Birth cohorts, the field's main instrument

A birth cohort recruits a sample of children born in a given period and follows them at regular intervals, collecting data on health, schooling, living conditions and family environment.

The United Kingdom has launched several since 1946, which allows comparisons between generations. France launched the Elfe cohort in 2011, with a sample of about eighteen thousand children recruited at maternity units.

The British cohorts of 1946, 1958 and 1970 were followed by the Millennium Cohort Study, launched with children born around the year 2000. Their succession makes it possible to compare, at the same age, children separated by several decades, using partly shared questionnaires.

It establishes temporal order, which a cross-sectional survey cannot. Observing that a characteristic measured at two precedes an outcome measured at ten eliminates a whole family of reverse explanations.

It also allows trajectories rather than states to be measured, a decisive distinction in a field where individual positions are unstable and a single measurement predicts the next one poorly.

Precedence is not causation. A family characteristic measured early may precede an outcome without producing it, if both depend on an unmeasured third factor, which is the ordinary situation.

Cohorts measure many variables precisely in order to adjust for such factors, but adjustment only covers what was measured. The most solid analyses exploit quasi-experimental situations inside the cohort.

Sibling comparison is a common example: it neutralises part of the unmeasured family factors, since these children share the same parents and often the same home. The approach has limits of its own, as children in the same family are never exposed to identical conditions.

Loss of participants is the central difficulty. It commonly reaches several tens of per cent after fifteen years, and it is not random: families in unstable circumstances, those who move often, drop out more.

The remaining sample becomes progressively less representative, precisely on the dimensions of greatest interest. Weighting methods correct part of the bias, under unverifiable assumptions. The problem is known from the design stage, and cohorts devote a large share of their resources to keeping in touch with families between waves. Attrition is thus reduced without disappearing, and its selectivity remains the main caveat in interpreting later results.

A cohort describes the generation it follows. Children born in 1958 grew up in a world whose conclusions do not transfer mechanically to those born in 2011, whether on screens, family structure or early schooling.

That is the main argument for launching new ones despite the cost, and it is also what makes comparisons between successive cohorts more instructive than any single cohort taken alone.