Saturday, April 4, 2015

The Certainty of Nothing

The 1986 EPA Cancer Risk Assessment Guidelines

Having been written on the heels of the influential 1983 NRC report, the EPA (1986) guidelines closely follow their recommendations.  Compared to later revisions, it is very brief.  However, it did confront the issue of model uncertainty:

Different extrapolation models, however, may fit the observed data reasonably well but may lead to large differences in the projected risk at low doses.

As the solution to the problem, the 1986 guidelines offered this:

The Agency will review each assessment as to the evidence on carcinogenesis mechanisms and other biological or statistical evidence that indicates the suitability of a particular extrapolation model. Goodness-of-fit to the experimental observations is not an effective means of discriminating among models (OSTP, 1985). A rationale will be included to justify the use of the chosen model. In the absence of adequate information to the contrary, the linearized multistage procedure will be employed. Where appropriate, the results of using various extrapolation models maybe useful for comparison with the linearized multistage procedure.

This short paragraph makes three important statements.  First, it recognizes that goodness-if-fit (i.e statistical evidence) and biological evidence (i.e. a biological rationale) are both important in selecting a suitable model for extrapolation.  Second, it introduces the linearized multistage model as the default option that will be used unless there is “adequate information to the contrary”.  Third, it is suggested that, at least under some circumstances, it may be desirable to make risk estimates using alternative models as well.  The second and third statements are further elaborated in the next paragraph:

It should be emphasized that the linearized multistage procedure leads to a plausible upper limit to the risk that is consistent with some proposed mechanisms of carcinogenesis. Such an estimate, however, does not necessarily give a realistic prediction of the risk. The true value of the risk is unknown, and maybe as low as zero. The range of risks, defined by the upper limit given by the chosen model and the lower limit which may be as low as zero, should be explicitly stated. An established procedure does not yet exist for making “most likely" or "best" estimates of risk within the range of uncertainty defined by the upper and lower limit estimates. If data and procedures become available, the Agency will also provide "most likely" or "best" estimates of risk. This will be most feasible when human data are available and when exposures are in the dose range of the data.

As the plausible worst case option, the justification for the default option is the same as the de minimus application used by the USFDA.  It is also acknowledged that lower estimates are also plausible.  In particular, it is suggested that the risk may be as low as zero.  It is also acknowledges that, even though there is no established procedure for doing so, providing a best estimate would also be desirable.

Adequate Evidence to the Contrary

The 1986 guidelines did not attempt to define what evidence would be considered sufficient to overturn the default option.  In the following decade, many chemical industry studies were sponsored to provide the necessary evidence, usually by showing, or attempting to show, that a nongenotoxic mechanism is responsible for the development of tumors at high doses.  But, as it turned out, no scientific argument fraught with uncertainty was ever found to be sufficient to overturn the bright and shiny default option that exhibited no uncertainty whatsoever.

Another NRC (1994) committee that took up this issue, a gave a discussion of the matter in one of the appendices.  The EPA (1999) guidelines summarized this discussion:

Appendix N of the report contains two presentations of alternative views held by some committee members on this issue.  One view, known as "plausible conservatism," suggested that departures from defaults should not be made unless new information improves the understanding of a biological process to the point that relevant experts reach consensus that the protective default assumption concerning that process is no longer plausible.  The same criterion was recommended where the underlying scientific mechanism is well understood, but where a default is used to address missing data.  In this case, the default should not be replaced with case-specific data unless it is the consensus of relevant experts that the proffered data make the default assumption no longer plausible.  Another view, known as the "maximum use of scientific information" approach, acknowledged that the initial choice of defaults should be protective, but argued that conservatism should not be a factor in determining whether to depart from the default in favor of an alternate biological theory or alternate data.  According to this view, it should not be necessary to reach expert consensus that the default assumption had been rendered implausible; it should be sufficient that risk assessors find the alternate approach more plausible than the default. 

The thing is, both these “views” are reasonable for some applications.  If the regulatory statute is intended to be conservative (e.g. the Delaney clause or a premarket approval process), then using the default as long as it remains plausible is sensible.  On the other hand, justifying a regulation with a cost-benefit analysis with a less-likely default option makes no sense at all; money will be spent to avoid health outcomes that probably won’t happen.   

The 1996 Proposed and 1999 EPA Interim Guidelines for Carcinogen Risk Assessment

Although these guidelines were never finalized, they exhibit the evolution that took place over the next 10 years.  With regard to model uncertainty, the key development was the division of the dose-response relationship into two zones that are separated by a “Point of Departure”.  The high-dose zone is defined by “the empirical data in the range of observation”.  As another “default option” the point of departure is to be defined by the LED10, which corresponds to the lower confidence limit of the estimated dose where 10% of the population exhibits the effect.  When used for effects other than cancer the Effective Dose (ED) is also known as the benchmark dose (BMD), which is roughly equivalent to the No Observed Adverse Effect Level.  The use of the LED was expressly motivated by a desire to harmonize methodologies used for cancer and non-cancer endpoints; “the LED10 can be regarded as an improved and harmonized estimate of the NOAEL”.  However, genotoxic carcinogens, the guidelines suggest that linear extrapolation from the POD is appropriate:

The use of straight line extrapolation for a linear default is a change from the 1986 guidelines which used the "linearized multistage" (LMS) procedure.  This change is made because the former modeling procedure gave an appearance of specific knowledge and sophistication unwarranted for a default.

This is the shell game, of course.  The linear model is no longer justified as a plausible worst-case scenario.  It is simply justified by agency policy.  It also seems rather clear that guideline authors wanted to restrict the choices presented to agency managers, who will by and large fail to notice that the scientific rug has been pulled from underneath the risk estimates.  When considering the possibility of characterizing model uncertainty associated with using different plausible models to estimate:

Discussion of the confidence in the extrapolation is appropriately done qualitatively or by showing results for alternatives that are equally plausible. It is not useful, for example, to conduct quantitative uncertainty analysis running multiple forms of linear models. This would obviate the function of the policy default.

Why, oh why, would obviating the function of the policy default be a bad idea?  As God and the 1983 NRC report intended, agency policy can always be applied AFTER the risk assessment is completed.   But, if you are playing the shell game, it isn’t.  This time, the shell game victims are the scientists who are trying to provide a theoretical basis for estimating risks.  Since inductive reasoning can’t be reliably controlled, keeping control of the decision process dictates that it be excluded.

The 90’s guidelines do promote the use of biological models or “mode-of-action” analyses that may influence the way the assessment proceeds.  If it can be shown that a carcinogen is nongenotoxic, then instead of a fake risk estimate, there will be no risk estimate.  
  

The 2005 Cancer Risk Assessment Guidelines

While the Weight of the Evidence discussion is much improved, there are no substantial procedural differences between the 2005 guidelines and the 1999 Interim guidelines with regard to dose-response modeling.  However, the shell game was ratcheted up another gear, presumably to put cancer risk assessment on the same footing as the ersatz noncancer variety.  Most notably, the procedure for “nonlinear extrapolation” is the same methodology used for the reference dose, which involves no extrapolation whatsoever.  The most vigorous shell game is played in the section on the “POD narrative”.  The problem is this:
As a single-point summary of a single dose-response curve, the POD alone does not convey all the critical information present in the data from which it is derived.  To convey a measure of uncertainty, the POD should be presented as a central estimate with upper and lower bounds. A POD narrative summarizes other important features of the database and the POD that are important to account for in low-dose extrapolations or other analyses.

It is not at all clear who the narrative is to be directed to.  Is the risk assessor writing the narrative for a risk manager who stands upon the POD gazing out across the vast, or not so vast, expanse of the Margin of Exposure towards the region where the actual exposures occur?  Or, is it some secondary risk assessor who is expected to pick up the pieces of this mess?  The most ironic directive is this one:

(d) Slope of the dose-response curve at the POD. How does response change as dose is reduced below the POD?  A steep slope indicates that risk decreases rapidly as dose decreases. On the other hand, a steep slope also indicates that errors in an exposure assessment can lead to large errors in estimating risk.  Both aspects of the slope are important.  The slope also indicates whether dose-response curves for different effects are likely to cross below the POD.

This narrative seems to instruct the reader to ignore the POD.  What a great idea; maybe we should actually do a risk assessment that, you know, actually provides risk estimates.

Really, the 1986 guidelines were pretty good.  It needed some tweaking and appendix N of the 1994 NRC report provided a pretty good start for doing just that.  But instead, we got the science-policy shell game designed to make the default option the only option.  Where the choice was between uncertain theory or nothing, we got the latter.  

References

USEPA (1986).  Guidelines for Carcinogen Risk Assessment.  EPA/630/R-00/004

USEPA (1996).  Proposed Guidelines for Carcinogen Risk Assessment.  EPA/600/P-
92/003C

USEPA (1999).  Guidelines for Carcinogen Risk Assessment, Review Draft, CEAF-0644
Office of Research and Development.

USEPA (2005).  Guidelines for Carcinogen Risk Assessment.  EPA/630/P-03/001F.

Official Post Soundtrack

Fixx, The (1998).  The Fatal Shore.  In: Elemental, Track 6.


Post Notes

Thesis post #26.  Second in the shell game series.  A longer than normal essay, but I think it all needs to be there.

Soundtrack: Not Australia; the fatal shore is the POD.

No comments:

Post a Comment