New primer on logistic models for research in decision neuroscience

The field of decision neuroscience—which aims to understand the neural mechanisms of choice—relies on logistic regression models to describe the factors that contribute to making a choice. In Neuron, Camillo Padoa-Schioppa, PhD, professor in the Department of Neuroscience at Washington University School of Medicine, published a guide to various logistic models and in what settings they’re most appropriately used. Here, Dr. Padoa-Schioppa explains the value of logistic regressions to his field and common mistakes in using them to explain behavior.

What is the utility of logistic models and, specifically, how do they apply to neuroscience?

Logistic models are a rather general and powerful tool to analyze choice behavior. They’re used extensively in the science that deals with choice first and foremost – namely, economics. They are also frequently used in a growing field called decision neuroscience, where neuroscientists study the mechanisms of decision making in general, and in particular economic decisions.

The goal of logistic analysis is to describe a function linking the characteristics of the options that can be chosen and the choice. For example, imagine that you are choosing between different ice cream options. Let’s say you have two flavors and the ice cream can be different sizes. You like chocolate more than vanilla, but you’ll pick vanilla if the cone is twice as large. As flavors and sizes vary, choices vary. The logistic analysis describes how choices relate to the options.

In some way, we neuroscientists think about logistic analyses in a slightly original way. I am trying to make this tool more intuitive for people in the field.

Camillo Padoa-Schioppa, PhD

The reason logistic analyses are useful for neuroscience is that they quantify the traits that determine choices. For example, they can tell us exactly how much you value vanilla ice cream compared to chocolate ice cream. This measure of value can then be used to interpret neural activity. Similarly, choices are often biased in all sorts of ways. For example, other things equal, people might be more likely to choose the option on the right, or perhaps the option that is offered last. These biases can be precisely quantified using logistic regressions. Once a bias is quantified, we can go on and study the neuronal mechanisms that induce that bias. Practically everything we have learned about how the brain generates choices between different goods has one way or another relied on a behavioral analysis that typically was done using logistic models.

What are the weaknesses or deficiencies in how logistic models are applied to economic choice studies in neuroscience, and how does your paper attempt to address them?

Because logistic analysis is a very general tool, it can be used in different ways – that is, one can construct different logistic models to analyze a given data set. But different models make different assumptions, and the results lend themselves to different interpretations. I have the impression that many authors use one model or the other without necessarily appreciating the underlying assumptions. So to some extent this article is a way to call attention to people already in the field to what each model means and how it should be interpreted. In another sense it’s meant to be a tool for students, postdocs, and scholars entering the field, to use as a training tool and as a reference when they design their own analysis.

For instance, let’s say you are choosing between ice creams. You might think the variable driving your choice is the difference in value between the two options. Alternatively, you might think the variable driving choices is the value ratio – or the log value ratio. The two models – value difference and log value ratio – make different assumptions and different predictions. It turns out that most studies in the field use the value difference model, but almost certainly the value ratio model is a better choice in most cases. In any case, one should be aware of what assumptions are made in one model versus another.

How would you like researchers in the field to use the resources in this paper?

My main motivation in writing this article is that every time I have a new student or postdoc in the lab, they don’t know about logistic regression and there’s not a good reference to give them. The literature on logistic analysis is extensive but it is mostly written by and for economists. For neuroscience scholars, it gets quickly mathematical and it is not very closely related to the problems we typically study. There wasn’t an easy-to-use tool for neuroscientists.

Putting it all in one resource could be useful for many people. And in some way, we neuroscientists think about logistic analyses in a slightly original way. I am trying to make this tool more intuitive for people in the field.

What features of economic choice have been illuminated by logistic models?

The field of neuroeconomics has emerged in the past 20 years and, as I said, much of what we have learned about the brain mechanisms underlying choices has relied on logistic analyses of behavior. Among other things, logistic analysis provides quantitative measures of the choice accuracy. Let’s say a person prefers chocolate ice cream to vanilla ice cream. They might typically choose one cone of chocolate to one cone of vanilla – but not always. Maybe ten percent or fifteen percent of the time they choose the vanilla. In other words, there is some variability in their choice. A lot of what we do is to research and make sense of the origins of this variability. Again, as we investigate the brain mechanisms, we rely on behavioral measures obtained through logistic analysis.