[Analysis of a multicenter clinical trial on pressure ulcer development by a marginal approach in the Cox model]

We exemplify the use of a marginal approach with proportional hazards model when failure times are correlated.

The marginal distribution for each failure time is formulated by the Cox proportional hazards model, while the dependence structure is unspecified. However, a correct variance-covariance estimate of the regression coefficients that takes into account the intra-group correlation is proposed. The program MULCOX2 which implements this statistical methodology is used to assess the effect of a nutritional supplementation intervention on pressure ulcer development on critically ill older patients from a multicentric trial (involving 19 wards). We compare the results obtained with those of the usual Cox regression.

The naive approach yields much smaller standard error estimates of the regression parameters than the robust approach.

In our example, the results obtained with the marginal approach do not modify the conclusions: a nutritional supplementation intervention tends to decrease significantly the formation of pressure ulcers. However in other situations, ignoring the intra-cluster dependence could lead to invalid statistical inference. The variability of the estimated effects by MULCOX2 can be quite sensitive to the number of clusters in the sample and to the clusters size.