Improved Detection of Event-Related Functional MRI Signals Using Probability Functions

Abstract

Selecting an optimal event distribution for experimental use in event-related fMRI studies can require the generation of large numbers of event sequences with characteristics hard to control. The use of known probability distributions offers the possibility to control event timing and constrain the search space for finding optimal event sequences. We investigated different probability distributions in terms of response estimation (estimation efficiency), detectability (detection power, parameter estimation efficiency, sensitivity to true positives), and false-positive activation. Numerous simulated event sequences were generated selecting interevent intervals (IEI) from the uniform, uniform permuted, Latin square, exponential, binomial, Poisson, χ2, geometric, and bimodal probability distributions and fixed IEI. Event sequences from the bimodal distribution, like block designs, had the best performance for detection and the poorest for estimation, while high estimation and detectability occurred for the long-decay exponential distribution. The uniform distribution also yielded high estimation efficiency, but probability functions with a long tail toward higher IEI, such as the geometric and the χ2 distributions, had superior detectability. The distributions with the best detection performance also had a relatively high incidence of false positives, in contrast to the ordered distributions (Latin square and uniform permuted). The predictions of improved sensitivities for distributions with long tails were confirmed with empirical data. Moreover, the Latin square design yielded detection of activated voxels similar to the χ2 distribution. These results indicate that high detection and suitable behavioral designs have compatibility for application of functional MRI methods to experiments requiring complex designs.