Types of Numerical Data

What kind of measurement a variable is, and why it matters for every summary computed from it. Nominal, ordinal, discrete and continuous data illustrated on one NHANES 2017-2018 participant's record and the 5,247 adults it comes from; Stevens's 1946 scales of measurement and the statistics each permits; a proof that group means of an ordinal rating can be reordered by recoding unless one group's cumulative distribution lies above the other's; ranks; how recording rules (even-numbered blood pressures, top-coded age and household size) shape continuous and discrete data; and the criticism that scale type depends on the question asked.

Here is one line of a health survey, the record of a man examined in the National Health and Nutrition Examination Survey (NHANES) of 2017 to 2018:

Variable Recorded value
Sex 1 (male)
Age 33
Race and Hispanic origin 3 (non-Hispanic White)
People in the household 3
Systolic blood pressure, three readings 134, 134, 134 mm Hg
Diastolic blood pressure, mean 66 mm Hg
Height 185.0 cm
Body mass index (BMI) 30.8 kg/m²
Health in general 3 ("Good")

Every entry is stored as a number, and a computer will happily average any column. Only some of the averages mean anything. Which ones, and how can we tell before computing them?

Consider four columns. The average of the sex codes is a proportion. The average of the race codes is nonsense. The average of the health ratings is a number whose meaning depends on an assumption nobody stated. And the average of the blood pressures is the most natural summary there is. The difference lies not in the numbers but in what they record. The type of a variable, in the sense of this article, is the kind of measurement its numbers come from, and it decides which summaries describe the variable faithfully.

First we meet the survey. Then we take the course's four types in turn: nominal, ordinal, discrete and continuous. On the way we prove when a comparison of average ratings can be trusted, and meet ranks. Finally we compare the four types with Stevens's scales of measurement, and see why the classification is a guide rather than a law.

The data

The examples in this unit use the 5,247 adults aged 18 and over who were examined in NHANES 2017 to 2018 and had at least one systolic blood pressure reading. They are the same adults whose blood pressures the later units of the course sample from, for instance in Confidence Intervals and Coverage. Besides blood pressure, the survey's public files record:

  • each person's sex, age, race and Hispanic origin and household size (file DEMO_J);
  • height, weight and BMI from the examination (BMX_J);
  • an answer to the question "Would you say your health in general is excellent, very good, good, fair, or poor?" (variable HUQ010 in HUQ_J).

All come from the National Center for Health Statistics (2020). As in the rest of the course, these adults are treated as a complete population to be described. NHANES is itself a weighted national sample, and nothing here is an estimate for the United States.

A variable is a characteristic recorded for each member of a group, and its values over the group are the data. The question this article asks of each variable is what its numbers can be trusted to say. We start with the variable whose numbers say the least.

Nominal data

The race and Hispanic origin variable assigns each adult one of six codes:

Code Category Adults Share
1 Mexican American 707 0.1347
2 Other Hispanic 479 0.0913
3 Non-Hispanic White 1,819 0.3467
4 Non-Hispanic Black 1,220 0.2325
6 Non-Hispanic Asian 744 0.1418
7 Other race, including multiracial 278 0.0530

The codes are names written as digits. The survey documentation notes that code 5 is not used at all (National Center for Health Statistics, 2020), which makes the point. The numbers could be replaced by letters, or by any other six distinct numbers, and no information would be lost. Data whose values are only labels for categories with no order are nominal.

What survives such a relabeling is all that nominal data can say: which category each person is in, how many people are in each, and therefore which category is most common. The mode, the most frequent category, is non-Hispanic White (code 3).

The mean of the codes, 3.509, describes nothing. Relabel the categories in order of size (3 becomes 1, 4 becomes 2, and so on), and the mean is 2.676 for exactly the same people. A statistic that changes when nothing about the people changes is a statistic of the labels, not of the data.

Check your understanding. Suppose the survey had used code 70 instead of 7 for the last category. Which of the summaries above would change: the counts, the mode, or the mean of the codes?

Only the mean, which would rise, though nothing about the 278 people has changed. The counts, the shares and the mode belong to the categories, not to the digits that name them.

Binary data and proportions

A nominal variable with two categories is binary (or dichotomous): sex as recorded here, or whether a person has hypertension. Binary data have one special property. If the two categories are coded 0 and 1, the mean of the codes is the proportion of ones.

Among the 5,247 adults, 2,700 are women. The mean of an indicator that is 1 for a woman and 0 for a man is $2700/5247 = 0.5146$, the proportion of women. That is why a proportion can be treated as a sample mean, the starting point of The Normal Approximation to the Binomial. The coding 0 and 1 is a choice, but a natural one, and the course uses it throughout.

So the sex codes can be averaged and the race codes cannot. The next variable sits between the two: its codes carry more than names, but less than quantities.

Ordinal data

The answers to "Would you say your health in general is ..." are recorded as 1 (excellent) to 5 (poor). Unlike the race codes, these numbers carry an order: a 2 is better health than a 3 in the respondent's own judgment.

What they do not carry is distance. Nothing says that the step from "excellent" to "very good" is the same size as the step from "fair" to "poor". Data whose categories are ordered but whose spacing is not defined are ordinal.

Rating Adults Share Cumulative share
1 Excellent 603 0.1150 0.1150
2 Very good 1,376 0.2625 0.3775
3 Good 1,959 0.3737 0.7512
4 Fair 1,095 0.2089 0.9601
5 Poor 209 0.0399 1.0000

These are the 5,242 adults who gave a rating; five refused or did not know. Because the categories are ordered, a cumulative share makes sense: 37.75% rated their health very good or better, and 75.12% good or better.

So does the median, the category that contains the middle person. The 50% point is crossed in "Good", so the median rating is good. Percentiles in general are meaningful for ordinal data, because they depend only on the order of the values.

The mean of the codes is 2.796. It is not meaningless in the way the race-code mean is, but it rests on the assumption that the five categories are equally spaced. Does that assumption matter? It depends on what the mean is used for, and it matters most when means are used to compare groups.

When recoding reverses a comparison

Any increasing set of numbers $c_1 < c_2 < c_3 < c_4 < c_5$ records the same order as 1 to 5. A researcher who thinks "poor" health is far worse than "fair" might score poor as 10. One who thinks "excellent" stands far above the rest might score it as $-5$. The table compares group means under three such codings. Since the codes rise toward "poor", a higher mean means worse rated health.

Group Adults Codes 1 to 5 Poor = 10 Excellent = $-5$
Ages 18 to 39 1,740 2.550 2.642 1.557
Ages 60 and over 1,923 2.987 3.278 2.547
Ages 60 to 69 1,021 3.005 3.250 2.570
Ages 70 and over 902 2.967 3.310 2.521

The younger adults rate their health better than the older ones under every coding. The second pair, the older group split at 70, behaves differently:

  • with codes 1 to 5, and with excellent scored as $-5$, the adults aged 60 to 69 have the higher (worse) mean;
  • with poor scored as 10, the order reverses, and the adults aged 70 and over have the worse mean.

All three codings respect the order of the categories, and all three are equally faithful to the data. So which group is in worse health? The difference between the two pairs is visible in their cumulative shares.

Cumulative shares of self-rated health (the share at each rating or better). Left: notice that the younger group's curve lies above the older group's at every rating, so every order-preserving coding gives the younger group the better mean. Right: the curves for ages 60 to 69 and 70 and over cross. The older group has more adults at "very good or better" (33.3% against 28.7%) and at "good or better", but fewer at "fair or better", because more of them rate their health poor (6.9% against 4.9%). So the order of the two means depends on the coding.
Cumulative shares of self-rated health (the share at each rating or better). Left: notice that the younger group's curve lies above the older group's at every rating, so every order-preserving coding gives the younger group the better mean. Right: the curves for ages 60 to 69 and 70 and over cross. The older group has more adults at "very good or better" (33.3% against 28.7%) and at "good or better", but fewer at "fair or better", because more of them rate their health poor (6.9% against 4.9%). So the order of the two means depends on the coding.

The pattern is a general fact, and it has a short proof. Here is the idea in plain words. A mean of codes can be rewritten as a sum over the cumulative shares, each weighted by a gap between neighboring codes. If one group's cumulative shares are larger everywhere, every term favors it, whatever the gaps. If the curves cross, a large enough gap at the right place can tip the sum either way.

The formal version. Let a rating have $K$ ordered categories. Let $p_k$ be the share of a group in category $k$, and let $F(k) = p_1 + \cdots + p_k$ be the cumulative share at category $k$ or better, with $F(0) = 0$ and $F(K) = 1$.

Proposition. For any increasing codes $c_1 < \cdots < c_K$, the mean code of a group is

$$\sum_{k=1}^{K} c_k p_k = c_K - \sum_{k=1}^{K-1} (c_{k+1} - c_k)\,F(k).$$

Consequently, suppose group A's cumulative shares are at least group B's at every category, $F_A(k) \ge F_B(k)$ for $k = 1, \ldots, K - 1$. Then:

  • A's mean is at most B's under every increasing coding;
  • it is strictly smaller under every increasing coding if the inequality is strict at some category;
  • the two means are equal under every coding if the cumulative shares are equal at every category.

If instead the curves cross, that is, $F_A(j) > F_B(j)$ at some category $j$ and $F_A(m) < F_B(m)$ at another category $m$, there are increasing codings that order the means either way.

Proof. Write $p_k = F(k) - F(k-1)$ and split the sum:

$$\sum_{k=1}^{K} c_k \bigl(F(k) - F(k-1)\bigr) = \sum_{k=1}^{K} c_k F(k) - \sum_{k=0}^{K-1} c_{k+1} F(k).$$

The term $k = K$ of the first sum is $c_K F(K) = c_K$, and the term $k = 0$ of the second is $c_1 F(0) = 0$. The remaining terms, $k = 1$ to $K - 1$, pair up as $(c_k - c_{k+1})F(k)$, which gives the formula.

Now write $g_k = c_{k+1} - c_k > 0$ for the gaps and $\delta_k = F_A(k) - F_B(k)$. Subtracting the formula for B from the formula for A, the common $c_K$ cancels:

$$\text{mean}_A - \text{mean}_B = -\sum_{k=1}^{K-1} g_k\,\delta_k.$$

If every $\delta_k \ge 0$, every term of the sum is nonnegative, so the difference is at most zero whatever the gaps are. It is negative if some $\delta_k > 0$, and zero if every $\delta_k = 0$.

If instead $\delta_j > 0$ and $\delta_m < 0$, take every gap equal to 1 except $g_j = G$. The difference is $-G\delta_j - \sum_{k \ne j}\delta_k$. It is negative once $G > \sum_{k \ne j}|\delta_k| / \delta_j$, so A has the smaller mean. Taking instead every gap equal to 1 except $g_m = G$, with $G > \sum_{k \ne m}|\delta_k| / |\delta_m|$, makes the difference positive, so A has the larger mean. $\square$

Check your understanding. The cumulative shares of the adults aged 18 to 39, at "excellent", "very good or better", "good or better" and "fair or better", are 0.166, 0.468, 0.834 and 0.982. Those of the adults aged 60 and over are 0.073, 0.308, 0.690 and 0.942. Could some increasing coding, however lopsided, give the older group the better mean?

No. The younger group's share is larger at every one of the four categories, so every $\delta_k$ is positive (taking A as the younger group). The proposition then says the younger group's mean is smaller, that is, better, under every increasing coding. The three codings in the table are three instances, and no fourth can disagree.

When one cumulative distribution lies entirely above another, the first group is said to be stochastically smaller (here, healthier), and every reasonable summary of an ordinal rating agrees on the comparison. When the distributions cross, the ordinal data alone do not say which group is better off. A mean of codes appears to settle the question only because the codes have quietly decided it.

This is the sense in which the mean is not a statistic of ordinal data. It is also why the comparisons of ordinal outcomes in Nonparametric Methods are built on ranks, which use the order and nothing else.

Ordinal categories made from measurements

Self-rated health is ordinal from the start. Ordinal data also often come from cutting a measured quantity into classes. The 2017 guideline of the American College of Cardiology and the American Heart Association sorts adults into four blood pressure categories (Carey and Whelton, 2018):

  • normal: systolic below 120 and diastolic below 80 mm Hg;
  • elevated: systolic 120 to 129 and diastolic below 80;
  • stage 1 hypertension: systolic 130 to 139 or diastolic 80 to 89;
  • stage 2 hypertension: systolic at least 140 or diastolic at least 90.

Because stage 1 and stage 2 are defined with "or", a person whose systolic and diastolic pressures point to different categories is placed in the higher one. The man on the first line of this article, at 134 over 66, is in stage 1. Applied to the mean NHANES readings:

Category Adults Share Cumulative share
Normal 2,095 0.3993 0.3993
Elevated 816 0.1555 0.5548
Stage 1 1,097 0.2091 0.7639
Stage 2 1,239 0.2361 1.0000

The median category is "elevated". These are categories of the pressure measured on the day of the examination, not diagnoses. The guideline asks clinicians to base their estimate of a person's blood pressure on the average of at least two readings obtained on at least two separate occasions (Carey and Whelton, 2018), and NHANES measures each person on one occasion. (The definition of hypertension used in Inference on Proportions also counts people taking blood pressure medication. That is why its prevalence is higher than the 0.4452 here in stage 1 or 2.)

The same blood pressure measurements at three levels of detail. Left: systolic pressure as measured, in classes of 10 mm Hg. Middle: the four ordered categories of the 2017 guideline. Right: a binary split at 130/80 mm Hg. Read from left to right: each panel is computed from the one before it, and each step discards information that cannot be recovered.
The same blood pressure measurements at three levels of detail. Left: systolic pressure as measured, in classes of 10 mm Hg. Middle: the four ordered categories of the 2017 guideline. Right: a binary split at 130/80 mm Hg. Read from left to right: each panel is computed from the one before it, and each step discards information that cannot be recovered.

The figure shows the price of the conversion. A person at 131 mm Hg and one at 180 mm Hg are both "stage 2" if their diastolic pressures are high enough. Once the categories are recorded, the difference is gone. Categories are indispensable for clinical decisions and for reporting. For analysis, it is usually better to keep the measurement and categorize at the end.

Ranks

The proposition above showed that comparisons of ordinal data are safe when they use only the order. Ranks are the tool that keeps only the order, for any data that have one.

Ranking a set of values replaces each by its position in order. Eight adults drawn from the survey (seed 3 in the script) have systolic pressures 120.0, 104.7, 124.0, 116.7, 117.3, 125.3, 118.7 and 134.0 mm Hg, rounded to one decimal. Their ranks, from smallest to largest, are 5, 1, 6, 2, 3, 7, 4 and 8. Tied values share the average of the positions they occupy (the midrank), so two values tied for third and fourth both get 3.5.

Ranks are sometimes listed as a separate type of data. They are better thought of as a transformation that turns any data with an order into ordinal data of a standard form: whatever the original scale, the ranks of $n$ values without ties are the numbers 1 to $n$.

That is their strength. A statistic computed from ranks is unaffected by:

  • any increasing change of scale;
  • the spacing of the original categories;
  • how extreme the largest values are.

Its distribution under a null hypothesis can also be found by counting. Nonparametric Methods builds the sign, signed-rank, rank-sum and Kruskal–Wallis tests on exactly this property.

Check your understanding. Suppose the eight pressures were converted to kilopascals, or the largest one, 134.0, were misrecorded as 1340. What would happen to the ranks?

Nothing. Both changes keep the order of the eight values, so the ranks stay 5, 1, 6, 2, 3, 7, 4 and 8.

The nominal and ordinal types are about categories. The other two types are about numbers that measure an amount, and they start with counting.

Discrete data

Discrete data are numerical values that can be listed one by one, usually counts: the number of people in a household, of previous pregnancies, of hospital admissions in a year, of new cases of a disease. Unlike ordinal codes, counts have meaningful distances and a meaningful zero. A household of 4 has two more people than a household of 2, and twice as many.

Means and standard deviations of counts are therefore interpretable. The two distributions most used for counts, the binomial and the Poisson, are the subject of Theoretical Probability Distributions.

People in household 1 2 3 4 5 6 7 or more
Adults 727 1,603 977 822 578 305 235

The mean household size is 3.148 people, the median 3 and the mode 2. The table also shows a common feature of real records. To protect privacy, the survey reports any household of 7 or more people as 7 (National Center for Health Statistics, 2020). The 235 adults in that column, 4.48% of the total, may live in households of 7, 9 or 12.

A value recorded as "at least 7" is said to be top-coded or censored. Its recorded value is a lower bound, and that has consequences for the summaries.

Check your understanding. Suppose we learned the true sizes of the 235 households recorded as 7. Which of the three summaries, mean, median or mode, could change, and in which direction?

Only the mean, and only upward. Every true size is 7 or more, so the true mean is at least 3.148. The median depends only on which side of 3 each person falls, and every top-coded household stays above 3. The mode, 2, keeps its 1,603 adults.

Continuous data

Continuous data come from measuring a quantity that can, in principle, take any value in an interval: blood pressure, height, weight, age, the concentration of a substance in the blood, the time until an event. The definition is about the quantity, not the record. Every recorded value has finite precision, and the recording rules leave marks on the data that a careful analyst should know about. The NHANES records show two such marks, and one pair of variables that is almost free of them.

Blood pressure is recorded in even numbers. The NHANES data-capture rules accept only even values of systolic and diastolic pressure (National Center for Health Statistics, 2020). All 15,146 systolic readings of the 5,247 adults are even, between 72 and 238 mm Hg, and only 80 distinct readings occur.

The person's value used in this course is the mean of up to three readings, so it can also be a multiple of 2/3. For instance, $(118 + 120 + 122)/3 = 120$ and $(118 + 120 + 124)/3 = 120.67$. In all, 237 distinct values occur among the 5,247 adults. The variable is continuous, but the data are a lattice of closely spaced values. Tables and Graphs for Data shows the lattice appearing in a histogram drawn with very narrow bars.

Age is recorded in whole years and top-coded. Age is continuous in time but recorded in completed years. The survey also records every participant aged 80 or older as 80, because single years of age above 80 were judged a disclosure risk. The documentation reports that the participants coded 80 had a (weighted) mean age of 85 (National Center for Health Statistics, 2020).

Among the 5,247 adults, 350 (6.67%) are coded 80. Their recorded ages pull the mean age, 49.87 years, down. If all 350 were in fact 85, the mean would be 50.20. The median, 51 years, is the same in both cases. As with household size, a summary based on order is immune to top-coding as long as fewer than half the values are affected. A summary based on the values themselves is not.

Height and weight behave almost ideally. Height is recorded to the nearest 0.1 cm and weight to the nearest 0.1 kg. That is fine enough that ties are rare and the data behave much like ideal continuous measurements.

Because continuous values are almost never exactly equal, a list of them is summarized in one of two ways. One is to group them into intervals, as a frequency table and its histogram do. The other is to compute numerical summaries of center and spread. Those are the subjects of the next two articles.

Interval and ratio scales

The course's four kinds of data merge two distinctions: whether the values are categories or numbers, and whether the numbers are counts or measurements. A finer classification of measurements, due to the psychologist S. S. Stevens (1946), separates numbers with a meaningful zero from numbers without one.

  • On an interval scale, differences are meaningful but the zero is a convention. Temperature in degrees Celsius is the standard example. The difference between 10 °C and 20 °C equals that between 20 °C and 30 °C, but 20 °C is not "twice as warm" as 10 °C, since the zero point is arbitrary. Such a scale keeps its form under any change of unit and origin, $x' = ax + b$ with $a > 0$, as when Celsius is converted to Fahrenheit.
  • On a ratio scale the zero means "none of the quantity", and ratios are meaningful. Blood pressure in mm Hg, height, weight and counts are ratio scales: a pressure of 160 is twice 80. Such scales keep their form only under multiplication by a positive constant, $x' = ax$, as when millimeters of mercury are converted to kilopascals.

Stevens (1946) arranged the four scales, nominal, ordinal, interval and ratio, in a hierarchy. Each is defined by the empirical operation it requires:

  • nominal: deciding equality;
  • ordinal: deciding greater or less;
  • interval: deciding equality of differences;
  • ratio: deciding equality of ratios.

Each is also defined by the transformations that leave it unchanged. His criterion for an appropriate statistic was invariance: a statistic is appropriate for a scale if its meaning survives every transformation the scale allows. His table lists, cumulatively, the number of cases, the mode and contingency correlation for nominal scales; the median and percentiles for ordinal scales; the mean, the standard deviation and correlation for interval scales; and the coefficient of variation for ratio scales. The demonstration with the race codes and the recoding of the health ratings above are instances of this criterion.

Check your understanding. The coefficient of variation is the standard deviation divided by the mean. Why does Stevens allow it for blood pressure but not for temperature in degrees Celsius?

Converting units on a ratio scale multiplies every value by the same $a$, so the standard deviation and the mean both scale by $a$ and their ratio is unchanged. Converting Celsius to Fahrenheit also adds 32. That shift moves the mean but not the standard deviation, so the ratio changes, and its value depends on where the arbitrary zero was put.

What the classification is for

Kind of data Examples in NHANES Summaries that respect it Graphs Where the course analyzes it
Nominal race and Hispanic origin, sex, hypertension yes or no counts, proportions, mode bar chart Inference on Proportions
Ordinal self-rated health, blood pressure category the above, median, percentiles, cumulative shares bar chart in category order Nonparametric Methods
Discrete people in the household the above, mean, standard deviation bar chart or histogram Theoretical Probability Distributions
Continuous blood pressure, height, BMI, age the above, and quantities that need a meaningful zero, such as ratios and the coefficient of variation histogram, box plot, scatter plot from Confidence Intervals on

The table is a starting point, and the statistics literature has argued about it since Stevens wrote. Stevens himself conceded the practical point. He wrote that in the strictest propriety means and standard deviations ought not to be used with ordinal scales. Yet for this "illegal" statisticizing "there can be invoked a kind of pragmatic sanction: In numerous instances it leads to fruitful results" (Stevens, 1946). He added that such means and standard deviations are in error to the extent that the intervals of the scale are unequal, and that conclusions drawn from them should be cautious.

Velleman and Wilkinson (1993) went further. They argued that the scale type of data is not fixed by the data alone, but may be determined in part by the questions asked of them and the purposes they serve. They recount Lord's (1953) story of football jersey numbers, nominal labels if anything is. A statistician happily averages them to test the accusation that the freshmen were sold unusually low numbers, and for that question the size of the numbers is exactly what matters. Velleman and Wilkinson also note that the four types do not exhaust the possibilities. Mosteller and Tukey's list of names, grades, ranks, counted fractions (such as percentages, bounded at both ends), counts, amounts and balances includes kinds of data that fit none of Stevens's scales cleanly.

The view this course takes follows from both. The type of a variable is a fact about how its values were produced and what they record. It should be the first thing an analysis establishes, because it decides which summaries and graphs describe the data faithfully and which inferential methods have their assumptions met.

But the type does not by itself dictate the analysis.

  • The race codes cannot be averaged meaningfully for any question.
  • The health ratings can be averaged when the result is checked against a coding-free comparison, like the cumulative shares above.
  • A household count can be treated as continuous when its values are many.
  • Blood pressure can be cut into categories when the question is clinical.

What matters is to know which assumption each summary adds, and to check whether the conclusion depends on it.

Corrections to the course notes

The unit's source notes list ranked data as a fifth type alongside nominal, ordinal, discrete and continuous. This article presents ranks instead as a transformation of ordered data into standard ordinal form, the way they are used in the course's nonparametric unit. The notes also say that the median is meaningful for ordinal data while the mean is not. The proposition above states the condition under which a comparison of means of ordinal codes does not depend on the coding.

What we have learned

Numbers in a data file are not all quantities. Nominal codes are names, and only counts, shares and the mode survive relabeling; a 0/1 code is the exception that turns a mean into a proportion. Ordinal codes carry order but not distance, so a comparison of average ratings is trustworthy only when one group's cumulative shares lie above the other's, as they do for the younger and older adults and do not for the two older groups. Counts and measurements support means and standard deviations, but recording rules such as top-coding and even-numbered readings leave marks that affect a mean more than a median.

Knowing the type tells us which summaries are allowed, but does not yet show us the data. How 5,247 values become a picture that reveals their shape without inventing features is the subject of the next article, Tables and Graphs for Data.

In Python

Most data files store categories as numbers, and a first step in any analysis is to tell the software which columns are categories. In pandas an ordered categorical keeps the order and refuses arithmetic:

import pandas as pd

levels = ['Excellent', 'Very good', 'Good', 'Fair', 'Poor']
health = pd.Categorical.from_codes(codes - 1, categories=levels, ordered=True)
pd.Series(health).value_counts(sort=False)   # counts in category order

Here codes holds the survey's values 1 to 5. Every number in this article is computed by tools/experiments/applied-statistics/types-of-numerical-data.py in the site's repository, which downloads the NHANES files it uses from the CDC and checks their checksums.

References

  • Carey, R. M., and Whelton, P. K. (2018). Prevention, detection, evaluation, and management of high blood pressure in adults: synopsis of the 2017 American College of Cardiology/American Heart Association hypertension guideline. Annals of Internal Medicine, 168(5), 351–358. record
  • Lord, F. M. (1953). On the statistical treatment of football numbers. American Psychologist, 8(12), 750–751. https://doi.org/10.1037/h0063675. Cited as Velleman and Wilkinson (1993) recount it. record
  • National Center for Health Statistics (2020). National Health and Nutrition Examination Survey 2017–2018: Demographic Variables and Sample Weights (DEMO_J), Blood Pressure (BPX_J), Body Measures (BMX_J) and Hospital Utilization and Access to Care (HUQ_J) data files and documentation. Centers for Disease Control and Prevention. https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2017 record · 45 pages
  • Pagano, M., Gauvreau, K., and Mattie, H. (2022). Principles of Biostatistics (3rd ed.). CRC Press. The course text. record · 35 pages
  • Stevens, S. S. (1946). On the theory of scales of measurement. Science, 103(2684), 677–680. record · 2 pages
  • Velleman, P. F., and Wilkinson, L. (1993). Nominal, ordinal, interval, and ratio typologies are misleading. The American Statistician, 47(1), 65–72. record · 2 pages