Jonathan Eisen already posted on this blog about a PLoS ONE paper by Mason et.. published on 23 October 2013. And he posted related comments on the PLoS ONE website. I also commented at this site, in reference to his comments and the authors' response. The purpose of my comments here are just to review those concerns and comment additionally on the PLos ONE response and what this means for the journal's publication model and the progress of science.
The paper by Mason and colleagues analyzes data on 48 people in each of 4 self-identified ethnic groups (African American, Caucasian, Chinese, and Latino). These study subjects are apparently volunteers, and the paper only states that they are non-smokers over 18 years old who are free of a list of diagnosed diseases and who have not recently had their teeth cleaned. Based on the text of the published paper, there is no consideration of their age, diet, social class, or even gender. The authors culture bacterial species from the study subjects and process the data through an algorithm that maximizes the prediction of racial group membership based on these measured data.
The prediction is moderately successful, but this could result from any number of unsurprising reasons. For example, if alcohol consumption affects some particular bacterial species, and whites drink more than Asians in central Ohio, then whatever species is diminished by alcohol exposure would help predict that a sample was from a white rather than from an Asian volunteer. And likewise for any of a million possible lifestyle, social class and demographic differences. In fact, this is a general problem with data mining exercises that Lazer et al describe in the current issue of Science.
The authors provide no information about how balanced this sample is with respect to any of these variables. Maybe the 48 Hispanics are younger than the 48 whites on average, or have more tooth decay or eat more refined sugar or any of a million other possibilities. The fact that these countless potentially imbalanced factors get represented in the oral bio-environment hardly seems surprising, and the fact that these behaviors and exposures might be differential by race is an observation that is completely trivial from a sociological perspective.
My concern here, however, the authors assert in the published text that these differences do not arise from any of these myriad environmental factors, but from some innate genetic characteristics of the groups. In the Discussion section on page 3 they state that "ethnicity exerts a selection pressure on the oral microbiome, and...this selection pressure is genetic rather than environmental, since the two ethnicities that shared a common food, nutritional and lifestyle heritage (Caucasians and African Americans) demonstrated significant microbial divergence." Here is a remarkable statement, that Caucasians and African Americans experience no differential dietary or lifestyle factors. It is directly contradicted by thousands of published papers in sociology, epidemiology and anthropology that document these differences for reasons of culture, geographic origin, social class and discrimination.
Jonathan Eisen's post directly confronted the authors on this point, and they responded with the following explanation:
"Subjects were selected based on extensive questionnaire surveys and clinical examinations to ensure homogeneity. These questionnaires evaluated educational level, socio-economic status, diet and nutritional history, systemic health status, oral hygiene habits and dental visits, among other things."
This is surely an important statement about the research design, but the problem is that it appears nowhere in the peer-reviewed text of the published paper. What exactly do the authors mean when they insist that the study subjects were perfectly balanced on factors such as socio-economic status and nutritional history? These complex social and lifestyle variables are notoriously difficult to define and measure. While the authors describe the laboratory techniques in baroque detail, they do not even mention in the published paper that they measured these factors, let alone how these variables were defined and considered in the analysis. This represents a profound limitation for the reader in assessing the validity of these measures and adjustments, and therefore the adequacy of the claimed "homogeneity". The complete omission of these crucial aspects of the analysis in the paper prevents the reader from investing much confidence in the boldly stated claim that observed differences are "genetic rather than environmental" in origin.
I expressed these concerns in my own post at the journal website on 14 November 2013, but the authors did not respond. Therefore, at Jonathan's suggestion, I addressed this concern to the PLoS ONE editors in an e-mail on 22 November 2013. I got passed along from one editor to another, and finally I got a very nice response from Elizabeth Silva on 4 December 2013. She wrote:
I wanted to let you know that I am discussing this article and your concerns with both the Academic Editor and the authors, as well as with my colleagues. We take such concerns very seriously and will ensure that appropriate measures are taken to correct any errors or discrepancies.
Then I waited. After 2 months I had heard nothing, so I wrote to Dr. Silva again asking for any word on progress, but received no reply. So I waited another month.
On 4 March it had been 3 months since the note from PLoS ONE promising appropriate measures to correct any errors or discrepancies, so I wrote again, this time a bit more insistently. This did message did finally generate a quick and reassuring response from Dr. Silva:
I really am very sorry for the extended delay in replying to you, and for my neglect in providing you with an update. Following your correspondence I contacted the authors to ask them for additional information relating to their statement that they corrected for confounding factors, and details of these methods. They promptly replied with a table of details of the baseline variables that they corrected for, and that they described in the comment on their article (see attached), as well as an additional correction to one of their figure legends. I then contacted the Academic Editor, Dr. Indranil Biswas with the full details of your concerns, as well as the table sent by the authors and the correction they requested for their figure legend. We asked Dr. Biswas to revisit the manuscript, in light of this new information, and he has informed us that he feels the conclusions of the manuscript are sound. We will now work with the authors to draft and issue a formal correction to the published article to update the methods to include the table, and to amend the figure legend in question.
The table that Dr. Silva forwarded displayed a list of variables and a p-value for some kind of test between the values in the 4 race groups. The test is not specified (t-test? chi-square test?) but presumably it is for any difference in means or proportions between the 4 groups. Most of the p-values are large, indicating little evidence for any difference between the groups in income, age, education, or frequency of tooth-brushing, etc. Based on this table, the populations differed only in their diets, which were characterized as "Asian Diet", "Hispanic Diet" and "American Diet". Not unsurprisingly, the Asians were significantly more likely to report an "Asian Diet" and the Hispanics were significantly more likely to report a "Hispanic Diet". The Blacks and Whites had similar reported consumption of the "American Diet", which presumably was the basis for the authors' assertion that these groups have identical social environments.
To date, there has been no correction made to the Mason et al paper at the PLoS ONE website. Therefore it is perhaps somewhat premature to speculate on how the authors will address the concern voiced by Jonathan Eisen's posted comment and blog post that balance across a handful of measured covariates does not in any way imply balance across all relevant factors except for genetics. Indeed, it has long been argued in the epidemiology literature that one cannot make indirect inferences about genes by measuring and adjusting for a few environmental exposures and attributing all remaining differences to genes. The argument that Blacks and Whites in Ohio experience identical environments is clearly false, even if a handful of measured covariates are not significantly different in their small convenience sample, the exact origin of which is still obscure.
There are many observations that can be made from this episode. I offer just a few:
- These authors are assiduous in describing their lab techniques, but regarding the study design and analysis they are quite cavalier. Presumably the reviewers were not population scientists, and so they failed to point out these embarrassing flaws. This raises the question of whether a multidisciplinary journal such as PLoS ONE has the relevant expertise to screen out scientifically invalid papers. The fact that Dr. Silva suggested that the authors' table of covariates and p-values solves the issue demonstrates a wide gap of understanding. Specialty journals that handle a narrow disciplinary range are not faced with this kind of crisis of competence.
- PLoS ONE is such a large operation with so many papers, that quality control seems to suffer. These beleaguered editors are responsible for an enormous publishing volume. Has quantity overwhelmed quality to the extent that gross errors of logic slip through? Months later, the Mason paper has been accessed thousands of times and generated a great deal of media attention, and yet no correction or erratum has appeared, despite the fact that the authors freely admit that the methods in their published paper are not accurate.
- The publishing model gives PLoS ONE a big incentive (almost $2000) to accept a paper, but once it is published, little incentive to correct or withdraw it.
Sadly, this is not an isolated example. This week, PLoS ONE published a paper by Wikoff et al which makes a similar logical gaffe about observed racial difference proving a genetic difference. We could post a comment online, but it seems that nobody (neither the authors nor the editors) has much time to spend monitoring such comments, nor much incentive to care about them. The authors have their publication, the journal has its $2000, and another tiny piece of horrific misinformation has been released into the world. The basic philosophy of PLoS ONE is to reduce the gate-keeper role of scientific publication. I am starting to become convinced that a little gate-keeping is not such a bad idea.