Usually
the loss of information from omission of true predictors is
larger than the gain by deleting noise variables. The correct
answer is (b) or (c), using a higher p-value for selection does
limit the loss of information due to omission of true predictors.
A full model, including all covariables, may often lead to the
best predictions from a model. This does require that the set
of included candidate covariables is carefully chosen, e.g.
based on empirical findings in other studies and clinical knowledge.