Pairs of goldfish retinal ganglion cells with overlapping receptive fields were recorded during stimulation with repeated light flashes. Cross-correlation histograms for 'maintained' discharge, 'on' responses, and 'off' responses were computed with a correction for the systematic responses to the stimuli; cross-covariances were derived from these. If stimulus-induced signals and noise combine linearly, then the cross-covariances are independent of differences in mean firing rate. Cross-covariances of pairs of cells with the same response polarity displayed a positive peak near zero lag; pairs with complementary responses showed a negative peak. 'On-off' cells could generally be classified as on-like or off-like, based on the plateau of firing during a prolonged flash and the relative magnitudes of the on and off peak responses; the cross-covariances of these cells were as one would predict if they were pure on- or off-centre neurones. The cross-covariances derived from the on period usually differed in magnitude from those derived in the dark (either maintained or off response). In general, cross-covariances for off responses were nearly identical to those for the maintained discharges of the same pair, although the mean rates at off were usually quite different from the maintained. The change in magnitude of the cross-covariances from on responses therefore appears to be a non-linear effect of light, and not of the changes in firing rate induced by the light. Other features of the cross-covariances were not affected by stimulation. The general shapes remained fairly constant, and the lags at which the peaks occurred were not consistently affected. We estimated the variance of the firing rate of each unit in three ways, and used two methods of portioning the variance implied by the cross-covariances; from these estimates, we obtained an upper bound for the proportion of the variance of firing of a cell which is due to the common noise that affects both members of a pair. We found that the common influence accounts for less than 20% of the total variance. During stimulation, both the magnitude of the cross-covariance and the variance of the rates change; however, the percentage of total variance contributed by the common noise source is constant. We conclude that light has the effect of changing the gain of the pathway after the introduction of both the common and unshared (private) noise sources but before the ganglion cell.(ABSTRACT TRUNCATED AT 400 WORDS)