[UPDATE: Jake Westfall points out in the comments that the paper discussed here appears to have made a pretty fundamental mistake that I then carried over to my post. I've updated the post accordingly.]
A new paper in Nature Neuroscience by Emmeke Aarts and colleagues argues that neuroscientists should start using hierarchical (or multilevel) models in their work in order to account for the nested structure of their data. From the abstract:
In neuroscience, experimental designs in which multiple observations are collected from a single research object (for example, multiple neurons from one animal) are common: 53% of 314 reviewed papers from five renowned journals included this type of data. These so-called ‘nested designs’ yield data that cannot be considered to be independent, and so violate the independency assumption of conventional statistical methods such as the t test. Ignoring this dependency results in a probability of incorrectly concluding that an effect is statistically significant that is far higher (up to 80%) than the nominal α level (usually set at 5%). We discuss the factors affecting the type I error rate and the statistical power in nested data, methods that accommodate dependency between observations and ways to determine the optimal study design when data are nested. Notably, optimization of experimental designs nearly always concerns collection of more truly independent observations, rather than more observations from one research object.
I don’t have any objection to the advocacy for hierarchical models; that much seems perfectly reasonable. If you have nested data, where each subject (or petrie dish or animal or whatever) provides multiple samples, it’s sensible to try to account for as many systematic sources of variance as you can. That point may have been made many times before, but it never hurts to make it again.
What I do find surprising though–and frankly, have a hard time believing–is the idea that 53% of neuroscience articles are at serious risk of Type I error inflation because they fail to account for nesting. This seems to me to be what the abstract implies, yet it’s a much stronger claim that doesn’t actually follow just from the observation that virtually no studies that have reported nested data have used hierarchical models for analysis. What it also requires is for all of those studies that use “conventional” (i.e., non-hierarchical) analyses to have actively ignored the nesting structure and treated repeated measurements as if they in fact came from entirely different subjects or clusters.
To make this concrete, suppose we have a dataset made up of 400 observations, consisting of 20 subjects who each provided 10 trials in 2 different experimental conditions (i.e., 20 x 2 x 10 = 400). And suppose the thing we ultimately want to know is whether or not there’s a statistical difference in outcome between the two conditions. There are three at least three ways we could set up our comparison:
- Ignore the grouping variable (i.e., subject) entirely, effectively giving us 200 observations in each condition. We then conduct the test as if we have 200 independent observations in each condition.
- Average the 10 trials in each condition within each subject first, then conduct the test on the subject means. In this case, we effectively have 20 observations in each condition (1 per subject).
- Explicitly include the effects of both subject and trial in our model. In this case we have 400 observations, but we’re explictly accounting for the correlation between trials within a given subject, so that the statistical comparison of conditions effectively has somewhere between 20 and 400 “observations” (or degrees of freedom).
Now, none of these approaches is strictly “wrong”, in that there could be specific situations in which any one of them would be called for. But as a general rule, the first approach is almost never appropriate. The reason is that we typically want to draw conclusions that generalize across the cases in the higher level of the hierarchy, and don’t have any intrinsic interest in the individual trials themselves. In the above example, we’re asking whether people on average, behave differently in the two conditions. If we treat our data as if we had 200 subjects in each condition, effectively concatenating trials across all subjects, we’re ignoring the fact that the responses acquired from each subject will tend to be correlated (i.e., Jane Doe’s behavior on Trial 2 will tend to be more similar to her own behavior on Trial 1 than to another subject’s behavior on Trial 1). So we’re pretending that we know something about 200 different individuals sampled at random from the population, when in fact we only know something about 20 different individuals. The upshot, if we use approach (1), is that
we do indeed run a high risk of producing false positives we’re going to end up answering a question quite different from the one we think we’re answering. [Update: Jake Westfall points out in the comments below that we won't necessarily inflate Type I error rate. Rather, the net effect of failing to model the nesting structure properly will depend on the relative amount of within-cluster vs. between-cluster variance. The answer we get will, however, usually deviate considerably from the answer we would get using approaches (2) or (3).]
By contrast, approaches (2) and (3) will, in most cases, produce pretty similar results. It’s true that the hierarchical approach is generally a more sensible thing to do, and will tend to provide a better estimate of the true population difference between the two conditions. However, it’s probably better to describe approach (2) as suboptimal, and not as wrong. So long as the subjects in our toy example above are in fact sampled at random, it’s pretty reasonable to assume that we have exactly 20 independent observations, and analyze our data accordingly. Our resulting estimates might not be quite as good as they could have been, but we’re unlikely to miss the mark by much.
To return to the Aarts et al paper, the key question is what exactly the authors mean when they say in their abstract that:
In neuroscience, experimental designs in which multiple observations are collected from a single research object (for example, multiple neurons from one animal) are common: 53% of 314 reviewed papers from five renowned journals included this type of data. These so-called ‘nested designs’ yield data that cannot be considered to be independent, and so violate the independency assumption of conventional statistical methods such as the t test. Ignoring this dependency results in a probability of incorrectly concluding that an effect is statistically significant that is far higher (up to 80%) than the nominal α level (usually set at 5%).
I’ve underlined the key phrases here. It seems to me that the implication the reader is supposed to draw from this is that roughly 53% of the neuroscience literature is at high risk of reporting spurious results. But in reality this depends entirely on whether the authors mean that 53% of studies are modeling trial-level data but ignoring the nesting structure (as in approach 1 above), or that 53% of studies in the literature aren’t using hierarchical models, even though they may be doing nothing terribly wrong otherwise (e.g., because they’re using approach (2) above).
Unfortunately, the rest of the manuscript doesn’t really clarify the matter. Here’s the section in which the authors report how they obtained that 53% number:
To assess the prevalence of nested data and the ensuing problem of inflated type I error rate in neuroscience, we scrutinized all molecular, cellular and developmental neuroscience research articles published in five renowned journals (Science, Nature, Cell, Nature Neuroscience and every month’s first issue of Neuron) in 2012 and the first six months of 2013. Unfortunately, precise evaluation of the prevalence of nesting in the literature is hampered by incomplete reporting: not all studies report whether multiple measurements were taken from each research object and, if so, how many. Still, at least 53% of the 314 examined articles clearly concerned nested data, of which 44% specifically reported the number of observations per cluster with a minimum of five observations per cluster (that is, for robust multilevel analysis a minimum of five observations per cluster is required11, 12). The median number of observations per cluster, as reported in literature, was 13 (Fig. 1a), yet conventional analysis methods were used in all of these reports.
This is, as far as I can see, still ambiguous. The only additional information provided here is that 44% of studies specifically reported the number of observations per cluster. Unfortunately this still doesn’t tell us whether the effective degrees of freedom used in the statistical tests in those papers included nested observations, or instead averaged over nested observations within each group or subject prior to analysis.
Lest this seem like a rather pedantic statistical point, I hasten to emphasize that a lot hangs on it. The potential implications for the neuroscience literature are very different under each of these two scenarios. If it is in fact true that 53% of studies are inappropriately using a “fixed-effects” model (approach 1)–which seems to me to be what the Aarts et al abstract implies–the upshot is that a good deal of neuroscience research is very bad statistical shape, and the authors will have done the community a great service by drawing attention to the problem. On the other hand, if the vast majority of the studies in that 53% are actually doing their analyses in a perfectly reasonable–if perhaps suboptimal–way, then the Aarts et al article seems rather alarmist. It would, of course, still be true that hierarchical models should be used more widely, but the cost of failing to switch would be much lower than seems to be implied.
I’ve emailed the corresponding author to ask for a clarification. I’ll update this post if I get a reply. In the meantime, I’m interested in others’ thoughts as to the likelihood that around half of the neuroscience literature involves inappropriate reporting of fixed-effects analyses. I guess personally I would be very surprised if this were the case, though it wouldn’t be unprecedented–e.g., I gather that in the early days of neuroimaging, the SPM analysis package used a fixed-effects model by default, resulting in quite a few publications reporting grossly inflated t/z/F statistics. But that was many years ago, and in the literatures I read regularly (in psychology and cognitive neuroscience), this problem rarely arises any more. A priori, I would have expected the same to be true in cellular and molecular neuroscience.