Home » Columns, Featured, Science Policy

Reproducible Research

1 January 2011 9,516 views 11 Comments

This month’s guest editors, Keith A. Baggerly and Donald A. Berry, make the case that journals have a key role to play in making research reproducible. Their call comes in the aftermath of attempts to reproduce the cancer research results of Duke’s Anil Potti and Joseph Nevins, whose seemingly promising 2006 work led to three clinical trials. Baggerly and colleague Kevin R. Coombes were the lead figures to uncover not only an inability to reproduce the research, but many obstacles in attempting to do so. In November, Duke terminated all three trials and Anil Potti resigned.
~ Steve Pierson, ASA Director of Science Policy, pierson@amstat.org

Contributing Editors

BaggerlyKeith A. Baggerly is professor of bioinformatics and computational biology at The University of Texas MD Anderson Cancer Center in Houston, Texas.

BarryDonald A. Berry is head of the division of quantitative sciences and chair and professor of the department of biostatistics at The University of Texas MD Anderson Cancer Center.

Research is reproducible if it can be reproduced by others. Of course, rerunning an experiment will give different results—an observation that gave rise to the development of statistics as a discipline. Our focus here is “reproducible research” (RR) in the sense of reproducing conclusions from a single experiment based on the measurements from that experiment.

In the 1990s, the geophysicist Jon Claerbout became frustrated with new students having great difficulty duplicating previous students’ research. Making further advances meant spending months, or even years, trying to reproduce previous advances. Reproducing computations can be agonizing, even with one’s own calculations. The problem is exacerbated by the passage of time. We suspect all applied statisticians have had such an experience.

Our definition of RR includes a complete description of the data and the analysis of that data—including computer programs—so the results can be exactly reproduced by others. RR is self-contained. This is obviously important in Claerbout’s motivating context. But it is as or more important in letting others judge the appropriateness of the analysis.

As data sets become more voluminous and complex, RR is increasingly important because our intuition fails in high dimensions. Investigators know how expression levels of some genes predict patient outcomes, but nobody understands the predictive efficacy of “signatures” involving hundreds of genes. Researchers may think they know, but we have (inadvertently) performed the null experiment of giving an investigator a list of random genes—purportedly prognostic—and having the investigator explain why they made sense. To use complex signatures, we must trust that the steps taken to produce them are accurate, or at least checkable.

Much scientific research cannot be reproduced, which can have severe consequences. We recently tried to reproduce reports that drug sensitivity “signatures” could predict patient response to cancer chemotherapy. Since the reported methods were incomplete, we employed “forensic bioinformatics” to reconstruct the methods used to derive the reported results. These reconstructions identified several basic errors, including drastic mistakes such as reversing “sensitive” and “resistant” labels. These errors initially went unnoticed because it was impossible to easily check the steps.

We published our reconstructions in September of 2009. Clinical trials in lung cancer at Duke University had begun two years before. Lack of clear reproducibility hampered corrective action. Trials were suspended in October of 2009 while Duke investigated our claims. Trials were restarted in January of 2010, although Duke didn’t release the investigation’s report or the data justifying the restart. A redacted report became available in May under the Freedom of Information Act. It indicated that important problems of reproducibility were not adequately addressed. In the interim, we publicized new problems with the analysis.

In July, it was revealed that a principal investigator had allegedly made false claims (including a Rhodes scholarship) on his CV and in grant applications. Several statisticians wrote to Harold Varmus, director of the National Cancer Institute, to request the trials be suspended again, and they were. Journals that had published the investigators’ results began independent investigations. In October of 2010, the senior author requested retraction of a paper on which the trials were based, citing problems we reported to Duke in November of 2009.

Journals must begin to demand RR. Convincing them will not be easy. Authors have legitimate concerns, including data ownership, intellectual property, and identifying where data can be accessibly stored. Despite these concerns, there is growing recognition that RR is important. Spurred by the forensic episode cited above, the Institute of Medicine announced a review of omics-based tests in clinical trials in October of 2010. Sessions on RR are more common at scientific and editorial board meetings. Journal editors are aware that irreproducibility can damage credibility and are open to suggestions regarding improvements.

Simple enforcement of rules many journals already have (e.g., that data be posted) at even a cursory level would help. The journal Biostatistics has an associate editor for reproducibility who can assign grades of merit to conditionally accepted papers: D: data are available, C: code is available, and R: the AE could run the code and reproduce the results without much effort.

This last constraint requires authors to make their code user friendly, but we see no alternative. Few reviewers have the time or expertise to assess reproducibility, though they can check whether data and code are provided. In a September Nature letter, we and others described information journals should elicit from authors. This includes data (raw, processed, and clinical, with indications of provenance), code, descriptions of nonscriptable steps, and prospectively defined analyses (in a protocol, for example), if any.

Motivated by the clear need, we have tried to make our work reproducible. We found the process less cumbersome than we feared. In bioinformatics and computational biology at MD Anderson Cancer Center, we now require that reports be prepared using Sweave. This software uses R and Latex. On running Sweave files through R, code is evaluated and results formatted for proper inclusion. Given the report, if we have access to the same raw data, we can run the report and confirm that we obtain the same values. (See example reports.) Our adoption of Sweave involved a few months of learning, but we now produce better reports faster because we have reusable templates and we are in the habit of planning for reproducibility from the outset. Alternatives to Sweave, including GenePattern, are available.

Finally, RR is necessary, but not sufficient, for good science. It needn’t contain the motivation for what was done, and the motivation may be data-dependent. For example, suppose we compare counts following two treatments, but the results are not statistically significant. So we take logarithms, after adding 0.01 to every count. Then, we tell the “complete” story, starting from taking logarithms. Or perhaps the data we used were “cleaned” before we got them. These potentially fatal biases will not be known by someone checking reproducibility, and they may not be known to the primary analyst. Difficulties with these and other “silent multiplicities” are described in a 2007 piece in Pharmaceutical Statistics written by Donald A. Berry.

We exhort statisticians to join in making RR more widespread. Click here for further reading.

1 Star2 Stars3 Stars4 Stars5 Stars (2 votes, average: 5.00 out of 5)


  • Russ Lenth said:

    Wouldn’t this more aptly be called reproducible ANALYSIS (RA)? Reproducible research would involve replicating the study.

    Both RA and RR are really important. See, most recently, the article “The Truth Wears Off: An odd twist in the scientific method” by Jonah Lehrer, _New_Yorker_, Dec 13, 2010.

  • Brad Huitema said:

    Lenth is exactly on target. Reproducible research is not demonstrated by a perfectly reproduced analysis. Jargon such as RR is sure to be misunderstood. I second RA.

  • Thomas Cox PhD RN said:

    Two situations come to mind.

    A biochemist friend of mine was asked, as part of his PhD work to develop a very volatile compound that had never been successfully synthesized though it had been worked on by a series of doctoral students over several years.

    After working on it for several months he actually succeeded and it disappeared by the time he turned around to jot down some notes. He was, over the next few weeks, able to stabilize the compound. But suppose that he had not been successful for several years. The compound was so volatile that many researchers may have failed to reproduce his results but they “could” be reproduced. While laudable, RR will certainly increase the lag time between cutting edge research and publication.

    Second situation: The obviously incorrect analysis and likely fabricated data that went into the April 1, 1998 JAMA article on Therapeutic Touch, supposedly crafted by a 9 year old co-author would have been immediately obvious to readers had the author’s data and analysis been provided along with the article. The data reported in the paper and on a website run by one of the authors actually differed on the report of the number of success in the same number of trials.

    So, on balance, some laudable research will be delayed but some really poor research will be less likely to slip past sleeping reviewers, or at least, if it does, the time it takes to bring it to light will be shorter.

  • Keith Baggerly said:

    Russ & Brad, I understand (and actually agree with) your point. It’s certainly true that my own efforts have focused more on the “reproducible analysis” end of things, in that I’m not regenerating raw data. However, there’s already a literature that has built up around “reproducible research”, so in this case we went with existing terminology.

  • Steve Pierson said:

    , Nature Magazine Blog.

  • Paul Thompson said:

    I don’t agree that RA is the correct term. I’ll stick with RR. That’s because in bioinformatics research and analysis are linked in a tight manner. In the article that the discuss in the piece, the issue lies in the manner in which a specific research project was unreproducible. In that case, the issue lay in the manner in which the analysis of a specific sample had been compromised by sloppiness and incompetence. The research project was inextricably linked to the analysis. See Section 1.4 of “Deriving chemosensitivity from cell lines…” by Baggerly and Coombes. I have a pre-print, so cannot give a more accurate citation.

    In other areas, RR is also correct. In biology, there is a huge emphasis on the keeping of accurate lab notebooks, and there are cases in which lab notebooks have been examined to determine if changes have been made in them. Thus, again, RR (reproducible research) is not different from RA (reproducible analysis).

  • Paul Thompson said:

    And a final thought: RR should be the term, since RA limits the concern to “the analysis part”. RR points out that “analysis” is an inextricable part of “research”, which those of us who perform statistical evaluations of the work of others would certainly hope is the case. RR additionally emphasizes the notion that both the physical manipulations involved in a research project (the culturing of cells, the collection of responses from persons, the observations of stars passing a fixed point of reference) and the analysis component of that information are inextricably linked together, and must both be done accurately and reproducibly for the project to have the ultimate value, that it be able to be done to the similar outcome by a different investigator.

  • Eric Sampson said:

    I encourage everyone to read “Statistical Analyses and Reproducible Research” by Gentleman and Temple Lang. It’s available for free download at


  • Juan Zuluaga said:

    As a lowly graduate student, I see this as extremely important. Sweave has helped me greatly to keep organized, and when working in a team it is absolutely key that everyone is on the same page about what was done to the data. I wish our institution would support the needed computer infrastructure.