They Kept Torturing the Data Until It Confessed to "Time Cells"

Behold the power of the modern neuroscientist. Like a magician making you believe in something that did not really happen, a neuroscientist can make you believe in something that's not really there. Part of the trick is to just use loaded language to describe particular cells. So if a neuroscientist wants you to believe that some cells store memories, he can just start calling any arbitrary cells he has selected "engram cells." And if a neuroscientist wants you to believe that some cells have something to do with time-related episodic memories, he can just arbitrarily pick some cells and start calling such cells "time cells." And if the neuroscientist wants to suggest that some cells store information about some place, he can just arbitrarily pick some cells and start calling such cells "place cells."  There are no generally agreed upon standards for identifying some cell as an engram cell or a "time cell" or a "place cell." A neuroscientist can make up any criteria he wishes for identifying some cell as an engram cell or a "time cell" or a "place cell."

Let us look at one of the studies claiming to supply evidence for so-called "time cells." The study is entitled, "Time cells in the human hippocampus and entorhinal cortex support episodic memory." At the very beginning of the paper we have a definition of time cells designed to make sure they will be found: "Time cells are neurons in the hippocampus and entorhinal cortex that fire at specific moments within a cognitive task or experience." It is well known that neurons are constantly firing. The page here (entitled "Neuron Firing Rates in Humans") states, "we expect average firing rates across the brain to be around 0.29 per second," meaning an average neuron would fire several times each second.  So if you have defined time cells as cells that "fire at specific moments within a cognitive task or experience," of course you will be able to find such cells, since neurons are constantly firing.  But in the next paragraph the text describes time cells as cells that "encode temporal information."  Of course, that's an entirely different definition.  The switch in definition does not inspire our confidence. 

The scientists describe below just a little bit of their convoluted and wildly unnatural method for trying to detect time cells:

"To identify time cells, we looked for an interaction between time and firing rate using a nonparametric ANOVA across time bins (Kruskal–Wallis test) after generating session-wide firing rate tuning curves with Gaussian convolution of the spike trains. Significance testing incorporated a permutation procedure, in which we repeated the ANOVA 1,000 times after circularly shuffling the original tuning curve."

This is not even a full description of the convoluted method the scientists have used to try to gin up some evidence for time cells. When you read the supplementary information of the paper, you will read about many other procedural twists and turns of their Byzantine method. For example, we read this:

"We first down sampled the spike train by a factor of 32 or 30, depending on the original sampling rate. We then compared the fits of two models that describe the likelihood of spiking activity at any given sample along the length of the encoding list (for encoding time cells) or retrieval list (for retrieval time cells)....A time field model, specified by a total of four parameters, included a Gaussian field of increased firing probability located somewhere along the length of the encoding list...The former was bound between 0 and 1, so that the mean of the field was located within the encoding list. To prevent excessively large Gaussian fields appearing as a flat line across the list, the standard deviation was bound at 1/6....We used matlab’s particleswarm with fmincon as a hybrid function to minimize the negative log-likelihood of these models to solve for their parameters.... We fit the model to data from all lists, only odd lists, and only even lists to avoid a single list driving the effect."

Reading about this labyrinthine methodology,  I'm reminded of a saying commonly stated among experimenters: if you torture the data long enough, it will confess to anything. 

Neuroscience experiments often go rather like this

The visual evidence the authors present as evidence for "time cells" are some "spike heat maps," not anything coming directly from any scientific instrument, but some visuals resulting from some convoluted arbitrary fiddling with the data.  Such "spike heat maps" don't look impressive at all, and look like what we would get from random data. 

There are two things we would like to see in a study such as this, in order to have any faith that it has actually discovered any evidence of cells that "encode temporal information":

(1) Evidence of pre-registration.  To have some faith that the scientists were not just playing around with data analysis until they found some faint effect they could call evidence of what they wanted to see, we would like to see the paper tell us that the study was a pre-registered study in which the scientists tested only a very specific hypothesis they had previously publicly committed themselves to testing (before collecting any data), using one and only one method of data gathering and data analysis they had previously publicly committed themselves to (before collecting any data).  We can assume the study was not pre-registered, since no claim is made of such a thing. 

(2) Evidence of blinding.  For a study like this to be credible, we would need to see a description of how an exact blinding protocol was followed, to reduce bias in data gathering and data analysis. No mention is made of any blinding protocol. 

The study provides no robust evidence at all that there are cells that "encode temporal information."