We use scattering theory, simulations, and empirical constraints on in
terstellar scintillations to discuss the intermittency of radio signal
s from extraterrestrial intelligence (ETI). The number of ETI sources
in the Galaxy has a direct influence on the expected dynamic range of
fluxes in a survey, through inverse square-law effects and, equally im
portantly, by the number of independent statistical trials made on mod
ulations caused by interstellar scintillations. We demonstrate that sc
intillations are very likely to allow initial detections of narrowband
signals, while making redetections extremely improbable, a result tha
t follows from the skewed, exponential distribution of the modulation.
This conclusion holds for relatively distant sources but does not app
ly to radio SETI toward nearby stars (less than or similar to 100 pc).
Recent SETI has found nonrepeating, narrowband events that are largel
y unexplained. We consider three models in order to assess these event
s and to analyze large surveys in general: (model I) radiometer noise
fluctuations; (model II) a population of constant Galactic sources tha
t undergo deep fading and amplification due to interstellar scintillat
ion, consistent with ETI transmissions; and (model III) real, transien
t signals (or hardware errors) of either terrestrial or extraterrestri
al origin. We derive likelihood and Bayesian tests of the models for i
ndividual events and globally on entire surveys. Applying them to The
Planetary Society/Harvard META data, we find that models II and III ar
e both highly preferred to model I, but that models II and III are abo
ut equally likely. In the context of model II, the likelihood analysis
indicates that candidate events above threshold (similar to 32 sigma)
are combinations of large amplitude noise fluctuations and scintillat
ion gains, making it highly probable that events seen once will only v
ery rarely be seen again. Ruling out model II in favor of model III is
difficult-to do so, many more reobservations (e.g., thousands) are ne
eded than were conducted in META (hundreds) or the reobservation thres
hold must be much lower than was used in META. We cannot, therefore, r
ule out the possibility that META events are real, intrinsically stead
y ETI signals. Our formalism can be used to analyze any SETI program.
We estimate the number of reobservations required to rule out model II
in favor of model III, taking into account that reobservations made p
romptly sample the same scintillation gain as in the original detectio
n, while delayed reobservations sample a decorrelated scintillation mo
dulation. The required number is a strong function of the thresholds u
sed in the original survey and in reobservations. We assess optimal me
thods for applying statistical tests in future SETI programs that use
multiple site and multiple beam observations as well as single site ob
servations. We recommend that results be recorded on many more events
than have been made to date. In particular, we suggest that surveys us
e thresholds that are far below the false-alarm threshold that is usua
lly set to yield a small number of noise-induced ''detections'' in a m
assive survey. Instead, large numbers of events should be recorded in
order to (1) demonstrate that background noise conforms to the distrib
ution expected for it; and (2) investigate departures from the expecte
d noise distribution as due to interference or to celestial signals. I
n this way, celestial signals can be investigated at levels much small
er than the false-alarm threshold. The threshold level for archiving c
andidate intensities and their corresponding sky positions is best def
ined in terms of the recording and computational technology that is av
ailable at a cost commensurate with other survey costs.