Лев Гунин - ГУЛаг Палестины

order to coat their stomachs and slow the absorption of the alcohol. They might do a

large number of things. What is important is that the knowledge of one's experimental

treatment can lead to one or more changes in behavior, and that it is these unintended

changes, and not the wine consumption itself, that could affect longevity, either in one

direction or the other.

Or, here is a particularly plausible confounding that might appear. Imagine that the

experiment attempts to control wine drinking, and no more than that, and that subjects

do faithfully follow the wine regimen that is imposed on them. Nevertheless, the less

wine that they were allowed to drink, the more beer and hard alcohol they would probably

end up drinking, but which would make the initially equal groups unequal on beer and

hard-alcohol consumption. And so then it would be impossible to tell if differences in

longevity should be attributed to differences in wine consumption, or to differences in

beer consumption, or to differences in hard-alcohol consumption.

But while we may choose to pause and speculate as to what confounding variables may

appear, scientific method does not obligate us to do so. We know that confounding

variables are possible in non-double-blind experiments, and the number that we are able

to imagine is limited only by the time that we allocate to trying. If I cared to spend

a few hours thinking about it, I could write several pages of possibilities. If I chose

to spend a few months thinking about it, I could write a book of possibilities. I am

able to imagine confounding variables either improving health or impairing it at the low

end of the alcohol-consumption continuum, and as well either improving or impairing

health at the high end of the alcohol-consumption continuum. Scientific method does not

require us to know for certain what and how many confounding variables may appear to

destroy the validity of an experiment which is not double-blind; rather, scientific

method assures us that it is so likely that one or more confounding variables will make

their appearance in a non-double-blind experiment, that such an experiment must be

considered to be fatally defective, and that no cause-effect conclusion can ever be

drawn from it with confidence.

Thus, no valid experiment exists. In short, we can be sure that no experiment has ever

been conducted to ascertain the effect of long-term alcohol consumption on longevity,

and that if such an experiment had ever been conducted, the impossibility of its being

double-blind, or even blind, would render it inconclusive.

The French Paradox Research

Must Have Been Correlational

But if the data featured in your 60 Minutes broadcast was not experimental, then what

was it? It must, by default, have been correlational. That is, rather than subjects

being assigned randomly to groups and being required to drink a given volume of alcohol

each day, it must have been merely observed what volume of alcohol they chose to drink

each day.

Alcohol consumption would be measured by self-report. Well, it is not quite true that

the experimenter would observe what volume of alcohol his subjects drank daily. It

would be impractical to follow subjects around and actually see how much alcohol they

consumed in restaurants, in bars, in their homes. Much more likely is that every once

in a long while, the subjects would be mailed a questionnaire asking them to report how

much alcohol they had been drinking lately. The inability to measure alcohol

consumption directly is already a weakness - subjects might not remember accurately how

much they had been drinking, or they might experience some pressure to distort how much

they had been drinking either upward or downward. However, this is not at all the big

weakness that I want to bring out, so let us get to that without further delay.

We have already seen that random assignment guarantees pre-treatment equality on all

dimensions. I first recapitulate that in the case of the random assignment of subjects

to groups in an experiment, we were guaranteed that the subjects in each group would be

initially equivalent on every conceivable dimension. The larger the random groups, the

closer to being precisely equal on every conceivable dimension would they become. Thus,

in a properly designed and executed double-blind experiment, any differences that

subsequently arose between groups would have to be attributed to the different

treatments that the experiment had administered to them - for example, if some groups

lived longer than others, nothing else would be able to explain this except that some

groups had consumed a different volume of wine than others.

Natural assignment guarantees pre-treatment inequality on many dimensions. But in a

correlational study, subjects are not assigned to groups randomly, they assign

themselves to groups naturally. A subject who is in a no-wine group, for example, is

one who has himself decided that he does not drink wine. Thus, the groups are referred

to not as randomly constituted, but as naturally constituted, as if nature had come

along and assigned each subject to one of the groups. Now here comes the really

important part. It is that experience teaches us that naturally-constituted groups are

capable of differing from each other on every conceivable dimension, and are highly

likely to differ from each other substantially on a number of dimensions. In other

words, people who drink no wine are likely to differ from people who drink several

glasses of wine in many ways. Perhaps the non-drinkers will have more females, and the

drinkers will have more males - or perhaps the opposite. Perhaps the drinkers will be

older or younger. Perhaps the drinkers will be richer or poorer. Perhaps the drinkers

will tend to be single and the teetotallers tend to be married, or vice versa.

Differences may readily be discovered in height, in weight, in education. Differences

could quite plausibly be discovered in smoking, in drug use, in exposure to industrial

pollutants, in diet. People who drink will tend to live in different parts of the city

from people who don't drink. People who drink may watch more television, use microwave

ovens more, spend more time breathing automobile exhaust - or less. As people of

different ethnic backgrounds, or religions, or races drink different amounts, it follows

that people who drink different amounts will differ in ethnic background, in religion,

and in race.

One can speculate about thousands of ways in which drinkers could differ from

teetotallers, and if one actually examined two such groups, one would find a few

dimensions on which such extraneous differences were large, several dimensions on which

such extraneous differences were moderate, and a large number of dimensions on which

such extraneous differences were present but small. The hurdle that the correlational

researcher is never able to overleap is that given that he is unable to look for every

conceivable difference, he will never know all the ways in which his

naturally-constituted groups did indeed differ from each other.

Natural groups may eat different amounts of broccoli. And so then, no cause-effect

conclusion will ever be possible from a correlational study. If the moderate drinkers

happen to live longer, we will never be able to conclude that this is caused by their

moderate drinking, because it might be caused by how close they live to high-voltage

lines or how often they wash their hands or how far they drive to work or how much

toothpaste they swallow or how much they salt their food or how close they sit to their

televisions or how many pets they keep or whether they sleep with their windows open or

whether they finish their broccoli. In an experiment, random assignment of subjects to

groups guarantees equality on all such extraneous dimensions, and this makes

cause-effect conclusions possible. In a correlational study, natural assignment of

subjects to groups guarantees inequality on many such extraneous dimensions, and this

makes cause-effect conclusions impossible.

Correlation does not imply causality. Every textbook on statistics or research

methodology underlines this same caveat, captured in the expression "correlation does

not imply causality," which warns that from correlational data, it is impossible to tell

what caused what. Science has developed only a single method for determining what

caused what - and that method is the experiment. No experiment, no cause effect

conclusion - it's that simple. Given correlational data, furthermore, there is no way

of extracting cause-effect conclusions by more subtle or more advanced analyses - no way

of equating the groups statistically, no way of matching subjects to achieve

statistically the pre-treatment equality that is needed to arrive at cause-effect

conclusions. Advanced methods of analyzing correlational data do exist, and are used by

naive researchers, and to the layman may appear to be effective, but the reality is that

all are fatally flawed, all have been demonstrated in the literature to be ineffective

and to lead to inconclusive results. The bottom line is that there is no way to extract

cause-effect conclusions from correlational data.

You overlooked that the causal direction might be reversed. In the case of The French

Paradox finding, I can readily see a plausible alternative interpretation as to how the

observed data could have arisen. The data do seem to show that as drinking declines

from a high to a moderate level, longevity increases. This accords with the notion that

alcohol is toxic, and that its effects are deleterious. What constitutes The French

Paradox, however, is that when one goes even farther along the drinking continuum from

moderate drinking all the way down to no drinking at all, instead of longevity

increasing still higher, the opposite happens - longevity shrinks.

What distinguishes the scientifically-trained mind from that of the layman in this case

is that the layman thinks of a single interpretation, and seizing on that as the only

one possible, stops thinking. That is, the layman thinks "Drinking not at all is

unhealthy, therefore I can improve my health by drinking." The scientifically-trained

mind, in contrast, recognizes that in correlational data a large number of

interpretations is possible, acknowledges the first interpretation that springs to mind

as one among the many that are possible, and keeps looking, and keeps finding, a number

of alternative interpretations, and ultimately acknowledges the impossibility of

choosing among them.

As illustrated in my own case. Specifically, I happen to find myself in a

naturally-constituted zero-alcohol group. That is, I drink not at all, or very close to

not at all. There is a reason for this, and that is that the effects of alcohol upon me

are toxic. Mainly, I get splitting headaches, even from the ingestion of small amounts

of alcohol, particularly if the alcohol comes in the form of wine. I take this to mean

that my constitution is weak, that I am unable to process alcohol efficiently, that I am

unable to detoxify my body of alcohol the way that others can, that my body chemistry is