Statistical mapping within a binary hypothesis testing framework is the most widely used analytical method in functional MRI of the brain. A common assumption in this kind of analysis is that the fMRI time series are independent and identically distributed in time, yet we know that fMRI data can have significant temporal correlation due to low-frequency physiological fluctuation (Weisskoff et al. [1993]; Proc Soc Magn Reson Med 9:7; Biswal et al. [1995]: Mag Reson Med 34:537-541). Furthermore, since the signal-to-noise ratio will vary with imaging rate, we should expect that the degree of correlation will vary with imaging rate. In this paper, we investigate the effect of temporal correlation and experimental paradigm on false-positive rates (type I error rates), using data synthesized through a simple autoregressive plus white-noise model whose parameters were estimated from real data over a range of imaging rates. We demonstrate that actual false-positive rates can be biased far above or below the assumed significance level alpha when temporal autocorrelation is ignored in a way that depends on both the degree of correlation as well as the paradigm frequency. Furthermore, we present a simple method, based on the noise model described above, for correcting such distortions, and relate this method to the extended general linear model of Worsley and Friston ([1995]: Neuroimage 2:173-181).