Archives of Iranian Medicine
, Volume 16, Number 5, May 2013
295
).DPDQJDU),VODPL
Introduction
“
H
ow many is enough?” is a question that epidemiologists
and clinicians ask themselves when they plan on conduct-
ing a new study. Researchers want to enroll a large
enough number of people such that statistical errors (type I and
type II) are minimized yet the cost, labor, and time to do the study
remain acceptable. Sample size calculations often remind us of
complex formulas. While we provide some formulas in the text,
the main aim of the article is not to offer a long list of such formu-
las. Rather, the main aim is to discuss the statistical principles
behind sample size calculation, issues that may make such calcu-
lations not-so-straightforward, and nonstatistical considerations in
sample size determination. Therefore, in this article, we discuss
the following topics:
1. The need to calculate sample size;
2. Principles;
3. Some formulas;
4. Factors that need to be determined for sample size calcula-
tions;
5. Assumptions made for sample size calculations;
6. Nonstatistical considerations;
7. Methods used to calculate sample size; and
8. Software used to calculate sample size.
7KH¿QDOSDUWRIWKHSDSHU6XPPDU\DQG&RQFOXVLRQVWLHVWKHVH
sections together.
1- The need to calculate sample size
When we would like to learn about an attribute of a population,
such as mean cholesterol of the people of China, it may not be
feasible for us to study the entire population due to cost or time
issues. Besides cost and time issues, it may not be ethical to study
the entire population if accurate enough results could be obtained
by studying a subgroup of all people. For these reasons, we need
to study a sample of the population.
However, results vary from sample to sample, and they may be
somewhat different from the true mean of the population. For
example, one sample of 100 Chinese people may have a mean
cholesterol of 182 mg/dL and another sample may have a mean
of 186 mg/dL. Nevertheless, as the sample size grows, the prob-
ability of obtaining a result that is close to the true mean for the
population increases. The question is how large the sample size
should be to make it very likely for the sample results to be within
a narrow distance from the true mean. The answer is discussed in
the next few sections. First, we start with some general principles,
and then we go into more details.
2- Principles
Although sample size depends on many factors, there are certain
principles that apply to nearly all sample size calculations, which
we discuss in this section. Sample size nearly always depends on
the factors discussed below.
2-1- Variation
The more variation there is in the variable of interest, the larger
is the required sample size. If there is no variation, even a sample
size of n = 1 is adequate.
Example 1: We want to determine the mean salary of all interns
in a hospital. If all interns receive exactly the same salary (e.g.,
$40,000 annually), knowing the salary of only one intern is ad-
HTXDWHWRNQRZWKHPHDQƔ
Example 2: There is a disease that is universally fatal, which
is equivalent to saying that there is no variation in its outcomes in
terms of death and life. Now if a new drug cures only one case of
this disease, assuming that the diagnosis is correct, that one single
FDVHLVHQRXJKWRDFFHSWWKDWWKHGUXJZRUNVƔ
2-2- Magnitude of error that we accept
The less the magnitude of the error that we accept, the larger is
the needed sample size. This is somewhat intuitive: larger sample
size is the price that we pay for less error.
Example 3: A researcher wants to determine the mean height
of a population. Any sample would most likely estimate the mean
Abstract
7KLVSDSHUGLVFXVVHVWKHVWDWLVWLFDOSULQFLSOHVPHWKRGVDQGVRIWZDUHSURJUDPVXVHGWRFDOFXODWHVDPSOHVL]H,QDGGLWLRQLWUHYLHZVWKH
SUDFWLFDOFKDOOHQJHVIDFHGLQFDOFXODWLQJVDPSOHVL]H:HVKRZWKDWEHFDXVHRIVXFKFKDOOHQJHVVWDWLVWLFDOFDOFXODWLRQVRIWHQGRQRWSURYLGH
XVZLWKDFOHDUFXWQXPEHUIRUWKHVWXG\VDPSOHVL]HUDWKHUWKH\VXJJHVWDUDQJHRIUHDVRQDEOHQXPEHUV7KHSDSHUDOVRGLVFXVVHVVHYHUDO
LPSRUWDQWQRQVWDWLVWLFDOFRQVLGHUDWLRQVLQGHWHUPLQDWLRQRIVDPSOHVL]HVXFKDVQRYHOW\RIWKHVWXG\DQGDYDLODELOLW\RIUHVRXUFHV
Keywords:3RZHUVDPSOHVL]HW\SH,HUURUW\SH,,HUURU
Cite this article as: Kamangar F , Islami F. Sample size calculation for epidemiologic studies: Principles and methods. Arch Iran Med. 2013; 16(5): 295 – 300.
Review Article
Sample Size Calculation for Epidemiologic Studies: Principles
and Methods
)DULQ.DPDQJDU0'3K'
1,2
, Farhad Islami MD PhD
3,2
$XWKRUV¶DI¿OLDWLRQV
1
School of Community Health and Policy, Morgan State
University, Baltimore, USA,
2
Digestive Disease Research Center, Tehran Univer-
sity of Medical Sciences, Tehran, Iran,
3
Institute for Translational Epidemiology,
Mount Sinai School of Medicine, New York, NY, USA.
&RUUHVSRQGLQJDXWKRUDQGUHSULQWVFarin Kamangar MD PhD, Department
of Public Health Analysis, School of Community Health and Policy, Morgan
State University, Portage Avenue Campus, Baltimore, MD 21251.
E-mail: [email protected].
Accepted for publication: 17 April 2013
Archives of Iranian Medicine
, Volume 16, Number 5, May 2013
296
6DPSOH6L]H&DOFXODWLRQIRU(SLGHPLRORJLF6WXGLHV
height with some error. For example, if the real mean height of
the entire population is 182.3 cm, a sample may estimate it as
182.5 mm, which has a 2 mm error. If we want to be relatively
certain that our sample mean has no more than 1 mm of error, the
required sample size is much larger than when we accept an error
RIPPFP)RUIXUWKHUGHWDLOVSOHDVHVHH([DPSOHƔ
2-3- Probability of making a certain magnitude of error
The smaller the probability of the error, the larger the sample
size should be. Consider Example 3. We are never sure that the
error is necessarily going to be less than 1 mm. In a large popula-
tion, there are many tall people. It could turn out that a random
sample, however large, have a mean height of 3 mm higher than
the entire population. We can only increase the sample size to the
extent that with a large probability, e.g., 95% or 99%, the sample
mean be within 1 mm from the true population mean. If we want
99% certainty, we would need a larger sample size than we accept
95% certainty. This is again somewhat intuitive, as larger sam-
ple size is the price we pay for more certainty. For more details,
please see Example 6.
3- Some formulas
Each study is unique and needs its own formula. However, to
provide some examples and to illustrate the principles mentioned
above, we will provide formulas for three cases: 1) estimating the
mean height in a population; 2) comparing the effects of two treat-
ments on mean blood pressure; and 3) comparing the effects of
two treatments on mortality.
3-1- Estimating mean height
We want to determine the mean height of men aged 18 or above
in a very large population. Here, the required sample size de-
pends on three factors: 1) the variance of height in men aged 18
RUDERYHı
2
); 2) the maximum magnitude of error that we accept
(d); and 3) the probability that our error will be higher than the
DFFHSWDEOHPDJQLWXGHĮ7KHIROORZLQJIRUPXODFRXOGEHXVHG
to calculate sample size for this study:
n=
d
2
(Z
1-
Į
/
2
)
2
(ı
2
)
Now, we discuss how each of these elements is related to the
principles discussed in Section 2.
9DULDQFHı
2
)
According to the formula, the more variation in height, the larger
our sample size should be. This is in line with Principle 1 in Sec-
tion 2 (Principle 2-1).
)LQGLQJWKHFRUUHFWıWRSXWLQWKHIRUPXODLVVRPHZKDWFKDOOHQJ-
ing. Determining the variance of height in the entire population
depends on knowing its mean. Since we don’t have the mean
(otherwise we wouldn’t do the study!), we cannot know the vari-
ance, so we can only estimate it, which is subject to some error.
Example 4: What number do we use for variance to estimate
the mean height of the population? If we assume that height is
normally distributed, 95% of the values of height in the popula-
tion will fall in the range of mean ± two standard deviations, in
other words in a range of four standard deviations. Therefore, if
the height of 95% of the people falls roughly between 160 cm and
200 cm, it is reasonable to assume that the population standard de-
YLDWLRQıLVDERXW· FP$OWKRXJKWKLVLVDUHDVRQDEOH
assumption, one can assume that the real standard deviation may
be 12 cm, which increases the required sample size. Researchers
PD\DOVRXVHSUHYLRXVOLWHUDWXUHLIDYDLODEOHWRHVWLPDWHıƔ
3-1-2- The acceptable maximum error (d)
According to the formula, d is in the denominator. Therefore, the
less error we accept, the larger the sample size should be. This is
consistent with Principle 2-2.
Example 5: The researchers may decide that they would like
WKH ¿QDO HVWLPDWH WR EH ZLWKLQ  FP RI WKH WUXH QXPEHU 7KLV
means that if our results show a mean height of 178 cm, we hope
that the true number is between 177 cm and 179 cm. To show the
effect of dZH¿[RWKHUIDFWRUVIRUH[DPSOH=DQGVWDQGDUG
GHYLDWLRQı FP8QGHUWKHVHFLUFXPVWDQFHVLIZHDJUHHWR
a maximum error of 1 cm, our sample size needs to be 400 (4 ×
·+RZHYHULIZHDFFHSWDPD[LPXPHUURURIFP
PPWKHQWKHVDPSOHVL]HQHHGVWREHî·Ɣ
Example 5, in addition to illustrating the effect of acceptable er-
ror, shows that determination of sample size is not entirely clear-
cut. One can substantially increase or decrease sample size by
changing one of these factors, particularly by changing the ac-
ceptable error.
3UREDELOLW\RIHUURUIDOOLQJRXWVLGHGĮ
As mentioned earlier, there is always a probability that our error
is larger than d7KLVSUREDELOLW\LVVKRZQDVĮDQGDIXQFWLRQRI
it, (Z
1-
D
/
2
) , or simply Z, appears in the formula. The smaller is
WKHĮWKHODUJHULVWKH=7KHUHIRUHDVPDOOHUSUREDELOLW\RIHUURU
needs a larger Z, and consequently a larger sample size. This is
consistent with Principle 2-3.
Example 6:0DQ\HSLGHPLRORJLFVWXGLHVFKRRVHWKHLUĮWREH
,QWKLVFDVH=ZKLFKLVDIXQFWLRQRIĮZLOOEH,I
RQHZDQWVWRUHGXFHĮWRWKHQ=ZLOOEH5HGXF-
ing the probability of error requires a larger Z and hence a larger
VDPSOHVL]HƔ
6LQFHFKRRVLQJ Į LV DWWKH GLVFUHWLRQRI WKHUHVHDUFKHU WKHUH-
quired sample size is not entirely clear-cut. This is a lesson that
we learnt from choice of d too.
If you feel that you have had enough of formulas, you can skip
the rest of this section and go to Section 4. However, if you are
interested in reading two more examples, go through Sections 3-2
and 3-3.
3-2- Comparing mean blood pressures
Suppose a study has one clearly-stated main objective: “To com-
pare mean systolic blood pressures between patients receiving six
months of treatment X versus those receiving six months of treat-
ment Y”. In this example, the formula is slightly more complex
than the formula shown in Section 3-1, but many of the elements
are common among the two. The sample size depends on four
IDFWRUVWKHYDULDQFHRIEORRGSUHVVXUHLQHDFKJURXSı
2
); 2)
the minimum difference that we would like to detect between the
WZRWUHDWPHQWVGWKHSUREDELOLW\RIW\SH,VWDWLVWLFDOHUURUĮ
DQGWKHSUREDELOLW\RIW\SH,,VWDWLVWLFDOHUURUȕ7KHIROORZLQJ
formula could be used to calculate sample size for each treatment
group:
n=
d
2
(Z
1-
Į
/
2
+
Z
1-E
)
2
(ı
2
1
ı
2
2
)
Archives of Iranian Medicine
, Volume 16, Number 5, May 2013
297
).DPDQJDU),VODPL
Now, we discuss the application and the intuitive meaning of
each of the factors in this formula.
9DULDQFHı
2
)
As discussed in Principle 2-1, sample size should be larger when
variance of the blood pressure reduction is larger. Conversely, if
there is no variation, for example if treatments X and Y decrease
blood pressure in each participant by exactly 25 mmHg and 20
mmHg, we could accurately estimate the difference with only one
person in each treatment group.
3-2-2- The minimum difference that we would like to
detect (d)
In general, detecting a large difference requires a small sample
size but detecting a small difference requires a large sample size.
This is in line with Principle 2-2.
Finding the right d to put into the formula is challenging. Before
the study, we don’t know what the difference between the two
treatments is. Therefore, we ask ourselves: “What is the small-
HVWGLIIHUHQFHWKDWPDWWHUVWRXV"´,IZHGHVLUHWR¿QGRQO\ODUJH
differences between the two treatments, e.g., 10 mmHg, then the
sample size wouldn’t need to be that large. However, if we want
to observe even very small differences, e.g., 1 mmHg, then sam-
ple size should be much larger, 100 times larger than that needed
for the former situation.
It is intuitively understandable that detecting smaller differences
requires larger sample sizes. Suppose you want to compare two
students for their English spelling. If the difference between the
two students is large, i.e., one student’s spelling is far better than
the other one, you could perhaps see the difference by asking only
10 questions. On the other hand, if both students are very strong
and the difference is minimal, you may need to test them with 200
questions before you learn who is better.
3UREDELOLW\RIW\SH,HUURUĮ
Type I error occurs when we erroneously reject the null hypoth-
esis. To make it simpler and more relevant to our own example,
type I error occurs if in truth (i.e., if one studies the entirety of our
target population) the two treatments affect the mean blood pres-
VXUHH[DFWO\WKHVDPHEXWLQRXUVWXG\VDPSOHZH¿QGWKDWWKH\DUH
GLIIHUHQW7KLVLVREYLRXVO\DQHUURUEHFDXVHZH¿QGDGLIIHUHQFH
where in reality there is no difference. Such errors may happen
due to sampling variation. Thinking about tossing a coin (rather
than blood pressure) may make it easier to understand.
Example 7: Let’s say we want to know whether two coins are
different with regards to their shape, such that the percentage of
the heads for each of the coins is different. To learn this, we toss
the two coins several times, compare the percentage of the heads,
and perform statistical tests. However, in a single study, two com-
SOHWHO\ VLPLODU FRLQV PD\ KDYH VWDWLVWLFDOO\ VLJQL¿FDQW GLIIHUHQW
results, which is a type I error.
To understand what was said above, let’s change the experiment.
7RVVWKH¿UVWFRLQWLPHV%\FKDQFH\RXPD\JHWWZRKHDGVRXW
of 10 (20%). Toss the same coin another 10 times, and you may
get nine heads out of 10 (90%). The difference between these
WZR QXPEHUV  DQG  LV VWDWLVWLFDOO\ VLJQL¿FDQW ZLWK D
two-sided Fisher exact P-value of 0.003. Since you used the same
coin, obviously the difference in the percentage of heads in the
two series of tosses was not due the design or shape of the coin; it
was merely due to random variation (chance). In statistical terms,
this was type I error, because while there was no difference, you
GHWHFWHGRQHƔ
,QVWXG\ GHVLJQ ZH XVXDOO\ ¿[WKH SUREDELOLW\RIW\SH, HUURU
For example, if we want a two-sided type I error probability of
0.05 (5%), its corresponding Z will be 1.96. For a type I error of
0.01, Z will be 2.58. If we want a smaller type I error, then our
Z, and consequently our sample size would be larger. In other
words, if we ask for a smaller probability of error, we need a larger
sample size, which is in line with Principle 2-3.
3UREDELOLW\RIW\SH,,HUURUȕ
Type II error occurs when in truth the two treatments are differ-
HQWEXWZHGRQRW¿QGWKHGLIIHUHQFHLQRXUVWXG\VDPSOH7KLV
happens quite commonly if the sample size is not large enough.
We obviously want to reduce the probability of such errors, i.e.,
we want to detect differences if they exist. Reducing type II error
is also called increasing the power of the study. If we want larger
power, or smaller type II error, our Z will increase, which requires
larger sample size. Intuitively, the smaller the probability of error,
the larger our sample size should be (Principle 2-3).
3-3- Comparing mortality
Let’s discuss the third case. We want to compare the effects of
treatments X and Y on reducing mortality. To do this, we ran-
domize subjects into two treatment groups, receiving X or Y. The
required sample size for each group could be obtained from the
formula below:
n=
d
2
(Z
1-
Į
/
2
+
Z
1-E
)
2
[(S
uS
S
uS
@
The elements used in this formula are exactly the ones used in
WKHSUHYLRXVIRUPXODH[FHSWWKDWYDULDQFHLVUHSODFHGE\ʌuʌ,
ZKHUHʌLVWKHSURSRUWLRQRISHRSOHZKRGLHLQWKHHQWLUHSRSXOD-
WLRQZLWKLQ¿QLWHVDPSOHVL]H,WFDQEHVKRZQWKDWWKLVODWWHUHOH-
ment plays the role of variance when the outcome is dichotomous
(died or did not die).
4- Factors that need to be determined and their impact
on sample size
A number of formulas were introduced in the previous section
to calculate the study sample size. The question is: “What is the
right formula for our study and what are the right numbers to put
into it?” A number of decisions should be made before we choose
the right formula. And after the formula is selected, we need to
decide what numbers to put into the formula. Such decisions,
which usually have enormous impacts on our calculated sample
size, are discussed here.
4-1- Deciding on the main objective of the study
The main objective of the study is the primary factor in selecting
the formula. For example, in Section 3, we chose three different
formulas for studies with three different objectives. The objec-
WLYHVKRXOGEHVSHFL¿HGYHU\FOHDUO\(YHQVOLJKWFKDQJHVWRWKH
objectives may have a substantial impact on sample size.
Example 8: The objective of a study is: “To compare the ef-
fects of treatments X and Y on serum cholesterol in a randomized
parallel design trial.” As simple as it sounds, the objective still
QHHGVWREHPRUHFOHDUO\GH¿QHGDVLWZLOOKDYHDPDMRULPSDFWRQ
sample size. For example, the required sample size would be dif-
Archives of Iranian Medicine
, Volume 16, Number 5, May 2013
298
6DPSOH6L]H&DOFXODWLRQIRU(SLGHPLRORJLF6WXGLHV
ferent if we chose to compare the effects of the treatments in a su-
periority trial – i.e., a trial that determines which drug is superior
– versus if compared them in a noninferiority trial – i.e., our plan
is to show that the new drug is not inferior to the standard treat-
ment by a certain amount. Also, it would make a big difference
if we decide to compare the mean cholesterol reduction versus
we decide to compare the proportion of people whose cholesterol
UHDFKWKHWDUJHWRIPJG/Ɣ
Unfortunately, determination of the main study objective is not
always straightforward, particularly in observational studies, such
as case-control and cohort studies. Consider the example of a
cohort study. During follow-up in a cohort study, there will be
a large number of possible outcomes, including overall mortal-
ity, ischemic heart disease mortality, and esophageal cancer inci-
dence. If the main outcome is a common event, such as overall
mortality or mortality from ischemic heart disease, the sample
size doesn’t have to be very large, whereas if the main outcome is
esophageal cancer incidence, a relatively uncommon cancer, then
sample size has to be quite large. Likewise, in a typical cohort
study, we collect information on a number of exposures, with each
exposure having its own distribution. Although protocols often
require that we determine sample size, it may not always be easy
to determine in advance what the main outcome and the main ex-
posure is. The usefulness of a cohort goes way beyond one out-
come and one exposure.
4-2- Selecting the design of the study
Design of the study may have a major impact on the sample
size. For example, it makes a large difference when we com-
pare the effects of treatments X and Y on serum cholesterol in
a parallel design trial versus a cross-over trial. Cross-over trials
often need much smaller sample sizes, as each person receives
both treatments. Also, each person serves as his or her own con-
trol, eliminating interpersonal variance, which again results in a
smaller sample size.
4-3- Deciding on the proportion of participants distributed into each
study arm
In the formulas discussed in Sections 3-2 and 3-3, we assumed
equal sample size in each of the two treatment groups. However,
we may decide otherwise. For example, in comparing treatments
X (an old treatment) and Y (the new treatment), we may decide to
randomize more people into receiving Y. This is because X is a
well-known and long-used treatment, but Y is a new one and we
want more information on it, particularly about its side effects.
Given equal variance in the two study arms, uneven distribution
of participants in the study arms requires higher sample size to
obtain the same power.
Example 9: We want our study participants to be distributed
with a ratio of 1 to 3 into X and Y treatment groups, respectively.
If so, a total sample size of 1333 (333 in X and 1000 in Y) will
give us the same power as randomizing 1000 people equally to
each study group (500 X and 500 Y). Here, we need a 33% in-
FUHDVHLQVDPSOHVL]HƔ
4-4- Deciding between Bayesian versus frequentist methods
All of the formulas and much of the discussion made in this pa-
per are based on the frequentist view of probability. This is be-
cause, at least thus far, much of the currently practiced statistics
is based on frequentist methods. For example, P-value, power,
type I error, and much of all other familiar statistics is rooted in
frequentist view.
However, Bayesian methods are gaining popularity. If Bayesian
analysis is considered, sample size calculations will be totally dif-
ferent. Sample size and power calculations for studies designed to
be analyzed using Bayesian methods heavily depend on the prior
distributions. Without going into any details, prior distributions
may come from various sources, including our beliefs. For exam-
SOHLIDSROLWLFDOOHDGHUEHOLHYHVWKDWKLVRSLQLRQLVGH¿QLWHO\FRU-
rect, no matter how much data you show him, he will stand by his
prior opinion. If so, even a huge sample size showing the contrary
would do no good! This one, of course, was an extreme example!
Using prior distributions could be very helpful in some cases.
4-5- Deciding the numbers to put in the formulas
Consider the study presented in Section 3-2. Assume the litera-
ture suggests that the variance of blood pressure in each group is
20 mmHg. If we choose a power of 0.90 and a type I error level
of 0.01, and we want to detect a d = 2 mmHg, the required sam-
ple size would be nearly 6000. However, if we choose a power
of 0.80 and a type I error level of 0.05, and we want to detect a
d = 3 mmHg, the required sample size would be approximately
1400. Therefore, with some minimal changes in requirements, all
perfectly reasonable and within the ranges used by clinicians and
VWDWLVWLFLDQVZHFDQ¿QGWKHUHTXLUHGVDPSOHVL]HWREHDVORZDV
1400 or as high as 6000.
5- The impact of assumptions
Several assumptions have been made for doing the calculations
made in Section 3. Departures from these assumptions may make
sample size calculations incorrect. Below, we provide a few ex-
amples of the assumptions and show their impact on sample size
calculations. To make it simple, in all examples we have assumed
that the calculated sample size under the assumption is 1000.
5-1- Independence of study samples
The formulas and methods discussed so far assume that indi-
viduals in the sample are independent. However, if they are not,
then the sample size must be larger to accommodate for lack of
independence (sometimes referred to as clustering).
Example10: Consider the extreme example that identical twins
always respond identically to a drug; i.e., the correlation between
response from identical twins is 1.00. If so, when a researcher re-
cruits 500 pairs of identical twins, although the sample size is 1000,
it only provides us with information equivalent to 500 people; once
we know the response from one twin, having the second one adds no
further information. Here we say the effective sample sizeLVƔ
Example 11: Assume that to reduce costs of enrolling study
participants, rather than selecting 1000 people randomly from an
entire population, we randomly select 20 villages from the popu-
lation and then randomly select 50 individuals from each village
(two-stage cluster sampling). Since the responses obtained from
each village can be correlated, the effective sample size may be
less than 1000. If so, the effective sample will fall somewhere be-
tween the number of independent units (here, number of villages
= 20) and the total number of study participants (here, 1000). In
other words, our sample selection is not quite as good as recruiting
1000 independent people, but it is not as poor as selecting only
SHRSOH:LWKRXWJRLQJLQWRGHWDLOVRIWKHIRUPXODZHVXI¿FH
to say that the effective sample size depends on the total number
of people, the number of units, and the intracluster correlation,
Archives of Iranian Medicine
, Volume 16, Number 5, May 2013
299
).DPDQJDU),VODPL
i.e., the correlation between responses from individuals in each
village. In this example, if the intracluster correlation is 0.10, the
effective sample size approximately 170, which is indeed between
DQGƔ
5-2- No attrition
The formulas shown in Sections 3 assumed no sample attri-
tion. If we assume an attrition of 20%, then the initial sample size
VKRXOGEHODUJHU· IRUH[DPSOHLQVWHDG
of 1000. However, it may be impossible to determine the extent of
sample attrition prior to conducting the study.
,Q¿QLWHO\ODUJHWDUJHWSRSXODWLRQ
The formulas in Section 3 assumed that the target population
ZDVLQ¿QLWH,IWKHWDUJHWSRSXODWLRQLV¿QLWHWKHUHTXLUHGVDPSOH
size may be slightly lower.
Example 12: If the target population is only 20,000 people, to deter-
PLQHDPHDQZHPLJKWQHHGDVDPSOHVL]HRILQVWHDGRIƔ
As illustrated by these numbers (975 versus 1000), as long as the
sample is relatively small compared to the target population (e.g.,
less than 5% of the entire population), the difference in sample
VL]HIRU¿QLWHDQGLQ¿QLWH SRSXODWLRQV LV TXLWH VPDOO7KHUHIRUH
size of the target population is usually not considered in sample
size calculations.
5-4- No adjustment for baseline characteristics
The formulas in Section 3 did not consider adjusting for base-
line characteristics. Multiple regression methods that adjust for
baseline characteristics usually result in reduced variance, thus we
obtain more power than we actually planned.
Example 13: Assume that the outcome of a study is depression
after six months of treatment with X or Y. If we measure depres-
sion at study baseline, and baseline depression is highly correlated
ZLWKWKH¿QDORQHWKHQDGMXVWLQJIRUEDVHOLQHGHSUHVVLRQVKRXOG
LQSULQFLSOHUHGXFHWKHYDULDQFHRI¿QDOGHSUHVVLRQDQGWKXVPDNH
the study more powerful. If the correlation between baseline and
¿QDO GHSUHVVLRQ VFRUH LV  WDNLQJ WKLV LQIRUPDWLRQ LQWR DF-
count, then a sample size of 1000 will actually give us a power
HTXLYDOHQWWRKDYLQJLQWKHVWXG\Ɣ
Note that a correlation of 0.50 is very high. With a correlation of
0.10, information from 1000 people provides with a power equal
to having 1010 people. Most correlations are around this size
(0.10 or so). Therefore, correlations are often ignored in sample
size calculations.
6- Nonstatistical considerations in determining sample
size
In addition to statistical calculations, there are other issues that
may matter is choosing our study sample size. Funding, time,
number of available patients, ethical issues, similar research being
done elsewhere, and novelty of the research topic may play a role
in determination of sample size.
6-1- Funding
As discussed in the previous sections, we can determine a rea-
sonable range of sample sizes (e.g., from 1400 to 6000) for a
study. If a researcher has funding to study only 20 subjects, then
he perhaps shouldn’t pursue that study. On the other hand, if he
has large resources and large number of participants available,
then he can determine a sample size between 1400 and 6000 for
his study, depending on how much error he is willing to accept.
6-2- Ethical issues
Conducting a study with 100,000 people, where at most 6000 is
needed, may be considered unethical, particularly if the study is a
randomized trial testing a new drug.
6-3- Fixed number of patients available to the researcher
6RPHWLPHVVDPSOHVL]HLVDOPRVW¿[HG)RUH[DPSOHDPHGL-
cal researcher may have been able to collect data from 200 cases
of a rare disease over his 20 years of experience (roughly 10 per
year). If the researcher plans to increase sample size to 500, he
may need to wait another 30 years (perhaps not feasible), or col-
laborate with other centers in the world, which again may or may
QRWEHIHDVLEOH7KHUHIRUHVDPSOHVL]HLVHVVHQWLDOO\¿[HGDW
In circumstances like this, sample size formulas can be used, but
not to determine sample size, rather to learn about the power to
detect a certain difference. For example, with 200 cases and 800
FRQWUROV¿[LQJW\SH,HUURUDWDVVXPLQJDSUREDELOLW\RIH[-
posure of 0.20 in controls based on previous research, we will
have 84% power to detect a difference (reject the null hypothesis)
if the true probability of exposure among cases is 0.30. Although
WKHVDPSOHVL]HLV¿[HGZHFDQHVWLPDWHSRZHUWRGHWHFWDFHUWDLQ
difference.
In some ways, determination of sample size is like buying a
KRPHSDUWLFXODUO\ZKHQVDPSOHVL]HLV¿[HG:KHQ\RXGHFLGH
to buy a home and you can afford only $300,000, you may be able
to buy a home with two bedrooms and a large living room, or a
home with three bedrooms and a small living room. Likewise,
LIIRU¿QDQFLDORUWLPHFRQVWUDLQWV\RXFDQFROOHFWGDWDIURPRQO\
300 patients, that is what you can afford; with that you can get a
VPDOOĮDQGDODUJHȕRUDODUJHĮDQGDVPDOOȕRUDVPDOOĮDQG
VPDOOȕEXWDODUJHd<RXQHHGWRPDNHVDFUL¿FHVVRPHZKHUH
6-4- Novelty of the study
Novelty of the topic is important in making a decision to do a study
RU WR SXEOLVK D SDSHU 7KH ¿UVW UHSRUW RQ ZKDW LVQRZ NQRZQ DV
DFTXLUHGLPPXQRGH¿FLHQF\V\QGURPH$,'6SXEOLVKHGLQ
GHVFULEHGRQO\¿YHFDVHVRIWKLVGLVHDVHDOOLQ\RXQJKRPRVH[XDOV
However, given that the results were novel and the disease was rare,
it was worth being published.
1
Today, a report of a far larger number
of such cases may not be interesting enough for publication.
6-5- Similar studies being underway
Similar studies being conducted in other places could encourage
or discourage conducting studies with relatively small sample sizes.
On the one hand, availability of results from many similar studies
may take away from novelty of the study. On the other hand, if mul-
tiple low-powered studies are conducted, then one could potentially
do a meta-analysis or a combined analysis to increase power. There-
IRUHDOWKRXJKHDFKVWXG\E\LWVHOIPD\QRWEHGH¿QLWLYHFRPELQHG
together, they would greatly contribute to our knowledge.
7- Methods used to calculate sample size
Sample size can be calculated using formulas or simulation
methods. In the example below, we will calculate the sample size
using a formula.
Example 14: We would like to compare, in a randomized paral-
lel design trial, the effect of treatments X and Y in reducing serum
cholesterol in a group of hypercholesterolemic patients. What is
Archives of Iranian Medicine
, Volume 16, Number 5, May 2013
300
6DPSOH6L]H&DOFXODWLRQIRU(SLGHPLRORJLF6WXGLHV
the required sample size?
7RDQVZHUWKLVTXHVWLRQZH¿UVWQHHGWRGHWHUPLQHZKDWSDUDP-
eter we are going to compare: the percentage of patients whose
cholesterol is reduced to target levels after treatment, or the mean
cholesterol after treatment? Let’s assume we are going to compare
means. If we want equal number of patients randomized to each
group, the formula for calculating sample size in each group is:
n=
d
2
(Z
1-
Į
/
2
+
Z
1-E
)
2
(ı
2
1
ı
2
2
)
Now, we need to determine each of the components. Let’s as-
sume that we accept a type I error of 0.05, for which Z is 1.96;
and a type II error of 0.20 (power of 0.80), for which the cor-
responding Z is 0.84. We need to provide an estimated variance
of cholesterol after treatment. After some literature review, we
determine that a standard deviation of 30 mg/dL is a reasonable
estimate for each of the treatments. Most importantly, we need to
determine the minimum mean cholesterol difference between the
two groups that is clinically useful and meaningful to us. Let’s
say a difference of 3 mg/dL is the minimum that we would like to
be able to detect; below that, if we don’t detect the difference, it
doesn’t matter, as it is a clinical tie. Plugging these numbers into
WKHIRUPXODZH¿QGWKDWWKHVDPSOHVL]HZRXOGQHHGWREH
IRUHDFKVWXG\DUPRUDWRWDORIƔ
Since manual calculations may be tedious, software programs
have been developed to calculate sample size. For example, us-
ing STATA’s sampsi command, we obtain a sample size of 1570
for each group, or a total of 3140 cases. The minimal difference
between manual and software calculations is due to rounding.
Prior to the wide availability of computers, tables and nomo-
grams were developed and used to calculate sample size. Again,
the idea was to reduce the pain of using formulas. Nomograms
can be found in books or on the Internet.
2
Although they are rela-
tively easy to use, nomograms may not be available for all study
designs, objectives, or for all levels of type I and type II errors.
Therefore, they are not as versatile as computers in calculating
sample size, and their use provides little advantage over other
methods. Tables have similar problems.
Simulation is another approach used to calculate sample size or
power. This method is highly versatile – more so than using for-
mulas – and can be used to calculate sample size under nearly all
circumstances. It is most useful when there are no commands
in our statistical package to calculate sample size of our study,
mostly when the design is complex. However, simulation usu-
ally requires programming and therefore needs to be done by a
statistician. As this method requires computer power, it has be-
come more commonly used with the increased availability of
faster computers. The idea is that we generate populations with
the given parameters over and over (for example normal popu-
lations with means of 200 and 197 for treatments X and Y and
standard deviations of 30 for each one), do the appropriate test
(e.g., t-tests), and determine the proportion of the tests that found
DVWDWLVWLFDOO\VLJQL¿FDQWGLIIHUHQFHKHUH3YDOXH7KLV
latter proportion gives us the power. We can change the sample
size to see which sample size gives us adequate power.
8- Software used to calculate sample size
Sample size can be calculated using almost all commercial sta-
tistical software, such as STATA and SAS. For example, STATA’s
sampsi command and SAS’s PROC POWER can do the work for
a variety of designs. There is also freely available and relatively
easy-to-use software designed for sample size calculation. One
example is the PS Power and Sample Size Calculation program,
written by Dupont and Plummer at Vanderbilt University.
3
The
program provides a step-by-step guide to calculate sample size.
Another example is the Power program, written by Lubin and
Garcia-Closas at the U.S. National Cancer Institute.
4,5
This pro-
gram is particularly useful to calculate sample size when the out-
come of interest is interaction. Yet another example is Epi Info,
a free software for statistical analysis and power calculation, de-
veloped by the U.S. Centers for Disease Control and Prevention.
Conclusions
Statistical calculations of sample size depend on a number of
factors including, but not limited to, the type of the study, the
parameter that is going to be estimated (e.g., a mean or a pro-
portion), the variance of the variable of interest, the acceptable
type I and type II errors, clustering of the samples, and correlation
among variables. Such calculations are to some extent subjective,
because it is usually not obvious which numbers we should put
in the formulas. The truth is that the number that comes out of
the formula is only one acceptable number within an acceptable
range. In addition, sample size may also depend on a number of
nonstatistical factors, such as novelty of the study. As Norman
and colleagues have put it,
6
“Sample size estimates are like the
HPSHURU¶VFORWKHVZHFROOHFWLYHO\DFWLQSXEOLFDVLIWKH\SRVVHVV
an impressive aura of precision, yet privately we (statisticians)
are acutely aware of their shortcomings and extreme impreci-
sion.” Having discussed all of these limitations, it is still prudent
to calculate sample size statistically, as the results provide us with
a range of reasonable sample sizes, as well as information on the
power to detect a certain difference.
Acknowledgments
The authors would like to thank Dr. Ashkan Emadi (School of
Medicine, University of Maryland, Baltimore, MD), Dr. Mahsa
Mohebtash (Union Memorial Hospital, Baltimore, MD), and Ms.
Gillian Silver (School of Community Health and Policy, Morgan
State University) for reading the paper thoroughly and providing
constructive comments.
References
1. Centers for Disease Control (CDC) and Prevention. Pneumocystis pneumo-
nia--Los Angeles. MMWR Morb Mortal Wkly Rep. 1981; 30: 250 – 252.
2. Altman DG. Statistics and ethics in medical research: III How large a
sample? Br Med J. 1980; 281: 1336 – 1338.
3. Dupont WD, Plummer WD, Jr. Power and sample size calculations for stud-
ies involving linear regression. Control Clin Trials. 1998; 19: 589 – 601.
4. Garcia-Closas M, Lubin JH. Power and sample size calculations in
case-control studies of gene-environment interactions: comments on
different approaches. Am J Epidemiol. 1999; 149: 689 – 692.
5. The US National Cancer Institute. Available from: URL: http://dceg.
cancer.gov/tools/design/POWER. (Accessed Date: 23 April, 2013).
6. Norman G, Monteiro S, Salama S. Sample size calculations: should the em-
peror’s clothes be off the peg or made to measure? BMJ. 2012; 345: e5278.