Surrogate endpoints in clinical trials: Some statistical perspectives
Lars Frisson
Astra Hässle, Mölndal
Abstract
Surrogate endpoints (SEs) are frequently employed in clinical trials,
for instance, lipid levels as a surrogate for arteriosclerosis and deep
vein thrombosis as a surrogate for pulmonary embolism. The benefits are
evident: smaller sample sizes, shorter trial durations, and sometimes
avoidance of expensive or uncomfortable (e.g. invasive) measurements.
Unfortunately, the other side of the coin is far from empty. In most
applications it is known that the SE is correlated with the true
endpoint (TE), however, whether a causal relationship exists, or simply
association, is often unknown. A further complication is the possibility
of an interaction causing the treatment effect to differ from SE
relative to TE. Many cases have appeared in the literature where
convincing positive results on SEs have translated to clearly negative
results on the corresponding TEs, one example occurring in cardiac
arrhythmia (e.g. the CAST trial). Obviously, the advantages gained by
using SEs are often counterbalanced by a loss of relevance of the new
underlying statistical hypothesis.
In this paper different statistical strategies for assessing the
validity and usefulness of SEs will be suggested and contrasted. It will
be shown that even strong statistical correlation between a SE and a TE
is far from sufficient for proving validity, evidence of causality is
required. For a valid SE displaying convincing results the degree of
confidence admissible in extrapolating conclusions to the TE will be
suggested. Special interest will focus on the use of composite endpoints
as SEs for rare TEs, issues addressed will include: how much treatment
effect must persist for less hard endpoints for them to be useful as
add-ins for a composite endpoint?