9By comparing computational model output to BOLD signal changes model-based fMRI has the potential to offer profound insight into what neural computations occur when. If this potential is to be fully realized, statistically significant outcomes must imply specific outcomes. That is, we must have a clear idea of how often a model not present in the BOLD signal but present in the predictor set will reach significance. We ran Monte Carlo simulations of reinforcement learning to examine this kind of specificity, focusing in on two aspects. One, to what degree can we tell related but theoretically distinct predictors apart. About 40% of the time the studied predictors were indistinguishable. Two, how well can we separate out different parameterizations of the same reinforcement learning terms. Nearly all parameter settings were indistinguishable. The lack of specificity between models and between parameters suggests a uncertain relation between significance and specificity. Follow up analyses suggest the temporally slow and prototyped nature of the haemodynamic response (HRF) can substantially increase correlations, ranging from -0.16 to 0.73 with an average of 0.27. Though we focused on a single case study, i.e., reinforcement learning, specificity concerns are potentially present in any design which does not account for the slow prototyped nature of the HRF. We suggest more specific conclusions can be reached by moving from null hypothesis testing approach to a model selection or model comparison framework. 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24onto each voxel's blood oxygen level dependent (BOLD) time course followed by a statistical contrast of 32 the two reward conditions, along with a multiple comparison correction. This is the standard statistical 33 parametric mapping (SPM) routine (Josephs et al., 1997) and it is relatively simple to understand and to 34 implement. It is also robust to noise and other natural (e.g. regional) HRF shape variation (Henson et al., 35 2001; Friston et al., 1998), as thousands of reports empirically demonstrate (Bandettini, 2007). 36 In reality however the neural response and the resulting HRF is not an all or none function. The
37HRF changes in size and shape as function of stimulus, e.g. as a function of visual contrast (Boynton 38 et al., 1996). In fact, stimulus and context dependent changes seem to be the norm. Reward valence and 39 magnitude (Delgado et al., 2003(Delgado et al., , 2000, motivation (Delgado et al., 2004), response accuracy (Seger and 40 Cincotta, 2005), strength of recall (Wais, 2008), degree of regret (Fujiwara et al., 2008), and many other 41 tasks and conditions all show distinct variations in HRF shape.
42Model-based fMRI tries to predict such shape variations by replacing impulse codes, where the HRF 43 shape representing each trial's response is identical, with varying trial-level estimates. We focus solely 44 here on estimates derived from computational modeling efforts. These model-based designs are therefore 45 specific hypotheses a...