In his article, Brette argues that the “coding metaphor” in neuroscience is inappropriate and misleading because it leads to false interpretations of causality. Brette states that “by postulating that neural codes are representations, we imply that these codes have a causal impact on the brain” (sect. 4.2, para. 1). However, this is implausible since “[in the] technical sense … the word code is used as a synonym for correlate” (sect. 1, para. 4). Restated, the coding metaphor is problematic because it can imply causal function where sufficient evidence to support causal inference does not exist. By relying on this criticism, Brette commits to a broader error: He interprets that isolated correlations, conditional correlations, and statistical inferences between neural activity and function support or refute causal inference. Isolated pieces of correlational or statistical evidence are insufficient to demonstrate a causal relationship between neural activity and functions, perceptions, or behaviors, and should be considered in aggregate to form the basis of causal inference. For this reason, it would be helpful for those seeking to design and interpret experiments to adopt a constructive framework for causal inference in neuroscience.

The correlational nature of individual studies in neuroscience has been explicit since the dawn of electrophysiology, when Caton (Reference Caton1875) stated that “[t]he electric currents of grey matter appear to have a relation to its functions.” Contemporary studies of neural activity and function are still strictly correlational, despite advances in recording and analysis methods. Traditional statistical techniques are agnostic to causal relationships between variables and thus cannot determine causality (Pearl et al. Reference Pearl, Glymour and Jewell2016). Experimental interventions that support causal inferences between brain (dys-)function and behavior have long been sought (Dodds Reference Dodds1878; Ferrier Reference Ferrier1886). However, even studies that use modern versions of these “causal” techniques (optogenetic, chemogenetic, electrical, and pharmacological modulation) provide correlations conditioned on perturbation. Causal inferences on the basis of single experimental results should be tempered because of plausible confounding and off-target effects (Jazayeri & Afraz Reference Jazayeri and Afraz2017).

Insights from other fields provide a clear path toward causal inference with individually circumstantial pieces of evidence. The most influential perspective may be that of medical statistician Austin Bradford Hill, who described nine “viewpoints” that guide causal inference in epidemiology when randomized controlled trials are not possible (Hill Reference Hill1965; see also Phillips & Goodman Reference Phillips and Goodman2004). Here these viewpoints are adapted to form a Bradford Hill-inspired framework for causal inference in neuroscience, where aggregated observational and interventional studies support causal inference:

1. 1. Correlational evidence: Relationships between measurements of neural activity and experimenter-defined responses (whether in downstream neural activity, other physiological or behavioral outcomes). These relationships can be characterized through a variety of forward and backward modeling techniques (see, e.g., Anderson Reference Anderson2019; Baayen et al. Reference Baayen, Davidson and Bates2008; Marinescu et al. Reference Marinescu, Lawlor, Konrad and Kording2018; Rougier Reference Rougier2019; Saxena and Cunningham Reference Saxena and Cunningham2019; Song et al. Reference Song, Langfelder and Horvath2013; Wang and Yang Reference Wang and Yang2016).

   1. i. Strength: Does the neural activity explain a reasonable amount of variability in the response?

   2. ii. Consistency: Does the neural activity reliably produce the outcome?

   3. iii. Specificity: Is the observed relationship between neural activity response unique or one of a vast array of potentially confounding correlations?

   4. iv. Relationship curve: Is there a clear geometric relationship between neural activity and the response?

   5. v. Temporality: Does the neural activity consistently precede the response in time?

   6. vi. Mechanistic plausibility: Is there a plausible mechanism whereby neural activity may produce response?

2. 2. Conditionally correlational evidence: The effect of direct or indirect modulation of neural activity on experimenter-defined outcomes. Modulation includes loss-of-function and gain-of-function “causal” manipulations that are under control of the experimenter.

   1. i. Strength: Does modulation of neural activity explain a reasonable amount of variability in the response?

   2. ii. Consistency: Does modulation of neural activity reliably produce the predicted outcome?

   3. iii. Specificity: Does modulation of neural activity lead to a prescribed outcome or one of a vast array of potential effects?

   4. iv. Relationship curve: Is there a predictable and replicable geometric relationship between modulation of neural activity and the response?

   5. v. Temporality: Does the predicted effect follow the perturbed neural activity at a reasonable delay?

   6. vi. Coherence: Is the predicted effect of modulation of neural activity coherent with other strong hypotheses?

   7. vii. Analogy: Does a modulation of closely related neural activity patterns produce similar effects?

With this framework in mind, one should reconsider Brette's claims related to neural codes and causal inference. For example, Brette states that “BOLD (blood oxygen level-dependent) signal … encodes visual signals in the same technical sense that the firing of neurons encodes visual signals” (sect. 4.2, para. 2). Functional magnetic resonance imaging and electrophysiology studies are both correlational, but Brette's assertion is deeply flawed in important ways. In fMRI and electrophysiology, fundamentally different biological activity is associated with stimulus or behavioral response of interest (Goense and Logothetis Reference Goense and Logothetis2008). Thus, the mechanistic plausibility of a link between neural activity and the experimental condition differs. Furthermore, the spatial specificity and temporality of visually evoked activity cannot be similarly addressed across techniques (Sejnowski et al. Reference Sejnowski, Churchland and Movshon2014). These factors are critically important in guiding causal inference, and therefore, each technique uniquely contributes toward causal inference. To suggest that BOLD signals and action potentials encode visual stimuli in the same technical sense is a conspicuous oversimplification. In this example, the proposed framework for causal inference aids in articulating the relative strengths and weaknesses of different experimental approaches. Furthermore, it provides guidelines for making causal inferences by aggregating individual pieces of evidence that are insufficient in isolation.

Regarding the causal relationship between the physical world and thought, Haugeland (Reference Haugeland1985, p. 106) stated, “If you take care of the syntax, the semantics will take care of itself.” This axiom presents a useful analogy: with a proper framework to describe the syntax (rules and criteria) of causal inference in neuroscience, Brette's claim – the coding metaphor perpetuates inappropriate causal inference – is reduced to an innocuous semantic debate. His further claim that metaphors perpetuate “semantic drift” (sects. 2.1, 2.2, and 3.1) should be addressed not by further semantic prescriptions, but by adhering to reasoned syntax. These semantic debates distract from the ultimate goal of discovering robust, causal relationships between the many levels of organization in the brain and behavior.