We define a block persistence probability p(l)(t) as the probability that t
he order parameter integrated on a block of linear size l has never changed
sign since the initial time in a phase-ordering process at finite temperat
ure T < T-c. We argue that p(l)(t) similar to l(-z theta O) f(t/l(z)) in th
e scaling limit of large blocks, where z is the growth exponent (L(t) simil
ar to t(1/z)), theta(0) is the global (magnetization) persistence exponent
and f(x) decays with the local (single spin) exponent theta for large x. Th
is scaling is demonstrated at zero temperature for the diffusion equation a
nd the large-n model, and generically it can be used to determine easily th
eta(0) from simulations of coarsening models. We also argue that theta(0) a
nd the scaling function do not depend on temperature, leading to a definiti
on of theta at finite temperature, whereas the local persistence probabilit
y decays exponentially due to thermal fluctuations. These ideas are applied
to the study of persistence for conserved models. We illustrate our discus
sions by extensive numerical results. We also comment on the relation betwe
en this method and an alternative definition of theta at finite temperature
recently introduced by Derrida [Phys. Rev. E55, 3705 (1997)].