Analysis of the feasibility outcomes will be by intention-to-treat (ITT), and will include all consented patients on whom an outcome is available, unless otherwise stated. A secondary analysis will include all consented patients with a recorded Hb below 12 g/dL during follow-up, and for whom an outcome is available (this secondary analysis will be restricted to RBC transfusion outcomes, Hb concentration outcomes, and adherence outcomes). A 5% significance level will be used. All analyses of feasibility outcomes will be unadjusted for baseline covariates.
Main analysis of the clinical outcomes will be by ITT, and will include all consented patients with a recorded Hb below 12 g/dL during follow-up, and for whom the outcome is available. Including only patients with a Hb <12 g/dL allows us to target those patients most likely to be affected by the treatment policy, resulting in a more powerful analysis on a more relevant patient population. A Hb of 12 g/dL was chosen as the cut-off point because it is likely that some patients will be transfused (against policy) above 10 g/dL; if the proportion of patients transfused above 10 g/dL differs between treatment arms, excluding these patients could lead to bias. Using a cut-off of 12 g/dL should allow the majority of transfused patients to be included in the analysis, leading to an unbiased comparison. A secondary analysis will include all consented patients, regardless of whether their Hb dropped below 12 g/dL.
Results will be considered statistically significant at the 5% level. Main analyses for clinical outcomes will be unadjusted for baseline covariates; however, a set of secondary analyses will be adjusted for patient age, the presence of shock, the number of major co-morbidities, and the presence of coagulopathy (defined as an international normalised ratio (INR) >1.5 or a prothrombin time (PT) >3 seconds greater than the control). Mean imputation within the centre will be used for patients with missing baseline covariates . Patient age and the number of major co-morbidities (encompassing ischaemic heart disease, cardiac failure, liver disease, renal disease, respiratory disease, malignancy and stroke) will be modelled using fractional polynomials to allow for the possibility of a non-linear association . Further bleeding up to Day 28 is regarded as the primary clinical outcome.
All analyses will account for clustering to ensure correct type I error rates and confidence intervals [13–15]. Many cluster randomised trials base their analysis on individual level patient data, and use appropriate statistical methods to account for clustering between patients in the same cluster (for example, mixed-effects models or generalised estimating equations). However, analysis methods based on individual level patient data may not perform well when the number of clusters is small [13, 14]. Analysis for TRIGGER will, therefore, be performed using cluster-level summaries, which performs well even with a very small number of clusters [13, 14].
Equal weight will be given to each of the six clusters. All analyses will compare the two treatment arms, unless otherwise stated. Binary outcomes will be presented as a difference in proportions.
Unadjusted analyses using cluster-level summaries can be performed by calculating a summary outcome from each centre, and fitting a linear regression model with the summary outcome as a response variable, and treatment arm as a covariate. For example, for the outcome of mortality, one might choose the proportion of patients who died as a summary measure. To perform the analysis, one would then need to calculate the proportion of patients who died in each centre. A linear regression model would then be fitted, with the proportion of patients who died in each centre as the outcome, and which treatment the centre was randomised to as a covariate (in the TRIGGER trial there would only be six data points, as there are only six centres).
Adjusted analyses using cluster-level summaries will be performed as follows [14
A regression model (linear for continuous outcomes and logistic for binary outcomes) will be fit to individual-level patient data, and will be adjusted for the baseline characteristics listed earlier (age, shock, presence of coagulopathy and the number of major co-morbidities). The model will not adjust for treatment effect, or for centre.
Predicted values based on the fitted regression model will be calculated for each patient (for binary outcomes this equates to the predicted probability of experiencing an event, for continuous outcomes this equates to the predicted mean).
The expected outcome in each cluster will be calculated. For binary outcomes, this is the expected number of events in each cluster, and is calculated by summing the predicted probabilities for each patient in that cluster. For continuous outcomes, the expected mean value is calculated by taking the mean of the predicted values in each cluster.
An appropriate residual is calculated for each cluster. For binary outcomes, this is the observed number of events minus the expected number of events, divided by the number of patients in the cluster. For continuous outcomes, this is the observed mean minus the expected mean.
A linear regression model will be fit using the residuals calculated above as cluster-level summaries, with only the treatment group as a covariate. No degrees of freedom correction will be made for performing an adjusted analysis, as only patient-level variables will be adjusted for.
Cluster level summaries for feasibility outcomes
Cluster-level summaries for feasibility outcomes will be calculated separately in each centre as follows:
Recruitment rate (patients consenting): the proportion of eligible patients providing consent.
Ineligible due to severity of bleeding: the proportion of screened patients who are ineligible due to severity of bleeding
Overall adherence: the mean adherence rate per patient
Adherence per patient: the proportion of patients who had no protocol deviations
Adherence per Hb count: the proportion of Hb counts that did not lead to a protocol deviation
Selection bias - baseline variables for consented patients: The mean age, baseline Hb, number of major co-morbidities, clinical Rockall score and Blatchford score, and the proportion of patients with shock will be calculated for consented patients.
Selection bias - difference between consented and non-consented patients in baseline variables: The difference in the mean Rockall score and Blatchford score, and mean baseline Hb between consented and non-consented patients will be calculated in each centre.
Red blood cell exposure (number of transfusions): the mean number of RBC units transfused per patient.
Red blood cell exposure (patients receiving at least one transfusion): the proportion of patients who receive at least one RBC transfusion.
Hb concentration up to Day 7, and over the entire in-hospital follow-up period, prior to discharge/death/Day28: the area-under-the-curve will be calculated for each patient
Hb concentration at discharge: the mean Hb will be calculated
Cluster level summaries for clinical outcomes
The cluster-level summary will be calculated as the proportion of patients in each centre experiencing the event of interest for the following outcomes: further bleeding, all-cause mortality, need for therapeutic intervention at index endoscopy, need for surgery or radiological intervention to control bleeding, any thromboembolic or ischaemic events (and each of the components separately), acute transfusion reactions, infections and SAEs.
A cluster-level summary for length of hospital stay will be calculated using the median length of stay in each centre, and a cluster-level summary for health related quality of life will be calculated using the mean EQ-5D  in each centre.
Subgroup analyses will be performed for two outcomes: further bleeding and all-cause mortality (both up to Day 28) using an interaction test, and considered statistically significant at the 5% level. The following subgroup analyses will be performed:
Interaction tests will be performed by calculating the difference in proportions (for the chosen outcome) between subgroups within each centre . A linear regression model will then be fit, with the difference in proportions between subgroups as the outcome, and treatment as a covariate. Interaction tests will be unadjusted for baseline covariates, and will be reported with a 95% confidence interval.
Missing data for each clinical outcome will be summarised by treatment arm. Sensitivity to missing data for further bleeding and all-cause mortality up to Day 28 will be assessed under a range of missing-not-at-random scenarios. This will be done by calculating the mean outcome in each cluster as follows:
is the observed cluster level summary, p
is the proportion of patients with missing data in that cluster, and x
is the proportion of patients with missing data who are assumed to have had the event of interest. x
will be varied between 0, 0.2, 0.4, 0.6, 0.8 and 1. The treatment effect and 95% confidence interval will be calculated as before.