Measurement vs Measurement
Clarifying neuroscience language
Neuroscience is already hard, so we shouldn’t make it harder with bad terminology.
That’s where fields like signal processing, control theory, and engineering can help drastically — but they need us to move away from some ingrained assumptions.
A central process in neuroscience is measurement, but we often take it for granted. In this post I’ll talk about a critical nuance in measurement and how ignoring this nuance causes confusion and unnecessary caution.
What is Measurement?
First, when I say measurement, what immediately comes to mind…?
Whatever we mean, we don’t all fully agree… So, first, we should get on the same page.
Measurement as Data
To some, measurement is the data. “I got the measurement,” pretty clearly refers to the numbers (or traces) we take down after something we did.
This is the “Result of [a] Process” — the thing we perform calculations on in an effort to study neural generators/systems.
Measurement as Process
To others, measurement is the process. “I did the measurement yesterday,” pretty clearly refers to an action we did.
This is “The Process” that is crucial for empirical neuroscience — something we spend a lot of effort tuning and perfecting.
The System Diagram
Let’s take the ambiguity we just saw and write it down in a clear mathematical formulation: a system diagram.
A — Ideally, whatever we measure is exactly equal to the neural phenomenon we care about. By finding patterns in the data we’re finding patterns in the neural generator.
B — Realistically, the things we measure are not equal to the neural generators we care about. There’s some distance there that (a) reduces the signal from the neural generator and (b) lets noise from other things into our data.
C — The best we can hope for is sensors that don’t overlap in what they’re measuring. That way, one sensor corresponds to one neural generator. Any correlation in the data measured can then be ascribed to the neural generators.
D — But sometimes we have the scenario where sensors overlap in what they’re measuring.
Across all these cases, things get tricky. And we need better language to disambiguate “measurements (data) are correlated” and “measurements (process) are correlated”.
The Problem
Measurements are correlated…
If we say “the measurements/recordings are correlated” — what specifically do we mean?
Which part(s) of our system are we implicating in whatever it is we’re saying? If we’re assuming different parts, we’ll almost certainly talk past each other.
Independent Neural Generators
Let’s say we’re measuring from two neural generators that are not linked in any way shape or form.
If your measurements (process) are correlated, then two uncorrelated neural generators can spuriously look like they’re correlated.
In this case, we know the neural generators are not linked, but the data we measured will have strong correlations because the sensors both record from both regions.
But if your measurements (process) are not correlated, then your data will also not be correlated. This is in the case where our measurement (process) is perfect and noiseless.
Dependent Neural Generators
In real neural systems, every generator in the system is somehow connected, directly or indirectly, to every other generator.
It’s important to minimize the correlation in your measurement (process) so that the only correlation in your data can be ascribed to dependencies in your neural generators.
But the real world is complicated and we’ll always have a combination of both. Having nuanced enough language that doesn’t conflate processes with results of processes is an important step to have meaningful discussion in the real world.
My Experience
I’m an engineer, so I spent a lot of time and effort reducing cross-talk in recording devices. I always called these “correlations in measurement”.
I’m also neuroscience adjacent, so I spent a lot of time trying to understand why my colleagues were so casual when saying “the hemispheric measurements are obviously correlated”.
I needed a way to disambiguate the generators being correlated (trivially true in any neural system) versus the process being correlated (crosstalk, EEG, etc.).
The first is a “yeah, ok” moment, while the second requires a deep-dive into hardware +/- non-standard analyses.
An Answer
If we change our language and agree on this convention, we can more succinctly convey more nuance — seems like a win-win!
Proposal
The proposal is simple:
- measurement and recording always refer to the process. It’s an operator that acts on neural generators, and is not necessarily a matrix.
- The results of this process are “data” or “timeseries” or “measured values”. This is a vector or matrix with numbers.
Example
This lets us more clearly talk about problematic correlations in our measurement process (from EEG) and utterly expected correlations in any data from realistic neural generators (from directly connected brain regions).
Take Away
Clarifying that “measurement” refers to a process and not the data that comes out for analysis will help clarify whether observed correlations are spurious or expected; especially in multidisciplinary settings involving engineers.