The noise in natural electromagnetic time series is typically non-stat
ionary. Sections of data with high magnetic noise levels bias impedanc
es and generate unreliable error estimates. Sections containing noise
that is coherent between electric and magnetic channels also produce i
nappropriate impedances and errors. The answer is to compute response
values for data sections which are as short as is feasible, i.e. which
are compatible both with the chosen bandwidth and with the need to ov
er-determine the least-squares estimation of the impedance and coheren
ce. Only those values that are reliable are selected, and the best sin
gle measure of the reliability of Earth impedance estimates is their t
emporal invariance, which is tested by the coherence between the measu
red and predicted electric fields. Complex demodulation is the method
used here to explore the temporal structure of electromagnetic fields
in the period range 20-6000 s. For periods above 300 s, noisy sections
are readily identified in time series of impedance values. The corres
ponding estimates deviate strongly from the normal value, are biased t
owards low impedance values, and are associated with low coherences. P
lots of the impedance against coherence are particularly valuable diag
nostic aids. For periods below 300 s, impedance bias increases systema
tically as the coherence falls, identifying input channel noise as the
cause. By selecting sections with high coherence (equivalent to the i
mpedance being invariant over the section) unbiased impedances and rea
listic errors can be determined. The scatter in impedance values among
high-coherence sections is due to noise that is coherent between inpu
t and output channels, implying the presence of two or more systems fo
r which a consistent response can be defined. Where the Earth and nois
e responses are significantly different, it may be possible to improve
estimates of the former by rejecting sections that do not generate sa
tisfactory values for all the response elements.