Re: Re: Arsenic carcinogenicity: numbers

[Guest guest]

On Mon, Mar 12, 2007 at 09:46:59PM -0000, wrote:

>

>> You'll notice that those numbers have the risk being proportional to dose

>> (within rounding error). I'll bet that if you look up the actual papers,

>> you'd find that this is an assumption, rather than an experimentally

>> proven fact, and that what they are really doing is extrapolating from

>> much higher doses (and much higher cancer rates), via the linear

>> no-threshold model of risk. That model doesn't make any sense -- it's

>> tantamount to assuming that the body has no detoxification or repair

>> mechanisms whatsoever -- but it's the usual model that's used in these

>> situations.

>

>Well, the full text NAS report is here. It has a looong table of

>borrowed data, several pages long. This URL will land you right in the

>middle of the table:

>

>http://books.nap.edu/openbook.php?record_id=10194 & page=42

>

>crane your neck there and take a look at the top line, Chiou et al

>2001. For two different tumors, there appear to be raw incidences taken

>from Chiou for four different ranges of As exposure. At least that's

>how I'm decoding it.

My decoding is the same. But notice that it's really the data for

patients with exposures >100 ug/L that make that page of the table

significant. Below that concentration, the number of cases is so small

that a single person not getting cancer would change the relative risk to

be something very close to one. And " greater than 100 " is not very

informative: how much greater? One might think that it was not very

much; but on the next page there's another study by Chiou, where the

high-arsenic range is " >0.71 mg/L " (710 ug/L). So " >100 " might mean a

lot greater than 100. In comparison, the report's risk estimates are for

levels of 20 ug/L and below.

>Also, I saw one assertion that there are 7 more or less plausibly

>significant avenues for the As carcinogenesis. (So on balance, the

>thing is a big tangle.) But one possible mech is inhibition of the DNA

>repair enzymes. Such inhibition might actually produce a linear no-

>threshhold assumptions, since even in the absence of any exogenous

>mutagen, DNA already undergoes significant decay at a constant rate,

>which has to be fixed continually.

>

>I'm really kind of stretching here, though. Surely there is somewhat of

>a surefeit/buffer of DNA repair activity at baseline, and furthermore I

>think that if and when DNA lesions of any kind do start building up,

>the SOS response will be triggered and DNA repair will intensify - a

>further buffer. Those facts certainly wouldn't be expected to

>correspond to a no-threshhold curve. But, assuming that Arsenic-

>sponsored oxidative damage and other mechs also contribute, it might

>happen to all come out linear-ish when you add it all up. (Not that I

>expect you to find that kind of talk very satisfying... but it could

>still be true!)

Oxidative damage would almost certainly have a threshold; and you can't

get a no-threshold curve by adding together lots of curves each of which

has a threshold. Even if there isn't any threshold per se, just about

any kind of repair mechanism works better at low levels of damage than at

high levels.

If you look earlier in the report, they explicitly confirm that they're

doing as I suspected, and extrapolating high-exposure data to lower

exposures via the linear no-threshold assumption, which they state to be

" partly a policy decision " -- as opposed, presumably, to a scientific

decision. They state the linear model to be the " default " model, and

require a reason for deviating from it, rather than requiring a reason

for adopting it in the first place.

I really wouldn't have much argument with this if they stated explicitly

that they're trying to err on the side of safety. But they don't: they

state that they are making " a maximum likelihood estimate " , not an " upper

bound " estimate. Erring on the side of safety can be very good policy.

With so many different dangers, and so much uncertainty in assessing

them, it's reasonable to want to keep each of them a considerable

distance away. But this doesn't excuse the act of pretending that one is

estimating the exact danger when one really isn't.

By the way, in terms of potential damage mechanisms, it's conceivable

even that there are mechanisms for which the damage increases at low

levels -- even in absolute terms. Indeed, some animal experiments have

suggested that arsenic is an essential trace nutrient.

[Guest guest]

On Tue, Mar 13, 2007 at 09:44:54PM -0000, wrote:

>But they are claiming the dose-dependency *trend* is significant at

>95% confidence. I lack the stats to really dig into that. I suppose

>it means something like, if you picked the best line for 4 random

>points on that x-y rectangle, the sum of squares of the points'

>differences from their best line would be larger than the

>corresponding value seen here, 95% of the time - ?? I pretty much

>made that up.

I don't think it's that. I think what they mean by the trend being

significant is that if you took four random points, and picked the

best-fit line through them, then the slope of that line would be less

than the slope found, 95% of the time.

> But assuming my guess is pretty close, it doesn't seem

>like a terribly fair way to decide between threshhold and no-

>threshhold models (not that Chiou asserted so), unless you use

>something more like 6 or 8 points. We know the null hypothesis (no

>correlation at all) is false at 95% confidence, but quite a bit of

>that confidence is deriving from the <10 and >100 points, which would

>behave the same whether there is any threshhold (any sub-100

>threshhold) or not.

>

>Which is the same thing you already said, if I grasp your meaning. I

>guess if you want to exclude existence of a threshhold above X, the

>only truly clear way to do that is to simply show that the cancer

>incidence at X is significantly different from that at zero - ??

Yes, that's my meaning.

With only 15 total cases of cancer for which they had data on arsenic

consumption, they really didn't have enough data for analyses like that.

Even the >100 point is only barely statistically significant.

It'd have been nice if they had simply listed the data points ( " patient 1

had exposure of X ug/L for Y years, patient 2 had...). That way we could

know what the " > " in " >100 " is hiding. But I don't see such a list

anywhere in the paper.

>> Oxidative damage would almost certainly have a threshold; and you can't

>> get a no-threshold curve by adding together lots of curves each of

>> which has a threshold. Even if there isn't any threshold per se, just

>> about any kind of repair mechanism works better at low levels of damage

>> than at high levels.

>

>I concede about ox damage. But say As inhibits enzyme XYZ that's

>somehow involved in anti-mutagenesis (maybe it functions in the

>apoptotic response to major DNA damage). And say that inhibition

>levels off around 30 ug/L. Add that to a threshhold curve and the

>result might look pretty linear. (Except in the region below the Kd

>affinity value between As and X, where inhibition of XYZ is a concave

>function of [As], but you might well have no data point in that

>region).

For there not to be any increased inhibition at levels greater than 30

ug/L, enzyme XYZ would need to be almost completely knocked out at that

dose. Otherwise, increasing the arsenic concentration further would

knock out more of it. Yet at the same time this completely knocked-out

enzyme could only be contributing a tiny fraction of the carcinogenic

effect at higher doses. Yes, you could suspect that sort of mechanism...

but you could also suspect it even if you didn't find any effect at

higher doses. While not at all impossible, this sort of effect is beyond

the reach of public health studies. One would have to find it by

studying mechanisms, not by epidemiology.

>> If you look earlier in the report, they explicitly confirm that they're

>> doing as I suspected, and extrapolating high-exposure data to lower

>> exposures via the linear no-threshold assumption

>

>What page do they say that on? I am surprised. Especially since the

>*press release* makes apparant flat statements about 5 ug/L being

>different from zero: " at 5 parts per billion, the risk is [...]. " I

>don't think too many reporters would go on to read the full report!

>The press release also acknowledges and discusses other contestible

>methods used - like comparison to incidences taken from a different

>study - making you think you're getting the full story....... no

>mention of a no-threshhold assumption. Potentially misleading, I can

>personally attest.

Page 6:

" Biostatistical approaches are required in a dose-response

assessment to extrapolate from the lowest concentration of

arsenic at which cancers are observed in a study population to

lower concentrations to which the study population of interest is

exposed. The mode of action by which a chemical causes cancer

can sometimes determine how ... data should be extrapolated ...

In the absence of definitive mode-of-action data, EPA's general

policy is to use a linear extrapolation ... After concluding that

the mode-of-action data were inadequate to define the shape of

the curve, EPA made a policy-based decision to use a default

assumption of linearity. "

The last five lines of that are just talking about the EPA, though; and

the people writing this are not the EPA. They state on their own behalf

(same page) that they ( " the subcommittee " )

" ... concludes that the available mode-of-action data on arsenic

do not provide a biological basis for using either a linear or

nonlinear extrapolation. "

Then, after discussion of possible mechanisms, they state on page 11:

" As discussed previously, there are no experimental data to

indicate the concentration at which any theoretical threshold

might exist. Therefore, the curve should be extrapolated

linearly from the ED_01 to determine risk estimates for the

potential concentrations of concern (3, 5, 10, and 20 ug/L). The

choice for the shape of the dose-response curve below the ED_01

is, in part, a policy decision. "

The reason that it is only 'in part' a policy decision is that:

" ... the Taiwanese and other human studies include data on

exposures at arsenic concetrations relatively close to some US

exposures. Consequently, the extrapolation is over only a

relatively small range of arsenic concentrations. "

That last sentence is quite true, by the standards of the business: a lot

of the projections one hears (e.g. cancer deaths from Chernobyl) rely on

much larger extrapolations (like by a factor of a thousand or so). This

extrapolation is more like a factor of ten (maybe a hundred at most), so

it is indeed " relatively small " .

--

Norman Yarvin http://yarchive.net