• Intraclass correlation coefficients;
  • injection practices;
  • sample size;
  • research design;
  • cluster analysis;
  • Pakistan
  • Coefficients de Corrélation Intra-classe;
  • pratiques de l'injection;
  • taille d’échantillon;
  • concept d’étude;
  • analyse en grappes;
  • Pakistan
  • Coeficientes de correlación intraclase;
  • inyecciones;
  • Paquistán;
  • prácticas con inyecciones;
  • tamaño de muestra;
  • diseño del estudio;
  • análisis de cluster;
  • Paquistán


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

Background  To assess injection practices and to test interventions aimed at reducing unsafe injections in developing countries, cluster surveys and cluster randomized trials are needed. The design of cluster-based studies requires estimates of intraclass correlation coefficients that have to be obtained from previous studies. This study presents such estimates.

Methods  Data were derived from a cross-sectional study of injection use and health seeking in Pakistan that used 34 clusters to select 1150 study subjects aged ≥3 months. We analysed variance to separate its components.

Results  Most of intraclass correlation coefficients were in the range of 0.01–0.05. For proportion of injections received during last 3 months, mean number of injections received and health seeking during the past 3 months the intraclass correlation coefficients were 0.02, 0.04 and 0.02, respectively.

Conclusion  These estimates can be useful in designing cluster surveys and cluster randomized trials for injection safety in Pakistan and other developing countries.

Données de base  Afin d’évaluer les pratiques d'injection et de tester les interventions visant à réduire les injections à risque dans les pays en développement, des études de surveillance et des essais randomisés en grappes sont nécessaires. La conception des essais en grappes nécessite une estimation des coefficients de corrélations intra-classe obtenus à partir d’études précédentes. Cette étude présente de telles estimations.

Méthodes  Les données proviennent d'une étude transversale sur les pratiques d'injections et le recours aux soins au Pakistan, basée sur 34 grappes pour la sélection de 1150 sujets âgés de ≥3 mois. Nous avons analysé la variance pour séparer les composants.

Résultats  La plupart des coefficients de corrélation intra-classe étaient dans la marge de 0,01 à 0,05. sur base de la proportion d'injections reçues durant les trois derniers mois, du nombre moyen d'injections reçues et des recours au traitement durant les trois derniers mois, les coefficients de corrélation intra-classe étaient respectivement de 0,02; 0,04 et 0,02.

Conclusion  Ces estimations peuvent être utiles dans la conception d’études de surveillance ou d'essais randomisés en grappes sur la sûreté des injections au Pakistan et dans d'autres pays en développement.

Antecedentes  Para evaluar las prácticas relacionadas con el uso de inyecciones y probar intervenciones enfocadas a reducir inyecciones no seguras en países en vías de desarrollo, se requieren estudios de grupos y ensayos aleatorizados y estratificados. El diseño de estudios basados en grupos precisa del cálculo del coeficiente de correlación intraclase que ha de obtenerse a partir de estudios previos. En este trabajo presentan dichos cálculos.

Métodos  Los datos se obtuvieron a partir de un estudio croseccional sobre el uso de inyecciones y la búsqueda de salud en Paquistán que utilizó 34 grupos para seleccionar 1150 sujetos con edades ≥3 meses. Se analizó la varianza con el fin de separar sus componentes.

Resultados  La mayoría de los coeficientes de correlación intraclase estaban dentro del rango de 0.01 a 0.05. Para la proporción de inyecciones recibidas durantes los últimos tres meses, el número medio de inyecciones recibidas y la búsqueda de salud durante los últimos tres meses, los coeficientes de correlación intraclase fueron 0.02, 0.04 y 0.02 respectivamente.

Conclusión  Estos cálculos pueden ser útiles para el diseño de estudios de grupos y ensayos aleatorizados por grupos sobre la seguridad de vacunas en Paquistán y otros países en vías de desarrollo.


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

Health outcomes of individuals in a same household or community tend to be more similar than those of individuals in other households and communities (Ukoumunne et al. 1999b). This phenomenon may be due to similar levels of exposures, similar behaviours or genetic pre-dispositions. Because of this similarity, the individuals in a group sharing some characteristics are unlikely to be independent with respect to their health outcomes, since responses of individuals in a group show positive intraclass correlation (Zucker et al. 1995).

In developing countries, sampling frames are often not available for epidemiological surveys; cluster sampling is recommended due to its logistic efficiency (Donner & Klar 2004). Another advantage of cluster sampling is the comparative ease in enumerating groups of households as clusters rather than individuals. In studies at places where responses are naturally clustered, such as patients in general practices or worksites, a similar sampling scheme is applied (Campbell et al. 2000). Similarly in community trials, the unit of randomization is a group of people rather than an individual because the intervention is applied to all people in a group; this lessens the risk of contamination and increases administrative efficiency (Donner & Klar 2004).

In the presence of positive intraclass correlation, application of standard statistical methods for calculating sample size will underestimate the error variance (Donner et al. 1990; Donner & Klar 1996; Ukoumunne et al. 1999b). In cluster-based studies, the between-clusters variation causes inflation of error variance. This needs to be accounted for at the time of designing a study, or it will reduce the power to give desired results. To have adequate power, the sample size has to be increased by a design effect or variance inflation factor (Kish 1965; Donner et al. 1981).

The design effect (D) is a function of average cluster size (inline image) and intraclass correlation coefficient (ICC, ρ) (Kish 1965):

  • image

The ICC quantifies the variation between the clusters and can be defined as proportion of total variation that is attributed to differences between the clusters:

  • image

where σb2 is between cluster component of variance and σw2 is within cluster component of variance.

Unsafe injection practices in developing countries

  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

Therapeutic injections are the most frequently performed medical procedures in developing countries (Hutin et al. 2003). Reuse of injection equipment is common. Reuse of syringes without sterilization transmits hepatitis B, C viruses and HIV (Luby et al. 1997; Quigley et al. 2000; Bari et al. 2001; Usman et al. 2003). The probability of transmission of infection increases with the increase in number of injections received (Hauri et al. 2004). Comprehensive preventive strategies are required to achieve the reduction in injection use and to promote safety. The evidence about the magnitude, the appropriateness of injection use and effective intervention provides a strong foundation for prevention programmes. Epidemiological studies to estimate magnitude of injection use that are representative of larger geographical areas will follow the cluster-sampling technique due to its cost and logistic feasibility. Similarly community randomized trials to evaluate effectiveness of intervention packages to achieve injection safety in community are most likely to use cluster randomization. Furthermore, the patients presenting for treatment at the clinics are naturally clustered at the clinics or hospitals and interventions aimed at changing behaviour of general practitioners (GPs) in clinics also need to randomize clinics rather than patients. Conduct of such studies has important design and analysis implications that should be considered for valid and precise estimates of effect measures. For calculating sample sizes for these studies, estimates of design effect or ICC are needed (Simpson et al. 1995). The available ICC estimates are for smoking prevention, hypertension and heart studies and have been calculated from studies in developed countries such as the UK and the USA (Hannan et al. 1994; Gulliford et al. 1999; Martinson et al. 1999; Murray et al. 2000, 2001, 2002). ICC estimates for various health problems in developing countries are not available and usually assumed. Availability of such estimates would enable the investigators to have more precise sample size and adequate power to estimate the outcome of interest.

Estimates of ICC should be presented in literature for use in future studies (Simpson et al. 1995; Campbell & Grimshaw 1998; Donner & Klar 2004; Varnell et al. 2004). This paper presents the estimates of ICC from a community-based survey that aimed at estimating the frequency of injection use in Pakistan for future cluster-based studies investigating injection use and health care seeking in developing countries.


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

We used data from a population based cross-sectional study conducted in an urban setting in Karachi and in a rural setting in Mirpurkhas district of Sindh province, Pakistan during July through September 2001. Public health care facilities, GP and private unqualified practitioners provide health care in both urban and rural area, but in the city there is a large number of health care providers, while in rural areas they are few and far between, usually in small towns. The study area consists of 237 clusters with 133 clusters in rural and 104 clusters in urban areas. We defined a cluster as a group of people living within specific geographic boundaries used by the district government for its administrative and service delivery purposes. These clearly demarcated areas are called sector in urban Karachi, and deh in rural Mirpurkhas. We interviewed 1150 respondents aged ≥3 months from 34 clusters with 17 clusters from each rural and urban setting based on probability proportional to size (PPS). The average number of houses in a cluster was 280 in rural areas and 650 in urban areas. From each cluster on average 34 households were selected through systematic sampling with a random start. One respondent per house was then selected randomly. We interviewed study participants or their guardians at home, using structured questionnaires in Urdu. Data were collected on number of visits to health care providers, injections received during past 3 months, type of health care provider and setting. The details of the original study are presented elsewhere (Janjua et al. 2005).

Statistical method

For analysis purposes, the individuals living in households were considered as clustered in communities. We obtained within- and between-components of variations using analysis of variance for continuous variables and binary variables (Ukoumunne et al. 1999b). We retained continuous variables as such and reduced categorical variables of more than two responses to dichotomy of yes and no. We performed anova in SPSS version 14.0 for unadjusted estimates. We obtained age and sex-adjusted estimates of variance components using general linear models in which clusters were modelled as random variables. We calculated the estimates of ICC and design effects through Microsoft Excel worksheet by using the following expressions (Fleiss 1981):

  • image

where ρ is intraclass correlation, BMS is between mean square, WMS is within mean square and inline image is average cluster size. We followed the scheme of ICCs presentation recommended by Campbell et al. (2004) for better clarity and use. The confidence interval (CI) can be estimated for continuous variables in case of unbalanced design but the method has been evaluated only for small clusters such as families (Donner & Wells 1986) and distributional assumptions are not met for binary outcomes (Donner & Klar 1993; Donner & Eliasziw 1994; Gulliford et al. 1999). Even with these limitations, the CI provides estimates that can help in sensitivity analysis during sample size calculation (Campbell et al. 2001). We used the following expression based on Fisher's method (Fisher 1970) described and compared with other methods by Ukoumunne for estimation of standard error for CI estimation (Ukoumunne 2002):

  • image

where m is the number of clusters; k the average number of observations per cluster. The estimates of parameters for ICC are presented (Table 3) in a format followed by previous studies to five decimal points (Hannan et al. 1994; Murray et al. 1994).

Table 3.   Estimates of intra cluster correlation coefficient and design effect for selected variables about health care seeking and injection use in Sindh province of Pakistan, 2001
VariablesPoint estim.nBetween varianceWithin varianceICC95 % CI for ICCDEFFAdj ICC95 % CI for Adj ICCDEFF
  1. Point estim, point estimate: mean for continuous and proportion for dichotomous variables.

  2. n, average cluster size.

  3. Between variance, mean sum of square of between the clusters variation.

  4. Within variance, mean sum of square of with in the clusters variation.

  5. ICC, intraclass correlation coefficient.

  6. DEFF, design effect.

  7. Adj, Adjusted.

  8. 95% CI, 95% confidence interval.

  9. *Dichotomous variable showing proportion.

  10. †Negative values truncated to zero.

Socio-demographic variables
 Mean number of rooms in the house2.1343.794131.427250.046510.01221–0.080812.50.046650.01229–0.081012.5
 Mean number of people in the house7.53440.5168611.986770.065420.02353–0.107323.20.066910.02443–0.109393.2
 Ownership of the house (%)*84.5340.684090.115260.126760.06253–0.190995.20.125490.06168–0.189295.1
 Presence of electricity in house (%)*84.3341.718460.087310.354630.23829–0.4709812.70.350050.23428–0.4658312.6
 Gas used as cooking fuel (%)*48.4347.375270.045850.824630.75441–0.8948428.20.821070.74972–0.8924128.1
 Wood used as cooking fuel (%)*50.3347.653370.038120.854550.79427–0.9148229.20.851570.79026–0.9128729.1
 Tap water in the house (%)*49.5341.360750.218370.133340.06692–0.199765.40.133160.06680–0.199525.4
Health care use variables
 Visited health care provider once in last 3 months (%)*84.1340.237240.130350.023550.00000†–0.048181.80.022860.00000†–0.047191.8
 Mean number of visits to providers during last 3 months1.73411.594131.490090.166270.08952–0.243036.50.157280.08325–0.231316.2
 Consulted a public health care facility at last visit (%)*20.4280.383690.154930.050090.01146–0.088732.40.057420.01586–0.098982.6
 Consulted a private health care facility at last visit (%)*79.6280.383690.154930.050090.01146–0.088732.40.057420.01586–0.098982.6
 Consulted public physician at last visit (%)*10.3340.213530.087910.040330.00859–0.072082.30.039280.00797–0.070592.3
 Consulted general practitioner at last visit (%)*61.8341.168370.204350.121840.05927–0.184425.00.121820.05926–0.184395.0
 Consulted public dispenser at last visit (%) *10.1340.430330.079620.114690.05457–0.174814.80.128060.06339–0.192735.2
 Consulted private dispenser at last visit (%) *15.0340.539960.113570.099450.04472–0.154174.30.099650.04485–0.154444.3
 Mean cost of the visit to patient at last visit (Pakistan rupees)90.028186320.90224210915.946620.00000†0.00000†–0.011220.90.00000†0.00000†–0.008830.9
Injection practices variables
 Received at least one injection in last 3 months (%)*73.7340.338000.188000.022930.00000†–0.047291.80.030140.00269–0.057602.0
 Mean number of injections in last 3 months2.834105.1886141.439490.043290.01031–0.076262.40.041940.00953–0.074352.4
 Mean number of therapeutic injections in last 3 months2.63494.3949437.600640.042540.00987–0.075202.40.044530.01104–0.078012.5
 Only oral drug prescribed during last visit (%)*12.6280.483770.116400.101300.04327–0.159343.70.100250.04259–0.157913.7
 Only injection prescribed during last visit (%)*7.7280.338270.074770.111800.05009–0.173514.00.125630.05923–0.192024.4
 Both oral and injection prescribed during last visit (%)*63.0280.371520.177900.037410.00396–0.070872.00.035610.00290–0.068312.0
 Injection prescriber public physician (%)*8.6250.130690.076810.027290.00000†–0.058531.70.053970.01183–0.096122.3
 Injection prescriber general practitioner (%)*63.0251.184260.196190.167670.08652–0.248835.00.127000.05842–0.195584.0
 Injection prescriber public dispenser (%)*11.6250.414650.090160.125840.05764–0.194044.00.158130.07979–0.236484.8
 Injection prescriber private dispenser (%)*16.0250.540220.119010.124010.05641–0.191614.00.129500.06010–0.198904.1
 Received therapeutic injection at last visit (%)*78.0280.432320.157690.058560.01655–0.100572.60.053870.01372–0.094022.5
 Last injection administered by physician (%)*14.1250.163610.119340.014620.08662–0.249011.40.013670.00000†–0.039091.3
 Last injection administered by dispenser (%)*85.9250.163610.119340.014620.09431–0.262841.40.013670.00000†–0.039091.3
 Use of freshly opened syringe (%)*51.0250.531640.237450.047220.00777–0.086662.10.045210.00658–0.083852.1


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

Sample description

We interviewed 575 respondents in urban and 575 in rural settings (total: 1150) from 34 clusters. The average cluster size was 34 households; however, it decreased to 28 for some variables of health care use and to 25 for some variables of injection use, as we used a subset of the overall sample in these analyses (Table 3). The mean age of the study participants was 25 years (SD: 18, median: 27 and range: 1/4–99 years); 75.9% (873) were females. The proportion of adults with no formal schooling was 63.4% (513/809). Most respondents belonged to the Sindhi (33%) and Punjabi (28%) ethnic groups. The median monthly family income was Rs.4000 (63 US$, mean: 4825, SD: 3538, Table 1).

Table 1.   Selected characteristics of the participants in injection use study in Sindh province of Pakistan 2001
  1. *Data collected as continuous variable and categorized later on.

  2. †Include these who originated from parts of India not presently included in Pakistan.

  3. ‡Income in Pakistani rupee, Rs.65 = 1 US$ at the time of study.

Age in years*
Adult (>18 years) education level (years of schooling)*
 Housewife/not employed33858.930453.164256.0
 Government/military servant71.2173.0242.4
 Employed in private firm274.710217.812911.2
Family income‡

Injection use

The original study assessed injection use, the details of which are presented separately (Janjua et al. 2005). The study showed that the crude ratio of injections per capita per year was 11.4 (median: 8, SD: 28). The ratio standardized for age and sex was 13.6 injections per person per year (Table 2). Frequency of injections did not differ between males and females (11.9 vs. 11.3 per year; P = 0.737). Of the 3585 injections administered during the last 3 months, 96% were for therapeutic purposes. GPs prescribed the majority of the injections received in urban (297; 75%) and rural areas (238; 53%) during the last visit. The dispensers were major providers of injections (n = 644, 76%).

Table 2.   Annual ratio of injections per capita by age groups in Sindh province, Pakistan 2001
Age (years)UrbanRuralOverall
nMean95% CInMean95% CInMean95% CI
  1. 95% CI, 95% confidence interval.

 > 456323.93.3–44.51018.24.7–11.716414.25.9–22.6
Age and sex standardized
 > 456318.85.1–32.610110.04.2–15.816417.75.0–30.4

ICC estimates

Point estimates and ICC for variables related with injection practices are described in Table 3. ICC estimates ranged from 0.0 for cost of a visit to a health care provider to 0.85 for use of wood as cooking fuel. Most of the estimates ranged between 0.01 and 0.1. The estimate for having received an injection during the past 3 months was 0.02293 and that for the number of injections during the last 3 months was 0.04329. The estimates for at least one visit to a health provider during the last 3 months and number of visit to health care provider during same period were 0.02355 and 0.16627. Use of new syringe taken from a closed packet was considered safe and its ICC estimate was 0.04722. Injections prescribed by GPs (0.16767) showed strong correlation compared with injections prescribed by a public physician (0.02729). This means that injection prescription is not dependent on the patient's complaint, but rather on the prescription behaviour of GPs.

We did adjust for age and sex but not for cluster level covariates. We did not collect information on cluster variables as this did not fall in the scope of the primary research. Overall, the adjusted estimates are smaller compared with unadjusted values for most of the variables but for some variables adjusted estimates of ICC are greater than the unadjusted estimates (Table 3).

Although most of the ICCs are small, for a large cluster size they can increase the design effect substantially, which in turn affects sample size. Hence even the small ICC should not be ignored while calculating the sample size.


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

This is the first study to present estimates of ICC for injection use. These estimates could be helpful for future conservative sample size estimation in cluster-based studies to assess injection use, and for calculating the population required for adequate power to detect differences among interventions for achieving safe injection practices in developing countries. Updates of global burden of disease in 2000 revealed that unsafe injections account for 21 million hepatitis B virus (HBV) infections, 2 million HCV infections and 260 000 HIV infections each year (Hauri et al. 2004).

Estimates for injection use variables ranged from 0.01462 to 0.05856, while estimates for health care seeking variables ranged from 0.02355 to 0.16627. General practitioners (0.16767) as injection prescribers showed more clustering than public physicians (0.02729). This indicates that injection prescription by GPs is independent of patients’ complaints. High ICCs for injection use by GPs have important implications for design of trials to bring reduction in injection prescription; design effect will be high and hence the required sample size. The large sample size has cost and time implications. Increasing the number of practitioners in the trial with a relatively small number of patients per practitioner will reduce over all sample size while retaining power. However, decisions regarding number of practitioners and number of patients per practitioner should rest on the cost benefit of increasing the practitioners and reducing overall sample size.

In Table 3, we also present design effects when cluster size is 34. The effect of ICC on value of design effect can be seen. For ICC values < 0.04, the design effect is < 2 and for values > 0.1 the design effect is > 4; with an increase in ICC the sample size will increase even if cluster size remains the same. Decreasing cluster size decreases values of design effect and thus required sample size. But recruitment from additional numbers of clusters also has cost implications. The optimal sample size and number of clusters can be achieved by using various values for cluster size.

It has been suggested that the estimates from studies with a number of clusters <40 are less reliable (Donner & Klar 2004). Increasing the number of clusters increases stability of estimates (Donner & Klar 2004). Our estimates come from 34 clusters, which is less than the recommended number of 40, but in the absence of more robust estimates, they can still be used for sample size estimation. However, ICC estimates from other studies investigating injection use would add to available information and give the researchers options for precise sample size calculation.

Confidence intervals presented for the ICCs should be interpreted with caution because of the skewed distribution of continuous variables and presence of dichotomous variables in the data. These conditions do not satisfy assumptions for robust CI estimation. It has been reported that performance of different methods of interval construction is poor for non-normal data (Ukoumunne 2002). Confidence limits for dichotomous outcome variables are more uncertain and application of methods that assume normality may not be appropriate (Donner & Eliasziw 1994). However, CIs provide some bound within which the users can do sensitivity analysis for sample size calculation. Investigators should use sensitivity analysis considering values of ICC and CIs for sample size calculation and use the conservative estimates according to prevailing circumstances (Campbell et al. 2000, 2001; Donner & Klar 2004).

The selection of two settings was based on logistic convenience; the results cannot be generalized to the country, although the health care system is similar, variations exist in the in different parts of Pakistan. But these estimates would help in the design and conduct of future intervention studies in the absence of more representative estimates.

Males in the age group 15–45 years and schoolchildren are under represented in the sample. This is a limitation of household surveys conducted during daytime, as at the time of visit the two groups were absent from most households. Due to logistic reasons only one visit was possible per household. This limits the generalisability of our findings to the general population.

For analysis of cluster studies, investigators should use methods that allow for clustering. The methods that have been recommended to be more appropriate are those that simultaneously adjust for cluster and individual level covariates. There are a variety of methods, generally called multilevel, hierarchical, or random-effects methods. They are based on a different of statistical models, such as the generalized linear, mixed model, generalized estimating equations, and hierarchical Bayesian models (Campbell & Grimshaw 1998; Ukoumunne et al. 1999a; Wears 2002).

The estimates presented in the study will be useful in design of injection use surveys and trials to assess intervention targeting reduction in injection overuse and unsafe injections. We further suggest presentation of estimates from the larger population based surveys and studies at the clinics; this will enhance our ability to better estimate the sample size for the future studies.


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References

Funding for the primary project was provided by USAID/ NVPO through Department of Essential Medical Technologies, World Health Organization, Geneva, Switzerland. NZJ is supported through International Training and Research in Environmental and Occupational Health grant no. 5D43TW05750 from the Fogarty International Centre at the National Institute of Health awarded to University of Alabama at Birmingham. We thank the anonymous reviewers for their comments that helped to improve our paper.


  1. Top of page
  2. Summary
  3. Introduction
  4. Unsafe injection practices in developing countries
  5. Methods
  6. Results
  7. Discussion
  8. Acknowledgements
  9. References
  • Bari A, Akhtar S, Rahbar MH & Luby SP (2001) Risk factors for hepatitis C virus infection in male adults in Rawalpindi–Islamabad, Pakistan. Tropical Medicine and International Health 6, 732738.
  • Campbell MK & Grimshaw JM (1998) Cluster randomised trials: time for improvement. The implications of adopting a cluster design are still largely being ignored. BMJ 317, 11711172.
  • Campbell MK, Mollison J, Steen N, Grimshaw JM & Eccles M (2000) Analysis of cluster randomized trials in primary care: a practical approach. Family Practice 17, 192196.
  • Campbell MK, Mollison J & Grimshaw JM (2001) Cluster trials in implementation research: estimation of intracluster correlation coefficients and sample size. Statistics in Medicine 20, 391399.
  • Campbell MK, Grimshaw JM & Elbourne DR (2004) Intracluster correlation coefficients in cluster randomized trials: empirical insights into how should they be reported. BMC Medical Research and Methodology 4, 9.
  • Donner A & Eliasziw M (1994) Statistical implications of choice between a dichotomous or contineous trait in studies of interobserver agreement. Biometrics 50, 550555.
  • Donner A & Klar N (1993) Confidence interval construction for effect measures arising from cluster randomization trials. Journal of Clinical Epidemiology 46, 123131.
  • Donner A & Klar N (1996) Statistical considerations in the design and analysis of community intervention trials. Journal of Clinical Epidemiology 49, 435439.
  • Donner A & Klar N (2004) Pitfalls of and controversies in cluster randomization trials. American Journal of Public Health 94, 416422.
  • Donner A & Wells G (1986) A comparison of confidence interval methods for the intraclass correlation coefficient. Biometrics 42, 401412.
  • Donner A, Birkett N & Buck C (1981) Randomization by cluster. Sample size requirements and analysis. American Journal of Epidemiology 114, 906914.
  • Donner A, Brown KS & Brasher P (1990) A methodological review of non-therapeutic intervention trials employing cluster randomization 1979–1989. International Journal of Epidemiology 19, 795800.
  • Fisher RA (1970) Statistical Methods for Research Workers. Hafner, New York.
  • Fleiss JL (1981) The measurement of interrater agreement. Statistical Methods for Rates and Proportions. John Wiley & Sons, Inc., New York, pp. 212236.
  • Gulliford MC, Ukoumunne OC & Chinn S (1999) Components of variance and intraclass correlations for the design of community-based surveys and intervention studies: data from the Health Survey for England 1994. American Journal of Epidemiology 149, 876883.
  • Hannan PJ, Murray DM, Jacobs DR & McGovern PG Jr (1994) Parameters to aid in the design and analysis of community trials: intraclass correlations from the Minnesota Heart Health Program. Epidemiology 5, 8895.
  • Hauri AM, Armstrong GL & Hutin YJ (2004) The global burden of disease attributable to contaminated injections given in health care settings. International Journal of STD & AIDS 15, 716.
  • Hutin YJ, Hauri AM & Armstrong GL (2003) Use of injections in healthcare settings worldwide 2000: literature review and regional estimates. BMJ 327, 1075.
  • Janjua NZ, Akhtar S & Hutin YJ (2005) Injection use in two districts of Pakistan: implications for disease prevention. International Journal for Quality in Health Care 17, 401408.
  • Kish L (1965) Cluster sampling and subsampling. Survey Sampling. John Wiley and Sons, Inc. London.
  • Luby SP, Qamruddin K, Shah AA et al. (1997) The relationship between therapeutic injections and high prevalence of hepatitis C infection in Hafizabad, Pakistan. Epidemiology and Infection 119, 349356.
  • Martinson BC, Murray DM, Jeffery RW & Hennrikus DJ (1999) Intraclass correlation for measures from a worksite health promotion study: estimates, correlates, and applications. American Journal of Health Promotion 13, 347357.
  • Murray DM, Rooney BL, Hannan PJ et al. (1994) Intraclass correlation among common measures of adolescent smoking: estimates, correlates, and applications in smoking prevention studies. American Journal of Epidemiology 140, 10381050.
  • Murray DM, Clark MH & Wagenaar AC (2000) Intraclass correlations from a community-based alcohol prevention study: the effect of repeat observations on the same communities. Journal of Studies on Alcohol 61, 881890.
  • Murray DM, Phillips GA, Birnbaum AS & Lytle LA (2001) Intraclass correlation for measures from a middle school nutrition intervention study: estimates, correlates, and applications. Health Education and Behavior 28, 666679.
  • Murray DM, Alfano CM, Zbikowski SM, Padgett LS, Robinson LA & Klesges R (2002) Intraclass correlation among measures related to cigarette use by adolescents: estimates from an urban and largely African American cohort. Addictive Behaviors 27, 509527.
  • Quigley MA, Morgan D, Malamba SS et al. (2000) Case-control study of risk factors for incident HIV infection in rural Uganda. Journal of Acquired Immune Deficiency Syndromes 23, 418425.
  • Simpson JM, Klar N & Donnor A (1995) Accounting for cluster randomization: a review of primary prevention trials 1990 through 1993. American Journal of Public Health 85, 13781383.
  • Ukoumunne OC (2002) A comparison of confidence interval methods for the intraclass correlation coefficient in cluster randomized trials. Statistics in Medicine 21, 37573774.
  • Ukoumunne OC, Gulliford MC, Chinn S, Sterne JA & Burney PG (1999a) Methods for evaluating area-wide and organisation-based interventions in health and health care: a systematic review. Health Technology Assessment 3, iii92.
  • Ukoumunne OC, Gulliford MC, Chinn S, Sterne JA, Burney PG & Donner A (1999b) Methods in health service research. Evaluation of health interventions at area and organisation level. BMJ 319, 376379.
  • Usman HR, Akhtar S, Rahbar MH, Hamid S, Moattar T & Luby SP (2003) Injections in health care settings: a risk factor for acute hepatitis B virus infection in Karachi, Pakistan. Epidemiology and Infection 130, 293300.
  • Varnell SP, Murray DM, Janega JB & Blitstein JL (2004) Design and analysis of group-randomized trials: a review of recent practices. American Journal of Public Health 94, 393399.
  • Wears RL (2002) Advanced statistics: statistical methods for analyzing cluster and cluster-randomized data. Academic Emergency Medicine 9, 330341.
  • Zucker DM, Lakatos E, Webber LS et al. (1995) Statistical design of the Child and Adolescent Trial for Cardiovascular Health (CATCH): implications of cluster randomization. Controlled Clinical Trials 16, 96118.