jt512
Philosopher
- Joined
- Sep 24, 2011
- Messages
- 5,523
The crux of Harcombe's argument is that the relationship between meat intake and mortality was confounded by a number of factors, such as total caloric intake, level of physical activity, smoking, etc., that were correlated with meat intake. And that, although the investigators attempted to adjust for these factors, they did not do so "satisfactorily." Her evidence for this conclusion is that if you compare the unadjusted results with the statistically adjusted results, the adjusted results change in what she believes is the wrong direction.
There are many problems with this argument, but the overarching one is that the unadjusted and adjusted figures she is comparing are not comparable. Specifically, for each quintile of meat intake, she computed the absolute risk of death, and compared it with the adjusted relative risk of death reported in the paper. She incorrectly states that "[t]he multivariate [results] should [due to the effects of the confounders] be substantially below my [raw] death rate [figures] for every quintile of intake..." But that is utter nonsense because there is no reason why the absolute rate should be above, below, or even close to its respective relative rate.
It is easy to see why. We need only to look at how these figures are constructed. As an example, let's look at Harcombe's table for total meat intake in the Health Professionals cohort:
Deaths [1]| 1713|1610|1679|1794|2130
Person-years [2]| 151,212|152,120|151,558|152,318|151,315
Raw death rate 100×[1]/[2]| 1.13|1.06|1.11|1.18|1.41
Adjusted relative rate| 1.00|1.12|1.21|1.25|1.37
To obtain her raw death rates, she divided the number of deaths in each quintile by the person-years of follow-up in the quintile (and multiplied the result by 100). That's fine; the result is the raw death rate per 100 person-years of follow-up. However, she compares those figures with the adjusted relative rates, which are the adjusted death rate for each quintile divided by the adjusted death rate in the first quintile.
To see the absurdity of Harcombe's claim that the adjusted relative rates should be less than the unadjusted absolute rates, we need only to examine the figures for the first quintile. The adjusted relative rate for Q1 will always be 1, because it is a number divided by itself. In contrast, the raw absolute rate could, in principle, be any non-negative number. In fact, it is a complete coincidence that the numbers in this case happen to be close, and indeed the only reason that they are even comparable to an order of magnitude is that Harcombe multiplied her figures by the arbitrary constant of 100. So there is no inherent relationship between the two figures, and the same goes for the other quintiles; so Harcombe's claim that the adjusted relative rates should be less than the unadjusted absolute rates is nonsensical.
Harcombe then proceeds to make an incoherent argument that the adjustment for confounding in the paper must have been deficient because the mortality trend by quintile in the unadjusted results is U-shaped, but linear in the adjusted results. She writes, "As meat consumption increases from Q1 to Q2 and Q1 to Q3, so the death rate falls. Only in Q4 and Q5 does this reverse and it is in these quintiles that we saw the highest levels of BMI, smoking, low activity, high calorie intake, high alcohol intake and so on and these have clearly not been adequately allowed for."
So what? The trend in most of the confounding factors that were adjusted for was monotonic across quintiles. If she thinks that the highest levels of the confounders are responsible for producing the highest risks attributed to red meat in the highest quintiles, then how does she explain that these same confounders, at intermediate levels, were responsible for producing an apparent protective effect for meat intake in the intermediate quintiles? That is, how does she explain a U-shape trend in the unadjusted risks in the presence of monotonic trend in the confounders?
In fact, with no justification, she bases her analysis on an arbitrary subset of the confounders adjusted for in the paper. The full set of confounders includes beneficial factors as well as detrimental ones, which would complicate any attempt to guess the direction that adjusting for the full set will have on the results. Furthermore, the actual adjustment occurs at the individual level. Attempting to ascertain the effect of adjustment from the quintile-level averages is therefore impossible (it is an example of the ecologic fallacy).
There were numerous other serious problems in her analysis, but hopefully, this explains why her main argument is way off base.
Last edited:
But seriously, we've been hearing so much about the supposedly ill effects of meat that the highest quintile MUST be risk-takers. They probably also drive faster, weigh more, go out in the midday sun without sun blocker, do not get regular lab tests, don't even own a scale. Maybe even eat less roughage. Lots of stuff to make correcting for confounding difficult. 
