Journal of Hospital Infection 76 (2010) 292e295
Available online at www.sciencedirect.com
Journal of Hospital Infection
journal homepage: www.elsevierhealth.com/journals/jhin
Opinion
Randomisation and sample size for clinical audit on infection control
I. Fournel*, M. Tiv, C. Hua, M. Soulias, K. Astruc, L.S. Aho
Hospital Hygiene and Epidemiology Unit, Hôpital du Bocage, Dijon, France
a r t i c l e i n f o
s u m m a r y
Article history:
Received 17 November 2009
Accepted 28 May 2010
Available online 6 August 2010
Clinical audit is both a part of clinical governance and an essential component of infection
prevention and control. It is frequently performed on a proportion of the target population.
The sample should represent the source population and be sufficient for statistical analysis. In
a hospital, infection control practices are likely to be quite similar within the same clinical area
(cluster effect). This must be taken into consideration when calculating the necessary number
of patients. Sample size is determined by the desired level of precision for estimating the
compliance rate, or by the difference between observed and expected rates, or on the difference before and after implementation of interventions. To estimate the hospital-wide
compliance rate without additional costs we suggest focusing the audit on a large number of
wards, even if fewer observations within each ward are obtained, rather than auditing a large
number of practices on a restricted number of wards.
Ó 2010 The Hospital Infection Society. Published by Elsevier Ltd. All rights reserved.
Keywords
Clinical audit
Cluster randomisation
Infection control
Representativeness
Sample size
Quality in healthcare
Introduction
Clinical audit, defined as ‘a quality improvement process that
seeks to improve patient care and outcomes through systematic
review of care against explicit criteria and the implementation of
change’, is a part of clinical governance.1 Evaluating professional
practices is also an essential component of infection prevention and
control.2e4 Indeed, infection control audits present an opportunity
to promote infection prevention as well as to suggest care
improvement strategies in partnership with an organisation’s
multidisciplinary teams.5 Usually these audits imply direct observation of hygiene practices. An external auditor (infection control
nurse and/or practitioner) observes hygiene practices in various
settings and for different healthcare professionals. Although several
guidelines on clinical audit are available, the specific modalities of
infection control audits have not been defined as far as we know.1,6
Indeed, the main feature of hygiene audits is that practices tend to
look similar within the same wards but differ between departments (cluster effect), which should be taken into account when
conducting an audit.
* Corresponding author. Address: Hospital Hygiene and Epidemiology Unit,
Hôpital du Bocage, BP 77908, 21079 Dijon Cedex, France. Tel.:þ33 380293394;
fax: þ33 380293497.
E-mail address: isabelle.fournel@chu-dijon.fr (I. Fournel).
The purpose of this article is to review the different modalities
in terms of sampling and of required number of subjects when
auditing hygiene at hospital level.
Sample representativeness
The comprehensive assessment of professional practices in all
heathcare units during a given period is both labour-intensive and
difficult to implement. Thus, in practice, to limit the number of
observations and resources needed to carry out the audit, we try to
select a sample of observations which is representative of the source
population.1 The target population needs to be defined first. This
depends upon the context, which induced an audit of infection
control practices. Audit can be carried out on request from health
authorities (national audit for instance) or on request from a specific
department such as, for instance, evaluating infection control
practices in case of an outbreak. In the first case, the practices will be
assessed in a large number of hospital departments, whereas in the
second case, the audit will be restricted to a specific ward.
Randomisation is required to obtain a representative population
sample, from which the results can be extrapolated to the target
population. If the audit covers the entire hospital or several wards,
healthcare practices tend to be much alike within the same clinical
area (cluster effect) due to health professionals’ working habits,
ward specialty, service policy, and due to standardised protocols
sometimes set up in specific services. This cluster effect must
0195-6701/$ e see front matter Ó 2010 The Hospital Infection Society. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.jhin.2010.05.025
I. Fournel et al. / Journal of Hospital Infection 76 (2010) 292e295
therefore be taken into account when sampling. Table I summarises
various sampling approaches.
Number of observations
In addition to its representativeness the sample must be of
a sufficient size to achieve statistical significance. Sample size
calculation is required prior to any study. It depends on the desired
level of precision for descriptive studies. In case of comparison with
another population or a reference value, the required sample size
depends on the power and precision of the statistical test and on
the size of the difference. To demonstrate a small difference in
compliance rates requires a larger sample number than showing
a large difference. Power and precision increase with sample size.
For clinical audits evaluating the compliance of the observed
practices with the reference standard the required number of
subjects will be calculated by the desired precision level of the
expected compliance rate. In this case, the calculation of the
number of required subjects depends on the expected compliance
rate and the desired precision of the estimate with an error a set
(that is for a degree of confidence 1a).7
For estimation of a single proportion:
n ¼
z2 pq
i2
293
of an attribute that is present in the population (here compliance
expected before the implementation of corrective actions).
Example 2: An audit on hand hygiene is performed after
corrective interventions. The compliance was 50% before and the
expected compliance is 70%. The sample size required to detect
such a difference with 80% power and 5% a risk is 93 per group.
Which statistical unit?
The sample size or the required number of subjects described
above refers to statistical units. In clinical audit this number can
correspond either to the number of observations or to the number
of different healthcare professionals. When the statistical unit is the
number of observations, the sample size refers to the number of
observations to be made. When the statistical unit is the number of
subjects, the sample size refers to the number of health professionals to be observed. In this case, the same subject can be
observed several times. If healthcare professionals have been
sampled from several wards, there will be two levels of clustering:
professionals on the same ward and observations on the same
professionals (Figure 1). A suitable statistical analysis is needed to
take this two-level clustering into account, using a random effects
model, for example.
(1)
Cluster effect
where n is sample size, z2 is the abscissa of the normal curve that
cuts off an area at the tails (1 is equal to the desired confidence
level, e.g. 95%; the value for z is found in software or in statistical
tables which contain the area under the normal curve); i is the
desired level of precision (e.g. i ¼ 0.05); p is the estimated proportion of an attribute that is present in the population (here
compliance expected); q is 1p.
Example 1: in the case of an audit dealing with hand hygiene, in
which z ¼ 1.96, i ¼ 0.05 and expected compliance ¼ 50%, the
required sample size is 384 (1.962 0.5 0.5/0.052). As mentioned
above, the assessment may also focus on comparing the observed
with the expected standard rate, or, in the case of repeated audits,
comparing the rate before and after corrective actions.1,8 A simple
formula for comparing two proportions (p1 and p0) is:
2
za=2 z1b
ðk þ 1Þ
n1 ¼
pffiffiffiffiffi
pffiffiffiffiffi 2 and n0 ¼ kn1
k
2jArcsin p1 Arcsin p0 j
(2)
where n1 is the sample size in group 1 and n0 the sample size in
group 2; p1 is the estimated proportion of an attribute that is present
in the population (here compliance expected after the implementation of corrective actions); and p0 is the estimated proportion
It is necessary to take into account the fact that practices appear
to be more similar within than between departments. This cluster
effect decreases power because individuals in the same cluster are
more homogeneous.9,10 To maintain the same power, we must
correct the number of subjects calculated with formula 1 previously
given by the design effect (DE), also called coefficient of inflation:11
ncorrected ¼
z2 pq
DE
i2
(3)
When the clusters are of equal size, the design effect is calculated as follows:
DE ¼ 1 þ rðm 1Þ
(4)
where m is the number of individuals in a cluster. The intracluster
correlation coefficient r is calculated as follows: 12
r¼
s2b
s2b
(5)
þ s2w
where s2b is the variance between clusters, and s2w the variance
within clusters.
Table I
Sampling procedures used in audits
Scope
Sampling methods
Benefits
Drawbacks
Only one hospital
department
All patients or all events during an entire
audit period within one department
e Suitable for individual problems
or resulting from a targeted request
e Completeness
Does not permit a comprehensive
hospital analysis
Several departments
Cluster randomisation: randomisation of
patients or events during the entire
period of audit
e To obtain data on practices
in various departments
Sample bias caused by choice of departments
All departments
at hospital-wide level
All departments are audited. For each
department, the audit period is randomised.
e Easier to organise
e A deficiency in a service could
be demonstrated
Period effect may occur, e.g. service auditing
during an influenza outbreak
Patients or care activities
chosen individually
at hospital-wide level
Randomisation of patients or events
individually at entire hospital level
e Optimum power at an equivalent
sample size
Difficult to implement:
e An exhaustive list of patients or events is required
e Services with low activity may never be audited
e Mobilise many staff members simultaneously
294
I. Fournel et al. / Journal of Hospital Infection 76 (2010) 292e295
Ward 1
Ward 2
Professional
A
Professional
B
Professional
C
Professional
A
Professional
B
Professional
C
Observation 1
Observation 2
Observation 3
Observation 1
Observation 2
Observation 3
Observation 1
Observation 2
Observation 3
Observation 1
Observation 2
Observation 3
Observation 1
Observation 2
Observation 3
Observation 1
Observation 2
Observation 3
Figure 1. Levels of clustering in hygiene audit.
The intracluster correlation coefficient measures the intensity of
the link between observations in the same cluster by comparing
intra- and intercluster variances.12 It can vary from 0 to 1. The closer
it is to 1, the more similar are hygiene practices within the same
departments and the more they differ from those in other departments. This method of calculation is appropriate when the clusters
are of the same size. When cluster sizes are unbalanced, the design
effect is underestimated, which thus leads to a decrease in study
power.13 Some authors have proposed calculating the design effect
weighting by cluster size.13 For a fixed total number of subjects the
number of clusters increases power.14,15
In order to obtain the best snapshot of infection control
compliance at hospital level, we suggest auditing a large number of
wards rather than a high number of care practices in a restricted
number of wards, even if we obtain fewer observations within each.
Table II summarises sample size for different compliance rates
including a correction for cluster effect.
When an audit is carried out on multiple wards, it is important
that the data are analysed in a way that takes this clustering into
account. Otherwise, the standard error of the final estimate will
tend to be too small and there will be a tendency to overstate the
precision of the findings.
Discussion
Voluntary participation in audit may lead to a selection bias
because volunteers are probably more compliant with good
hygiene practices. Invitation to participate in hospital audit may
lead to greater involvement of the various services. The ideal
would be to embed the audit into as many different wards as
possible. The calculation of the required number of subjects is
rarely reported in published audits and seems to be more related
to practicability than to a-priori sample size calculation. The
French National Authority for Health considers that a number of
observations between 30 and 50 observations is a good compromise.16 Other authors state that 20e50 observations should be
sufficient for single audits.8 To our knowledge, there is no justification for these numbers, except that numbers of observations
exceeding 30 maximise the conditions for the application of
parametric statistical tests. Indeed, the distribution of a continuous variable can sometimes be approximated by the normal
distribution.17 For the British National Audit Office, a sample of
fewer than 30 subjects is a case series, for which it is not possible
to extrapolate the results to the whole target population.18 The
French group for evaluation of hospital hygiene practices
(GREPHH) website stipulates 30 as the minimum number of
observations for analysis. However, the cluster effect is rarely
taken into account in this type of study. Estimating the variability
between clusters is the main difficulty. Table II demonstrates that
the sample size for a compliance rate of 90% can range between
138 and 1383, if 10 subjects or events are observed in each cluster
(e.g. ward). The development of national hygiene audits could
facilitate obtaining the data necessary for calculating the intraclass correlation coefficient and thus the required number of
subjects could become predictable. When the statistical unit is the
number of subjects and when this number per cluster is small,
Kerry estimates that the design effect can be ignored.11 Nevertheless the threshold from which he estimated that the number of
subjects per cluster is small has not yet been defined.
In conclusion, to obtain the best snapshot of overall compliance
rates hospital-wide at constant means, we recommend that future
audits on infection control should focus on a large number of wards
even if the observations within each ward are few, rather than
auditing a high number of care activities in a small number of
services. Randomisation is required to ensure a good representation of results. The choice of wards and departments should be
adjusted to meet the purpose of the audit as well as for the nature
of corrective interventions.
Table II
Sample size corrected for intracluster correlation coefficient (ICC) for different compliance rates at 95% confidence level and 10 subjects per cluster
Compliance (%)
10
20
30
40
50
60
70
80
90
ICC
0
0.01
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1
138
246
323
369
384
369
323
246
138
151
268
352
402
419
402
352
268
151
263
467
613
701
730
701
613
467
263
387
688
904
1033
1076
1033
904
688
387
512
910
1194
1365
1421
1365
1194
910
512
636
1131
1484
1696
1767
1696
1484
1131
636
761
1352
1775
2028
2113
2028
1775
1352
761
885
1574
2065
2360
2459
2360
2065
1574
885
1010
1795
2356
2692
2804
2692
2356
1795
1010
1134
2016
2646
3024
3150
3024
2646
2016
1134
1259
2237
2937
3356
3496
3356
2937
2237
1259
1383
2459
3227
3688
3842
3688
3227
2459
1383
The closer ICC (Formula 5) is to 1, the more similar are hygiene practices within the same departments and the more they differ from those in other departments.
I. Fournel et al. / Journal of Hospital Infection 76 (2010) 292e295
Acknowledgements
We thank M. Fournel for assistance in editing.
Conflict of interest statement
None declared.
Funding sources
None.
References
1. National Institute for Clinical Excellence. Principles for best practice in clinical
audit. Oxford: Radcliffe Medical Press; 2002.
2. Scottish Office Department of Health Advisory Group on Infection. Scottish
infection manual. Edinburgh: SODH; 1998.
3. Hajjar J. Healthcare associated infection control in France: 2005e2008 national
program. J Hosp Infect 2008;70:17e21.
4. Millward S, Barnett J, Thomlinson D. Evaluation of the objectivity of an infection
control audit tool. J Hosp Infect 1995;31:229e233.
5. Bryce EA, Scharf S, Walker M, Walsh A. The infection control audit: the
standardized audit as a tool for change. Am J Infect Control 2007;35:271e283.
6. Copeland G. A practical handbook for clinical audit. London: National Clinical
Governance Support Team; 2005.
295
7. United Bristol Healthcare NHS Trust Clinical Audit Central Office. How to select
an audit sample. UK: UBHT Clinical Audit Central Office; 2005.
8. Cochran WG. Sampling techniques. 2nd ed. New York: Wiley; 1963.
9. Cosby RH, Howard M, Kaczorowski J, Willan AR, Sellors JW. Randomizing
patients by family practice: sample size estimation, intracluster correlation and
data analysis. Fam Pract 2003;20:77e82.
10. Donner A. An empirical study of cluster randomization. Int J Epidemiol 1982;11:
283e286.
11. Kerry SM, Bland JM. The intracluster correlation coefficient in cluster
randomisation. BMJ 1998;316:1455.
12. Killip S, Mahfoud Z, Pearce K. What is an intracluster correlation coefficient?
Crucial concepts for primary care researchers. Ann Fam Med 2004;2:204e208.
13. Eldridge SM, Ashby D, Kerry S. Sample size for cluster randomised trials: effect
of coefficient of variation of cluster size and analysis method. Int J Epidemiol
2006;32:1292e1300.
14. Guittet L, Ravaud P, Giraudeau B. Planning a cluster randomised trial with
unequal cluster sizes: practical issues involving continuous outcomes. BMC
Med Res Methodol 2006;6:17.
15. Guittet L, Giraudeau B, Ravaud P. A priori postulated and real power in cluster
randomised trials: mind the gap. BMC Med Res Methodol 2005;5:25.
16. Agence Nationale pour le Développement de l’Evaluation Médicale. L’évaluation des pratiques professionnelles dans les établissements de santé. L’audit
clinique. Paris: ANDEM; 1994.
17. Jekel JF, Katz DL, Elmore JG. Epidemiology, biostatistics and preventive medicine.
Philadelphia: WB Saunders; 2001.
18. National Audit Office. Statistical and technical team. A practical guide to
sampling. London: National Audit Office; 1998.