 9
tive risk. Although the odds ratio is
always a valid measure of association, chances “no”). When the chance of
it is not always a good substitute for rain is low, the probability (0.10)
the relative risk. Because of the diffi- ization, and a deck of playing and odds (0.11) of rain are similar,
culty in understanding odds ratios, cards, we will explain odds and and one could reasonably use these
their use should probably be limited odds ratios and show their relation- terms interchangeably.
to case-control studies and logistic ship to the better-understood prob- But what if a nor’easter is ap-
regression, for which odds ratios are proaching? In this situation, proba-
ability and relative risk.
the proper measures of association. bility and odds of rain diverge.
(Obstet Gynecol 2008;111:423–6) Assume that, with a major storm
BASIC TERMS blowing in, the weather forecast
From Family Health International, Research Trian- Probability calls for a 90% probability of rain
gle Park, North Carolina. Probability is the proportion (ie, (ie, 0.90). The odds of rain would
Supported in part by Family Health International percentage) of times an event be 0.90/(1– 0.90) or 0.90/0.10⫽9.0.
(FHI) with funds from the U.S. Agency for Interna- With the storm approaching, the
tional Development.
would occur if an observation were
repeated many times. Probabilities odds of rain are 9/1 (nine chances
The views expressed in this article do not necessarily
reflect those of Family Health International or the range from 0.0 to 1.0. For example, “yes” to one chance “no”). The
funding agency. the weather forecast today, a typi- probability of rain is 0.90, while the
Corresponding author: David A. Grimes, MD, Fam- cal day, might call for a probability odds of rain are 9.0, a tenfold dif-
ily Health International, P.O. Box 13950, Research
of rain of 10%, or 0.10. A medical ference. (Fortunately, the conver-
Triangle Park, North Carolina 27709; e-mail: sion between probability and odds
dgrimes@fhi.org. example would be that the cumu-
lative 10-year probability of preg- is simple [see the Box, “Converting
Financial Disclosure
The authors have no potential conflicts of interest to nancy after tubal sterilization (for Probability and Odds”]).
disclose.
all methods combined) in the U.S.
© 2008 by The American College of Obstetricians Collaborative Review of Steriliza- Relative Risk
and Gynecologists. Published by Lippincott Williams
& Wilkins. tion study was 19 per 1,000 women Both clinicians and patients readily
ISSN: 0029-7844/08 or 0.019.3 understand relative risk. It is simply a

VOL. 111, NO. 2, PART 1, FEBRUARY 2008 OBSTETRICS & GYNECOLOGY 423
 pingectomy. Relative risks will 10% chance of rain (820%) is non-
Converting Probability and sense,4 since the probability of rain
vary, depending on the referent
Odds
group used. cannot exceed 100%. The relative
Probability to Odds risk of rain in this situation is lower:
The formula for converting from Odds Ratio probability with nor’easter/proba-
probability to odds is Odds ratios are more difficult to bility on average day⫽0.90/
probability/(1–probability)⫽odds understand than are relative risks. 0.10⫽9.0. Thus, the likelihood of
As the name implies, an odds ratio rain with a storm approaching is
Downloaded from http://journals.lww.com/greenjournal by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0h

Example is the odds of the outcome in one nine times higher than on an aver-
With a probability of 0.30, the group divided by the odds of the age day (relative risk), not 82 times
CywCX1AWnYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC1y0abggQZXdgGj2MwlZLeI= on 09/19/2024

odds would be calculated as outcome in the other group (analo- higher (the odds ratio). Since the
gous to relative risk). As a ratio, it chance of rain is high, the odds
0.30/(1–0.30)⫽0.30/0.70⫽0.43 ranges from zero to infinity. The ratio greatly exaggerates the rela-
odds ratios formula is more cum- tive risk; hence, the two terms can-
Odds to Probability
bersome, so a few abbreviations not be used interchangeably.
In the reverse direction, the for- may help. Here, p1 refers to the This same exaggeration occurs
mula for converting odds to probability of the outcome in when the odds ratio is inappropri-
probability is group 1, and p2 is the probability of ately used as a proxy for the rela-
odds/(odds⫹1)⫽probability the outcome in group 2. tive risk in clinical research.5 For
example, a study of the familial
p 1 /(1⫺p 1 )
Example Odds ratio⫽ recurrence of dystocia reported
p 2 /(1⫺p 2 )
With an odds of 4, the probabil- (based on an odds ratio of 24) that
ity would be calculated as In the sterilization example, the “the risk is increased more than
4/(4⫹1)⫽4/5⫽0.80 relative risk of pregnancy with the 20-fold . . . .” Since the baseline risk
Hulka spring clip was 4.9 com- was 11%, a 20-fold increase in 11% is
pared with postpartum partial sal- impossible because the probability
ratio of probabilities. For example, in pingectomy. The odds ratio of of dystocia cannot exceed 100%.2
a cohort study, the relative risk pregnancy with the Hulka clip
would be the probability of the out- compared with postpartum partial EXAMPLES WITH CARDS
come in those exposed divided by salpingectomy would be 5.1, simi- We will next illustrate these four
the probability of the outcome in lar to the relative risk. Because the terms in sequence, using a deck of
those not exposed. Because this ex- frequency of pregnancy was low 52 playing cards (jokers removed).6
pression is also a ratio, its values with both methods (the “rare dis-
range from zero to infinity. ease assumption”), the odds ratio
Probability
The tubal sterilization study and relative risk are close, and the
above provides some examples. two can be used interchangeably. What is the probability of draw-
The probability of pregnancy after The odds ratio and relative risk are ing a diamond card from a ran-
sterilization with the Hulka spring similar when the outcome is un- domly shuffled deck? Since the
clip was 37 per 1,000 women, or common; the odds ratio exagger- deck contains 13 diamond cards
0.037. The comparable figure for ates the relative risk when the out- out of a total of 52, the probability
postpartum partial salpingectomy come is common. is 13/52⫽1⁄4⫽0.25. Thus, with many
was 7.5 per 1,000 women, or In the weather example above, repetitions (and with replacement
0.0075.3 Hence, the relative risk of the odds of rain with a nor’easter of the drawn card each time), the
pregnancy with the Hulka clip approaching are 9.0, whereas the probability of drawing a diamond
compared with postpartum salpin- odds on an typical day are 0.11. will be 0.25.
gectomy was (0.037 per 1,000)/ Thus, the odds ratio of rain with an
(0.0075 per 1,000)⫽4.9. The risk of approaching nor’easter compared Odds
pregnancy was nearly five times with a usual day is 9.0/0.11⫽82. A What are the odds of drawing a
higher with this clip than with the weather forecaster might excitedly diamond card from the same shuf-
postpartum operations. Relative proclaim on the evening news that fled deck? The odds would be the
risks always relate to some bench- rain is 82 times more likely with the probability of drawing a diamond
mark (sometimes called the “refer- impending storm compared with card divided by the probability of
ent”), here, postpartum partial sal- an average day! Not so: 82 times a drawing a card of another suit. Here,

424 Grimes and Schulz Odds and Odds Ratios OBSTETRICS & GYNECOLOGY
 this would be 13/(52–13)⫽13/39⫽1/ baseline probabilities of the out- relative risk and odds ratio if a new
3⫽0.33. In other words, one can come exceed 0.10 to 0.20. treatment boosted fertility to 0.99?
expect to draw one diamond for However, authors sometimes The relative risk of fertility with the
every three cards of another suit. use the odds ratio as a proxy for the new treatment compared with
The odds (0.33) exaggerate the prob- relative risk even when the disease baseline would be 0.99/0.85, or
ability (0.25) of drawing a diamond is not rare, which is inappropriate 1.2. Thus, the new treatment in-
because the event is not uncommon and can lead to misinterpretation. creased the fertility rate by only
(the probability of diamonds is 25%). A study of odds ratio use in two 20%. The highest possible increase
Downloaded from http://journals.lww.com/greenjournal by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0h

obstetrics and gynecology journals in fertility rate can lead to only a

Relative Risk revealed that, in 44% of 107 arti- small increase in the relative risk
CywCX1AWnYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC1y0abggQZXdgGj2MwlZLeI= on 09/19/2024

What is the relative risk of drawing cles, the odds ratio exaggerated the because the baseline rate (the refer-
a diamond card compared with relative risk by more than 20%.2 ent) is so high.12
drawing the ace of spades? The Other investigators have mistaken In epidemiology, relative risks
probability of a diamond is 0.25; the odds ratio for the relative risk, in the range of 1.2 to 1.5 (“weak
the probability of drawing the ace leading to the claim that a twofold associations”) are difficult to inter-
of spades is 1/52, or 0.019. Hence, difference in cesarean delivery fre- pret; bias can easily account for
the relative risk of drawing a dia- quency was a threefold difference.10 them.13 In contrast, the odds ratio
mond compared with drawing the of the fertility treatment in this ex-
WHY BOTHER WITH ODDS ample would be 99/5.7, or 17. The
ace of spades is 0.25/0.019⫽13.
RATIOS? relative risk of 1.2 is not much
Thus, a diamond card is 13 times
more likely than the ace of spades Given these inherent problems different from 1.0 (no effect), in
(the referent). with odds ratios, why use them at contrast to the odds ratio of 17.
all? Odds ratios have several key Here, the wide range of possible
roles to play. Although they are odds ratios is more informative
Odds Ratio always a valid measure of associa- than the narrow possible range of
What would be the odds ratio of tion, they are not always a good relative risks.
drawing a diamond compared with substitute for relative risk. First, Third, odds ratios are the output
drawing the ace of spades? As cal- odds ratios are the appropriate of logistic regression, a technique
culated above, the odds of drawing measure of relative effect in case- often used to control for confound-
a diamond are 0.33. The odds of control studies.11 In these studies, ing bias in research analysis.14 Ad-
drawing the ace of spades are researchers have a numerator (cases) justment of odds ratios in logistic
1/(52–1)⫽1/51⫽0.020. So the odds but no denominator, so rates and regression analysis is widely avail-
ratio would be 0.33/0.020⫽17. relative risks cannot be determined. able through software packages
Once again, the odds ratio of draw- Instead, investigators compare the and is easily understood. Properly
ing a diamond compared with the frequency of exposure among cases addressing confounding in re-
ace of spades (17) is higher than the with the frequency of exposure search usually trumps concerns re-
relative risk (13), so the two terms among the controls. In case-control garding odds ratios.
cannot be used interchangeably. studies, the odds ratio is a good
proxy for the true relative risk when SUMMING UP
PITFALLS OF ODDS RATIOS the “rare disease assumption” is met. Odds ratios are hard for many clini-
Odds ratios have important draw- Examples of rare conditions include cians to understand and susceptible
backs.7 First, clinicians think in ovarian cancer, systemic lupus ery- to misinterpretation.2 Under the
probabilities, not odds. Second, as thematosus, and uterine rupture; “rare disease assumption,” when the
noted above, odds ratios exagger- these have all been investigated with probability of an outcome is less than
ate the effect size compared with a case-control studies. 0.10 or 0.20, the odds ratio and rel-
relative risk, especially for com- Second, odds ratios are com- ative risk can be used interchange-
mon outcomes. If the odds ratio is monly used in meta-analysis, which ably. This is not true when the prob-
greater than 1.0, it is larger than the aggregates research studies to in- ability is higher.
relative risk. Conversely, if the crease the power to find differ- Methods are available to ap-
odds ratio is less than 1.0, it is ences. With frequent outcomes, rel- proximate the relative risk from an
smaller than the relative risk.8,9 ative risks are constrained.12 For adjusted odds ratio.9,15 This may
This discrepancy becomes clini- example, assume that baseline fer- provide more reader-friendly arti-
cally important only when the tility is 0.85. What would be the cles.2 The formula proposed by

VOL. 111, NO. 2, PART 1, FEBRUARY 2008 Grimes and Schulz Odds and Odds Ratios 425
 Zhang and Yu15 is useful: REFERENCES 9. Schechtman E. Odds ratio, relative
1. Bland JM, Altman DG. Statistics notes. risk, absolute risk reduction, and the
OR number needed to treat: which of
RR ⫽ The odds ratio. BMJ 2000;320:1468.
关共1 ⫺ P0兲⫹共P0⫻OR兲兴 these should we use? Value Health
2. Holcomb WL Jr, Chaiworapongsa T,
2002;5:431–6.
where P0 is the proportion of Luke DA, Burgdorf KD. An odd mea-
sure of risk: use and misuse of the odds 10. Xenakis EM, Langer O, Piper JM,
those unexposed who develop the ratio. Obstet Gynecol 2001;98:685–8. Conway D, Berkus MD. Low-dose ver-
outcome, OR is the odds ratio, and 3. Peterson HB, Xia Z, Hughes JM, sus high-dose oxytocin augmentation of
RR is the relative risk estimated from Wilcox LS, Tylor LR, Trussell J. The labor: a randomized trial. Am J Obstet
Downloaded from http://journals.lww.com/greenjournal by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0h

the odds ratio. As the background risk of pregnancy after tubal steriliza- Gynecol 1995;173:1874–8.
tion: findings from the U.S. Collabora- 11. Schulz KF, Grimes DA. Case-control
rate of the outcome gets low (ie, P0 tive Review of Sterilization. Am J studies: research in reverse. Lancet
CywCX1AWnYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC1y0abggQZXdgGj2MwlZLeI= on 09/19/2024

approaches zero in the “rare disease Obstet Gynecol 1996;174:1161–8. 2002;359:431–4.

assumption”), the denominator in 4. Dewdney AK. 200% of nothing: an eye- 12. Cook TD. Advanced statistics: up with
brackets approaches 1.0, and the rel- opening tour through the twists and turns
odds ratios! A case for odds ratios when
of math abuse and innumeracy. New
ative risk approaches the odds ratio. York (NY): John Wiley & Sons; 1993.
outcomes are common. Acad Emerg
The safest strategy, however, Med 2002;9:1430–4.
5. Schwartz LM, Woloshin S, Welch HG.
may be to limit the use of odds ratios Misunderstandings about the effects of 13. Grimes DA. Progestins, breast cancer,
and the limitations of epidemiology.
to case-control studies and logistic race and sex on physicians’ referrals for
Fertil Steril 1992;57:492–4.
regression (Deeks J. When can odds cardiac catheterization. N Engl J Med
1999;341:279–83. 14. Grimes DA, Schulz KF. Bias and causal
ratios mislead? [letter] BMJ 1998; associations in observational research.
6. Katz KA. The (relative) risks of using
317:1155). We suggest that, in the odds ratios. Arch Dermatol 2006;142: Lancet 2002;359:248–52.
interest of better understanding, re- 761–4. 15. Zhang J, Yu KF. What’s the relative
searchers should report results as 7. Sackett DL, Deeks JJ, Altman DG. risk? A method of correcting the odds
Down with odds ratios! Evidence Based ratio in cohort studies of common out-
probabilities and relative risks (with comes. JAMA 1998;280:1690–1.
Med 1996;1:164–6.
confidence intervals)16 whenever
8. Davies HT, Crombie IK, Tavakoli M. 16. Grimes DA. The case for confidence
possible.6 For most clinicians, odds When can odds ratios mislead? BMJ intervals. Obstet Gynecol 1992;80:
ratios will remain . . . well, odd.2 1998;316:989–91. 865–6.

426 Grimes and Schulz Odds and Odds Ratios OBSTETRICS & GYNECOLOGY