Andrew Reid PhD

This post continues my exploration of causal inference, focusing on the type of problem an empirical researcher is most familiar with: where the underlying causal model is not known. In this case, the model must be discovered. I use some Python code to introduce the PC algorithm, one of the original and most popular approaches to causal discovery. I also discuss its assumptions and limitations, and briefly outline some more recent approaches. This is part of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology.

In this post, I attempt (as a non-expert enthusiast) to provide a gentle introduction to the central concepts underlying causal inference. What is causal inference and why do we need it? How can we represent our causal reasoning in graphical form, and how does this enable us to apply graph theory to simplify our calculations? How do we deal with unobserved confounders? This is part of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology.

In a previous series of posts, I discussed simple and multiple linear regression (MLR) approaches, with the aid of interactive 2D and 3D plots and a bit of math. In this post, I am sharing a series of short videos aimed at psychology undergraduates, each explaining different aspects of MLR in more detail. The goal of these videos (which formed part of my second-year undergraduate module) is to give a little more depth to fundamental concepts that many students struggle with. This is part of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology.

In this post, I explore multiple linear regression, generalizing from the simple two-variable case to three- and many-variable cases. This includes an interactive 3D plot of a regression plane and a discussion of statistical inference and overfitting. This is part of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology.

In this post, I focus on the logic underlying statistical inference based on fMRI research designs. This consists of (1) modelling the hemodynamic response; (2) "first-level" within-subject analysis of time series; (3) "second-level" population inferences drawn from a random sample of participants; and (4) dealing with familywise error. This is part of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology.

In this post, I introduce the concept of simple linear regression, where we are evaluating the how well a linear model approximates a relationship between two variables of interest, and how to perform statistical inference on this model. This is part of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology.

In our new preprint, we describe a novel methodology for (1) identifying the most probable "core" tract trajectory for two arbitrary brain regions, and (2) estimating tract-specific anisotropy (TSA) at all points along this trajectory. We describe the outcomes of regressing this TSA metric against participants' age and sex. Our hope is that this new method can serve as a complement to the popular TBSS approach, where researchers desire to investigate effects specific to a pre-established set of ROIs.

This is the first of a line of teaching-oriented posts aimed at explaining fundamental concepts in statistics, neuroscience, and psychology. In this post, I will try to provide an intuitive explanation of (1) the Pearson correlation coefficient, (2) confounding, and (3) how partial correlations can be used to address confounding.

One important caveat when working with large datasets is that you can almost always produce a statistically significant result when performing a null hypothesis test. This is why it is even more critical to evaluate the effect size than the p value in such an analysis. It is equally important to consider the distribution of your data, and its implications for statistical inference. In this blog post, I use simulated data in order to explore this caveat more intuitively, focusing on a pre-print article that was recently featured on BBC.

In neuroscience, the conversation around the term "functional connectivity" can be confusing, largely due to the implicit notion that associations can map directly onto physical connections. In our recent Nature Neuroscience perspective piece, we propose the redefinition of this term as a causal inference, in order to refocus the conversation around how we investigate brain connectivity, and interpret the results of such investigations.

Functional connectivity is a term originally coined to describe statistical dependence relationships between time series. But should such a relationship really be called connectivity? Functional correlations can easily arise from networks in the complete absence of physical connectivity (i.e., the classical axon/synapse projection we know from neurobiology). In this post I elaborate on recent conversations I've had regarding the use of correlations or partial correlations to infer the presence of connections, and their use in constructing graphs for topological analyses.

In the movie Limitless, Bradley Cooper stars as a down-and-out writer who happens across a superdrug that miraculously heightens his cognitive abilities, including memory recall, creativity, language acquisition, and action planning. It apparently also makes his eyes glow with an unnerving and implausible intensity. In this blog entry, I explore this intriguing possibility from a neuroscientific perspective.

The term "connectome" was introduced in a seminal 2005 PNAS article, as a sort of analogy to the genome. However, unlike genomics, the methods available to study human connectomics remain poorly defined and difficult to interpret. In particular, the use of diffusion-weighted imaging approaches to estimate physical connectivity is fraught with inherent limitations, which are often overlooked in the quest to publish "connectivity" findings. Here, I provide a brief commentary on these issues, and highlight a number of ways neuroscience can proceed in light of them.

Our article proposing a threshold-free method for comparing seed-based connectivity estimates was recently accepted to Brain Structure & Function. We compared two structural covariance approaches (cortical thickness and voxel-based morphometry), and two functional ones (resting-state functional MRI and meta-analytic connectivity mapping, or MACM).

Announcing four new submissions to the ANIMA database, which brings us to 30 studies and counting. Check them out if you get the time!

Exaptation refers to the tendency across evolution to recycle existing mechanisms for new and more complex functions. By analogy, this is likely how episodic memory — and indeed many of our higher level neural processes — evolved from more basic functions such as spatial navigation. Here I explore these ideas in light of the current evidence.

Publishers of scientific articles have been slow to adapt their business models to the rapid evolution of scientific communication — mostly because there is profit in dragging their feet. I explore the past, present, and future of this important issue.

Machine learning approaches to neuroimaging analysis offer promising solutions to research questions in cognitive neuroscience. Here I reflect on recent interactions with the developers of the Nilearn project.

Announcing a new submission to the ANIMA database: Hu et al., Neuroscience & Biobehavioral Reviews, 2015.

Musings on who I am, where I came from, and where I'm going as a Neuroscientist.