Statistical Analysis of Fear Conditioning Data Using Mixed Models: Methods and Applications

Fear conditioning is one of the most widely used paradigms in neuroscience to study mechanisms underlying learning and memory and to better understand post-traumatic stress disorders and phobias. Fear conditioning is traditionally achieved by pairing a conditioned stimulus, such as a cue or a context, with an unconditioned stimulus, which is usually a mild foot shock. There are different temporal ways to pair those stimuli generating different contingencies. Learning can occur in one or several sessions, followed by testing sessions in which memory can be assessed at different time points. These testing sessions can also elicit distinct brain mechanisms that take control of the behavioral output, such as reconsolidation or extinction. This implies that protocols which evaluate the course of the memory subsequent to the training session may vary in number. Moreover, different physiological responses can be assessed to measure fear, albeit the most used way by far is to analyze freezing behavior, which involves absence of movement except those related to breathing. In fact, different labs and companies share or sell distinct scripts or softwares for measuring freezing, some of them even involve machine learning. Despite the common knowledge on how to perform experiments using the fear conditioning paradigm, the way in which results are analyzed vary tremendously among papers. This results in lack of reproducibility and comparability between labs, which, altogether, highlights the need of a statistical guideline. This protocol offers a detailed pipeline to perform an array of statistical analyses on freezing behavior in a context of fixed and random, crossed and nested effects, covering classic experimental situations such as repeated behavioral sessions and more than one animal cohort. Although this guide has been designed for fear conditioning experiments, it could be useful to analyze data arising from other behavioral experiments. Analyzing correlated data using tools such as mixed effects models, which account for natural dependence and correlation, can contribute to more reliable and reproducible findings in basic neuroscience research.