Distribution-free results beyond Gauss-Markov theory-re found under we
ak assumptions regarding the errors. Symmetry, unimodality, and locati
on-scale families are studied in estimation; nonstandard versions of G
auss-Markov results are given; and distribution-free confidence sets a
re tightened under symmetry and unimodality of errors. Normal-theory a
pproximate tests are seen to exhibit monotone power in certain classes
of symmetric unimodal errors. (C) 1997 Academic Press.