Roost counts may be a useful method for assessing and monitoring parrot populations as long as counting regimes can detect real differences in abundance above the noise of daily variability in roost size. We studied a roost of up to 85 Red-tailed Amazons (Amazona brasiliensis) for 28 consecutive mornings and evenings from 12 July to 7 August 2001, and recorded bird behavior and associated weather data. The roost declined significantly over the survey period as the breeding season drew nearer. There was no significant difference between evening and morning roost counts, but we suggest that as long as misty mornings are avoided, morning roost counts were more effective as birds left more quickly and predictably. It took longer for birds to arrive on evenings when the roost was large, but birds left quicker in the morning when there were large numbers in the roost. Weather influenced both roost size and timing of arrival, with larger than expected numbers in the roost, and birds arriving later in the afternoons of sunny, warm days. We tested the reliability of four roost counting regimes: counts from five consecutive nights, five counts, one every fourth night, five nights picked at random, and 10 randomly picked nights. Counts from every fourth day performed significantly worse than all the other regimes in estimating the mean numbers in the roost. The 10-d random sampling regime performed significantly better than the 5-d regime in detecting very large roosts which occurred occasionally through the month.