Confounders, observation, caution.

A confounder is a factor, generally not measured a priori, that may be causing an observed correlation but was not measured or allowed for. The usual method to deal with this in observational studies is to correct for as many factors as present using multivariate analysis. These analyses may have to extend to include familial ones. This example is from a paper considering in utero exposure to nicotine with risk of psychiatric disorder.

Associations Between Maternal Smoking During Pregnancy and Offspring Severe Mental Illness (SMI)
Hazard ratios from Cox proportional hazards regression models for any SMI (A), bipolar disorder (B), schizophrenia spectrum disorders (C), and SMI with substance use disorder (D). Population models (model 1) were adjusted only for offspring sex and parity. Adjusted models (model 2) additionally included maternal and paternal covariates. Cousin fixed-effects models (model 3) compared discordant cousins and included all covariates. Sibling fixed-effects models (model 4) compared discordant siblings and included offspring and paternal covariates and maternal age at childbirth; maternal covariates that could not differ among siblings were excluded. Error bars indicate 95% CIs. Moderate smoking during pregnancy: 1-9 cigarettes per day; high smoking during pregnancy: ?10 cigarettes per day. Y-axes are natural log-scaled.

A couple of confounders are obvious to anyone who works with the mad. The mad smoke. The mad die early: an analogy used last week was that the mad have the same life expectancy as the general population — born in 1900, not the three decades greater life expectancy of those born now. The authors comment correctly.

Sibling comparisons have several limitations that we could not fully address. First, they are susceptible to confounding from unmeasured factors that make siblings different from one another. The within-family associations may, therefore, overestimate or underestimate the true SDP effect. However, to explain the weak and nonsignificant sibling differences found here, unmeasured within-family confounders would have to be positively associated with SDP and negatively associated with offspring SMI (or vice versa)

But this does not consider the behaviours that smoking hunts with. In a time when smokers are the new pariahs, the marginalized, the dysocial, and the rebellious light up. These things are bad for you. Kendler, writing a review on this paper, noting correlation is not causation, concludes wisely.

I close with these cautions. With observational data, we can never be certain about causal processes. We can only seek for increased confidence that causal effects are likely present. It is this confidence that can help guide planned prevention and intervention efforts. However, this issue has a flip side. Finding associations in observational data are too easy. Researchers who report such results are obliged to avoid causal language and wherever possible, use available methods to provide some insight into the possible causal relationship between their exposure and their outcome.

I advise people not to smoke, because the data on it as a risk factor for lung cancer and cardiovascular disease is as certain as we can get in medicine. However, I have yet to have a trainee include in their formulation for a person with first episode psychosis that they were exposed to maternal smoking. We are more interested in their substance abuse, particularly cannabis.

For significant links are too easy to find. Some are noise. And sorting out the confounding factors is not merely conceptually difficult, it adds expense.