Open-cohort designs for cluster-randomised trials in institutional settings: A methodology bolt-on to DCM-EPIC

A randomised clinical trial (RCT) is an accepted research method for testing interventions. Recruited participants are often randomly assigned to receive the intervention or to a control group. Calculations are made at the start to work out how many participants are needed to get a precise result. However, many interventions in care homes or hospital wards work at a setting (unit, ward or care home) level. This means the care home or hospital ward, and individual participants in them, are randomised as a cluster. This is called a cluster RCT. In cluster RCTs the number of clusters as well as the number of participants is important.

The applicants are currently carrying out a cluster RCT that is looking at how effective an intervention called Dementia Care Mapping (DCM) is at reducing agitation in people with dementia. Staff members use DCM to identify improvements they can make in the care home. Therefore, any resident living in the care home while DCM is being used, will potentially benefit.

There are two accepted ways of designing a cluster RCT. A closed-cohort design follows the same individuals over time, collecting data on them at the start, end and sometimes at middle time points during a trial. This allows the effects of the intervention to be analysed on individuals and the group. Our DCM trial uses this design. A problem in care homes, hospitals and other settings is the high turnover of individuals due to discharge/moving or death. Additional individuals also enter the setting over this time and are exposed to the intervention but are not included in the RCT. The longer the RCT is, the more individuals are likely to drop out. To manage this many more clusters and individuals have to be recruited at the start of the RCT to be sure of having enough individuals still there at the end. This makes the research very expensive. In our trial of DCM, around 50% of residents will have left the trial by our final 16-month follow-up, mainly due to death. A cross-sectional design is the current alternative. This design collects data on individuals in a cluster at the start and at end of the RCT. It does not track individuals over time, instead assuming that individuals differ at each data collection point. This again reduces the number of individuals that can be recruited at later time points.

A different design that could be used is an open-cohort cluster RCT. In this design, all newly eligible individuals are recruited and data collected on them either continuously or at set data collection time points after randomisation. Potential advantages of this design are it may be more (i) economical because fewer clusters need to be recruited, (ii) flexible as it can allow research questions relating to individuals exposed to the intervention throughout an RCT to be answered. However, as this design is not widely used, there is currently no standard guidance for those carrying out open-cohort cluster RCTs and there remain questions about how statisticians should work out the sample size needed and how to best analyse the data. This is a barrier to using an open-cohort design.

In this study, we will address these barriers. We will review the literature on open-cohort RCTs. We look at the statistical properties of open-cohort RCTs by reanalysing data from our DCM trial and from two further RCTs we identify in the literature review. This will include how to work out the sample size, how to analyse the data and advice for researchers on the situations when an open-cohort RCT will be the best choice design. We will also consult researchers who carry out RCTs, through expert group meetings and on-line surveys to make sure our findings are seen as acceptable by those who will use them. We will develop practical guidelines about open-cohort cluster RCTs that can be used by researchers who are planning to carry future trials. These will be shared through academic papers, conferences and briefing documents for research funders.

Our aim is to improve the information from, and efficiency of, randomised trials of complex health and social care interventions in care homes, inpatient wards and clinics treating patients with chronic conditions. We will develop practical guidelines and supporting documents (including training material) for trialists wishing to design an open-cohort cluster-randomised trial (OC-CRT) in an institutional setting.

Although there are examples of its use, the statistical literature on OC-CRTs is extremely sparse. To facilitate comparisons, Feldman and McKinlay suggested a "unified model", encompassing closed-cohort and cross-sectional CRTs. However, this makes strong assumptions, the impact of which is unclear, particularly if applied to open-cohorts with informative entry and exiting. For example, it assumes constant correlation among repeated measurements, one time scale (time from randomisation), independent Gaussian random effects with common variance, stability of intervention effects across time, uninformative cluster sizes and a target population with a steady state. Methods for handling missing data are also unclear.

Building on literature relating to epidemiological studies, which tends to assume survival outcomes, we will analytically derive multilevel linear models that progressively relax the assumptions made by Feldman and McKinlay, exploring the inferences possible from these models. We will consider different methods for handling missing data arising from discharge or death (unconditional, pattern-mixture or joint outcome models or regressions conditioning on being alive) for OC-CRTs where data collection is continuous or at discrete follow-ups. We will derive formulae for calculating the sample size required for a range of illustrative examples. Next, we will perform simulation studies evaluating our methods in terms of bias and precision. Finally, we will address potential barriers to the uptake of these designs, alongside methods developments.

The impact of this research on potential beneficiaries is summarised as follows:

1. Patients will benefit from the increased information obtained from trials of complex interventions delivered to staff within institutions providing health and social care. Also, if efficiency gains are made, new interventions will reach patients quicker than at present. Given the time-scale of trials, it may take 20 years for these benefits to be fully realised, although specific groups of patients will benefit from participating in the trials on which we collaborate, or within about 5 years of the publication of trial results.
2. Providers of health and social care, including care home, hospital and clinic managers, will benefit from better understanding of interventions that affect their culture, as well as better quantification of the uncertainties attached to them. The time-scale will be as for patients. In the even longer term, an improved evidence-base could see a step-change in the culture of these providers.
3. Health and social care funders will benefit from the interventions in institutions being evidence-based. This will mean less funding is spent on interventions that ultimately have no chance of success. Knowledge-based interventions will also allow greater certainty in required future funding. The time-scale again corresponds to that of patients.
4. The UK taxpayer will benefit from the reduced costs on the same time-scale. Society in general will benefit from improvements to health and social care, including improved care of the elderly and inpatient psychiatric care, etc., which will improve quality of life and allow people to return to work and contribute to society more generally more quickly than is currently the case. This is again a long-term impact.
5. Trial funders (e.g. NIHR) will benefit from the reduced cost of finding evidence-based conclusions from well-designed trials, and the increased success in translating trial results into practice, in line with an increasing focus on efficient trial design from NIHR. This will start to have an impact shortly after the life of the project, but the full impact will only come as successful trials using these methodologies are completed and reported in around 10 years' time.

Ultimately, the most convincing metric of success will be the number of trials that are successfully reported using, or influenced by, the methodology recommended by this research, which in turn lead to improvements in staff training and patient care. However, this will take considerably longer than two years to accrue. During the course of the proposed project, we aim to have members of the team contribute to 2 trial grant applications. We anticipate that one of these will be funded, which will have an impact on the NHS. The number of users willing to collaborate with us will give an early indication of the perceived value of our work; we aim to have two by Year 2. We aim to have at least three papers accepted for publication in clinical and medical statistics journals, which will indicate a level of acceptance by the clinical trials research community. Finally, the number of people who attend our training course will give an indication of the level of interest. We aim to have a total of 20 registrations.