Mounting a community-randomized trial: sample size, matching, selection, and randomization issues in PRISM

https://doi.org/10.1016/j.cct.2003.12.002Get rights and content

Abstract

This paper discusses some of the processes for establishing a large cluster-randomized trial of a community and primary care intervention in 16 local government areas in Victoria, Australia. The development of the trial in terms of design factors such as sample size estimates and the selection and randomization of communities to intervention or comparison is described. The intervention program to be implemented in Program of Resources, Information and Support for Mothers (PRISM) was conceived as a whole community approach to improving support for all mothers in the first 12 months after birth. A cluster-randomized trial was thus the design of choice from the outset. With a limited number of communities available, a matched-pair design with eight pairs was chosen. Sample size estimates, adjusting for the cluster randomization and the pair-matched design, showed that with eight pairs, on average, 800 women from each community would need to respond to provide sufficient power to determine a 3% reduction in the prevalence of maternal depression 6 months after birth—a reduction deemed to be a worthwhile impact of the intervention to be reliably detected at 80% power. The process of selecting suitable communities and matching them into pairs required careful collection of data on numbers of births, size of the local government areas (LGAs), and an assessment of the capacity of communities to implement the intervention. Ways of dealing with boundary issues associated with potential contamination are discussed. Methods for the selection of feasible configurations of sets of pairs and the ultimate allocation to intervention or comparison are provided in detail. Ultimately, all such studies are a balancing act between selecting the minimum number of communities to detect a meaningful outcome effect of an intervention and the maximum size budget and other resources allow.

Section snippets

Background

This paper discusses some of the processes for establishing a cluster-randomized trial of a community and primary care intervention in 16 local government areas (LGAs) in Victoria, Australia. It traces the development of the trial in terms of design factors such as sample size estimates and the selection and randomization of communities to intervention or comparison prior to trial commencement. The protocol for the trial is presented and discussed elsewhere [1].

In brief, prior population-based

Number of clusters and matching in pairs

The intervention program to be implemented in PRISM was conceived as a whole community approach to improving support for all mothers in the first 12 months after birth. A cluster-randomized trial was thus the design of choice from the outset. In the Victorian context, LGAs were chosen as the appropriate cluster unit because of their statutory responsibility for community service provision to mothers and babies post-hospital discharge and for their data systems based on birth notifications from

Sample size estimates

Sample size estimates where the intervention is assigned at the individual level depend on the primary health outcomes, a baseline or comparison group value, and a minimum program effect to be reliably detected at 80% power.

The primary health outcomes for mothers were prevalence of maternal depression determined by a score of 13 or more on the Edinburgh Postnatal Depression Scale (EPDS) [21] and maternal physical health functioning and mental health summary scores of the health status measure,

Estimates of components of variance

There are two components of variance: the assumed binomial variation within clusters (mothers' depression) (σw2) and a between-cluster (normal) variation (σc2), which comprise within-cluster and between-pair variation, respectively. The proportion of cluster variation reduced by the matching is the correlation ρM. The usual intraclass correlation is also reduced by ρM (see Appendix A).

The effect on sample size then depends on the interplay of assumptions about σc2 and ρM with σw2 estimated as

Missingness

The impact of missingness was considered in two ways: missingness due to nonresponse from women within the communities and missingness due to the possible loss of a cluster. Nonresponse was estimated to be around 33% [27]; this was allowed for by inflating the numbers of births required by 50%. Since numbers required corresponded to the average annual number of births in the communities, the data collection time was increased to 18 months in the modified sample size calculations of 1998. By

Criteria for determining possible matched pairs and potential sets of LGAs : defining the final set of participating LGAs

Pairing between communities was undertaken to minimize potential imbalance between the comparison and intervention arms of the study in the baseline level of the primary outcomes and in associated risk factors, such as the size of each community, the size of the population of interest, and community capacity to implement the intervention.

Reliable maternal depression prevalence data by LGA were not available, so a range of data was obtained to assist with determining useful matching criteria,

Minimizing potential contamination between LGAs

A grid was used to assess possible matches within metropolitan and rural strata. The factors on the margins were data available at a community level (number of births, number of non-English-speaking background mothers, area, population size, estimates of population change, community infrastructure rating, and geographical north/south and east/west grid references). By sorting rows and columns on any of these factors, close matches could be identified and, from the map grids, close communities

Selection of LGAs

With communities enrolled, possible pairs determined, and contamination issues assessed, a two-stage process for selection of the communities was employed to assign the intervention using an appropriate randomization process, designed to promote goodwill and continuing cooperation of all groups involved.

Stage 1 entailed the selection of one final LGA set from all possible sets of metropolitan and rural pairs. Only one metropolitan set of four independent pairs was possible resulting in all

Discussion

At the time this study was first designed, relatively few studies in which the logical or economic unit of randomization was the group or community had been published, and of those, some had quite severe design limitations. Since then, the interest in such studies has burgeoned so that not only are there many published design and results papers available [31], [32], [33], [34], [35], there exist commentaries [20], [36], [37], cautionary tales [38], [39], and texts devoted to the methodology

Acknowledgements

This study was funded from a range of sources: La Trobe University and Victorian Department of Human Services Collaborative Industry grant 1997, National Health and Medical Research Council project grants 1997–1999 and 1999–2001, 2002–2004, Victorian Department of Human Services program implementation and evaluation grants 1998–2000, Victorian Health Promotion Foundation grants for program resources and implementation 1998–2000, Felton Bequest grant 1998, Sidney Myer Fund grant 1999, and

References (56)

  • L. Hancock et al.

    Effect of a community action intervention on cervical cancer screening rates in rural Australian towns: the CART Project

    Prev. Med.

    (2001)
  • S.B. Green

    The advantages of community-randomized trials for evaluating lifestyle modification

    Control. Clin. Trials

    (1997)
  • J. Lumley et al.

    PRISM (Program of Resources, Information and Support for Mothers). Protocol for a community-randomised trial [ISRCTN03464021]

    BMC Public Health

    (2003)
  • R. Small et al.

    Missing voices: what women say and do about depression after birth

    J. Reprod. Infant Psychol.

    (1994)
  • S. Brown et al.

    Maternal health after childbirth: results of an Australian population based survey

    Br. J. Obstet. Gynaecol.

    (1998)
  • J. Gunn et al.

    Visits to medical practitioners in the first 6 months of life

    J. Paediatr. Child Health

    (1996)
  • Department of Human Services, Victoria

    Maternal and Child Health Program Report 1995-6

    (1997)
  • S. Brown et al.

    Physical health problems after childbirth and maternal depression at six to seven months postpartum

    Br. J. Obstet. Gynaecol.

    (2000)
  • J. Gunn et al.

    The role of the general practitioner in postnatal care: a survey from Australian general practice

    Br. J. Gen. Pract.

    (1998)
  • R. Small et al.

    Depression after birth. Does social context matter?

    Med. J. Aust.

    (1994)
  • S. Brown et al.

    Missing Voices. The Experience of Motherhood

    (1994)
  • A. Donner et al.

    A methodological review of non-therapeutic intervention trials employing cluster randomisation, 1979–1989

    Int. J. Epidemiol.

    (1990)
  • J.M. Simpson et al.

    Accounting for cluster randomisation: a review of primary prevention trials, 1990 through 1993

    Am. J. Public Health

    (1995)
  • D.M. Murray

    Design and analysis of community trials: lessons from the Minnesota Heart Health Program

    Am. J. Epidemiol.

    (1995)
  • J.T. Salonen et al.

    Analysis of community-based cardiovascular disease prevention studies—evaluation issues in the North Karelia Project and the Minnesota Heart Health Program

    Int. J. Epidemiol.

    (1986)
  • S.P. Fortmann et al.

    Community intervention trials: reflections on the Stanford Five-City Project Experience

    Am. J. Epidemiol.

    (1995)
  • J. Lomas et al.

    Opinion leaders vs audit and feedback to implement practice guidelines. Delivery after previous cesarean section

    JAMA

    (1991)
  • D.C. Martin et al.

    The effect of matching on the power of randomized community intervention studies

    Stat. Med.

    (1993)
  • Cited by (18)

    • Comparison between computer-assisted-navigation and conventional total knee arthroplasties in patients undergoing simultaneous bilateral procedures: A randomized clinical trial

      2011, Journal of Bone and Joint Surgery
      Citation Excerpt :

      This study was a prospective randomized controlled trial of patients who underwent simultaneous bilateral total knee arthroplasty with one knee treated with conventional total knee arthroplasty and the other treated with computer-assisted-navigation total knee arthroplasty. On the basis of our previous experience and a literature review12,13, a power analysis was conducted to estimate the minimum sample size needed to assess a significant difference between the two groups. The type-I error was set at 0.05 (α < 0.05) and the type-II error, at 0.2 (80% power).

    View all citing articles on Scopus
    View full text