Sunday Science Lesson: Misconceptions about the gateway effect

by Carl V Phillips

Disturbingly close to 100% of writings about whether there is a gateway effect among tobacco products, particularly about e-cigarettes being a gateway to smoking, are nonsense. That includes most metas on the topic, where someone tries to explain what mistakes others are making. Perhaps addressing a few simple misconceptions can clear some of this up (and it is a favor to a few of my tweeps — I hope I have addressed all the points our conversation left hanging). I have written about all of this much more extensively and comprehensively (see this in particular), but not in soundbite form.

Is there anything particularly unusual about the gateway claim?

No. It is a straightforward claim of causation of the sort that epidemiologists and other social scientists investigate all of the time. Does exposure E (which can be a behavior or something else) cause outcome O (which can be a disease, a behavior, or something else). There are lots of cases where one behavior causes another (e.g., attending university causes someone to get a professional job, because they are then qualified for it). A great deal of harm to understanding has probably been done by giving a cutesy name (which happens to be an incorrect metaphor) to the phenomenon of exposure to one drug causing the use of another. If everyone just said “does exposure to e-cigarettes cause some people to be smokers?” rather than “is there a gateway effect?”, many of the stupid statements would be self-evidently stupid.

Does that mean that “gateway” really refers to a well-defined concept and is not just rhetoric?

Well, yes and no. Yes, it is clearly a statement about an exposure causing an outcome, and any attempt to mis-define it otherwise is either rhetoric or confusion. On the other hand, any specific claim about it requires more specifics. The phrase “exposure to e-cigarettes” could, quite legitimately, be defined as regular vaping, trying an e-cigarette once, or infinite other intensities. Someone could even refer to the effect of merely being in a world with e-cigarettes (seeing use and advertising), though this would be pushing the definition of “gateway” a bit (though it would still be a perfectly valid question about causation). There are also issues of quantity: If the claim is “caused at least one person ever to smoke”, then it is inevitably true but not very interesting. Social sciences like epidemiology are about measurement (quantity) not the mere existence of a phenomenon. Similarly timing: If the claim is “causes smoking within a week of exposure” then the number for whom it is true is probably vanishingly small, but that is also not very interesting. If the question is “…ten years after initial use” then it is interesting but we have no data yet.

So which of those possible specific definitions is right?

No one of them is right. Some might be further from the spirit of the concept than others, but not wrong. Many candidate definitions are equally right. Note in particular that this means that someone is speaking utter nonsense if they say something like (and, yes, this has been said), “this study did not show a gateway effect because they only looked at whether someone had ever tried an e-cigarette rather than measuring whether someone was a regular user.” It is no less valid a valid gateway hypothesis to say that trying an e-cigarette causes smoking, rather than being a regular user causes it.

Now if you want to actually do some science, rather than just engaging in semantic rhetoric as most people “debating” this issue do, you might be inclined to predict that any gateway effect would be stronger for those who used e-cigarettes more, and then check to see if the observed data followed that pattern. If it did not, it would suggest that perhaps whatever association was observed was not actually a gateway effect.

Is this one of those extraordinary claims that requires extraordinary evidence?

No. It is an extraordinary claim to suggest that e-cigarettes are not actually causing people to quit smoking. It is an extraordinary claim that smokers do not really want to smoke or derive no benefits from it. Those are claims that fly in the face of huge bodies of evidence and theory, and anyone who wants to claim otherwise needs to produce something a lot more compelling than the flat-earther straws they grasp at. The gateway claim is a perfectly normal and prima facie plausible scientific hypothesis.

However, if someone is trying to make a claim beyond the simple “there is inevitably at least one case of this”, they need to specify what they are claiming (in particular, quantify their claim) and provide some evidence. The mere observation of some association that would exist with or without a gateway effect is obviously not such evidence. E.g., observing that people who use e-cigarettes are more likely to become smokers than those who do not is no more informative than observing that the sun rises. Both of those would be true in a world with a substantial gateway effect, and both would be true in a world without one.

That, of course, captures the real and valid arguments against all the gateway claims that are being made: 1. The writers are not actually making a concrete claim of any sort. 2. The supposed evidence is merely the observation that people who inclined to use one tobacco product are more likely than those who do not to be inclined to use another tobacco product, which would clearly be true whether or not there was any gateway effect. I have explained this at length. That is the alpha and omega of refutation of the gateway “evidence”. Yes, this point gets made. But it also gets buried in a dozen incorrect — nay, utterly innumerate — claims that are rather more common.

I have this vision of a serious scientific thinker from another field stumbling into this debate and deciding she would rather spend her time on a forum that is debating creationism or racial supremacy, because there those on both sides are making more scientifically defensible arguments.

What is this “cause” concept that seems to be important here?

Yes, rather important. The best practical way to think of it is, “if someone does/has/is E, then O will occur, whereas if you could somehow run a different version of the world such that everything else was the same but he did not do E then O would not occur.” At the population level, it means that this is true for some people. Note in particular that this does not just mean that E precedes O in time (though as far as we understand the universe, that is necessary, just not sufficient). It also does not mean that E merely predicts O (that is, O is more likely to occur if E than if not E). A predictor that is not a cause might be a confounder (if you are not familiar with what that means, you can search the archives here) or it might be a proxy for the actual causal exposure (e.g., spending a lot of time on a university campus predicts getting a professional job, because it is strongly associated with attending university, but that does not mean that just hanging around will get you that job). So in particular, the fact that vaping among nonsmokers predicts future smoking (as we would expect) does not mean it causes it.

But how can you know for sure if an exposure caused an outcome, such as the gateway effect, if it requires imagining an alternative running of the world?

You can’t! Not ever. About anything. Causation can never be observed or proven, only inferred with varying degrees of confidence. If you ever read someone making claims about proving causation (often found in attempts to claim that some study types can do it and others cannot), you would be wise to stop reading because Dunning-Kruger has a nasty habit of being contagious. You have encountered someone who thinks the dumbed-down understanding of science he learned in some (bad) intro class is sufficient grounds to pontificate as an expert, and so he probably has other wrong convictions too. (Note: Sometimes “proven” is used colloquially to mean something like “established beyond a reasonable doubt.” That is not what I am talking about. I refer to cases where someone seems to think there is a bright line between proof and non-proof in science. There isn’t.)

So how can you ever be sure there is not a gateway effect?

You can’t! And even more so than the previous “You can’t!” You cannot observe that the exposure is not causing some cases of the outcome, as explained the previous paragraph (assuming there is anyone who has both the exposure and outcome, as there will be). But even worse, even if there has never existed a circumstance under which the gateway effect has occurred, even once, that does not mean that it will not happen under the new circumstances that come tomorrow. Again, there is nothing special here — that is true for any causal claim.

Wait, “circumstances”? That seems to complicate things.

Indeed. There are no universal laws or constants is social sciences (or in biology or many other sciences). Our educational system does students a great disservice by presenting physics as the canonical science, when it is actually very odd among the sciences in having universal laws (we think). In social sciences, everything depends on which people you are talking about, and when, where, etc. To put that in practical terms for present purposes, anyone who says, “X shows there is no gateway effect” is speaking such utter nonsense that nothing he says about the topic should be taken seriously.

But what if only smokers are using e-cigarettes? Surely that means there is no gateway effect.

Nope. This is the most common error among those who are making erroneous claims that there is no gateway effect. They cite an observation that most exposed individuals in some population are not at risk of becoming smokers and then — in an utterly unforgivable act of scientific illiteracy — claim that this means there is no gateway effect. Even in the most extreme (fictitious) scenario, where no at-risk individual was ever exposed to e-cigarettes, there could still be a gateway effect. (Note: “at risk” is the technical term for someone who could possibly later have an outcome — in this case, it is everyone who was not already a smoker, and thus could become a smoker. This is not to be confused with the misuse of that phrase as if it meant “high risk” or as a euphemism that roughly means “someone whose life is very screwed up.”)

At the very least, this is an incorrect claim because it is still useful to consider what will happen if a time comes when someone is at risk, which is inevitable. Consider an analogy: We can be very confident that gestating a baby outside Earth’s ionosphere would cause risks of bad birth outcomes, based on what we know about cosmic radiation and the effects of radiation on fetuses. The fact that it has not happened yet does not negate this causal relationship.

The question of whether such radiation causes bad birth outcomes right now, even though literally no one is at risk, can be left to esoteric philosophy and need not concern us here. That is because all such claims in the gateway case are based on evidence that few people in a population are at risk, not none. Thus, even as these commentators are saying this is evidence there is no gateway effect, they are actually pointing out the evidence that there is a population in which it might be happening.

As for those cases along the lines of “this study found that all of the kids who were vaping in this one Scottish village had already smoked” leading to “this is evidence that no gateway effect exists, anywhere, any time”, I think a simple *facepalm* is a sufficient response.

But surely if very few at-risk individuals have the exposure, the gateway effect cannot really matter much.

I agree, so long as you add an “as of now” caveat at the end of each clause — recall that circumstances matter and they can change. However, that is never what anyone claims. No one ever says “yes, there is undoubtedly some gateway effect, but its magnitude is so small that it is not worth worrying about, at least for now.” That would be a very defensible claim. Sadly, the fact that you never see that statement (except here) does not seem attributable to mere political games (“ignore nuance! speak in absolutes regardless of their obvious falsehood! rally the base!”), but to genuine innumeracy.

But not only are few at-risk individuals exposed, meaning very few could be caused to smoke by the exposure, but many would-be smokers are taking up vaping instead of smoking. So the net effect on smoking is clearly negative. Doesn’t this mean there is no gateway effect?

No. These are different effects and the existence of one does not affect the existence of the other. Vaccines save lives, but that does not change the fact that they kill a few people also. It is perfectly valid to claim that the net effect of something is in a particular direction despite a few cases that are caused to go the other way. It is pure deceit to use this as a basis for denying the few cases. Note that this means that the second-most common “there is no gateway” argument is also patently invalid: The fact that smoking rates are dropping in, say, a population of teenagers that is adopting e-cigarettes, does not mean there is no gateway effect, but merely that other factors are having a bigger effect in the other direction.

Well at the very least, doesn’t this mean we should not care if there is a gateway effect because it is swamped by the effect in the other direction?

So long as this is not confused with the claim that there is no effect, a case could be made for that. Indeed, most people would probably agree. However, such questions are far from uncontroversial.

This delves into the realm of ethics rather than the materialistic science, but it comes up in the context so it is worth briefly addressing. Let us start with the simplification that causing non-smoking is good and causing smoking is bad (which glosses over other ethical questions). There are still defensible arguments that causing bad for some people cannot be justified by the good caused to others. To take the extreme example that is often used, most everyone agrees that it would not be ok to capture and kill one healthy innocent person and harvest his organs to save six other people who are dying for lack of transplants. There are libraries worth of literature debating these points (search “trolley dilemma” for a toehold). Regulations in practice (like requiring vaccinations) are generally based on a “needs of the many outweigh the needs of the few” rule. Still, it cannot just be treated as a given, let alone as obvious, that the net effect is all that matters.

So this sounds pretty hopeless to assess. How are we supposed to do it?

Well the first step is to concede that we cannot hope to distinguish a tiny effect from zero effect (as is pretty much always the case in most sciences), and to reassure ourselves that that difference does not matter much. Then we need to at least roughly specify how big an effect we care about, what the timing of it is, and so forth. With that valid and testable hypothesis, there are options available (I go into detail in the above-linked paper).

There are various things we can look at, such as whether there are particular patterns of one behavior preceding another. Of course, the most common (willful?) error of those who claim a gateway effect is conflating mere precedence with causation. Obviously someone who vaped and then became a smoker might have smoked anyway. So we need do a lot more than that.

For example, if we had a really good propensity score that predicted someone’s probability of later becoming a smoker and discovered that vaping increased the probability substantially above this prediction, that would be a good clue. (Note that I am not talking about the mindless throwing-in of whatever covariates are conveniently available, which is the risible standard in public health research, with an unsupported claim that they control for propensity. One problem with most of the anti-gateway commentary is that those writing it themselves have no clue about how to properly deal with covariates and propensity.)

Another tactic is to make predictions about comparisons that would go a particular way under the gateway claim. For example, someone might hypothesize that the gateway effect is primarily due to nicotine, and thus expect to see vapers who use nicotine, but not those who do not use nicotine, taking up smoking more often than predicted.

Note, however, that this emphasis on timing does not support another somewhat common “argument” that is offered against the studies that claim to show a gateway effect: The claim is that since timing is so important, you therefore cannot make any inferences from a study that does not follow people over time (i.e., the claim that you need a cohort study, of which there are only one or two in this space, rather than the more common cross-sectional study). This claim is another Dunning-Kruger example of people who half-understand scientific inference. Scientists, quite legitimately, draw causal inferences all the time without following their subjects (people, ecosystems, molecules, whatever) over time. When you are talking about behaviors, it is particularly easy to just ask retrospective questions (as is done to identify people who quit smoking by switching to e-cigarettes). Though that is not even strictly necessary (for reasons that would make this post even longer).

(Note that the gateway claims made by those cross-sectional study authors are still all wrong. But this is not the reason they are wrong. It is for the reasons given above.)

Could we maybe learn more with brain scans or something, about whether vaping causes changes that make people more inclined to smoke?

In theory such things could be done. (In practice, that brain porn “science” is a joke — little more than latter-day phrenology, for now anyway.) However, it would not actually answer the question. I have discussed this with length in the context of “addiction”. If the phenomenon of interest is behavior and not biology, then a particular biological change is neither necessary nor sufficient for the behavior to have been caused. The biomarker might be associated with it, in which case it might be an explanation for the observed effect, but it is not the effect itself. This is easy to understand: Imagine someone who evidenced whatever the supposed marker is on a PET scan, but he never started smoking; obviously he would not be a gateway case.