3 Look at the distribution before the mean
The mean and standard deviation are summaries that work well when the distribution is symmetric with a single peak. Real epidemiological data are skewed, contain a handful of extreme values, include numbers that people rounded when answering, or are truncated at an upper limit. In this chapter, you will experience what you can notice by looking at figures before computing summary statistics.
We use the baseline data base (one row per person) created in the previous chapter (already loaded).
3.1 Looking only at summary statistics
The mean of blood glucose GLUCOSE is 82 and the standard deviation is about 24 mg/dL. Can we assume that “about 95% of people fall within mean ± 2SD (34–131)”? What shape do you think the distribution has?
3.2 A skewed distribution: blood glucose
- Most people are concentrated between 60 and 110, with a long right tail (maximum 394)
- The mean (82) is pulled by the tail and is larger than the median (78)
- Nobody actually has hypoglycaemia like mean − 2SD = 34 mg/dL (minimum 40)
Looking at who is in the tail explains the shape of the distribution.
Those in the right tail are mainly the 121 people with diabetes (2.7% of the total) (all 36 people with glucose of 200 or more have diabetes). In other words, the glucose distribution was a mixture of populations with different characteristics: “the healthy majority” and “a minority with diabetes”.
- Describe with the median (interquartile range, IQR) (also commonly used in Table 1 of papers)
- If the tail consists of a particular group (here, diabetes), describe that group separately
- When entering the variable in a regression model, consider log transformation, categorisation, and the influence of outliers
3.3 Extreme values: entry error or real value?
There are values such as total cholesterol of 696 mg/dL and systolic blood pressure of 295 mmHg. When you find extreme values, do not delete them mechanically; check whether they are really plausible.
The person with 696 mg/dL also had 614 mg/dL at the examination 5 years later. Since it is repeatedly high, it is more likely a real value, e.g. familial hypercholesterolaemia, than an entry error. On the other hand, a value that appears only once and looks off by one digit (e.g., height 1700 cm) suggests an entry error.
In epidemiology, people with extreme values are often the ones most likely to develop disease (hypertension, hypercholesterolaemia). Removing them mechanically, as in “exclude values outside mean ± 3SD as outliers”, can mean removing the highest-risk people from the analysis. If you exclude, decide the criteria (e.g., physiologically impossible ranges) before the analysis, and report how many were excluded.
3.4 Rounded numbers: cigarettes per day
With self-reported numbers, people tend to round to convenient numbers when answering.
35.6% of smokers answered “exactly 20 (one pack)”, and 78% answered a multiple of 5. This kind of heaping (numbers piling up) commonly occurs in self-reported numeric values such as age, weight, alcohol intake, sleep duration, and duration of symptoms.
- The true number of cigarettes for people who answered “20” is scattered around 15–25 (measurement error)
- If you look at a dose–response relationship with finely divided cigarette counts, an unnatural step can appear at 20
- Place cut-points (categorisation) where they do not straddle the heaps (e.g., 1–9, 10–19, 20, 21 or more …)
3.5 Values truncated at an upper limit: age in NHANES
Let us look at the distribution of age in NHANES.
There are 87 people aged 79, but 788 aged 80. It is not that the number of 80-year-olds suddenly jumped. In NHANES, to prevent identification of individuals, everyone aged 80 or over is recorded as “80” (top-coding).
- “80” means “80 or older”. This data cannot tell you the mean age of those aged 80+, or the characteristics of people in their 90s
- If you enter age as a continuous variable in a model, everyone aged 80+ is treated as 80
The same thing happens with income (“$XX or more”), the upper and lower limits of the measurement range of laboratory values (“below the detection limit”), upper limits in self-reports (“10 times or more”), and so on.
3.6 Whose values are mixed in: height in NHANES
“The mean height of Americans is 156 cm”? Looking at the distribution reveals why.
NHANES also includes children (from age 0), and 42% are under 20. The long left tail was children. The mean itself is not wrong, but if you interpret it without checking “whose mean it is”, you get it wrong. If you want to know adult height, restrict to adults before describing.
- Look at histograms and bar charts before summary statistics
- Describe skewed distributions with the median (IQR), and check who is in the tail
- Do not mechanically delete extreme values; check whether they are really plausible. Decide exclusion criteria in advance and report them
- Self-reported numbers have rounding (heaping), and public data have truncation at upper limits
- Before computing a mean, check who is included in it (children, sick people)