Table 5 Other grouping rules when randomisation is stratified by (a) centres or (b) treatment providers (Question 6)

 ID4

Would normally analyse together but adjust for stratification factors (which normally include centres) in analysis.

 ID7

There will be instances where we have combined centres at the analysis stage due to small numbers.

 ID8

Different statisticians/trials do different things. Often site = fixed effect and course within site = random effect. If too few within site, then would combine.

 ID14

Retain structure at analysis.

 ID15

Have grouped by region/country where numbers are small. Any adjustment should be documented in the Statistical Analysis Plan, and final decision regarding appropriateness can be discussed during blind review of data.

 ID30

Have used both pre-specified and ad hoc approaches (due to recruitment issues).

 ID32

We either include as a stratification factor (small number of centres, large patient numbers) or by including centre/provider as a cluster.

 ID7

Thinking about complex intervention studies, we don’t usually allow for a ’provider’ effect in the primary analyses, although not necessarily explicitly stated in protocol—many of these studies effectively have partial clustering. We’ve had recent interesting discussions regarding provider effect in such trials, with Chief Investigators strongly feeling that with standardised/manualised intervention and training, it isn’t relevant.

 ID15

Any adjustment should be documented in the Statistical Analysis Plan, and final decision regarding appropriateness can be discussed during blind review of data.

 ID24

Experience with multiple treatment providers is in oncology trials with different doctors delivering protocol treatment, e.g. chemotherapy/radiotherapy. The actual treating doctor has not been recorded on the Case Report Forms, hence all providers implicitly combined within a centre.

 ID30

Have used both pre-specified and ad hoc approaches (due to recruitment issues).

 ID39

Treatment providers combined by default—as we don’t routinely distinguish them in the analysis.