Share

Cluster Unit Randomized Trials

6. Unit of Inference

Specifying the Unit of Inference

A key feature of cluster randomization trials is that the unit of inference is often at the individual level while randomization is performed at a higher level of subject aggregation. This was the case in the hypertension screening trial reported by Bass et al., 1986, which aimed to evaluate the impact of screening on cardiovascular outcomes in individual patients. Although medical practices were chosen as the unit of randomization, this choice was driven entirely by practical considerations, including administrative convenience and the desire to avoid experimental contamination. Similar considerations applied in the design of the Vitamin A trial described in Example A, where villages were randomized to either the experimental group or a control group. However, studies of Vitamin A supplementation have also been carried out using several other units of randomization, including individuals, households, neighborhoods, and entire communities (West et al., 1991). In each of these trials it was the individual that was the unit of inference.

In some trials the unit of randomization and the unit of inference are both defined at the cluster level, which removes the need to adjust for clustering effects.

Example 2

Althabe et al., 2004 report on a matched-pair trial aimed at reducing the rate of caesarian section deliveries in Latin American maternity hospitals. The intervention in this trial required the obstetrician to seek a second opinion from a senior colleague before proceeding with the c-section, with outcomes recorded at the hospital level only.

Likewise, Diwan et al., (1995) evaluated a policy of “group detailing” on the prescribing of lipid-lowering drugs in a trial randomizing community health centers. A primary endpoint in this study was the number of appropriately administered prescriptions per month, with the health center serving as the unit of analysis.

In both these trials outcomes on any one individual are not of direct interest. Therefore from the perspective of sample size assessment and choice of analysis, the challenges involved are essentially the same as those that apply to individually randomized trials.

These distinctions imply it is important for investigators to clearly specify the primary unit of inference at the planning stage of their trial. Unfortunately this issue has sometimes been referred to as the “unit of analysis problem” (e.g., Whiting-O’Keefe et al., 1984; Divine et al., 1992). Although intended to emphasize the need for accounting for clustering effects when the unit of analysis is at the individual level, this terminology has sometimes been interpreted to imply that all CRTs should use cluster level analyses. On the contrary, it is the unit of inference that determines the level at which the analysis is conducted.

<a href="/longdesc/Donner_Sec6_Exer3.html">Flash Description</a>

Flash is not available on mobile devices. Please view the Flash Description.

Example A
450 villages in Indonesia were randomly assigned to either participate in a Vitamin A supplementation scheme or serve as a control. One-year childhood mortality rates were compared in the two groups (Sommer et al., 1986).
Example B
90 families were randomly assigned to receive either treated nasal tissues or standard tissues. 24-week incidence of respiratory illness was compared in the two groups (Farr et al., 1988).
Example C
One member of each pair of 11 matched maternity hospitals in Belarus was randomly assigned to receive a breastfeeding promotion strategy, with the other member of the pair receiving a control condition based on usual practice (PROBIT trial). The rate of breastfeeding at 12 months was compared between the two groups (Kramer et al., 2001).
Example D
207 general practices were randomized to receive either a structured group education program or standard care offered to patients with newly diagnosed type 2 diabetes. A variety of response variables, including biomedical, lifestyle, and psychosocial measurements were collected over a one-year follow-up period (Davies et al., 2008).
Example E
One member of each pair of 11 matched communities was randomly assigned to a city-wide intervention that promoted the hazards of smoking with the other member serving as a control. Five-year smoking cessation rates were compared in the two groups (COMMIT Research Group, 1995).
Diwan V.K., Wahlström R., Tomson G., Beermann B., Sterky G., Eriksson B. (1995). Effects of ‘Group Detailing’ on the prescribing of lipid-lowering drugs: A randomized controlled trial in Swedish primary care. Journal of Clinical Epidemiology, 48: 705-711.
Alexander F., Roberts M.M., Lutz W., Hepburn W. (1989). Randomization by cluster and the problem of social class bias. Journal of Epidemiology and Community Health, 43:29-36.