While very powerful, fMRI analysis packages can produce results that are easily misinterpreted or, more problematic, have advanced features that can be misused. To ensure that you and your reader exactly understand the model, it is essential that the approach be described in detail. Although most fMRI studies now report analyses using the general linear model (GLM), there remain substantial differences in how these models are specified and estimated. To a great degree these differences can be captured by knowing which software package was used to perform the analysis, but there can be substantial variability within packages depending upon which options are chosen. Whenever possible, provide a rationale for the user-specified parameters of the software. Some of the important details that may vary even within a package include how the error covariance structure is modeled (e.g., temporal autocorrelation in fMRI timeseries, or correlation induced by repeated measures across subjects). Even within the framework of GLM-based analyses, there are many different approaches to building models. For task-related regressors, it is important to be clear about how the task was modeled (e.g., for a blocked design, was the model based on a boxcar or a series of impulses for each trial within a block?) and how the BOLD impulse response was modeled (e.g., a single or dual-gamma canonical hemodynamic response, or a finite impulse response basis set?). If other regressors such as motion parameters or behavioral covariates are included these should also be described, as should any measures to orthogonalize these regressors. One increasingly common way to present GLM-based design matrices is as an image, which is available from most statistical packages. It is also important to describe the how group effects, as opposed to those in individual subjects, were analyzed and, finally, what precise statistical tests formed the basis for inferences reported. The comparisons that have been performed should be clearly specified in terms of which regressors were included in the contrast and be related to the hypotheses that these comparisons are meant to test.
The majority of published studies today use methods that are part of established software packages and have been described in methodological publications. However, it is not uncommon for a paper to present results using a method that has not been previously described in a methodological publication. In this case, it is critical that the method be described in algorithmic detail so that it can be reproduced by others. We encourage researchers to do this by making their code available with their publication as the most complete description of the procedure. It may also be useful to attach an appendix that describes the method, either mathematically or with pseudocode.
The best test of reproducibility is allowing others to directly reproduce the analysis on your own data. We strongly encourage researchers to make their raw data publicly available with their publication, e.g., via a central database or local web site.