6  Describe follow-up and dropout — Who remained until the end?

In cohort studies, as time passes, participants die, move away, and stop coming. Those who remain until the end are not the same population as those originally recruited. In this chapter, you will experience what you can misjudge if you do not describe follow-up and dropout.

6.1 Summarise by examination round

Summarise the characteristics of the participants at each of the 3 Framingham examinations (about 6 years apart).

NoteMake a prediction

From the 1st to the 3rd examination, the proportion of smokers fell from 49.2% to 34.4%. Is this because “the participants quit smoking”? What other reasons can you think of?

6.2 Who came to the 3rd examination?

Split the 4434 participants of the 1st examination by whether they came to the 3rd examination, and compare their characteristics at the 1st examination (baseline).

  • By the 3rd examination, 590 people had died. At baseline, they were older, had higher blood pressure, and more often had diabetes or a history of coronary heart disease
  • 581 people were alive but did not come. Smokers were slightly more common among them
  • The 3263 people who came to the 3rd examination were younger and healthier from the start

In other words, the 3rd-examination population is not “the original population 12 years later”, but the people who survived and were able to come to the examination. The changes in the per-round summaries mix both changes in the participants themselves (quitting smoking, ageing) and who dropped out.

WarningSurvivor bias

If you “examine the relationship between blood pressure and XX using the 3rd examination data”, you are looking at a population from which high-risk people have already dropped out. For example, if people who die early from hypertension drop out, the harm of hypertension may look smaller among those who remain. When you analyse only data from the last time point, always describe who dropped out up to that point.

To separate the two, summarise by following the same people. Look only at people who came all 3 times.

Even among the same 3206 people, the smoking rate falls from 47.6% to 34.2%, so most of the fall in the smoking rate was real quitting (was your prediction right?). On the other hand, at the 1st examination these 3206 people had a slightly lower smoking rate than the whole (49.2%), and less diabetes and prior coronary heart disease. “Which reason it is” cannot be known without describing and checking.

6.3 Describe the follow-up period

Cohort study papers first report for how long and how many people were followed up, and how many events occurred.

  • All survivors are censored at exactly 24 years (the end of the study’s follow-up period; administrative censoring)
  • There is nobody who was lost to follow-up partway (censoring due to dropout)

This data is an example of extremely complete follow-up. In real cohorts (especially health check-up and claims data), there are many people whose follow-up is interrupted by moving, changing jobs, or switching insurance, and how likely follow-up is to be interrupted may be related to health status. Describing the reasons for and timing of censoring lets you assess the assumption of survival analysis (that censoring is unrelated to the event).

6.3.1 Describe with rates (per person-year)

When the follow-up period differs between people, describe with rates per person-year rather than proportions.

6.4 Analysing onset: are people who were already ill at the start excluded?

Let us count new onsets of coronary heart disease (CHD).

28.0% … but let us check for people who already had coronary heart disease at baseline (PREVCHD == 1).

All 194 people with a prior history are recorded with ANYCHD = 1 and 0 days to onset. They did not “newly develop” the disease; they were ill from the start.

Excluding those with a prior history, new onset is 24.7%. If you include people with a prior history in an analysis of the risk of onset, the risk in groups with more prior disease (older, male, hypertensive) looks inflated. Restricting to “people in whom the event has not yet occurred at the start of analysis” is a basic principle of cohort studies.

6.5 Count events by group

Finally, check the number of events in each group you want to compare. Even with many people, estimates are unstable if there are few events.

Even in a cohort of 4240 people, there are only 109 people with diabetes and 55 onsets. If you further split the people with diabetes by sex or age, or adjust for many variables, the estimates quickly become unstable.

ImportantSummary of this chapter
  • The population in the later part of follow-up consists of people who survived and kept participating. Compare the baseline characteristics of those who dropped out and those who remained
  • Describe the follow-up period, censoring, number of events, and person-years
  • In analyses of onset, exclude people who already had the event at baseline
  • Check the number of events in each group before deciding on stratification or model complexity