Don’t Double Dip Data

Meta-analyses have become commonplace in epidemiology and in other sciences. When well conducted and transparently reported, meta-analyses can be extremely helpful. In several litigations, meta-analyses determined the outcome of the medical causation issues. In the silicone gel breast implant litigation, after defense expert witnesses proffered meta-analyses[1], court-appointed expert witnesses adopted the approach and featured meta-analyses in their reports to the MDL court[2].

In the welding fume litigation, plaintiffs’ expert witness offered a crude, non-quantified, “vote counting” exercise to argue that welding causes Parkinson’s disease[3]. In rebuttal, one of the defense expert witnesses offered a quantitative meta-analysis, which provided strong evidence against plaintiffs’ claim.[4] Although the welding fume MDL court excluded the defense expert’s meta-analysis from the pre-trial Rule 702 hearing as untimely, plaintiffs’ counsel soon thereafter initiated settlement discussions of the entire set of MDL cases. Subsequently, the defense expert witness, with his professional colleagues, published an expanded version of the meta-analysis.[5]

And last month, a meta-analysis proffered by a defense expert witness helped dispatch a long-festering litigation in New Jersey’s multi-county isotretinoin (Accutane) litigation. In re Accutane Litig., No. 271(MCL), 2015 WL 753674 (N.J. Super., Law Div., Atlantic Cty., Feb. 20, 2015) (excluding plaintiffs’ expert witness David Madigan).

Of course, when a meta-analysis is done improperly, the resulting analysis may be worse than none at all. Some methodological flaws involve arcane statistical concepts and procedures, and may be easily missed. Other flaws are flagrant and call for a gatekeeping bucket brigade.

When a merchant puts his hand the scale at the check-out counter, we call that fraud. When George Costanza double dipped his chip twice in the chip dip, he was properly called out for his boorish and unsanitary practice. When a statistician or epidemiologist produces a meta-analysis that double counts crucial data to inflate a summary estimate of association, or to create spurious precision in the estimate, we don’t need to crack open Modern Epidemiology or the Reference Manual on Scientific Evidence to know that something fishy has taken place.

In litigation involving claims that selective serotonin reuptake inhibitors cause birth defects, plaintiffs’ expert witness, a perinatal epidemiologist, relied upon two published meta-analyses[6]. In an examination before trial, this epidemiologist was confronted with the double counting (and other data entry errors) in the relied-upon meta-analyses, and she readily agreed that the meta-analyses were improperly done and that she had to abandon her reliance upon them.[7] The result of the expert witness’s deposition epiphany, however, was that she no longer had the illusory benefit of an aggregation of data, with an outcome supporting her opinion. The further consequence was that her opinion succumbed to a Rule 702 challenge. See In re Zoloft (Sertraline Hydrochloride) Prods. Liab. Litig., MDL No. 2342; 12-md-2342, 2014 U.S. Dist. LEXIS 87592; 2014 WL 2921648 (E.D. Pa. June 27, 2014) (Rufe, J.).

Double counting of studies, or subgroups within studies, is a flaw that most careful readers can identify in a meta-analysis, without advance training. According to statistician Stephen Senn, double counting of evidence is a serious problem in published meta-analytical studies. Stephen J. Senn, “Overstating the evidence – double counting in meta-analysis and related problems,” 9, at *1 BMC Medical Research Methodology 10 (2009). Senn observes that he had little difficulty in finding examples of meta-analyses gone wrong, including meta-analyses with double counting of studies or data, in some of the leading clinical medical journals. Id. Senn urges analysts to “[b]e vigilant about double counting,” id. at *4, and recommends that journals should withdraw meta-analyses promptly when mistakes are found,” id. at *1.

Similar advice abounds in books and journals[8]. Professor Sander Greenland addresses the issue in his chapter on meta-analysis in Modern Epidemiology:

Conducting a Sound and Credible Meta-Analysis

Like any scientific study, an ideal meta-analysis would follow an explicit protocol that is fully replicable by others. This ideal can be hard to attain, but meeting certain conditions can enhance soundness (validity) and credibility (believability). Among these conditions we include the following:

  • A clearly defined set of research questions to address.

  • An explicit and detailed working protocol.

  • A replicable literature-search strategy.

  • Explicit study inclusion and exclusion criteria, with a rationale for each.

  • Nonoverlap of included studies (use of separate subjects in different included studies), or use of statistical methods that account for overlap. * * * * *”

Sander Greenland & Keith O’Rourke, “Meta-Analysis – Chapter 33,” in Kenneth J. Rothman, Sander Greenland, Timothy L. Lash, Modern Epidemiology 652, 655 (3d ed. 2008) (emphasis added).

Just remember George Costanza; don’t double dip that chip, and don’t double dip in the data.

[1] See, e.g., Otto Wong, “A Critical Assessment of the Relationship between Silicone Breast Implants and Connective Tissue Diseases,” 23 Regulatory Toxicol. & Pharmacol. 74 (1996).

[2] See Barbara Hulka, Betty Diamond, Nancy Kerkvliet & Peter Tugwell, “Silicone Breast Implants in Relation to Connective Tissue Diseases and Immunologic Dysfunction:  A Report by a National Science Panel to the Hon. Sam Pointer Jr., MDL 926 (Nov. 30, 1998)”; Barbara Hulka, Nancy Kerkvliet & Peter Tugwell, “Experience of a Scientific Panel Formed to Advise the Federal Judiciary on Silicone Breast Implants,” 342 New Engl. J. Med. 812 (2000).

[3] Deposition of Dr. Juan Sanchez-Ramos, Street v. Lincoln Elec. Co., Case No. 1:06-cv-17026, 2011 WL 6008514 (N.D. Ohio May 17, 2011).

[4] Deposition of Dr. James Mortimer, Street v. Lincoln Elec. Co., Case No. 1:06-cv-17026, 2011 WL 6008054 (N.D. Ohio June 29, 2011).

[5] James Mortimer, Amy Borenstein & Laurene Nelson, Associations of Welding and Manganese Exposure with Parkinson’s Disease: Review and Meta-Analysis, 79 Neurology 1174 (2012).

[6] Shekoufeh Nikfar, Roja Rahimi, Narjes Hendoiee, and Mohammad Abdollahi, “Increasing the risk of spontaneous abortion and major malformations in newborns following use of serotonin reuptake inhibitors during pregnancy: A systematic review and updated meta-analysis,” 20 DARU J. Pharm. Sci. 75 (2012); Roja Rahimi, Shekoufeh Nikfara, Mohammad Abdollahic, “Pregnancy outcomes following exposure to serotonin reuptake inhibitors: a meta-analysis of clinical trials,” 22 Reproductive Toxicol. 571 (2006).

[7] “Q So the question was: Have you read it carefully and do you understand everything that was done in the Nikfar meta-analysis?

A Yes, I think so.

* * *

Q And Nikfar stated that she included studies, correct, in the cardiac malformation meta-analysis?

A That’s what she says.

* * *

Q So if you look at the STATA output, the demonstrative, the — the forest plot, the second study is Kornum 2010. Do you see that?

A Am I —

Q You’re looking at figure four, the cardiac malformations.

A Okay.

Q And Kornum 2010, —

A Yes.

Q — that’s a study you relied upon.

A Mm-hmm.

Q Is that right?

A Yes.

Q And it’s on this forest plot, along with its odds ratio and confidence interval, correct?

A Yeah.

Q And if you look at the last study on the forest plot, it’s the same study, Kornum 2010, same odds ratio and same confidence interval, true?

A You’re right.

Q And to paraphrase My Cousin Vinny, no self-respecting epidemiologist would do a meta-analysis by including the same study twice, correct?

A Well, that was an error. Yeah, you’re right.


Q Instead of putting 2 out of 98, they extracted the data and put 9 out of 28.

A Yeah. You’re right.

Q So there’s a numerical transposition that generated a 25-fold increased risk; is that right?

A You’re correct.

Q And, again, to quote My Cousin Vinny, this is no way to do a meta-analysis, is it?

A You’re right.”

Testimony of Anick Bérard, Kuykendall v. Forest Labs, at 223:14-17; 238:17-20; 239:11-240:10; 245:5-12 (Cole County, Missouri; Nov. 15, 2013). According to a Google Scholar search, the Rahimi 2005 meta-analysis had been cited 90 times; the Nikfar 2012 meta-analysis, 11 times, as recently as this month. See, e.g., Etienne Weisskopf, Celine J. Fischer, Myriam Bickle Graz, Mathilde Morisod Harari, Jean-Francois Tolsa, Olivier Claris, Yvan Vial, Chin B. Eap, Chantal Csajka & Alice Panchaud, “Risk-benefit balance assessment of SSRI antidepressant use during pregnancy and lactation based on best available evidence,” 14 Expert Op. Drug Safety 413 (2015); Kimberly A. Yonkers, Katherine A. Blackwell & Ariadna Forray, “Antidepressant Use in Pregnant and Postpartum Women,” 10 Ann. Rev. Clin. Psychol. 369 (2014); Abbie D. Leino & Vicki L. Ellingrod, “SSRIs in pregnancy: What should you tell your depressed patient?” 12 Current Psychiatry 41 (2013).

[8] Julian Higgins & Sally Green, eds., Cochrane Handbook for Systematic Reviews of Interventions 152 (2008) (“7.2.2 Identifying multiple reports from the same study. Duplicate publication can introduce substantial biases if studies are inadvertently included more than once in a meta-analysis (Tramèr 1997). Duplicate publication can take various forms, ranging from identical manuscripts to reports describing different numbers of participants and different outcomes (von Elm 2004). It can be difficult to detect duplicate publication, and some ‘detectivework’ by the reviewauthors may be required.”); see also id. at 298 (Table 10.1.a “Definitions of some types of reporting biases”); id. at 304-05 ( Duplicate (multiple) publication bias … “The inclusion of duplicated data may therefore lead to overestimation of intervention effects.”); Julian P.T. Higgins, Peter W. Lane, Betsy Anagnostelis, Judith Anzures-Cabrera, Nigel F. Baker, Joseph C. Cappelleri, Scott Haughie, Sally Hollis, Steff C. Lewis, Patrick Moneuse & Anne Whitehead, “A tool to assess the quality of a meta-analysis,” 4 Research Synthesis Methods 351, 363 (2013) (“A common error is to double-count individuals in a meta-analysis.”); Alessandro Liberati, Douglas G. Altman, Jennifer Tetzlaff, Cynthia Mulrow, Peter C. Gøtzsche, John P.A. Ioannidis, Mike Clarke, Devereaux, Jos Kleijnen, and David Moher, “The PRISMA Statement for Reporting Systematic Reviews and Meta-Analyses of Studies That Evaluate Health Care Interventions: Explanation and Elaboration,” 151 Ann. Intern. Med. W-65, W-75 (2009) (“Some studies are published more than once. Duplicate publications may be difficult to ascertain, and their inclusion may introduce bias. We advise authors to describe any steps they used to avoid double counting and piece together data from multiple reports of the same study (e.g., juxtaposing author names, treatment comparisons, sample sizes, or outcomes).”) (internal citations omitted); Erik von Elm, Greta Poglia; Bernhard Walder, and Martin R. Tramèr, “Different patterns of duplicate publication: an analysis of articles used in systematic reviews,” 291 J. Am. Med. Ass’n 974 (2004); John Andy Wood, “Methodology for Dealing With Duplicate Study Effects in a Meta-Analysis,” 11 Organizational Research Methods 79, 79 (2008) (“Dependent studies, duplicate study effects, nonindependent studies, and even covert duplicate publications are all terms that have been used to describe a threat to the validity of the meta-analytic process.”) (internal citations omitted); Martin R. Tramèr, D. John M. Reynolds, R. Andrew Moore, Henry J. McQuay, “Impact of covert duplicate publication on meta­analysis: a case study,” 315 Brit. Med. J. 635 (1997); Beverley J Shea, Jeremy M Grimshaw, George A. Wells, Maarten Boers, Neil Andersson, Candyce Hamel, Ashley C. Porter, Peter Tugwell, David Moher, and Lex M. Bouter, “Development of AMSTAR: a measurement tool to assess the methodological quality of systematic reviews,” 7(10) BMC Medical Research Methodology 2007 (systematic reviews must inquire whether there was “duplicate study selection and data extraction”).

Print Friendly, PDF & Email

Comments are closed.