For example, a very carefully conducted study that was published last year presented time-series data for a whooping 709125 Swedish teenagers in order to investigate if violent behavior among these teenagers was related to whether their mothers had smoked during their pregnancy or not. One of the things the Swedish researchers found was that after controlling for various socio-economic variables and other potentially confounding factors, the risk of conviction for violent crimes increased by a staggering 300 percent for teenagers whose mothers had smoked at least one cigarette a day during their pregnancy.

Another thing they found from their analysis of the very same data, however, seemed to strongly negate the first finding: When the researchers restricted their analysis to sibling-pairs for which the mother had smoked only during her pregnancy with one of the siblings, but not the other, the difference in risk for violent behavior disappeared. If anything, those siblings who were born at a time at which their mothers weren't indulging their smoking habit, were at greater risk for violent behavior.
Both findings are illustrated in the graph below.

Before I offer some additional thoughts on the Swedish study, I will briefly introduce a different, but related study that also considers the effects of smoking during pregnancy. The finding is often referred to as the "low birth-weight paradox":

Children who are below a certain weight at birth are found to be at a higher risk of death during infancy. The link between low birth-weight and infant mortality is so strong, that birth-weight has essentially become the primary unit of measurement to express this risk. In fact, I would guess that, long before scientists established ant statistical correlations between the weight and the health of a new-born, concerned parents around the world were assessing the "strength" of their newborns by asking "how much does s/he weigh?".

Here is the paradox then: Mother's smoking is believed to reduce the health of the infant, but infant mortality among babies from smoking mothers is much lower than it is in the general population. This is not an isolated finding, but one that has been observed and confirmed on various occasions, and has led even behemoths in the field of statistics (like R.A. Fisher), to conclude that smoking during pregnancy may indeed have beneficial effects for babies; at least for those with low birth-weight, who are exactly those at a greater risk of death.

The point I am trying to make by citing these kinds of empirical findings is a simple one; an often overlooked truism: Causal inference from observations demands critical thinking. Regardless of whether particular data can be handled in a statistically straight-forward manner, or whether it requires sophisticated modeling techniques, the final conclusions we would draw from our scientific inquiries do not appear as the unequivocal consequence of the data. There is always the need to reason and justify the explicit and implicit assumptions that went into the construction of the sample and the collection of the data, and that governed the choice of methods employed. Underlying all scientific/and non-scientific inquiries, there are always implicit or explicit theories, models, or simply ideas about how the world works; and these ideas are inevitably invoked and deployed whenever we try to "make sense" of something.

Getting back to the examples of smoking mothers, statistical approaches offer some additional guidance on how to make sense of the observations. In the case of the low birth-weight paradox, I am convinced (and it is generally accepted among today's scientists) that what is actually being observed is a statistical treasure that is itself often referred to as Simpson's paradox. It results from the way our data samples arise in non-experimental settings. It is further influenced by the fact that low birth-weight is not exactly the same thing as the risk of infant mortality (it is only an indicator, and there exist other factors that determine the risk of death, but those may not be related to weight), and that we choose a somewhat arbitrary cut-off for what qualifies as "low birth-weight".

Simpson's paradox achieved pop-culture acclaim as part of the so-called Berkeley Sex-Bias case of 1973 in which it was argued (from empirical data) that women were discriminated against when it came to admission to this prestigious university. To support this claim, plaintiffs showed that out of 8442 male applicants to the university, 44% had been admitted, while only 35% of the 4321 female applicants found admission. A closer look at admissions figures, however, revealed that admissions for any given degree program were actually more in favor of women, and that the aggregate numbers did not present convincing (or even reasonable) evidence of discrimination. Instead, the university-wide 9% discrepancy in admissions had resulted from the fact that women were disproportionately applying to programs with stricter admissions standards. This made women less likely to be accepted to the school as such, but for any given program that a woman applied to, her chances of being admitted were observably at least as high as they were for a man applying to the same program.

In the low-birth weight case, Simpson's paradox plays out more subtly, but it still plays a role: Here it helps to consider the fetus that possess the characteristics that are relevant for survival, and would typically have normal birth-weight. Smoking , by this fetus' mother may reduce the child's health so much that it is born underweight (i.e. it is counted as "low birth-weight to a smoking mother"), but not so much that it dies during infancy (this is especially true, when not all health-related qualities are tied to weight). Likewise there will be children that possess qualities that increase their risk of death during infancy, and that would be born with low birth-weight. Smoking may reduce their health so much that mothers might experience a miscarriage (which also implies that there will not be a "birth-weight" measurement for this particular child; it will not be counted).

As a result of these two health-reducing effects of smoking, our sample of low birth-weight infants born to smoking mothers will include a disproportionate amount of children that have above average other health-related qualities. Additionally, those with extremely poor survival chances to begin with, are entirely absent from the sample of children born to smoking mothers, while they will be just healthy enough to be included in the sample for the non-smoking mothers.

As a combined effect of these two biasing trends, we will see increased survival rates among children born to smoking mothers. Paradoxically this bias in observed survival rates comes about not "despite" the health-reducing effects of smoking, but precisely BECAUSE of the health reducing effects of smoking.

As far as the Swedish study is concerned, I still have to wrap my head around it as much as you might have to. Which piece of evidence is more convincing? In this regard, it is worth noting, that both approaches attempt to control for confounders that rely on their own set of assumptions about what is "unobserved":

The first approach, which shows the dramatic increase in violent behavior, assumes that the most important contributors to violent behavior, are captured in the researcher's control-variables. This does not mean that EVERYTHING is included in the analysis, but only that those things which would correlate systematically with a mother's decision to smoke, and which also influence a child's risk for violent behavior are included. In the Berkley Sex Case, the choice of department correlated highly with the sex of the applicant, and also was a determinant of the admissions rate. Including the department in our considerations presumably solved our problem of sample selection bias...

The second approach of the Swedish study, which looked only at siblings, assumes that all important determinants of the teenagers' violent behavior are balanced among siblings. That is, they are not systematically related to whether you consider the sibling for which the mother was smoking during her pregnancy, and the sibling for which the mother didn't smoke while pregnant. One may argue, that this is plausible only alongside a host of additional assumptions:

For example, it is less common for parents to take up smoking habits after their first child has been born, and so one may assume that the siblings that count for the lion's share of the "mother smoked during pregnancy sample" are elder siblings to those that made the comparison group. Possibly elder siblings have a greater formative effective on younger siblings than vice versa. Maybe younger siblings are more likely to adopt violent tendencies of their older siblings than the other way around. Also, this bias will lead to mothers being older (on average) for the group that constitutes the "did not smoke during pregnancy" part of the sample. One may even argue that a mother's decision to stop smoking may tend to coincide with the experience of health problems, so that as a result, one of the groups would be biased towards having been born while their mother's were comparatively less healthy (despite not smoking). These types of systematic influences may still have an effect on the children's development which could manifest itself in the lower of the two graphs presented earlier.

As it stands, even without posting pictures of mutilated lungs, it seems safe to say that smoking is terribly bad for your health. To smoke while pregnant is unhealthy to mother and child (although whether smoking during pregnancy causes(!) violent behavior in children is far less certain). The conclusion that smoking is unhealthy rests on a broad foundation of converging evidence, which combines theoretical understanding of (chemical) pathways with experimental results as well as observational data.

It is the property of convergence accross multiple, disparate modes of inquiry from which scientific results derive their strength, and I believe that this is something that is worth keeping in mind when navigating today's data-oriented, science-driven world. 

After all, many of the decisions we make throughout the day (especially regarding our health), and certainly many of our civic responsibilities, require that we think like scientists and interpret the evidence that is presented to us. We should be able to tell apart converging evidence for human-caused climate change, from data patterns that may make isolated weather patterns look like natural fluctuations. Similarly, we need to be able to distinguish correlations between increases in diagnosed autism cases and number of vaccinations from what would qualify for evidence of a causal relation between the two. In fact, the list of questions that require us to think like scientists is long:

Should you keep your child from being vaccinated? From thinking like a scientist you might conclude that: Not unless you want to take the risk of the child dying from something as preventable as the measles.

Is it still safe to eat fish-fingers produced from catches in the Pacific Ocean? From thinking like a scientist, you might conclude that: It was NEVER safe to eat fish fingers to begin with.

Can we draw causal inferences from empirical observations without critically engaging with the data and making up my own mind about how convincing I find the evidence? From thinking like a scientist you might conclude that: the answer can be derived from this blog post.

Main References:

D, B. M. (2010-05-01) Familial Confounding of the Association Between Maternal Smoking During Pregnancy and Offspring Criminality: A Population-Based Study in Sweden. Archives of General Psychiatry, 67(5), 529-538. DOI: 10.1001/archgenpsychiatry.2010.3

Neovius, M. (2009-02-24) Combined effects of overweight and smoking in late adolescence on subsequent mortality: nationwide cohort study. BMJ, 338(feb24 2), b496-b496. DOI: 10.1136/bmj.b496