Monday, January 18, 2016

That editorial in Cladistics

On the weekend, a colleague made me aware of the recent editorial in Cladistics clarifying that all papers in the journal need to use parsimony methods, and she indicated that it had already produced a lot of harsh criticism. Today I read the editorial and several responses including the ones on the Twitter hash tag #parsimonygate. At the moment I am somewhat in two minds about the affair.

To summarise the Cladistics editorial:
  • Authors should always show parsimony trees. Alternative phylogenies can be shown in addition if they are different and there is "a pertinent reason" to do so. This is not really new, as when I published a paper in Cladistics a few years ago I was already asked to do the same. In other journals I have seen the reviewers or editors demand likelihood analyses. I am a methods pragmatist and can live with either. If the data are strong, the results will be the same anyway.
  • The editors consider parsimony analysis to be preferable to other methods for "philosophical" reasons. As far as I can tell, most biologists would be more impressed by biological and pragmatic reasons; I know I am. They also stress repeatability, intelligibility and clarity, implying somewhat that model based methods don't have those. This I find odd except perhaps in the case of certain Bayesian analyses suffering from an over-abundance of unknowable priors.
  • The editors point out that all methods have their weaknesses, and that we have to live with not knowing the truth for certain. Seems obvious.
  • The editors think that they are doing something right because the journal has a high impact factor. Well... let's just say that I have very mixed feelings about impact factors.
To summarise the criticism, let's go through #parsimonygate as of today 10:49 Sydney time:
  • 8x Equating the Cladistics editors to religious fundamentalists or calling them otherwise dogmatic. 
  • 3x Mocking Cladistics as too traditional and/or stuck in the 1980ies. 
  • 1x Some analogy to physics that I don't quite understand. 
  • 3x Ridicule without argument. 
  • 1x Call for boycott. 
  • 1x A different journal advertises itself as accepting all phylogenetic approaches. 
  • 5x Meta, as in expressing amusement at somebody else's comment.
Elsewhere I have seen another call for boycott and somebody mocking the sentence about how we cannot know The Truth with complete certainty, strangely equating this stance with creationism. I think that is exactly backwards, because it is creationists who crave certainty, and it is good scientists who accept that conclusions are tentative and open to future falsification. But okay, it is at least clear the main arguments are that parsimony is outdated and that accepting only papers using a certain approach is dogmatic.

As for parsimony being outdated, that belief is just the flip-side of what the editors are being criticised for. Dogmatic rejection of any and all parsimony analysis is not actually any more rational than dogmatic rejection of everything but parsimony. The belief that parsimony shouldn't be used any more is also generally based on a very selective perception of the literature, e.g. the mistaken assumption that likelihood analyses do not suffer from long branch attraction. These are all tools, they all have their advantages and disadvantages, and they all have their use cases. Often it is quite simply a good idea to use two different ones to see if the results agree, or the one that you would otherwise consider to be the best is computationally too demanding for your dataset.

So now to the second criticism. If several methods all have their use cases, isn't it then dogmatic to reject papers that don't use parsimony, merely on that basis? Well, in principle I really see no problem with a journal deciding to serve any given niche, including a methodological one. We do, after all, also have journals like Molecular Phylogenetics and Evolution, and I do not see any of the people tweeting at #parsimonygate calling its editors anti-science because they would presumably reject all morphology-only manuscripts. They just decided that they were going to serve the molecular niche. It's in the title!

Similarly, I have a hunch that few, if any, of the outraged tweeters would see a problem if somebody founded a new journal called Bayesian Phylogenetics that accepted only studies using the eponymous methodology. They would probably say that there are still plenty of other phylogenetics journals, and this is just serving a niche. And again, they'd be correct!

Maybe I am mistaken - maybe such a Bayesian-only journal would draw the same denunciations from these tweeters - but given the fact that we have lots of other focused journals that nobody complains about and given the proud use of additional hash tags such as "Bayesian4Life" I suspect that a lot of the criticism really feeds at least partly on the feeling that parsimony analyses shouldn't be allowed at all at this stage.

In other words, at least some of the criticism seems hypocritical (dogmatic opponents of parsimony calling its proponents dogmatic) or hyperbolic. The charges of dogmatism, ignoring evidence, and quasi-religiosity in particular seem to confuse a preference for a method with a preference for wrong results.

But yes, that editorial could have easily been written to be less of a target. It doesn't just say that while there are several alternative, equally valid tools all with their own advantages and disadvantages the journal is going to break a lance for parsimony, although it does say that too; but it also at least implies that all other methods are inferior. It is unsurprising that this would raise the hackles of those who are convinced of the exact opposite. Where it could have referred to the active development of new parsimony methods even in the last few years (e.g. Most Parsimonious Reconciliations, Doyon et al. 2010) to showcase their continued relevance, its youngest reference is from 1983. It was obvious that this would be mocked as backwards.

But even if somebody is really convinced that Bayes is the only way to go, calling for a boycott seems ... a bit ... of an overreaction. Should we all boycott the Botanical Journal of the Linnean Society since they decided to stop accepting purely alpha-taxonomic papers? Or American Journal of Botany after it stopped accepting primer notes? Or that journal where I saw a reviewer demand likelihood inference of ancestral character states although one glance at the tree made clear that the results would be the same anyway? Can of worms.

15 comments:

  1. I think for me, the most troubling aspect is that they argue for parsimony being more preferable based on philosophical grounds, and imply that they aren't open to counter-arguments, especially statistical grounds. That doesn't sound like... science.

    That said, although I fully admit I did post a gif of Hennig's face on Immortal Joe's body, I agree with Brian O'Meara who said that this is a good thing, in that by clarifying this policy, as it will perhaps save the time of potential authors and editors. Authors know what they need to do prior to submission, or choose to submit elsewhere.

    By the way, 'unknown priors' affects every analysis. The point of Bayesian phylogenetics is to make those decisions explicit. Just because you can't toggle that switch in an ML analysis, like the prior on state transition frequencies (http://sysbio.oxfordjournals.org/content/early/2015/12/22/sysbio.syv122.abstract) doesn't mean it wouldn't impact the analysis. Forcing us to be explicit about our choices, I think, is a very good thing.

    Me, I've got a paper right using parsimony, as well as Mk in Bayesian phylogenetics (for brachiopod morphology). The former ends up being just a side-car, because the latter allows me to test things about our assumptions about heterogeneity of rates across characters, etc, and how that impacts our inferences of topology. And so, now thanks to this editorial, I know not to publish in Cladistics, because I don't really foresee any need to show a parsimony tree that is just slightly different from one of several Bayesian trees that tell a more interesting story.

    ReplyDelete
    Replies
    1. That is as far as I can tell the real reason why so few people use parsimony as their preferred method today: all the things you can only do if you have a time-calibrated tree, or only if you have a posterior distribution to evaluate.

      Delete
  2. I laughed so much when I saw this. In 2016 !

    In the Internet era, I can't understand why a journal would decide to serve only a restricted methodological niche.

    But the most problematic is the claim that parsimony is always superior to other methods. Through it is perfectly understandable for historical reasons, since this very journal is an organ of the Willi Hennig Society which was founded in 1980 after the 1979 Harvard split. The superiority of parsimony was one of its a founding dogma and those who were developping statistical methods remained outside. See Felsenstein (2001) "The troubled growth of statistical phylogenetics".

    ReplyDelete
    Replies
    1. I understand your mirth, but please note that most of the critics of the editorial would probably think the same of 'evolutionary' systematics as they do of parsimony analysis, except it would be about being stuck in the 1950s instead of the 1980s.

      Delete
    2. The problem is, that when talking about systematics, we talk about historical science. We search for explanations in the past for our observations in the presence. This kind of reasoning is called abduction and works different from statistical reasoning, which is induction. I can recommend Works bei Charles Sanders Peirce or if you think, that something older than 10 years cannot be good, works by Kirk Fitzhugh. The latter did a good job explaining the philosophical flaws of ML and Bayesian inference when applied for abductive reasoning.
      While Bayesian inference is simply a misinterpretation of Bayes' theorem, ML is a statistical method, which works totally fine in statistical induction. These methods have been applied "just to see what happens" and people were satisfied with the results. Unfortunately, they have not clarified the theoretical foundations (in fact, there is only a weak justification for algorithms with miraculously derived prior propabilities...).
      To put it in a nutshell, speaking of philosophy of science (which somehow is ignored by most biologists) parsimony has a good point. When using other methods of reasoning and saying that these are superior to parsimony, because they yield better results (how do you know??? support values are meaningless numbers as they are derived from the data you used for inference itself...) this is just as dogmatic as you calim the oppisite to be...

      Delete
    3. I have to disagree with all of this. It is parsimony which is philosophically flawed since it rely on an implicit model. Explicit models are testable because they aren't hidden assumptions. Furthermore bayesian posteriors aren't "support values" but probabilities. The most probable tree is not always the most parsimonious tree, but it is always the most "parsimonious" hypothesis since you would need an ad hoc assumption to prefer a tree not supported by evidences. I also disagree that the choice between these methods rely only on philosophical grounds, simulation studies exist.

      Delete
    4. I never said, Bayesian posteriors are support values. But things such as bootstrap and so on are meaningless numbers which lack logical legitimation as "test" evidence.
      You cannot use the data, that you have used for the inference itself as support evidence as well. This is because your hypothesis ought to explain your observations (i.e. your character data), that is circular reasoning. It (simplified) is like tossing a coin ten times with 8x tails and 2x heads and you thus assume, that your propability for tails is 80%...
      We don't know propabilities in evolution. There are statistical frequencies for certain substitutions to happen more often than others, right. But this cannot be a general rule as selection does not work on the sequence level. It is the phenotype in its environment, that is affected by selection. So these models cannot tell anything about establishment of certain mutations as they are selection-driven.
      It is just generally interesting, that e.g. in physics, reasoning has to follow rules of philosophy of science, while in biological systematics, noone cares whether he does his reasoning on deduction, induction or abduction. And thus the logical prerequisites of the different types of reasoning are simply ignored...

      Delete
    5. "It is like tossing a coin ten times with 8x tails and 2x heads and you thus assume, that your propability for tails is 80%..."

      But without any prior knowledge of the probability of tails, this is the best estimate. If we assume an informative prior distribution on this probability (a reasonable assumption since we have probably tossed many coins in our lives), then this small amount of data will have little impact on our estimate. Is this abductive reasoning according to you?

      I fail to see how phylogenetic inference is any different. We assume that evolution proceeds by a stochastic process and we wish to infer the parameters of that process. Instead of coin tosses we have related DNA sequences as our data, and instead of a single parameter determining the probability of heads/tails, we have many parameters that specify the behavior of the evolutionary process.

      Yes, the particular model we assume is necessarily false. Statistical inference always acknowledges that our estimates are only an approximation of the truth. The same is true of maximum parsimony - it is equivalent to maximum likelihood when we assume rates of change are infinitesimally small, or when we assume no-common-mechanism. Both assumptions are only approximations of the truth.

      Delete
    6. Sorry the previous comment is from me.

      Delete
  3. There is a categorical difference between the Cladistics editorial and journals who will only accept papers in a certain niche. These editors aren't saying that they'll take papers about only a certain topic, but rather that within a topic you MUST use a particular method, otherwise they will not accept it. That's a sort of authoritarian methods control that differs from just not accepting primer notes or not accepting alpha-taxonomic papers.

    ReplyDelete
    Replies
    1. Maybe, but I am not so sure that there is a categorical difference between rejecting papers that don't use parsimony analysis and rejecting papers that don't use molecular data, for example. And there are plenty people out there who reject papers if they don't use likelihood, which for some reason is considered to be acceptable behaviour. Perhaps because they don't make that policy explicit in an editorial?

      But again, I use all of them depending on what seems best or computationally feasible in each case. If I were an editor I would not exclude any approach a priori.

      Delete
    2. Yeah, that seems reasonable to me, too - really, it's any a priori rejection of any valid method that seems antiscientific. And perhaps, then, Cladistics deserves credit for being transparent!

      Delete
  4. The editorial in Cladistics specifies parsimony as primary method, and requests special attention for justification of methods when using other approaches than parsimony. It is explicitly mentioned that authors are welcome to present results obtained with such additional methods, provided that their use is properly argued for. I wouldn't call that a priori rejection.

    ReplyDelete
  5. I think the crux of the issue is not that different methods have their strengths and weaknesses and that parsimony has philosophical justification. Parsimony is 'identical' as a tree selection procedure to ML using a no common mechanism model (each character gets its own set of edgelengths) and a Jukes+Cantor subustitution process. This ML model is clearly over-parameterized (2T-3 parameters per site where T is the number of taxa) and suffers from inconsistency as a result of the infinitely many parameters problem (that the number of parameters to be estimated grows with the number of sites). As a statistical procedure it is nearly absurd as it breaks all the common sense rules of statistical modeling and fails to have statistical consistency (SC) with a correctly specified model. One may wish to discount SC as being very important...but surely a method that has SC should be preferred to one that doesn't have SC all other things being equal. I think the philosophical justifications for parsimony are far from accepted by philosophers of science. Simulation studies show that ML and Bayesian methods do seem to be efficient and perform well, often under conditions where the model is misspecified. This should be the reason why we use them and not parsimony. Long branch attraction is a real problem and, while it can occur in ML and Bayesian analyses with small sample sizes (under all conditions) OR large sample sizes under model misspecification, we can at least try to improve our models to avoid it.

    ReplyDelete
    Replies
    1. I must say I am always a bit sceptical of arguments that everything boils down to one's preferred approach, as with those Bayesians who claim that all reasoning is Bayesian anyway.

      That being said, for phylogenetic analyses of DNA sequence data I personally would prefer likelihood anyway.

      That now being said, however, there are also other applications of parsimony beyond simple gene tree inference, for example TCS haplotype networks, morphological data, MRP supertrees, and gene tree parsimony. In particular likelihood reconstruction of ancestral (morphological) character states has yet to convince me due to its often odd results, and while I am aware of the danger of GTP being mislead it does on the other hand seem more robust than ML/Bayesian/distance methods when faced with missing data.

      Delete