Abstracts

Cecchetti, L. "On the issue of statistical power and error rates in brain-behavior associations" (with G. Handjaras)  It is undeniable that many scientific disciplines are grappling with a reproducibility crisis. Neuroimaging, and particularly studies on the association between brain and behavior, represent no exception in this regard. Among the potential contributors to this crisis, are the inaccurate application of the Neyman-Pearson approach to null hypothesis significance testing (NHST) and the relatively small sample size of most studies. Interestingly, rather than emphasizing the detrimental impact of questionable research practices, there is a growing tendency in the neuroimaging field to see larger samples as a miracle cure. In this talk, we will present evidence on the rather unsurprisingly fact that, when correctly employed, the NHST ensures the validity of statistical inference regardless of the sample size. We will also contend that educating the neuroimaging community about the proper use of statistical methods is more effective in enhancing reproducibility than simply relying on continually increasing sample sizes.

Dorémus, N. "Causal inference by nonlinearity" (with F. Cordoni and A. Moneta)  We propose a statistical identification procedure for recursive structural vector autoregressive (VAR) models that present a nonlinear dependence (at least) at the contemporaneous level. By applying and adapting results from the literature on causal discovery with continuous additive noise models, we show that, under certain conditions, a large class of structural VAR models is identifiable. We spell out these specific conditions and propose a scheme for the estimation of structural impulse response functions in a nonlinear setting. We assess the performance of this scheme in a simulation experiment. Finally, we apply it in a study on the effects of monetary policy on the economy.

Fuentes M., R. "Non-linear dependence and Granger causality: A vine copula approach" (with I. Crimaldi and A. Rungi)   This talk addresses the controversial debate about the dependence and causal relation between energy consumption and economic growth in Italy. In particular, we focus on the analysis of the Granger-causal relationships between prices of oil and gas, energy consumption and GDP growth. The main contribution is the use of novel parametric non-linear models such as vine copulas to study the dependence between these variables as well as a Granger causality test based on this recent family of time series models. Our results show that vine copula models tend to provide a better fit to energy and economic data, in terms of their log-likelihood, than linear models such as Vector Autoregressive models even when allowing for multiple lags of the considered variables. Furthermore, the application of the tests provide evidence of a unidirectional flow of causality from energy consumption to GDP, supporting the growth hypothesis in contexts where traditional linear Granger causality tests find no causal relationship between these variables. 

Martinoli, M. "Statistical inference in macroeconomic simulation models" (with A. Moneta and G. Pallante)   We provide an interpretation of the concepts underlying the estimation and validation of simulated models, focusing on the dependence of the simulated data on the parameters being estimated. We start by considering the classical approaches for simulation-based econometric models (e.g., indirect inference, method of simulated moments and simulated maximum likelihood) and we propose novel inferential tools designed for the estimation of complex simulation models.

Van de Ville, D. "Signals, graphs, and brains: An interdisciplinary tale"   State-of-the-art neuroimaging such as magnetic resonance imaging (MRI) provides unprecedented opportunities to non-invasively measure human brain structure (anatomy) and function (physiology). To fully exploit the rich spatiotemporal structure of these data and gain insights into brain function in health and disorder, novel signal processing and modeling approaches are needed, instilled by domain knowledge from neuroscience and instrumentation. I will highlight some exciting research avenues that leverage graph signal processing by integrating a brain graph (i.e., the structural connectome determined by diffusion-weighted MRI and tractography) and graph signals (i.e., the spatial activity patterns obtained by fMRI). This methodology offers a rich repertoire of tools to quantify properties of signals and graphs relevant for brain sciences.