1  Missing-data mechanisms

1.1 Missingness in real data

Using data on 10990 people aged 20–79 from NHANES 2009–2012 (adults), let us look at the proportion missing for each variable.

Each individual variable is only about 4–10% missing, but dropping everyone with at least one missing value reduces 10990 people to 8662 (79%).

Let us count which variables are missing together (the missing-data patterns).

  • Common patterns are total cholesterol missing alone (800 people), income missing alone (547), and blood pressure missing alone (331)
  • Those with BMI, blood pressure, and cholesterol all missing together (276) are probably people who did not undergo the physical examination and blood draw

1.2 Who are the people with missing values?

Let us compare people with and without missing blood pressure.

People with missing blood pressure tend to be slightly more often female and to have slightly lower income. The people with missing values are not necessarily the same as those who were observed. If you ignore missingness and analyse only those observed, the results may be biased.

1.3 The three missing-data mechanisms

We classify what causes missingness into the following three types (Rubin’s classification).

Mechanism Whether a value is missing depends on Example (missing blood pressure)
MCAR (missing completely at random) Nothing The measuring device failed at random
MAR (missing at random) Observed variables Younger people are less likely to attend the check-up (age is known)
MNAR (missing not at random) The missing value itself (or unobserved variables) People with higher blood pressure avoid being measured

The names are confusing, but MAR is not “random”. It means “random once you condition on the observed variables”.

1.4 Checking with data whose true values are known

In real data, the true values of missing observations are unknown. So we use data on 2000 people with no missing values (full), create missingness ourselves, and check what happens.

NoteAbout full

full consists of 2000 people selected from those aged 20–79 in NHANES with no missing values. Age, sex, BMI, etc. are real values, but only systolic blood pressure BPSysAve has been regenerated, mimicking its NHANES relationship with age, BMI, and sex, so that the true relationship is known. In addition, home blood pressure HomeBP (correlation 0.87 with the check-up blood pressure) has been added as an auxiliary record (used in Chapter 6).

We make 30% of blood pressure values missing under each of the three mechanisms (make_missing() and miss_prob() are functions defined in R/setup.R).

Computing the mean using only those observed:

  • MCAR: almost the same as the mean before creating missingness (those observed are a random sample of everyone)
  • MAR: too high (young people with lower blood pressure are missing)
  • MNAR: too low (people with high blood pressure are missing)

1.5 Can the data tell you the mechanism?

Under MAR, the missing people are skewed toward the young. This can be seen by comparing age, which is observed.

  • Under MCAR, the ages of observed and missing people are the same
  • Under MAR, the missing people are much younger
  • Even under MNAR, the missing people are older. Because blood pressure is related to age, “people with high blood pressure are missing” looks like “older people are missing”

By comparing observed variables, you can tell that the data are not MCAR. But whether they are MAR or MNAR cannot be distinguished from the data, since the missing values themselves are unknown.

ImportantThink about the missing-data mechanism with knowledge, not data
  • Consider “why values are missing” from the study procedures, the participants, and the circumstances of measurement
  • Record the reasons for missingness (did not attend, device failure, participant refused…)
  • If you assume MAR, observe variables likely to be related to missingness (Chapter 6)
  • If MNAR remains possible, check with a sensitivity analysis how much the conclusions change (Chapter 8)

1.6 Try it

NotePractice
  1. In adults, compare age, sex, and BMI between people with and without missing total cholesterol (TotChol)
  2. If the proportion of MAR missingness is increased from 30% to 50%, how does the mean blood pressure of those observed change?
  1. People with missing cholesterol are only slightly older (48.1 vs. 47.0 years), and sex and BMI are almost the same. As far as the observed variables show, there is no large imbalance (though this does not mean the data are MCAR)
  2. The mean becomes even higher (around 126 mmHg). The more missingness, the larger the bias
ImportantChapter summary
  • Even if each variable has little missingness, the number of people with all variables complete drops considerably
  • The missing-data mechanisms are MCAR, MAR, and MNAR. MAR means “random once you condition on the observed variables”
  • Except under MCAR, the mean of only those observed is biased
  • The data can show that missingness is not MCAR, but cannot distinguish MAR from MNAR. Think with knowledge, and check with sensitivity analyses