In this article, a general theory for the construction of confidence i
ntervals or regions in the context of heteroskedastic-dependent data i
s presented. The basic idea is to approximate the sampling distributio
n of a statistic based on the Values of the statistic computed over sm
aller subsets of the data. This method was first proposed by Politis a
nd Romano (1994b) for stationary observations. We extend their results
to heteroskedastic observations, and prove a general asymptotic valid
ity result under minimal conditions. In contrast, the usual bootstrap
and moving blocks bootstrap are typically valid only for asymptoticall
y linear statistics and their justification requires a case-by-case an
alysis. Our general asymptotic results are applied to a regression set
ting with dependent heteroskedastic errors. (C) 1997 Elsevier Science
S.A.