Apples, Arsenic, and Risk - Part 10: Type 2 diabetes - 1.26 based on what?

In my last post I ended with the question of how much confidence should we have in this Odd Ratio: 1.26 (1.02-1.56)?

The reason for our confidence in it, how solid it is, how much it actually reflects the population it measures, is because it is used to generate the conclusion:

This finding supports the hypothesis that low levels of exposure to inorganic arsenic in drinking water, a widespread exposure worldwide, may play a role in diabetes prevalence.

And with that conclusion, Consumer Reports was able to tell its readers that there is a link between:

Low-level arsenic exposure with the prevalence of type 2 diabetes in the United States.

Although it may seem like I am beating a dead horse with all of this, I mean c'mon 10 posts?!?, what I am trying to do is take a complicated subject and break it down so one can separate fact from spin.

Step one is being skeptical, even if the work is performed by a respectable organization like Consumer Reports with help from a Johns Hopkins University’s Bloomberg School of Public Health physician - epidemiologist.

Step two is reading the work that is used to support the claim.  Consumer Reports and their scientists claim that low levels of arsenic is harmful.  We know from their 84 samples analyzed for inorganic arsenic the mean level of inorganic arsenic is 4 ug/L or 4 ppb.  We know that the EPA has established an MCL for arsenic in drinking water of 10 ppb based on a 70 kg person consuming 2 liters per day for a lifetime as a "safe" level.  Is 4 ppb arsenic in the apple juice a concern?  Is it a concern for contributing to type 2 diabetes?

Step three is looking at the data.  Looking at it closely to see how it was derived and concurring with the author or disregarding the findings as unsupported.

Here's the thing I want the reader to know.  If her work supported her conclusion I would still write about it and show why.  It doesn't, which is why it takes this direction.  If it did, I would go in that direction. I've got no dog in this hunt as well.

So let's look at how those ORs we derived.  Starting with Table 2:

Source

Here is how they got their "n":

• They randomly selected a one-third random sample of NHANES 2003-2004 study participants aged 6 years and older (n=2673)
• They then selected 1027 participants from those 2673 who had fasted 8 to 24 hours before venipuncture.
• They then excluded 38 pregnant women, 34 participants missing total urine arsenic or urine arsenic species, 24 participants without prior diagnosis of diabetes missing serum glucose, 4 participants missing glycated hemoglobin, 
• They then restricted their analyses to participants who did not report seafood intake in the past 24 hours.

That last bullet was done for this reason:

Our primary assessment of exposure was total urine arsenic concentration adjusted for objective biomarkers of seafood consumption.

So far, so good.  This gave them a total population of 788 (n = 788).

Of that 788, 95 had type 2 diabetes and 695 did not.  All of the OR and percentages they now report will be based on that population.  They are comparing one group to another - those with diabetes to those without.

The hypothesis here is that you will see a higher urinary arsenic level for those with diabetes compared to those without.

Doh!  That's not what they saw when the compared the mean total arsenic in the diabetes group with that of the non-diabetes.  Seems the non-diabetes group had more 7.4 ug/L compared to 6.2.

Now there could be a logical reason for this based on the makeup of the two groups being compared.  We want to compare a similar group with a like similar group.  Instead we are dealing with a random group we pulled from a larger group.  Enter the "Model."

Model 1:

First, we adjusted for sex, age, race and ethnicity (known determinants of diabetes that may be related to arsenic exposure), and urine creatinine.

Okay, even with that adjustment there is still less total arsenic in the diabetes group compared to the non-diabetes group.  0.94 OR, but the range contains a "1" so we can conclude nothing there, although it is consistent with what we see with the raw data.

Model 2:

Second, each model was further adjusted for education, body mass index, serum cotinine, and use of antihypertensive medication.  This model represents the association of total arsenic exposure with diabetes independent of the source but adjusted for traditional diabetes risk factors.

Okay, now there is a shift, based on this model, 1.01, but again, the range contains a "1" so this model statistically shows no association.

Model 3:

Third, each model was further adjusted for urine arsenobetaine and blood mercury levels. This model provides estimates for the association of inorganic arsenic not derived from seafood and for arsenobetaine.

So after adjusting for 1 & 2, they further adjust the population by excluding those who were found to have arsenobetaine and blood mercury levels.  The assumption here is that these folks ate fish and their total arsenic is tainted by organic arsenic from the fish.

Aha!  Now we get a positive association, the 1.26 with a range of 1.02-1.56.  Woot!  Now they can claim a positive association!  There ain't a "1" in that range is there!

Now before we continue, you might be wondering about other risk factors for type 2 diabetes:

Further adjustment for smoking status and alcohol intake, as well as exclusion of participants showing levels below the limit of detection, did not modify the observed associations.

With all that adjusting going on our populations being compared get smaller and smaller allowing the influence of sample error to have a greater role in the accuracy and precision of the number reported.  This is why it is hard for me to accept a positive association where the range is 1.02-1.56.

You'll notice how at the beginning of their paper - the only part 99% of all journal paper readers read - it does not report these ORs, instead it reports a percentage:

The prevalence of type 2 diabetes was 7.7%. After adjustment for diabetes risk factors and markers of seafood intake, participants with type 2 diabetes had a 26% higher level of total arsenic (95% confidence interval [CI], 2.0%-56.0%) and a nonsignificant 10% higher level of dimethylarsinate (95% CI, −8.0% to 33.0%) than participants without type 2 diabetes, and levels of arsenobetaine were similar to those of participants without type 2 diabetes.

I wonder why?

Well for starters, 26% higher sounds much more significant than 1.26.  And 2.0% - 56% sounds more pronounced than 1.02-1.56.  Yeah it means the same thing, but the perception is different, that's why we sell stuff at $9.99 and not $10.00.  Kind of sneaky isn't it?  It's not lying, but it's not being forthright either.  It's spin.  Ahh the games you must play when your results are weak.

But wait...we can make support for our hypothesis even stronger!  Let's compare percentiles!  Now we can report an OR that's got a bit of meat on its bones!

After similar adjustment, the odds ratios for type 2 diabetes comparing participants at the 80th vs the 20th percentiles were 3.58 for the level of total arsenic (95% CI, 1.18-10.83), 1.57 for dimethylarsinate (95% CI, 0.89-2.76), and 0.69 for arsenobetaine (95% CI, 0.33-1.48).

Much...much better wouldn't you conclude?  That's brings us back to Figure 2 in my last post, the one with the graphic showing how many "favors association."

Rule number one: Always read the report the findings/recommendations were based on.


Next Post: Apples, Arsenic, and Risk - Part 11: Type 2 diabetes - 80th vs the 20th percentiles


.