A longitudinal study follows 2,000 patients for three years to examine whether a treatment is associated with cardiovascular outcomes.
At the end of follow-up, outcome data are available for 1,650 participants.
What should happen to the remaining 350?
The simplest option is to remove them from the analysis.
That may be reasonable in some studies. It may also change the estimate substantially.
Consider two different scenarios.
Scenario 1
The 350 participants with missing outcomes are distributed similarly to those with complete data. Their probability of having missing data does not appear to depend on relevant characteristics.
Excluding them may have relatively little effect beyond reducing the sample size.
Scenario 2
Participants who experience early adverse effects are more likely to stop attending follow-up visits.
Now the missingness is related to the health outcome being studied.
The 350 excluded participants are no longer an arbitrary group of observations. Their absence may change the composition of the analysed population.
This distinction is central to the epidemiology of missing data.
Three common mechanisms
MCAR — Missing Completely At Random
The probability of missingness is unrelated to both observed and unobserved information.
MAR — Missing At Random
Missingness may depend on information that has been observed, but after accounting for that information, it does not depend on the missing value itself.
MNAR — Missing Not At Random
Missingness depends on information that remains unobserved, potentially including the value that is missing.
These mechanisms cannot always be established directly from the dataset. They are assumptions about the process that generated the missing observations.
That matters when choosing an analytical strategy.
Complete-case analysis, inverse probability weighting, multiple imputation and other approaches all rely on assumptions about why data are missing.
Multiple imputation, for example, can be useful when its underlying assumptions are plausible and the imputation model is appropriately specified. It is not a universal solution for missing data.
The percentage of missing observations is therefore only part of the story.
A study with 5% missing data can face a serious problem if the missingness is strongly related to the outcome.
Another study with 20% missing data may be less affected if the mechanism is sufficiently well understood and appropriately addressed.
For researchers reviewing a dataset, three questions are particularly useful:
How much information is missing?
Which participants or measurements are missing?
What process could have produced those missing observations?
The third question is often the hardest — and the most important.
Because missingness is part of the data-generating process, and that process can influence the estimate we ultimately report.