Chapter 1 Variation and covariation

1.1 Accounting for variation

The 560,000 murders-per-year statistic surprisingly tells us nothing at all about how to cut the murder rate or what are the causes of high and low murder rates. For that, instead of refering to a summary we need to examine the variability itself. For example, we might look at the variation from country to country of the number of murders per year, as in Table 1.1.

Table 1.1: The number of murders in each of 290 countries in 2004

murders
2596
253
63
235
8
1007
164
73
57250
235
… and so on for 290 rows altogether.

The variation from country-to-country in the number of murders is huge. Just by scanning the 10 values displayed, you can see the range is at least 8 to 57,250. The high end of the range is 7000-times larger than the low end.11 In Chapter 5 we’ll see why obscure measures of variation like the standard deviation or summary interval are to be preferred to min-to-max range.

It’s natural when looking at data like Table 1.1 to try to account for the variation, to explain why some values are so large and others so small. For data to give us any traction in accounting for variation, we need more data – not necessarily more countries but more information about each country. Table 1.2 identifies the countries by name.

Table 1.2: Identifying the countries in the 2004 number-of-murders data.

country	murders
Argentina	2596
Australia	253
Austria	63
Azerbaijan	235
Bahrain	8
Belarus	1007
Belgium	164
Bosnia and Herzegovina	73
Brazil	57250
Bulgaria	235
… and so on for 290 rows altogether.

If you are familiar with the various countries, you may start to form hypotheses about the variation. For instance, Brazil is very large in both area and population, while Bahrain is small. There are many other possible explanations. The Economist article listed a few: “fragile government; guns and fighters left over from wars; families broken up and forced into the city by rural violence and poverty; drugs and organised crime that police cannot or will not confront; and large numbers of unemployed young men.” These explanations are not mutually exclusive. Each factor might contribute in its own way.

An important statistical task in data science is untangling the various influences so that they can be examined individually. For this, we need to know additional characteristics of each country, for instance population, land area, and so on. These additional characteristics with which we seek to explain the country-by-country variation in the number of murders are called explanatory variables.22 Other terms in common use include predictor variables and independent variables. The characteristic that we are trying to explain, in this example the annual number of murders, is called the response variable. The word “variable” refers to the characteristic showing variability. Unlike mathematical algebra, where a variable means an “unknown” (e.g. x) whose value is to be found, in statistics, a variable is a characteristic to be measured and stored as data.

Table 1.3 shows the number of murders in each country along with the country’s population and surface area (in square km). As would be expected, both the population and surface area vary from country to country. In accounting for the variation in murders, we want to look at the covariation of the response variable (murders) and the explanatory variables (here, population and surface area).

Table 1.3: The number of murders in 2004 along with two variables that might account for the country-by-country variation in the number of murders.

country	murders	population	surface_area
Argentina	2596	38728778	2780400
Australia	253	19985475	7741220
Austria	63	8196624	83870
Azerbaijan	235	8466304	86600
Bahrain	8	807989	730
Belarus	1007	9695791	207600
Belgium	164	10492643	30530
Bosnia and Herzegovina	73	3825872	51210
Brazil	57250	186116363	8514880
Bulgaria	235	7742740	111000
… and so on for 290 rows altogether.

Later chapters will introduce many techniques for looking at covariation between response and explanatory variables. Here, we’ll look at two simple methods.

First, we’ll adjust the number of murders for the population size. Here, the adjustment is as simple as dividing the number of murders by the population, to get the “murder rate” per person. This number will be very small – thankfully only a very small fraction of the population is murdered! So, as a matter of convention, we’ll report the murder rate as the number of murders per 100,000 population.

Table 1.4: The murder rate – murders per 100,000 people – shows murders adjusted for the size of the population.

country	murders	population	surface_area	murder_rate
Argentina	2596	38728778	2780400	6.70
Australia	253	19985475	7741220	1.27
Austria	63	8196624	83870	0.77
Azerbaijan	235	8466304	86600	2.78
Belarus	1007	9695791	207600	10.39
Belgium	164	10492643	30530	1.56
Bosnia and Herzegovina	73	3825872	51210	1.91
Brazil	57250	186116363	8514880	30.76
Bulgaria	235	7742740	111000	3.04
Canada	446	31918582	9984670	1.40
… and so on for 167 rows altogether.

Clearly the murder rate also varies from country to country. And while it’s common sense that the number of murders depends in some sense on the population size, it’s helpful to have a way to quantify the extent to which population does (or does not) explain the number of murders. One way to do this is to look at the variation in the murder rate. Among countries with a population of at least a million, the lowest murder rate in 2004 is 0.2 per 100,000 people (Cyprus) and the highest rate is 85.7 per 100,000 people (Colombia). Thus, the highest rate is more than 400 times bigger than the smallest rate. This also is a huge amount of variation, but its much less variation than shown by the murder count, where the high rate was more than than 7000 times as large as the small rate. The reduction in variation with population adjustment is a statistical sign that population does indeed account for at least some of the variation in the number of murders. (If population accounted for all of the variation, the murder rate would be the same in all countries. It’s obviously not.)

Consider another technique for examining the covariation of a response and explanatory variable. In this example, we’ll look at whether a surface area helps to account for the variation in murder rate. One of the panels in Figure 1.1 shows the murder rate graphed against surface area. Each country is one dot in the graph.

The other five panels in Figure 1.1 show the same data, but makes use of a statistical technique called permutation. In permutation, the data are shuffled in a specific way: the response variable is shuffled and the explanatory variables are not. This shuffling breaks any possible relationship between the response and explanatory variables. To use Figure 1.1 to examine covariation, look for a pattern in the cloud of dots in each panel. Five of the panels, showing shuffled data, have no obvious pattern.

If there is significant covariation in the actual data, the panel showing the actual data will show a clear pattern and you will be able to pick that panel out from the other five. Can you? If you can’t, that’s a sign the covariation in the actual data is not any larger than created by the luck of the draw in the shuffled data. (Curious which is the actual data? It’s the panel on the lower left.)

Let’s look again at the relationship between number of murders and population that we examined earlier using the technique of adjustment. Now we’ll plot number of murders vs population and judge whether there is covariation in the actual data by whether we can pull the corresponding panel out of the line-up in 1.2.

The actual data is in the top row, center panel. If you could spot this easily, then you have identified covariation between the number of murders and the population size.

1.2 Untangling covariation

Recall that the Economist hypothesized several causes of a high murder rate, including “large numbers of unemployed young men.” Suppose we want to test this hypothesis. We can look at the murder rate as explained by the fraction of the population that is young males. Using techniques that will be introduced in Chapters 10 and @ref(effect_size), we find that a one-percent increase in the fraction of the population that is young (15-19) males is associated with about a 20% increase in the population- and age-adjusted murder rate for men. Before we conclude that the Economist’s hypothesis is right, let’s check some other possibilities. For instance, let’s look at how the murder rate is explained by the fraction of the population that is young females. The result? Almost exactly the same as when we tried to explain the murder rate by the young-male fraction of the population.

How could the Economist, a respected publication, get it so wrong? They didn’t. Common sense suggests that the fraction of the population that is young males will be almost exactly the same as the fraction that is young females. The two go hand-in-hand, as it were; they are entangled. Since the variables are almost the same, either variable can be used to account for the murder rate. Indeed, another measure of the youth of the population, the birth rate, is just as strongly associated with the murder rate. But we hardly think that babies are murderers!

To get meaningful results, we need to untangle the variables. The strategy for doing so is to use the two variables simultaneously to explain the murder rate, maybe including birth rate as an explanatory variable as well. Later chapters will introduce the methods for doing this and, importantly, show how to tell whether the data are up to the task. Rather than leave you in suspense, here’s what we find:

• Variation in birth rate accounts for a lot of the variation in murder rate: higher birth rates are associated with higher murder rates.
• One percentage point increase in the young-male population is associated with a roughly 20% increase in the murder rate.
• In contrast, a one percentage point increase in the young-female population is associated with a decrease in the murder rate by roughly the same 20%.

How you look at data can have a big impact on what you see in data. As a result, it’s important to examine data from many perspectives, to be sensitive to the ambiguities introduced when there is covariation among explanatory variables, and to seek to explain data with a solid understanding of how the world works.

1.3 Covariation and causation

In the previous example, we calculated a population-adjusted murder rate – murders per 100,000 population – and demonstrated that this has much lower variability from country to country than does the raw number of murders. This great reduction in variability tells us that the varying population of countries explains much of the variability in the number of murders.

Suppose we had done things the other way around: divided population by the number of murders and, seeing a reduced country-to-country variability, concluded that much of variation in population is explained by variation in the number of murders. This is clearly nonsense, but why? It’s because we share an understanding of what causes what. Murders are caused by murderous people. The more murderous people, the more murders. And the number of murderous people tends to be larger when the population is larger. So population causes murderous people causes murders.

What leads to reasonable conclusions is not just the analysis of data, but the analysis of data in the context of a sensible application of an understanding of what causes what.

Traditionally, statistics textbooks have suggested that the data tells all, and have avoided talking about what causes what. Indeed, one of the most famous statistical ways of quantifying covariation, the correlation coefficient is by definition exactly the same whether we’re talking about murders versus population or about population versus murders. For traditional statistics, the only setting in which it’s considered legitimate to talk about causation is an experiment, where the experimenter herself sets the value of one variable A and observes the consequent value of another variable B. Then you get to say, statistically, that A causes B.

Over the past half century, statisticians have started to move away from this very narrow perspective. The reasons are familiar to data scientists. An important role for data science is inform decision making and sensible decision making relies on getting causation right. Raising the price of a produce will lower sales, but lowering sales (by, say, reducing quality) will not raise prices. Perhaps the most famous dispute that led to embracing a broader role for causal theories in data analysis was that about smoking and lung cancer, an example we will return to in later chapters. For epidemiologists and public health officials, the question was whether to intervene to discourage smoking with, for instance, age restrictions, taxes, banning advertisements, and so on. This makes sense since smoking causes cancer. But in the 1940s and 1950s, leading statisticians disputed the causal direction of the observed relationship. Some of the techniques we’ll cover in this book were originally motivated by the need to make a strong, justifiable causal case. This was not just a matter of common sense, mathematical and statistical arguments were brought to bear to demonstrate, without doing the impossible-to-do experiment, that there had to be a direct causal connection from smoking to cancer. One of the most important developments in statistics in the past quarter century has been the development of mathematical approaches to causation, which we’ll introduce in later chapters.

Finally, a brief note about data …. In the example used so extensively in this chapter, we used data aggregated at the level of a country: murders, population, murder rates, etc. measured country-by-country. In the past, this was the only practical way of working with global data. But today it’s technically easy to look at individual events. For example, it would not be difficult to work with details about each of the half-million murders in a year: who, when, where, what and why. We might discover that such details provide great insight into the conditions that lead to murder and that it’s not so much country that matters as neighborhood.