COMMIT-CCS2 logo ClOpidogrel and Metoprolol in Myocardial Infarction Trial
Data analysis plan for the assessment of the effects of metoprolol in the COMMIT/CCS-2 study
General principles
The overall, blinded, event rate (8% for mortality and 10% for the combined outcome of death, reinfarction or cardiac arrest) actually observed in the study was lower than that anticipated (10% for mortality and 14% for the combined outcome) in the published study protocol. Hence, the planned sample size was increased during the course of the study to be at least 45,000. However, the final completion date for patient recruitment will depend not on the exact number of patients accumulated, but on a particular date (the start of the Chinese New Year festival on 9 February 2005) that was chosen in 2002 to precede a large international cardiac conference (March 2005 meeting of the American College of Cardiology) at which the provisional results will be presented. Any patients inadvertently allocated a treatment pack after 9 February will not be considered part of the study in any way.
Table 1 describes the characteristics of the patients and Table 2 describes their adherence to treatment and use of other treatments. All analyses will be based on the randomly allocated study treatments, irrespective of adherence (i.e. "intention-to-treat" analyses). For each particular outcome, or group of outcomes, the analysis will be of the number of patients suffering such an outcome at least once from the time of randomisation to whichever comes first of death in hospital, first discharge alive from hospital or day 28 (i.e. during the scheduled trial treatment period). Although patients in COMMIT/CCS-2 are randomized between four treatment groups in a 2 x 2 factorial design with placebo control, the main analyses will involve two-way comparisons of each treatment modality, not four-way comparisons of all possible combinations. Reliable assessment of the effects of metoprolol will not be interfered with by reliable assessment of the effects of clopidogrel (or vice versa), as outcomes among all patients allocated active metoprolol can still be compared with those among all those allocated placebo metoprolol (even though some of both groups will have received clopidogrel and some will not, analogously to the use in both groups of various non-trial treatments). Hence, in assessing the main effects of metoprolol, no formal statistical test is needed of the possible interaction between the effects of clopidogrel and metoprolol, though information will be provided about whether the combined proportional effects of these two treatments are approximately multiplicative (as was the case with aspirin and streptokinase in ISIS-2).
Primary comparisons
For the metoprolol comparison there are two pre-specified primary end-points: (1) death, reinfarction or cardiac arrest; and (2) death. The main comparisons will involve "log-rank" analyses of these two end-points (Table 3: see also Table 4), and all time-to-event analyses will be based on the first relevant event. If a patient is discharged alive before day 28 without a relevant event then it is, by definition, impossible for them to have a relevant event (even if they are known to have died soon after discharge), so the log-rank analysis will keep that patient in until day 28 (rather than censoring on the day of discharge). If (which rarely happens) the time of an event is unknown for a particular patient, then it will be assumed to have been as early as possible, given whatever else is known about that patient. In Table 5, however (where time to event is not generally available), the analyses involve ordinary odds ratio calculations.
No adjustment for multiple comparisons will be made to the calculation of the two-sided P-value (2P) for the first of the two primary end-points. However, if the P-value attributed to the second primary end-point (death) appears more extreme than for the combined end-point, then only the P-value for the combined outcome will be used in discussing the significance of the effects of treatment on mortality. Conventionally, in the final analyses of primary comparisons, 2P<0.05 results are often described as "significant". But, the more extreme the P-value, the more reliable the comparison and, hence, the more definite any finding will be considered to be. Moreover, in interpreting the primary (and the subsidiary) mortality analyses, it is appropriate to consider not only the P-values but also the extent to which they are supported by evidence on relevant non-fatal events.
During the course of this study, interim analyses of the primary end-points (and of other information, such as any serious adverse events thought possibly to be related to the trial treatments) were performed at regular intervals for the Data Monitoring Committee (DMC). In general, the DMC would probably have required a difference of at least three standard deviations in an interim analysis of mortality to justify halting or modifying such a study prematurely. This criterion has the practical advantage that the exact number of interim analyses is of little importance and will not affect the determination of what constitutes a "significant" P-value in the final analysis of the primary outcomes.
Principal subsidiary comparisons: subdivision by time
The principal subsidiary comparisons will, as specified in the published protocol, be of the effects of metoprolol (1) on the combined end-point and (2) on death, during days 0-1, days 2-7 and days 8-28 of the scheduled treatment period (Figures 2 and 3).
Other subsidiary comparisons
The other subsidiary comparisons will involve (A) log-rank analyses of the effects of metoprolol on the combined outcome in certain pre-specified subgroups (listed below) and (B) log-rank analyses of certain other outcomes (listed below) among all randomised patients.
(A) Combined outcome of death, reinfarction or cardiac arrest among the following protocol-specified subgroups, as recorded (after resolution of any data queries) on the entry form (Figure 4):
  • sex: men, women
  • age: <60, 60-69, 70+ years
  • delay from symptom onset: <6, 6 to <13, 13 to 24 hours
  • systolic blood pressure: <120, 120-139, 140-159, 160+ mmHg
  • heart rate: <70, 70-89, 90-109, 110+ bpm
  • presence or absence of fibrinolytic therapy before randomisation
  • presence or absence of allocation to active clopidogrel
  • 3 similar-sized groups with respect to the size of the absolute risk, estimated from a prognostic model using Cox regression analysis. (N.B. The Cox regression will be used to derive an overall prognostic index based on baseline characteristics [excluding allocated treatments] for the overall risk of death, reinfarction or cardiac arrest. Once the best fit is found for a set of variables, the prognostic index will be calculated for each patient using the derived coefficients of those prognostic variables. Three similar-sized groups based on absolute risk [i.e. good, average, and poor prognosis] will then be constructed according to this index.
If, for a particular patient, the value is missing for certain variables (e.g. age or systolic blood pressure), then that patient will be included with the most common subgroup, so each set of subgroup analyses will include all patients. (N.B. It had been intended that sex be included in this list of pre-specified subgroups, but it was inadvertently omitted from the protocol.)
Tests for heterogeneity of the proportional effect observed in subgroups will be used (with appropriately cautious allowance for multiple comparisons) to determine whether the effects in specific subcategories are clearly different from the overall effect. If, however, patient categories can be arranged in some meaningful order (e.g. systolic blood pressure: <120, 120-139, 140-159, 160+) then a test for trend on 1 degree of freedom in the proportional effects would be made, rather than a test for heterogeneity.
In interpreting these subsidiary comparisons, substantial allowance will be made for the very large number of hypotheses involved, for plausibility and consistency, and for evidence from other studies. The very large number of patients in this trial may allow reasonably reliable direct assessment of the effects of the treatments on common outcomes in some major subcategories of patient. But, if there is a clearly positive overall result then, even if there really is benefit in each category of patients, false negative results should be expected in many subgroups (as chance alone may well lead to there being no apparent effect in several subgroups in which treatment really is effective). In such circumstances, "lack of direct evidence of benefit" would not be good "evidence of lack of benefit", and clearly significant overall results would provide strong indirect evidence of benefit in subgroups where the results, considered in isolation, are not conventionally significant (or even, perhaps, appear slightly adverse). Hence, unless the proportional effect of treatment in some specific subcategory is clearly different from that observed overall (including, for example, in the presence and absence of the other study treatment), the effect in that subcategory is likely to be best estimated indirectly by applying the proportional effect observed among all patients in the trial to the absolute risk of the event observed among control patients in that category.
In addition to the planned subgroup analyses described above, many other exploratory analyses could also be performed, including any considered to be necessary by the regulatory agencies for submission purposes. But, in interpreting such results, even greater caution will be exercised about the plausibility of any unexpected findings.
(B) Other outcomes to be analysed among all patients
  • Any reinfarction (separating fatal and non-fatal)
  • Any cardiac arrest (separating VF and other cardiac arrest, and fatal or not)
  • Cardiogenic shock (separating fatal and non-fatal)
  • Severe persistent hypotension requiring treatment
  • Severe bradycardia (heart rate <45 bpm)
  • Other major clinical events in hospital during the scheduled treatment period that were explicitly recorded (i.e. bleeding, heart failure requiring persistent treatment, presumed cardiac rupture, pulmonary embolus)
As in the analyses of primary comparisons, 2P<0.05 results on these other outcomes will be described as being conventionally "significant". But, in interpreting any such findings that are of only marginal statistical significance, it may well be necessary to consider whether they are supported by evidence from related studies. Moreover, the larger the number of events on which a comparison is based and the more extreme the P-value, the more reliable the comparison and, hence, the more definite any finding will be considered to be.
Annex
The study protocol, which pre-specified the main analyses, the principal subsidiary analyses and other subsidiary analyses, was published during the first year of randomisation (J Cardiovasc Risk 2000; 7:435-441), and is appended to this data analysis plan which has been revised blind to the treatment specific results.