Focusing on non-ergodic macroscopic systems we reconsider the variances δO 2 of time averages O[x] of time-series x. The total variance δO 2 tot = δO 2 int + δO 2 ext (direct average over all time-series) is known to be the sum of an internal variance δO 2 int (fluctuations within the meta-basins) and an external variance δO 2 ext (fluctuations between meta-basins). It is shown that whenever O[x] can be expressed as a volume average of a local field Or the three variances can be written as volume averages of correlation functions Ctot(r), Cint(r) and Cext(r) with Ctot(r) = Cint(r) + Cext(r). The dependences of the δO 2 on the sampling time ∆τ and the system volume V can thus be traced back to Cint(r) and Cext(r). Various relations are illustrated using lattice spring models with spatially correlated spring constants.