Multivariable risk prediction can greatly enhance the statistical power of clinical trial subgroup analysis

Table 1 Results of conventional vs. risk-stratified analyses when treatment decreases pre-treatment risk by 50% but at a cost of 3 serious adverse events per year of treatment (6 independent risk factors (RF's) exist, each with a prevalence of 25%)

	True Control Event Rate (CER)	True Relative Risk Reduction (RRR)	True Number Needed to Treat (NNT)	Statistical Power of Subgroup Comparison*
N = 8,800 (% of study population)	For 5-Year Follow-up			P < 0.05
Conventional Subgroup Comparison
Risk factor absent (75%)	2.2	-.19†	-239†	.23
Risk factor present (25%]	4.2	.13	183
Risk Index (Dichotomized measure)
0–1 Risk factors (53.4%)	1.4	-.57†	-125†	.72
≥ 2 Risk factors (46.6%)	4.4	.16	143
Risk Index (continuous measure)‡
0 Risk factors (17.8%)	0.75	-1.59†	-88†	.83
1 Risk factors (35.6%)	1.5	-.51†	-132†
2 Risk factors (29.7%)	3.0	-.02	-1936
3 Risk factors (13.2%)	6.0	.21	83
≥ 4 Risk factors (3.7%)	12.8	.35	24

* For the subgroup comparisons, the statistical comparison tests whether the subgroup with the risk factor receives more or less benefit (two-tailed testing) than the subgroup without the risk factor (testing for an interaction between the risk factor and intervention [treatment vs. control] in a logistic regression model. ²¹ For example, the conventional subgroup comparison had a statistical power of 23% for detecting that those with the risk factor had a greater relative benefit from treatment than those without the risk factor.
† The minus sign denotes that treatment had net harm, rather than benefit.
‡ Area Under the Receiver Operator Characteristic (AUROC) curve for the Risk Index was 0.65.

ISSN: 1471-2288