P-values

.

The following is from Nathan Schachtman’s legal blog, with various comments and added emphases (by me, in this color). He will try to reply to comments/queries.

“Courts Can and Must Acknowledge Multiple Comparisons in Statistical Analyses”

Nathan Schachtman, Esq., PC * October 14th, 2014

In excluding the proffered testimony of Dr. Anick Bérard, a Canadian perinatal epidemiologist in the Université de Montréal, the Zoloft MDL trial court discussed several methodological shortcomings and failures, including Bérard’s reliance upon claims of statistical significance from studies that conducted dozens and hundreds of multiple comparisons.[i] The Zoloft MDL court was not the first court to recognize the problem of over-interpreting the putative statistical significance of results that were one among many statistical tests in a single study. The court was, however, among a fairly small group of judges who have shown the needed statistical acumen in looking beyond the reported p-value or confidence interval to the actual methods used in a study[1].

.

A complete and fair evaluation of the evidence in situations as occurred in the Zoloft birth defects epidemiology required more than the presentation of the size of the random error, or the width of the 95 percent confidence interval.  When the sample estimate arises from a study with multiple testing, presenting the sample estimate with the confidence interval, or p-value, can be highly misleading if the p-value is used for hypothesis testing.  The fact of multiple testing will inflate the false-positive error rate. Dr. Bérard ignored the context of the studies she relied upon. What was noteworthy is that Bérard encountered a federal judge who adhered to the assigned task of evaluating methodology and its relationship with conclusions.

*   *   *   *   *   *   *

There is no unique solution to the problem of multiple comparisons. Some researchers use Bonferroni or other quantitative adjustments to p-values or confidence intervals, whereas others reject adjustments in favor of qualitative assessments of the data in the full context of the study and its methods. See, e.g., Kenneth J. Rothman, “No Adjustments Are Needed For Multiple Comparisons,” 1 Epidemiology 43 (1990) (arguing that adjustments mechanize and trivialize the problem of interpreting multiple comparisons). Two things are clear from Professor Rothman’s analysis. First for someone intent upon strict statistical significance testing, the presence of multiple comparisons means that the rejection of the null hypothesis cannot be done without further consideration of the nature and extent of both the disclosed and undisclosed statistical testing. Rothman, of course, has inveighed against strict significance testing under any circumstance, but the multiple testing would only compound the problem.

Second, although failure to adjust p-values or intervals quantitatively may be acceptable, failure to acknowledge the multiple testing is poor statistical practice. The practice is, alas, too prevalent for anyone to say that ignoring multiple testing is fraudulent, and the Zoloft MDL court certainly did not condemn Dr. Bérard as a fraudfeasor[2]. [emphasis mine]

I’m perplexed by this mixture of stances. If you don’t mention the multiple testing for which it is acceptable not to adjust, then you’re guilty of poor statistical practice; but its “too prevalent for anyone to say that ignoring multiple testing is fraudulent”. This appears to claim it’s poor statistical practice if you fail to mention your results are due to multiple testing, but “ignoring multiple testing” (which could mean failing to adjust or, more likely, failing to mention it) is not fraudulent. Perhaps, it’s a questionable research practice QRP. It’s back to “50 shades of grey between QRPs and fraud.”

  […read his full blogpost here]

Previous cases have also acknowledged the multiple testing problem. In litigation claims for compensation for brain tumors for cell phone use, plaintiffs’ expert witness relied upon subgroup analysis, which added to the number of tests conducted within the epidemiologic study at issue. Newman v. Motorola, Inc., 218 F. Supp. 2d 769, 779 (D. Md. 2002), aff’d, 78 Fed. App’x 292 (4th Cir. 2003). The trial court explained:

“[Plaintiff’s expert] puts overdue emphasis on the positive findings for isolated subgroups of tumors. As Dr. Stampfer explained, it is not good scientific methodology to highlight certain elevated subgroups as significant findings without having earlier enunciated a hypothesis to look for or explain particular patterns, such as dose-response effect. In addition, when there is a high number of subgroup comparisons, at least some will show a statistical significance by chance alone.”

I’m going to require, as part of its meaning, that a statistically significant difference not be one due to “chance variability” alone. Then to avoid self contradiction, this last sentence might be put as follows: “when there is a high number of subgroup comparisons, at least some will show purported or nominal or unaudited statistical significance by chance alone. [Which term do readers prefer?] If one hunts down one’s hypothesized comparison in the data, then the actual p-value will not equal, and will generally be greater than, the nominal or unaudited p-value.”

So, I will insert “nominal” where needed below (in red).

Texas Sharpshooter fallacy

Id. And shortly after the Supreme Court decided Daubert, the Tenth Circuit faced the reality of data dredging in litigation, and its effect on the meaning of “significance”:

“Even if the elevated levels of lung cancer for men had been [nominally] statistically significant a court might well take account of the statistical “Texas Sharpshooter” fallacy in which a person shoots bullets at the side of a barn, then, after the fact, finds a cluster of holes and draws a circle around it to show how accurate his aim was. With eight kinds of cancer for each sex there would be sixteen potential categories here around which to “draw a circle” to show a [nominally] statistically significant level of cancer. With independent variables one would expect one statistically significant reading in every twenty categories at a 95% confidence level purely by random chance.”

The Texas sharpshooter fallacy is one of my all time favorites. One purports to be testing the accuracy of his aim, when in fact that is not the process that gave rise to the impressive-looking (nominal) cluster of hits. The results do not warrant inferences about his ability to accurately hit a target, since that hasn’t been well-probed. Continue reading →