STATISTICS AND RESEARCH DESIGN
Simple randomization may lead to
unequal group sizes. Is that a problem?
Wendel Minoru Shibasakia and Renato Parsekian Martinsb
Araraquara, S~ao Paulo, Brazil
I
t is virtually impossible to control all the factors that
may confound results in research. Factors unknown
to the researcher can create differences between
groups that are not related to the primary factor under
analysis. Randomization is perhaps one of the best
ways to try to overcome such unknown factors.1 It ho-
mogenizes groups so that each specimen, participant,
or intervention has the same chance of being allocated
to each experimental group. Therefore, it is important
to include randomization in experimental designs of or-
thodontic studies when possible.
Imagine a research project where 2 adhesive systems
from 2 manufacturers are being compared for shear
bond strength. The researcher may collect data of the
adhesive system from manufacturer A in the morning
and from manufacturer B in the afternoon. It is possible
that the researcher might be tired toward the end of the
day and therefore could, unconsciously, be less precise
when collecting data. Thus, if differences were found be-
tween the groups, would those be attributed only to the
adhesive system or to the researcher's fatigue as well?
Similarly, imagine a clinical study aiming to compare
the bracket debonding rates using 2 composites. To
compare them in the same patient in a split-mouth
design, the researcher could bond the right-side brackets
with composite A and the left-side ones with composite
B. Because the orthodontist might have a different
perspective between the patient's right and left sides,
the bonding quality could be slightly different between
the sides, and this may result in differences that are
not really attributed to the composites. This could un-
dermine the results because it is expected that differ-
ences between 2 interventions are explained only by
the factors being evaluated.
a
Graduate Program of Orthodontics, Faculdade de Odontologia de Araraquara,
Universidade Estadual Paulista, Araraquara, S~ao Paulo, Brazil.
b
Private Practice and Graduate Program of Orthodontics, Faculdade de Odonto-
logia de Araraquara, Universidade Estadual Paulista, Araraquara, S~ao Paulo,
Brazil.
All authors have completed and submitted the ICMJE Form for Disclosure of Po-
tential Conflicts of Interest, and none were reported.
Am J Orthod Dentofacial Orthop 2018;154:600-5
0889-5406/$36.00
Ó 2018 by the American Association of Orthodontists. All rights reserved.
https://doi.org/10.1016/j.ajodo.2018.07.005
600
In these 2 examples, one might think that the biases
mentioned could be controlled if the adhesive systems
(in the in-vitro test) or composite types (in the in-vivo
test) were evaluated or applied in an alternate manner.
Even though this idea may seem interesting, it could
also lead to selection bias: eg, if the operator knows
which adhesive is being evaluated, a systematic bias
might be introduced. Besides eliminating biases and giv-
ing all specimens the same chance of allocation to any of
the groups, randomization is a fundamental premise
that justifies most of the statistical procedures used in
data analysis.2
Thus, any attempt at randomization that has a logical
pattern of allocation and that deviates from pure chance
(eg, chart identification numbers, day of the week, date
of birth) can possibly, even if unconsciously, introduce
unknown factors into the groups being studied. This
could confound the intervention investigated and
compromise the research results.
In orthodontics, these problems are usually circum-
vented by generating random numbers that do not follow
any pattern by a method called simple randomization.
Specific softwares or functions, such as RANDBETWEEN
(Excel; Microsoft, Redmond, Wash), are normally used to
produce a list of random numbers that leave all expected
or unforeseen biases to chance alone.
Nevertheless, simple randomization has a significant
problem in studies with small sample sizes, because
there is a high probability that the groups will be in a
state of imbalance3; in most of the current orthodontic
literature, sample sizes are usually small. Orthodontic
randomized controlled trials published from 1992
through 2012 were shown to have a median sample
size of 46 to obtain adequate power, and a median sam-
ple size of 60 was necessary to produce results.4 Howev-
er, the percentages of these studies with groups in a
state of imbalance, or whether they used any methods
to restrict randomization, decreasing the chance of pro-
ducing groups in a state of imbalance, are unknown.
This imbalance can produce differences in distribution
and variances, decreasing the statistical power.5,6
One way to overcome this problem is to use block
randomization, which produces equal numbers of
Statistics and research design
specimens in each group. In such a method, the spec-
imens are distributed into blocks of multiple numbers
related to the number of groups under study, contain-
ing all possible combinations of allocation but main-
taining a 1:1 balance. Thus, in the aforementioned
clinical research example, when determining which
composite, A or B, will be tested on the right or left
side of 24 patients in a split-mouth design, 4 alloca-
tions are arranged into 6 blocks (AABB, BBAA, BABA,
ABAB, ABBA, and BAAB). These blocks will then un-
dergo simple randomization to determine the sequence
in which they will be applied until all patients are
bonded. The investigator must be careful, though,
when using blocks of the same size, which are easier
to manage, because that could lead to a prediction
of which treatment will be allocated next. Different
block sizes can be used to overcome that issue.
Therefore, the aim of this article was to determine
when block or simple randomization is necessary, based
on the probability of imbalance between groups and on
the influence of the statistical power.
MATERIAL AND METHODS
Four hypothetical research designs were analyzed,
varying the numbers of subjects (20, 30, 60, and 90)
allocated into 2 groups, using an independent 2-tailed
t test, with a 5 0.05, by simple randomization using
the RANDBETWEEN function of the Excel 2011 soft-
ware.
A total of 100 allocation simulations were made for
each research design to describe the differences between
specimens in the groups, their frequencies, and the bal-
ance ratios. Statistical power was also calculated for each
balance ratio using the G*Power software (http://www.
gpower.hhu.de/en.html).7 The effect sizes were varied
from small and medium to large using Cohen's
d (d 5 0.2, d 5 0.5, and d 5 0.8, respectively)8 and
the aforementioned parameters.
The values obtained with block randomization simu-
lations, with a fixed equilibrium ratio of 1:1, were
compared with those using simple randomization.
RESULTS
When simulating with 20 specimens, 17% of the
simulations showed a ratio of 1:1 between the samples.
On the other hand, 70% of the simulations showed im-
balances from 1.2:1 to 1.9:1, which caused a maximum
reduction of 10% of the test power to evaluate large
effects, but less than 1% when medium or small effects
were evaluated. The remaining 13% of the simulations
American Journal of Orthodontics and Dentofacial Orthopedics
601
had larger imbalances (from 2.3:1 to 5.7:1), with grad-
ually decreasing test power. This decrease was marked
especially when a large effect was used in the simula-
tion, causing a 17% decrease (Table I). The decrease
in test power with imbalanced samples occurred in all
effect sizes, with greater reductions for larger effect
sizes. In block randomization, all groups have a 1:1
equilibrium rate; thus, the test power will be 40%
with large effect sizes, 19% with medium effect sizes,
and 7% with small effect sizes.
Only 8% of the simple randomization simulations
with 30 specimens showed balanced groups (1:1). In
the remaining simulations, imbalance ranged from
1.1:1 to 2.3:1, with maximum decreases of 7% in the
test power with a large effect size, 3% with a medium ef-
fect size, and less than 1% with a small effect size. In all
effect sizes, the test power dropped, but the drop was
more pronounced when a large effect size was used.
As in the previous simulation, block randomization
groups had a 1:1 equilibrium, but the test power was
higher for these comparisons. Test powers were 56%,
26%, and 8% for large, medium, and small effect sizes,
respectively (Table II).
With 60 specimens, only 10% of the simulations
had balanced groups, which resulted in an 86% power
in a test with a large effect size. In the remaining sim-
ulations of large effect size, test power was above 80%
even when the imbalance was 2.2:1. When a medium
effect size was used, the maximum decrease of the
test power was 5% (from 1:1 to 2.2:1); in small effect
sizes, the test power decreased by only 1%. In the block
randomization, the groups had a 1:1 equilibrium rate,
and the test power values were 86%, 48%, and 12%
with large, medium, and small effect sizes, respectively
(Table III).
When the sample size was 90, 4% of the simulations
produced balanced groups, whereas 15% of the groups
were similar (1:1) “on average.” In the other 77% of
the simulations, where imbalances ranged from 1.1:1
to 1.6:1, the test power (96%) did not change for the
large effect size, but it fell by 2% and 1% for the medium
and small effect sizes, respectively. The largest imbal-
ance (1.8:1) had a frequency of 3% and was responsible
for decreasing the test power by 1%, 4%, and 1% in
large, medium, and small effect sizes, respectively. In
block randomization, the groups had a 1:1 equilibrium,
and the test power values were 96%, 65%, and 16% for
the large, medium, and small effect sizes, respectively
(Table IV).
The decreases in test power became less as the sample
size increased, and the effect sizes were larger.
October 2018  Vol 154  Issue 4
602 Statistics and research design

Table I. Results from 100 simulations of the allocation of 20 patients into 2 groups by simple randomization
Power (%)

G1 G2 Imbalance Imbalance ratio Frequency d 5 0.8 d 5 0.5 d 5 0.2

10 10 1:1 17% 40 19 7
11 9 0.8 1.2:1 25% 39 18 7
12 8 0.7 1.5:1 25% 38 18 7
13 7 0.5 1.9:1 20% 37 17 7
14 6 0.4 2.3:1 6% 34 16 7
15 5 0.3 3:1 4% 31 15 7
16 4 0.3 4:1 2% 27 14 6
17 3 0.2 5.7:1 1% 23 12 6
G, Group.

Table II. Results from 100 simulations of the allocation of 30 patients into 2 groups by simple randomization
Power (%)

G1 G2 Imbalance Imbalance ratio Frequency d 5 0.8 d 5 0.5 d 5 0.2

15 15 1:1 8% 56 26 8
16 14 0.9 1.1:1 30% 56 26 8
17 13 0.8 1.3:1 18% 55 26 8
18 12 0.7 1.5:1 19% 54 25 8
19 11 0.6 1.7:1 12% 53 25 8
20 10 0.5 2:1 9% 51 24 8
21 9 0.4 2.3:1 4% 49 23 8

G, Group.

Table III. Results from 100 simulations of the allocation of 60 patients into 2 groups by simple randomization
Power (%)

G1 G2 Imbalance Imbalance ratio Frequency d 5 0.8 d 5 0.5 d 5 0.2

30 30 1:1 10% 86 48 12
31 29 0.9 1.1:1 14% 86 48 12
33 27 0.8 1.2:1 18% 86 47 12
34 26 0.8 1.3:1 23% 86 47 12
35 25 0.7 1.4:1 7% 85 47 12
36 24 0.7 1.5:1 7% 85 46 12
37 23 0.6 1.6:1 3% 84 46 11
39 21 0.5 1.9:1 4% 83 44 11
41 19 0.5 2.2:1 1% 81 43 11

G, Group.

DISCUSSION showed.10 As the imbalance increases, regardless of sam-

The probability of groups being balanced after simple
randomization is small, regardless of the sample size.
This does not happen with block randomization, where
the groups are balanced, as has already been shown in
the literature.9 The problem with group imbalance is
that the power of the test decreases, increasing the likeli-
hood of a type II error, which is what our results
ple size, the test power decreases. This may lead to the
need of having groups balanced to maintain a high po-
wer in the statistical tests. Thus, researchers may use a
block type of randomization or erroneously try to bal-
ance the groups by improperly interfering with the sim-
ple randomization process. This improper balancing may
occur due to a simple randomization imbalance or even

October 2018  Vol 154  Issue 4 American Journal of Orthodontics and Dentofacial Orthopedics
Statistics and research design 603

Table IV. Results from 100 simulations of the allocation of 90 patients into 2 groups by simple randomization
Power (%)

G1 G2 Imbalance Imbalance ratio Frequency d 5 0.8 d 5 0.5 d 5 0.2

45 45 1:1 4% 96 65 16
46 44 1.0 1:1 15% 96 65 16
47 43 0.9 1.1:1 16% 96 65 16
48 42 0.9 1.1:1 13% 96 65 15
49 41 0.8 1.2:1 11% 96 65 15
50 40 0.8 1.3:1 11% 96 64 15
51 39 0.8 1.3:1 11% 96 64 15
52 38 0.7 1.4:1 3% 96 64 15
53 37 0.7 1.4:1 7% 96 64 15
54 36 0.7 1.5:1 2% 96 63 15
55 35 0.6 1.6:1 3% 96 63 15
56 34 0.6 1.6:1 1% 95 62 15
58 32 0.6 1.8:1 3% 95 61 15

G, Group.

by losing some specimens of 1 group. Let us analyze the
following example: in a sample size calculation for an
independent 2-tailed t test, with a large effect size
(0.8), with the type I error probability set to 5% and
type II error set to 80%, the result will be a sample of
26 participants in each group. If 3 patients drop out of
1 group and 3 are removed from the other group (in
an attempt to recover a 1:1 balance), 2 groups of 23 pa-
tients will be left. This would decrease the power of the
test by 4.4%, to 75.6%. On the other hand, if the patient
drop is not compensated for and the imbalance is main-
tained, resulting in 23 patients in 1 group and 26 in the
other, the test power would only drop by 2%, to 78%,
with a smaller decrease than when trying to balance
the groups.
The innocent attempt to remove specimens to bal-
ance groups will reduce the test power because it is
more sensitive to the overall sample size than to the bal-
ance between the groups (Fig). Although this procedure
is not considered common, it happens more often than
expected. A search conducted in the PubMed database
on July 17, 2017, using the terms "orthodontic [All
Fields] AND (Controlled Clinical Trial[ptyp] AND
("2016/07/17"[PDAT]: "2017/07/17"[PDAT]) AND "hu-
mans"[MeSH Terms] AND English [lang])" returned 30
studies. A more in-depth analysis of their methodologies
showed that 15 articles applied the simple randomiza-
tion method, and surprisingly, 10 of them obtained
identical groups.
As the sample size increases, larger imbalance ratios
are less frequent, and as mentioned, these large ratios
will significantly affect the test power.10 In sample sizes
larger than 60, the highest observed imbalance ratio
was 2.2:1, reducing the test power by 5%, regardless
of the effect size. If 80% test power is considered to
be adequate for an orthodontic research, simple
randomization can be used safely in samples larger
than 60 if the effect size is large, because the test po-
wer will not be lower than 80%. In small samples
(n 5 30) or experiments that intend to demonstrate
small differences between groups, block randomization
is suggested as a safe way to maintain balance between
groups and to preserve the test power. If a simple
randomization is used instead, not only would it pro-
duce a small test power (based on the experimental
design itself), but it could also decrease the test power
even further due to an almost certain imbalance be-
tween groups.
Even though a sample of 46 has been shown to be
adequate to obtain enough power to produce results in
the orthodontic literature,4 it is always wise to restrict
randomization, decreasing the chance of producing
imbalanced groups, because this sample size is considered
too small to produce balanced results. In addition to block
randomization, there are other methods to overcome
some of the problems of simple random allocation: eg,
stratification and minimization. They can be used when
known factors could confound the results and need to
be controlled, especially in small sample sizes. A good
example may be the influence of sex on compliance. If
women comply more in using their retainers, a study
comparing relapse between 2 appliances may need to
have a balance between the sexes within each group;
otherwise, the outcome can be confounded.11
The first method, stratification, involves dividing the
groups studied into subgroups, called strata, where

American Journal of Orthodontics and Dentofacial Orthopedics October 2018  Vol 154  Issue 4
604
Fig. Graph depicting the decrease of test power in relation to the group imbalance in different sample
sizes.
known factors that could confound the results can be
controlled. Then the features of block randomization
are used in each stratum to balance them; therefore,
it is a combination of both methods. It is good for
small sample sizes when imbalances of prognostic pre-
dictors are more likely to occur.1 However, when there
are too many prognostic factors that could confound
the results, the division of the sample into many strata
will result in a large number of subgroups with fewer
specimens, which could imbalance the treatment allo-
cation. The alternative method for the problem of hav-
ing several prognostic factors in small samples is
minimization.12
Minimization works by randomizing the sample into
strata dynamically during the experiment. In contrast to
stratified randomization, where the specimens are allo-
cated before the experiment, in minimization, the treat-
ment allocated to the next participant enrolled in the
trial depends on the characteristics of participants
already enrolled.12 Besides being slightly more compli-
cated to use, it is best performed with the help of soft-
ware; the main issue with minimization is that it may
be possible to predict the next allocation by looking at
the past assignment.
Another issue with small sample sizes that should be
discussed is the situation where the covariates may
remain imbalanced even when the groups are numeri-
cally balanced. Normally, this will not occur in large
samples where randomization was performed correctly,
but it can affect the results of small samples, particularly
with fewer than 50 specimens.13 When baseline covari-
ates are imbalanced, a multiple linear regression, where
the baseline variable of the dependent variable is also
October 2018  Vol 154  Issue 4 American Journal of Orthodontics and Dentofacial Orthopedics
Statistics and research design
included in the model (analysis of covariance), should
be used.14 This will account for possible bias because
of the baseline values of covariates between groups.15
However, the interpretation of this postadjustment
approach is often difficult because the imbalance of co-
variates frequently leads to unanticipated interaction ef-
fects, such as unequal slopes among subgroups of
covariates.16
CONCLUSIONS
For independent parallel group studies, the following
conclusions were made.
1. Simple randomization almost always produces im-
balances between groups, which is a problem in
studies with fewer than 60 specimens. Therefore,
simple randomization can be used for studies with
samples larger than 60.
2. Block randomization or another method that re-
stricts randomization and manages imbalance,
such as stratified randomization or minimization,
should be used in studies with less than 60 samples.
3. The sample size influences the test power more than
the imbalance rate.
REFERENCES
1. Pandis N, Polychronopoulou A, Eliades T. Randomization in clin-
ical trials in orthodontics: its significance in research design and
methods to achieve it. Eur J Orthod 2011;33:684-90.
2. Hahn GJ, Meeker WQ. Assumptions for statistical inference. Am
Stat 1993;47:1-11.
3. Lachin JM. Properties of simple randomization in clinical trials.
Control Clin Trials 1988;9:312-26.
Statistics and research design
4. Koletsi D, Pandis N, Fleming PS. Sample size in orthodontic ran-
domized controlled trials: are numbers justified? Eur J Orthod
2014;36:67-73.
5. Chen Z, Zhao Y, Cui Y, Kowalski J. Methodology and application of
adaptive and sequential approaches in contemporary clinical trials.
J Probab Stat 2012;2012:1-20.
6. Suresh K. An overview of randomization techniques: an unbiased
assessment of outcome in clinical research. J Hum Reprod Sci
2011;4:8-11.
7. Faul F, Erdfelder E, Lang AG, Buchner A. G*Power 3: a flexible sta-
tistical power analysis program for the social, behavioral, and
biomedical sciences. Behav Res Methods 2007;39:175-91.
8. Cohen J. A power primer. Psychol Bull 1992;112:155-9.
9. Lachin JM, Matts JP, Wei LJ. Randomization in clinical trials: con-
clusions and recommendations. Control Clin Trials 1988;9:
365-74.
American Journal of Orthodontics and Dentofacial Orthopedics
605
10. Lachin JM. Statistical properties of randomization in clinical trials.
Control Clin Trials 1988;9:289-311.
11. Pratt MC, Kluemper GT, Lindstrom AF. Patient compliance with
orthodontic retainers in the postretention phase. Am J Orthod
Dentofacial Orthop 2011;140:196-201.
12. Altman DG, Bland JM. Treatment allocation by minimisation. BMJ
2005;330:843.
13. Grizzle JE. A note on stratifying versus complete random assign-
ment in clinical trials. Control Clin Trials 1982;3:365-8.
14. Kirkwood BR, Sterne JA, Malden MA. Essential medical statistics.
2nd ed. Oxford, United Kigdom: Blackwell; 2003.
15. Pandis N. Analysis of covariance. Am J Orthod Dentofacial Orthop
2016;150:200-1.
16. Kang M, Ragan BG, Park JH. Issues in outcomes research: an over-
view of randomization techniques for clinical trials. J Athl Train
2008;43:215-21.
October 2018  Vol 154  Issue 4