December 19, 2005
On modeling the human response time function; Part 3.
Much to my surprise, this morning I awoke to find several emails in my inbox apparently related to my commentary on the Barabasi paper in Nature. This morning, Anders Johansen pointed out to myself and Luis Amaral (I can only assume that he has already communicated this to Barabasi) that in 2004 he published an article entitled Probing human response times in Physica A about the very same topic using the very same data as that of Barabasi's paper. In it, he displays the now familiar heavy-tailed distribution of response times and fits a power law of the form P(t) ~ 1/(t+c) where c is a constant estimated from the data. Asymptotically, this is the same as Barabasi's P(t) ~ 1/t; it differs in the lower tail, i.e., for t < c where it scales more uniformly. As an originating mechanism, he suggests something related to a spin-glass model of human dynamics.
Although Johansen's paper raises other issues, which I'll discuss briefly in a moment, let's step back and think about this controversy from a scientific perspective. There are two slightly different approaches to modeling that are being employed to understand the response-time function of human behavior. The first is a purely "fit-the-data" approach, which is largely what Johansen has done, and certainly what Amaral's group has done. The other, employed by Barabasi, uses enough data analysis to extract some interesting features, posits a mechanism for the origin of those and then sets about connecting the two. The advantage of developing such a mechanistic explanation is that (if done properly) it provides falsifiable hypotheses and can move the discussion past simple data-analysis techniques. The trouble begins, as I've mentioned before, when either a possible mechanistic model is declared to be "correct" before being properly vetted, or when an insufficient amount of data analysis is done before positing a mechanism. This latter kind of trouble allows for a debate over how much support the data really provides to the proposed mechanism, and is exactly the source of the exchange between Barabasi et al. and Stouffer et al.
I tend to agree with the idea implicitly put forward by Stouffer et al.'s comment that Barabasi should have done more thorough data analysis before publishing, or alternatively, been a little more cautious in his claims of the universality of his mechanism. In light of Johansen's paper and Johansen's statement that he and Barabasi spoke at the talk in 2003 where Johansen presented his results, there is now the specter that either previous work was not cited that should have been, or something more egregious happened. While not to say that this aspect of the story isn't an important issue in itself, it is a separate one from the issues regarding the modeling, and it is those with which I am primarily concerned. But, given the high profile of articles published in journals like Nature, this kind of gross error in attribution does little to reassure me that such journals are not aggravating certain systemic problems in the scientific publication system. This will probably be a topic of a later post, if I ever get around to it. But let's get back to the modeling questions.
Seeking to be more physics and less statistics, the ultimate goal of such a study of human behavior should be to understand the mechanism at play, and at least Barabasi did put forward and analyze a plausible suggestion there, even if a) he may not have done enough data analysis to properly support it or his claims of universality, and b) his model assumes some reasonably unrealistic behavior on the part of humans. Indeed, the former is my chief complaint about his paper, and why I am grateful for the Stouffer et al. comment and the ensuing discussion. With regard to the latter, my preference would have been for Barabasi to have discussed the fragility of his model with respect to the particular assumptions he describes. That is, although he assumes it, humans probably don't assign priorities to their tasks with anything like a uniformly random distribution and nor do humans always execute their highest priority task next. For instance, can you decide, right now without thinking, what the most important email in your inbox is at this moment? Instead, he commits the crime of hubris and neglects these details in favor of the suggestiveness of his model given the data. On the other hand, regardless of their implausibility, both of these assumptions about human behavior can be tested through experiments with real people and through numerical simulation. That is, these assumptions become predictions about the world that, if they fail to agree with experiment, would falsify the model. This seems to me an advantage of Barabasi's mechanism over that proposed by Johansen, which, by relying on a spin glass model of human behavior, seems quite trickier to falsify.
But let's get back to the topic of the data analysis and the argument between Stouffer et al. and Barabasi et al. (now also Johansen) over whether the data better supports a log-normal or a power-law distribution. The importance of this point is that if the log-normal is the better fit, then the mathematical model Barabasi proposes cannot be the originating mechanism. From my experience with distributions with heavy tails, it can be difficult to statistically (let alone visually) distinguish between a log-normal and various kinds of power laws. In human systems, there is almost never enough data (read: orders of magnitude) to distinguish these without using standard (but sophisticated) statistical tools. This is because for any finite sample of data from an asymptotic distribution, there will be deviations that will blur the functional form just enough to look rather like the other. For instance, if you look closely at the data of Barabasi or Johansen, there are deviations from the power-law distribution in the far upper tail. Stouffer et al. cite these as examples of the poor fit of the power law and as evidence supporting the log-normal. Unfortunately, they could simply be due to deviations due to finite-sample effects (not to be confused with finite-size effects), and the only way to determine if they could have been is to try resampling the hypothesized distribution and measuring the sample deviation against the observed one.
The approach that I tend to favor for resolving this kind of question combines a goodness-of-fit test with a statistical power test to distinguish between alternative models. It's a bit more labor-intensive than the Bayesian model selection employed by Stouffer et al., but this approach offers, in addition to others that I'll describe momentarily, the advantage of being able to say that, given the data, neither model is good or that both models are good.
Using Monte Carlo simulation and something like the Kolmogorov-Smirnov goodness-of-fit test, you can quantitatively gauge how likely a random sample drawn from your hypothesized function F (which can be derived using maximum likelihood parameter estimation or by something like a least-squares fit; it doesn't matter) will have a deviation from F at least as big as the one observed in the data. By then comparing the deviations with an alternative function G (e.g., a power law versus a log-normal), you get a measure of the power of F over G as an originating model of the data. For heavy-tailed distributions, particularly those with a sample-mean that converges slowly or never at all (as is the case for something like P(t) ~ 1/t), sampling deviations can cause pretty significant problems with model selection, and I suspect that the Bayesian model selection approach is sensitive to these. On the other hand, by incorporating sampling variation into the model selection process itself, one can get an idea of whether it is even possible to select one model over another. If someone were to use this approach to analyze the data of human response times, I suspect that the pure power law would be a poor fit (the data looks too curved for that), but that the power law suggested in Johansen's paper would be largely statistically indistinguishable from a log-normal. With this knowledge in hand, one is then free to posit mechanisms that generate either distribution and then proceed to validate the theory by testing its predictions (e.g., its assumptions).
So, in the end, we may not have gained much in arguing about which heavy-tailed distribution the data likely came from, and instead should consider whether or not an equally plausible mechanism for generating the response-time data could be derived from the standard mechanisms for producing log-normal distributions. If we had such an alternative mechanism, then we could devise some experiments to distinguish between them and perhaps actually settle this question like scientists.
As a closing thought, my interest in this debate is not particularly in its politics. Rather, I think this story suggests some excellent questions about the practice of modeling, the questions a good modeler should ponder on the road to truth, and some of the pot holes strewn about the field of complex systems. It also, unfortunately, provides some anecdotal evidence of some systemic problems with attribution, the scientific publishing industry and the current state of peer-review at high-profile, fast turn-around-time journals.
References for those interested in reading the source material.
A. Johansen, "Probing human response times." Physica A 338 (2004) 286-291.
A.-L. Barabasi, "The origin of bursts and heavy tails in human dynamics." Nature 435 (2005) 207-211.
D. B. Stouffer, R. D. Malmgren and L. A. N. Amaral "Comment on 'The origin of bursts and heavy tails in human dynamics'." e-print (2005).
J.-P. Eckmann, E. Moses and D. Sergi, "Entropy of dialogues creates coherent structures in e-mail traffic." PNAS USA 101 (2004) 14333-14337.
A.-L. Barabasi, K.-I. Goh, A. Vazquez, "Reply to Comment on 'The origin of bursts and heavy tails in human dynamics'." e-print (2005).