Upstream Thinking

April 14th, 2020

Dan Heath, who has written several fine books with his brother Chip (Made to Stick, Decisive, Switch, and The Power of Moments), recently wrote a book on his own that is every bit as good and important: Upstream: The Quest to Solve Problems Before They Happen.

Here’s an excerpt from the description that appears on the inside flap of the book’s dust cover:

So often in life, we get stuck in a cycle of response. We put out fires. We deal with emergencies. We stay downstream, handling one problem after another, but we never make our way upstream to fix the systems that caused the problems. Cops chase robbers, doctors treat patients with chronic illnesses, and call-center reps address customers complaints. But many crimes, chronic illnesses, and customer complaints are preventable. So why do our efforts skew so heavily toward reaction rather than prevention?

Upstream probes the psychological forces that push us downstream…Dan Heath introduces us to the thinkers who have overcome these obstacles and scored massive victories by switching to an upstream mind-set…

Upstream delivers practical solutions for preventing problems rather than reacting to them. How many problems in our lives and in society are we tolerating simply because we’ve forgotten that we can fix them?

Essentially, Upstream explains and promotes the importance of systems thinking (a.k.a. systems science). Systems thinking involves synthesis, seeing the whole and how its many parts interact to form a system. It’s the flipside of analysis, which focuses on the parts independently and too often gets lost in the trees with little understanding of the forest. When we move upstream to examine problems at their source, we become immersed in the study of complex systems. Understanding root causes—often many and complexly interrelated—is hard work, so it involves considerable time, effort, and upfront costs, but the return on that investment is considerable and the potential benefits are great.

If you work with data to solve real-world problems, I highly recommend this book. It will open your eyes to greater understanding.

Comparing COVID-19 Mortality Rates Over Time By Country

April 13th, 2020

As COVID-19 spreads its deadly effects around the world, many data analysts are struggling to track these effects in useful ways. Some attempts work better than others, however. Comparing these effects among various countries is particularly challenging. Some attempts that I’ve seen are confusing and difficult to read, even for statisticians. Here’s an example that was brought to my attention recently by a statistician who found it less than ideal:

I believe that the objectives of displays like this can be achieved in simpler, more accessible ways.

Before proposing an approach that works better, let’s acknowledge that country comparisons of deaths from COVID-19 are fraught with data problems that will never be remedied by any form of display. Even here in the United States, many deaths due to COVID-19 are never recorded. If someone with COVID-19 suffers from pneumonia as a result and then dies, what gets recorded as the cause on the death certificate: COVID-19 or pneumonia? Clear procedures aren’t currently in place. Medical personnel are focused on saving lives more than recording data in a particular way, which is understandable. This problem is no doubt occurring in every country. The integrity of the data from country to country differs to a significant degree and does so for many reasons. It’s important to recognize whenever we display this data that country comparisons will never be entirely reliable. Nevertheless, working with the best data that’s available, we must do what we can to make sense of it.

If we want to compare the number of deaths due to COVID-19 per country, both in terms of magnitudes and patterns of change over time, the following design choices seem appropriate:

  1. Assuming that we want to understand the proportional impact on countries, use a ratio such as the number of deaths per 1 million people rather than the raw number of deaths, to adjust for population differences.
  2. Aggregate the data to weekly values to eliminate the noise of day-to-day variation.
  3. Use rolling time (i.e., week 1 consists of days 1 through 7, week 2 consists of days 8 through 14, etc.) rather than calendar time, beginning with the date on which the first death occurred in each country.

The following line graph exhibits these design choices. To keep things simple for the purpose of illustrating this approach, I’ve included four countries only: the U.S., China, Italy, and Canada. Also, for the sake of convenience, I’ve relied on the most readily available data that I could find, which comes from

Most people in the general public could make sense of this graph with only a little explanation. It’s important to recognize, however, that no single graph can represent the data in all the ways that are needed to make sense of the situation. Perhaps the biggest problem with this graph is the fact that the number of weekly deaths per 1 million people per country varies so much in magnitude, ranging from over 90 at the high end in Italy to less than 1 at its peak in China, the blue line representing China appears almost flat as it hugs the bottom of the graph, which makes its pattern of change unreadable. Assuming that the number of deaths in China is accurate (not a valid assumption for any country), this tells us that COVID-19 has had relatively little effect on China overall. The immensity of China in both population and geographical space is reflected in this low mortality rate. The picture would look much different if we considered Wuhan Province alone.

Obviously, if we want to compare the patterns of change among these countries more easily, regardless of magnitude, we must solve this scaling problem. Some data analysts attempt to do this by using a logarithmic scale, but this isn’t appropriate for the general public because few people understand logarithmic scales and their effects on data. Another approach is to complement the graph above with a series of separate graphs, one per country, that have been independently scaled to more clearly feature the patterns of change. Here’s the same graph above, complemented in this manner:

With this combination of graphs, there is now more that we can see. For instance, the pattern of change in China is now clearly represented. Notice how similar the patterns in China and Italy are. From weeks 1 through 7, which is all that’s reflected in Italy so far, the patterns are almost identical. Will their trajectories continue to match as time goes on? Time will tell. Notice also the subtle differences in the patterns of change in the U.S. versus Canada. In the beginning, mortality increased in Canada at a faster rate but started to decrease from the fourth to fifth week while the pattern in the U.S. does not yet exhibit a decrease as of the sixth week. Will mortality in the U.S. exhibit a decline by week 7 similar to China and Italy? When another complete week’s worth of data is added to the U.S. graph, we’ll be able to tell.

Clearly, there are many valid and useful ways to display this data. I propose this simple set of graphs as one of them.

Display New Daily Cases of COVID-19 with Care

April 9th, 2020

Statistics are playing a major role during the COVID-19 pandemic. The ways that we collect, analyze, and report them, greatly influences the degree to which they inform a meaningful response. An article in the Investor’s Business Daily titled “Dow Jones Futures Jump As Virus Cases Slow; Why This Stock Market Rally Is More Dangerous Than The Coronavirus Market Crash” (April 6, 2020, by Ed Carson) brought this concern to mind when I read the following table of numbers and the accompanying commentary:

U.S. coronavirus cases jumped 25,316 on Sunday [April 5th] to 336,673, with new cases declining from Saturday’s record 34,196. It was the first drop since March 21.

The purpose of the Investor’s Business Daily article was to examine how the pandemic was affecting the stock market. After the decline in the number of reported new COVID-19 cases on Sunday, April 5th, on Monday, April 6, 2020, the stock market surged (Dow Jones gained 1,627.46 points, or 7.73%). This was perhaps a response to hope that the pandemic was easing. This brings a question to mind. Can we trust this apparent decline as a sign that the pandemic has turned the corner in the United States? I wish we could, but we dare not, for several reasons. The purpose of this blog post is not to critique the news article and certainly not to point out the inappropriateness of this data’s effects on the stock market, but merely to argue that we should not read too much into the daily ups and downs of newly reported COVID-19 case counts.

How accurate should we consider daily new case counts based on the date when those counts are recorded? Not at all accurate and of limited relevance. I’ll explain, but first let me show you the data displayed graphically. Because the article did not identify its data source, I chose to base the graph below on official CDC data, so the numbers are a little different. I also chose to begin the period with March 1st rather than 2nd, which seems more natural.

What feature most catches your eye? For most of us, I suspect, it is the steep increase in new cases on April 3rd, followed by a seemingly significant decline on April 4th and 5th.

A seemingly significant rise or fall in new cases on any single day, however, is not a clear sign that something significant has occurred. Most day-to-day volatility in reported new case counts is noise—it’s influenced by several factors other than actual new infections that developed. There is a great deal of difference between the actual number of new infections and the number of new infections that were reported as well as a significant difference between the date on which infections began and the date on which they were reported. We currently have no means to count the number of infections that occurred, and even if we tested everyone for the virus’s antibodies at some point, we would still have no way of knowing the date on which those infections began. Reported new COVID-19 cases is a proxy for the measure that concerns us.

Given the fact that reported new cases is probably the best proxy that’s currently available to us, we could remove much of the noise related to the specific date on which infections began by expressing new case counts as a moving average. A moving average would provide us with a better overview of the pandemic’s trajectory. Here’s the same data as above, this time expressed as a 5-day moving average. With a 5-day moving average the new case count for any particular day is averaged along with the four preceding days (i.e., five-days-worth of new case counts are averaged together), which smooths away most of the daily volatility.

While it still looks as if the new case count is beginning to increase at a lesser rate near the end of this period, this trend no longer appears as dramatic.

Daily volatility in reported new case counts is caused by many factors. We know that the number of new cases that are reported on any particular day do not accurately reflect the number of new infections. It’s likely that most people who have been infected have never been tested. Two prominent reasons for this are 1) the fact that most cases are mild to moderate and therefore never involve the medical intervention, and 2) the fact that many people who would like to be tested cannot because tests are still not readily available. Of those who are tested and found to have the virus, not all of those cases are recorded or, if recorded, are forwarded to an official national database. And finally, of those new cases that are recorded and do make it into an official national data base, the dates on which they are recorded are not the dates on which the infections actually occurred. Several factors determine the specific day on which cases are recorded, including the following:

  1. When the patient chooses or is able to visit a medical facility.
  2. The availability of medical staff to collect the sample. Staff might not be available on particular days.
  3. The availability of lab staff to perform the test. The sample might sit in a queue for days.
  4. The speed at which the test can be completed. Some tests can be completed in a single day and some take several days.
  5. When medical staff has the time to record the case.
  6. When medical staff gets around to forwarding the new case record to an official national database.

There’s a lot that must come together for a new case to be counted and to be counted on a particular day. As the pandemic continues, this challenge will likely increase because, as medical professionals become increasingly overtaxed, both delays in testing and errors in reporting the results will no doubt increase to a corresponding degree.

Now, back to my warning that we shouldn’t read too much into daily case counts as events are unfolding. Here’s the same daily values as before with one additional day, April 6th, included at the end.

Now what catches your eye. It’s different, isn’t it? As it turns out, by waiting one day we can see that reported new cases did not peek on April 3rd followed by a clear turnaround. New cases are still on the rise. Here’s the same data expressed as a 5-day moving average:

The trajectory is still heading upwards at the end of this period. We can all hope that expert projections that the curve will flatten out in the next few days will come to pass, but we should not draw that conclusion from the newly reported case count for any particular day. The statistical models that we’re using are just educated guesses based on approximate data. The true trajectory of this pandemic will only be known in retrospect, if ever, not in advance. Patience in interpreting the data will be rewarded with greater understanding, and ultimately, that will serve our needs better than hasty conclusions.

Ordinal Malpractice

March 19th, 2020

We love to put things in order. “Which college is best, second best, third best, etc., and how can I get my kid into one near the top of the list?” “I love God, Mom, America, and apple pie, in that order.” “Formal education consists of elementary school, middle school, high school, undergraduate school, and finally graduate school, if you’re lucky.” “Our best salesperson is John, second best is Mary, Sally is third, and poor Harold is at the bottom of the list.” We sometimes forget, however, that when we sequence things, even when that sequence is based on a quantitative measure (e.g., salespeople ranked by sales revenues), the order itself is merely ordinal, not quantitative. The company’s top salesperson, John, might be mediocre at best, and the second-best salesperson, Mary, might sell but a smidgen less than John or perhaps only half as much. A #2 ranking merely reveals that Mary sells less than John, not how much less.

An ordered list that appears along the axis of a graph, such as the ranked list of salespeople below, is called an ordinal scale.

An interval scale, like the one below, is quite different.

An interval scale subdivides a continuous range of quantitative values into equal intervals, in this case a range extending from 0 to 500, subdivided into intervals of 100 each. One of the most common interval scales that we use in quantitative data analysis involves ranges of time (e.g., from years 1950 through 2020) subdivided into equal periods (e.g., the 1950s, 1960s, etc.). Interval scales, by definition, are quantitative in nature; ordinal scales are not. In general, all we can say about ordinal scales is that the items have a meaningful order, nothing more. Along an interval scale, quantitative distances between adjacent items are always equal, but along an ordinal scale, distances between items typically vary.

A Likert scale is an example of an ordinal scale that is often used in social science research and surveys. Likert scales allow people to respond to questions, such as “How often do you drink more than a single serving of an alcoholic beverage in a day?”, by selecting from an ordered list such as the following:

    1. Never
    2. Seldom
    3. Occasionally
    4. Frequently
    5. Always

Notice that the items have a meaningful order, but the scale itself is not quantitative. The difference in the frequency of occurrence between “Never” and “Seldom” is not necessarily the same as the difference between “Seldom” and “Occasionally.” Item #5 (“Always”) is not five times greater than item #1 (“Never”). Even though items in ordinal scales are often labeled with numbers (e.g., “1 – Never” and “5 – Always”), the numbers only indicate a sequence (item #1, item #2, etc.), not quantities.

Ordinal scales often provide meaningful and useful ways to arrange items in a list. I often arrange the values that appear in a graph in order from low to high or high to low because it is easier to compare values when those that are close to one another in magnitude are near one another in the graph. You can see this by comparing the two graphs below: one ordered by SAT scores from high to low and the other arranged alphabetically by student names.

As I’ve already mentioned, when values are ranked, the ranking itself is not and shouldn’t be treated as quantitative. Shouldn’t, but often is.

Here are three examples of salespeople ranked by sales revenues, displayed graphically. Notice how differently the salespeople vary in sales performance among these three graphs even though the rankings are the same.

For most purposes, it is the sales revenues themselves in these examples that are important. They deserve far more attention than the rankings.

Let’s get back to Likert scales for a moment. Assigning numeric values to items in a Likert scale is not appropriate, in my opinion, but it is routinely done. For example, social science research that uses a questionnaire to measure depression among a population, based on the following Likert scale, could simply add up the numbers associated with the items to produce an overall 5-point depression score.

    1. Never
    2. Seldom
    3. Occasionally
    4. Frequently
    5. Always

Measuring people’s depression in this manner, however, does not qualify as truly quantitative.

If ordinal quantification is misleading, why is it done? Mostly, for convenience. It allows people to represent something that is difficult to measure with a simple number, but that number is inherently misleading. Sometimes this is also done for another reason: to lend Likert scales an inflated sense of accuracy, precision, significance, and objectivity when making research claims. A great deal of social science and survey-based performance reporting (e.g., customer satisfaction surveys) is based on quantified Likert scales. In my opinion, this renders any claims that are based on them suspect.

In science and data sensemaking (including data visualization), it is important to understand the difference between interval scales and ordinal scales: the former are quantitative; the latter are not. Both play a role, but their roles should not be conflated.

A Thinker’s Guide to Artificial Intelligence

March 5th, 2020

I just finished reading the book about Artificial Intelligence (AI) that I’ve been craving for years: Artificial Intelligence: A Guide for Thinking Humans, by Melanie Mitchell. More than any other book on this hot but largely misunderstood topic, this book describes AI in clear and accessible terms. It cuts through the hype to present a sane assessment with no agenda apart from a desire to inform. Reading this book, you’ll likely discover that AI is quite different from what you imagined.

Melanie Mitchell qualifies as a second-generation pioneer in the field of AI. Beginning in the mid-1980s she earned her Ph.D. in the field under the supervision of Douglas Hofstadter, the previous generation pioneer whose book Gödel, Escher, Bach: an Eternal Golden Braid inspired many to pursue AI. She continues to do research and development in AI as a professor at Portland State University and at the Santa Fe Institute. I’d wager that few people, if any, understand AI in general better than she does. In this book she explains what AI is, covers its history from inception through today, describes the approaches that have been pursued (symbolic AI, neural networks, machine learning, etc., including explanations for how these approaches work), and presents the strengths and limitations of AI in unvarnished terms. She does all of this with a practical eloquence that is rare among technology writers.

Should we be concerned about AI? You bet, but probably not for the reasons that you imagine. AI has never exhibited anything that qualifies as general intelligence (i.e., thinking as humans do), despite years of diligent effort. Will it ever? Nobody knows. In the meantime, however, we do know that computers can perform “narrow AI” tasks that are quite helpful. We should make sure that AI is only applied in ways that are truly useful and understood. If we can’t understand how AI’s results are produced, we can’t trust those results. We must also make sure that AI applications are designed in ways that are both effective and ethical. Current applications exhibit worrisome flaws. As AI researcher Pedro Domingos has said: “People worry that computers will get too smart and take over the world, but the real problem is that they’re too stupid and they’ve already taken over the world.” I agree. We can and should produce better, more useful AI technologies. Knowing that people like Melanie Mitchell are involved in the effort gives me hope—a glimmer, at least—that we just might head in that direction.

Stories in Our Time

March 3rd, 2020

We are surrounded by stories. We spend our lives navigating among and by means of stories. In our ancient past, everyone in a given place and time shared the same stories, and those stories were simple and few. They informed people’s lives and gave them meaning. When circumstances changed, which seldom happened, the stories changed as well.

Today, the stories that inform our lives are countless, often confusing, frequently confused, and locked in vicious combat with one another. This battle is tearing us apart. Stories are not intrinsically good or true. Stories are just narratives that people develop to explain, and to an increasing degree, to persuade. We must reject stories that fail to explain the world and our lives meaningfully, usefully, and truthfully. We must not tolerate stories that incite misunderstanding and hatred. We must choose our stories carefully. If we don’t, someone else will choose them for us.

Proportionally Speaking

March 1st, 2020

As data sensemakers, we spend a great deal of time examining quantitative relationships. Along with distribution, correlation, and time-series relationships, proportion is the other quantitative relationship that plays a significant role in data sensemaking. A proportion is just a relationship between two quantities. If we compare the number of friends that Sally and John each have, Sally’s 20 friends compared to John’s 10 friends is a proportion. It’s really that simple, but confusion often occurs when we communicate proportions.

Much of the confusion probably stems from the fact that proportions can be expressed in several ways: as ratios, fractions, rates, and percentages. The proportion of Sally’s 20 friends compared to John’s 10 can be written as a ratio in either of the following ways: “20 to 10” or “20:10”. This same proportion can also be expressed as “2 to 1” or “2:1”, for these ratios represent the same proportion in which the first value is double the second. This proportion can also be expressed as the fraction “20/10”. The symbol for division (i.e., /) that appears in the fraction indicates that a proportion can also be expressed as the result of division, which is called the rate. In this case, the rate of Sally’s friends to John’s is 2, because 20 divided by 10 equals 2. A rate of 1, expressed as a percentage, is 100%, so the proportion of Sally’s friends to John’s, expressed as a percentage, is 200% (i.e., the rate of 2 multipled by 100%).

All of these expressions of the proportion reveal that Sally has twice as many friends as John. Expressed as a percentage, we could also say that Sally has 100% more friends than John, for Sally’s 200% minus John’s 100% results in a 100% difference. We should express it this way cautiously, however, for people often find “less than” or “greater than” expressions of proportions confusing. When we express a greater than or less than proportion, we must remember to express only the difference between the two values.

Each expression of the proportion above treats Sally’s friends as the point of reference. To get the rate of 2, we began with Sally’s number of friends and divided that by John’s number of friends (i.e., 20 / 10 = 2). The order matters. If we instead treated John’s number of friends as the point of reference, we could express the proportion of John’s 10 friends to Sally’s 20 in each of the following ways: the ratio 10 to 20, 10:20, 1 to 2, or 1:2; the fraction 10/20 or 1/2; the rate of 0.5; the percentage 50%. We could also say that John has 50% fewer friends than Sally (i.e., John’s 50% minus Sally’s 100% equals -50%).

If John lost 8 of his friends, leaving only 2, while Sally maintained all 20 of her friends, we could say that John has 0.1 or 10% the proportion of friends that Sally has. This is fairly straightforward and clear to anyone who understands the basic concepts of rates and percentages. On the other hand, would could also say that John has 90% fewer friends than Sally (10% minus 100% equals -90%), but, as I warned previously, this isn’t nearly as straightforward and clear for many people.

In the Oxford English Dictionary (OED), the first two definitions of “proportion” are:

1. A portion, a part, a share, especially in relation to a whole; a relative amount or number.
2. A comparative relation or ratio between things in size, quantity, number, etc.

Both of these definitions fit what we’re discussing here. The sixth definition that appears in the OED, however, which is particular to mathematics, can lead to confusion.

6. MATH. A relationship of equivalence between two pairs of quantities, such that the first bears the same relation to the second as the third does to the fourth.

In other words, when comparing the ratio 1:2 to the ratio 10:20, mathematicians would not just say that they are in proportion but that they actually are a proportion. I mention this only to point out that, if you’re talking to a mathematician about proportions, you might be using the term differently, so be careful. I’ve encountered this problem myself a few times.

A few months ago, I ran across an example of a proportion gone wrong. It appeared on the PBS Newshour broadcast on September 23, 2019 in a segment titled “Judges weigh Trump’s family planning finding rule.” Near the end of the broadcast the following text appeared on the screen:

This is an example of a proportion that has been expressed as the difference between two values (i.e., the average income of Title X patients minus the income that’s defined as the poverty level) rather than more clearly as the relationship between them, but that’s not the only problem here. An income that is 150% below the poverty line makes no sense. An income can’t be more than 100% below the poverty line, for that would produce a negative value and negative income doesn’t make sense in this context. The person who wrote this text must not understand proportions, and apparently the program’s hosts were confused as well, which is all too common. I suspect that they meant to say that 78% of Title X patients have incomes that fall below 150% of the poverty line. In 2019, the U.S. federal poverty level for a family of one was $12,140, so 150% of that is $18,210. It seems reasonable that 78% of people who take advantage of Title X—people who tend to have low incomes and thus are in need of Title X assistance—made less than $18,210 for a family of one, $24,690 for a family of two, $31,170 for a family of three, and so on.

When dealing with proportions, a rate of 1 and a percentage of 100% play an important role. They both express the same equal, one-to-one proportion. In other words, the two values that are being compared are the same. For this reason, we tend to think of proportions as being less than, equal to, or greater than 1 when expressed as a rate or less than, equal to, or greater than 100% when expressed as a percentage. Consistent with the importance of 100%, there is a special type of proportional relationship that is based on 100% of something: the part-to-whole relationship. The whole is 100% of some measure (e.g., total sales revenues) and the parts are lesser percentages into which the whole has been divided (e.g., sales revenues in separate geographical regions, consisting of North, South, East, and West), which add up to 100%. When examining parts of a whole, we spend much of our time comparing the parts to one another. As such, graphical displays of part-to-whole relationships are only effective if they make it easy to compare the parts. Unfortunately, the most popular part-to-whole display—the pie chart—does this job poorly. It is difficult to compare slices of a pie. If you don’t already know why this is so, I recommend that you read my article “Save the Pies for Dessert.” As it turns out, this problem with pie charts is well understood but routinely ignored.

The ways that changes in proportions are expressed are another common source of confusion. Let me illustrate. According to a recent survey, the obesity rate among U.S. adults is now 42%. If we’re told that the obesity rate has increased 40% in the last 20 years, what was the rate in the year 2000? Think about this for a moment. In the year 2000, was the obesity rate 2% (i.e., 2% + 40% = 42%) or was it 30% (i.e., 30% * 140% = 42%)? It depends on how you interpret the words “increased 40%.” People sometimes mistakenly interpret this as a percentage point increase rather than a percentage increase. In this particular case, common sense suggests that the obesity rate must have been 30% in the year 2000, for there’s no way that only 2% of U.S. adults were obese just 20 years ago. Without this context, however, people might be confused.

This increase may be expressed in any of the following ways: “From the year 2000 to the year 2020, the obesity rate among U.S. adults…”

  • “…increased from 30% to 42%.”
  • “…increased 12 percentage points to 42%.”
  • “…increased 40% to 42%.”

Which of the three expressions above would least likely result in confusion? I suspect that the first, “…increased from 30% to 42%”, is the clearest. We could, of course, state the change more thoroughly by saying “…increased 12 percentage points from 30% to 42%” or “increased 40% from 30% to 42%.” When communicating with the general public, extra care in expressing changes in proportions works best.

Communicating proportions isn’t terribly difficult if we’re aware of how people might misinterpret them and take care to express them clearly. If you’re proportionally challenged, I hope this helps.

Logarithms Unmuddled

February 21st, 2020

I often write about topics that I myself have struggled to understand. If I’ve struggled, I assume that many others have struggled as well. Over the years, I’ve found several mathematical concepts confusing, not because I’m mathematically disinclined or disinterested, but because my formal training in mathematics was rather limited and, in some cases, poorly taught. My formal training consisted solely of basic arithmetic in elementary school, basic algebra in middle school, basic geometry in high school, and an introductory statistics course in undergraduate school. When I was in school, I didn’t recognize the value of mathematics—at least not for my life. Later, once I became a data professional, a career that I stumbled into without much planning or preparation, I learned mathematical concepts on my own and on the run whenever the need arose. That wasn’t always easy, and it occasionally led to confusion. Like many mathematical topics, logarithms can be confusing, and they’re rarely explained in clear and accessible terms. How logarithms relate to logarithmic scales and logarithmic growth isn’t at all obvious. In this article, I’ll do my best to cut through the confusion.

Until recently, my understanding (and misunderstanding) of logarithms stemmed from limited encounters with the concept in my work. As a data professional who specialized in data visualization, my knowledge of logarithms consisted primarily of three facts:

  1. Along logarithmic scales, each labeled value that typically appears along the scale is a consistent multiple of the previous value (e.g., multiples of 10 resulting in a scale such as 1, 10, 100, 1,000, 10,000, etc.).
  2. Logarithmic scales make it easy to compare rates of change in line graphs because equal slopes represent equal rates of change.
  3. Logarithmic growth exhibits a pattern that goes up by a constantly decreasing amount.

If you, like me, became involved in data sensemaking (a.k.a., data analysis, business intelligence, analytics, data science, so-called Big Data, etc.) with a meagre foundation in mathematics, your understanding of logarithms might be similar to mine—similarly limited and confused. For example, if you think that the sequence of values 1, 10, 100, 1,000, 10,000, and so on is a sequence of logarithms, you’re mistaken, and should definitely read on.

Before reading on, however, I invite you to take a few minutes to write a definition for each of the following concepts:

  • Logarithm
  • Logarithmic scale
  • Logarithmic growth

In addition to definitions, take some time to describe how these concepts relate to one another. For example, how does a logarithmic scale relate to logarithmic growth? Give it a shot now before reading any further.

Regardless of how much you struggle to define these concepts and their relationships to one another, it’s useful to prime your brain for the topic. Now that you have, let’s dive in.


The logarithm (a.k.a., log) of a number is the power that the log’s base must be raised to equal that number. I realize this definition might not seem clear but hang in here with me. I promise that greater clarity will emerge. Logarithms always have a base (i.e., a number on which it is based). The most common base is 10, expressed as log10, but any number may serve as the base. To determine the log10 value of the number 100, we must determine the power of 10 that equals 100. What this means will become clear in a moment through an example, but before getting to that, let’s review what raising the power of a number means in mathematics.

Raising a number to a power involves multiplying the number by itself a specific number of times. The power indicates how many instances of the number are multiplied. For example, 10 to the power of 3, written as 103 (the 3 in this case is called the exponent), involves multiplying 10 * 10 * 10, which equals 1,000. Raising a number to the power of 1 involves only one instance of that number—there is nothing to multiply—so the number remains unchanged. For example, 101 remains 10. Raising a number to the power of 2 involves multiplying two instances of that number, so 102 is 10 * 10, which equals 100. In these examples so far, the only time multiplication wasn’t involved was with the power of 1. Multiplication is also not involved whenever the exponent is zero or negative. In those cases, raising the power of a number involves division. For example, with the power of 0, rather than multiplying instances of the number by itself, we divide the number by itself, so 100 is equal to 1, for 10 / 10 = 1. Here’s a list of values that result from raising the number 10 to the powers of 0 through 6:

Now that we’ve reviewed what it means to raise a number to a particular power, we can get back to logarithms. Remember that the log of a number is the power that the log’s base must be raised to equal that number. So, to find the log2 value of the number 8, we must determine the power of 2 (the log’s base) that is equal to 8. In other words, we must determine how many times 2 must be multiplied by itself to equal 8. Since 21 = 2 and 22 = 4 (i.e., 2 * 2 = 4) and 23 = 8 (i.e., 2 * 2 * 2 = 8), we know that log2 of 8 is 3. Given this procedure, what is the log10 value of 100? It is 2, for 10 must be raised to the power of 2 (i.e., 10 * 10) to equal 100. What’s the log2 of 64? It is 6, for 2 must be raised to the power of 6 (i.e., 2 * 2 * 2 * 2 * 2 * 2) to equal 64.

So far, we’ve only dealt with logs that result in nice, round numbers, but that isn’t always the case. For example, what is the log2 of 100? The log2 of 64 is 6 and the log2 of 128 is 7, so the log2 of 100 is somewhere between 6 and 7. When expressed to eight decimal places, the log2 of 100 is precisely 6.64385619. What is the log10 of 5? It must be less than 1, because 5 is less than 10. The precise answer is 0.698970004.

Have you ever examined a list of the logarithms associated with an incremental sequence of numbers? Doing this is revealing. Here’s a list of the log2 values for the numbers 1 through 32, with an additional column that shows the proportional relationship between log2 values and the numbers on which they’re based:

Notice that, other than the log2 value of the number 3 (i.e., 1.584963, which is 52.832% of 3), as we read down the list, each log is a decreasing percentage of the number on which it is based. Keep this fact in mind. It will come in handy as we examine logarithmic scales and logarithmic growth.

Logarithmic Scales

A logarithmic scale (a.k.a., log scale) is one in which equal distances along the scale correspond to equal logarithmic distances. Because of the nature of logarithms, each number that typically appears along the scale is a consistent multiple of the previous number. The example below includes a log10 scale along the Y axis.

Along a log10 scale, because the base is 10, each number is 10 times the previous number. The example above begins at 1, but it could begin at any number. For example, a log10 scale could begin at 40 and continue with 400, 4,000, 40,000, and so on, each ten times the previous. A log2 scale that begins with 1 would continue with 2, 4, 8, 16, and so on, each two times the previous. Unlike a linear scale in which the intervals from one number to the next are always equal in value, such as 0, 10, 20, 30, 40, etc., along a log scale the intervals (i.e., the quantitative distances between the labeled values) consistently increase in value, each time multiplied by the base.

The numbers 1, 10, 100, 1,000, 10,000, 100,000, and 1,000,000 in the graph above correspond to logarithms with a base of 10, but those numbers are not themselves logarithms. Instead, they are the numbers from which the logarithms were derived. Here’s the scale that appears along the Y axis of the graph above, this time with the actual log10 values 0 through 6 labeled in addition to the numbers 1 through 1,000,000 from which those logarithms were derived.

We usually label the log scales with the numbers from which the logarithms were derived rather than the logarithms themselves because the former are typically more familiar and useful.

Another characteristic of a log scale that reinforces its nature bears mentioning, which I’ll illustrate below by featuring a single interval only along the Y axis of the graph shown previously.

Notice that the minor tick marks between 1,000 and 10,000 get closer and closer together from bottom to top. This is easier to see if the scale is enlarged and the minor tick marks are labeled, as I’ve done below.

Each interval from one tick mark to the next (1,000 to 2,000, 2,000 to 3,000, etc.) consistently covers a numeric range of 1,000, but the spaces between the marks get smaller and smaller because the differences in the logarithms corresponding to those numbers get smaller and smaller. To illustrate this, I included a column of the log10 values that correspond to each tick mark in the example above. The decreasing distances between the tick marks correspond precisely to the decreasing differences between the log values.

Logarithmic Growth

Because the numbers that typically appear as labels on a log scale are each a consistent multiple of the previous number, if you didn’t already understand logarithms, you might assume that logarithmic growth involves a series of values that are each a consistent multiple of the previous value. Here’s an example of how that might look as a series of values:

In this example, each daily value doubles the previous value. This, however, is not an example of logarithmic growth. It is instead an example of exponential growth (a.k.a., compound growth). With exponential growth, the amount of increase from one value to the next is always greater. Compound interest earned on money in a savings account is an example of exponential growth. As the balance grows, even though the rate of interest remains constant, the amount of growth in dollars consistently increases because of the growing balance. For example, 10% interest on $100 (i.e., $10) would increase the balance to $110 during the first period, and then during the next period, it would be based on $110 resulting in $11 of interest, a dollar more. Even though the interest rate remains constant, because the balance grows from one period to the next, the amount of increase grows as well.

Contrary to exponential growth, logarithmic growth (a.k.a., log growth) exhibits a constant decrease in the amount of growth from one value to the next. In other words, it always grows but it does so to a decreasing degree over time. A simple example is the distance that a bullet travels when you shoot it straight up into the air from the moment it leaves the gun until it reaches its apex, before beginning its descent. The height of the bullet starts off by increasing quickly but those increases constantly decrease in amount from one interval of time to the next due to the pull of gravity.

So, how does log growth relate to log scales? It’s not at all obvious, is it? Good luck finding an explanation on the web that’s understandable if you’re not fluent in mathematics. Here’s a graphical example of log growth, based on the log2 values for the numbers 1 through 64:

I’ve annotated this graph with lines connected to points in time when the logarithm has increased by a whole unit (i.e., from 0 to 1, 1 to 2, etc.). Starting on day 1 the log value is zero and whole-unit increases are subsequently reached on days 2, 4, 8, 16, 32, and 64. Do you recognize this pattern of days along the X axis? It matches the numbers that would appear along a log2 scale that begins with 1. In other words, the intervals between the days on which the logarithms increased by a whole unit consistently grew by a multiple of 2.

Have you noticed that the pattern formed by log growth is the inverse (i.e., flipped top to bottom and left to right version) of the pattern formed by exponential growth? To illustrate this, the graph below displays three different patterns of growth: logarithmic, linear, and exponential.

This inverted relationship between patterns of logarithmic and exponential growth visually confirms the inverted relationship that exists between logarithms and the exponential powers that are used to produce them.

Given the nature of logarithms, what do you think would happen to the shape of the blue exponential line above if I changed the scale along the Y axis from linear to logarithmic? If your answer is that the blue line would now take on the shape of logarithmic growth similar to the orange line above, you’re thinking in the right direction, but you went too far. The nature of logarithms to progressively decrease in the amount that they grow from one to the next would cancel out the nature of exponents to progressively increase in amount that they grow from one to the next, resulting in a linear pattern similar to the gray line in the graph above.

I hope you agree that these concepts actually make sense when they’re explained with clear words and examples. You still might not have much use for logarithms unless your work involves advanced mathematics, but now you’re less likely to embarrass yourself by saying something dumb about them, as I’ve done on occasion.

Context Is for Kings

February 20th, 2020

In season 1, episode 3 of the television series “Star Trek: Discovery,” when faced with a particularly wicked problem the captain of the starship Discovery speaks these words: “Universal law is for lackeys; context is for kings.” I suspect that the writers of this show consciously crafted these words for quotability. They rise to the heights of wisdom that Star Trek occasionally reaches. When I heard these words, I quickly paused the show and ran to my computer to record them because they eloquently expressed an important truth that I’ve been teaching for many years. Simple rules can serve as guides for novices, but experts operate in the more subtle realm of context.

In my work in the field of data visualization, I teach many simple rules of thumb to encourage best practices, but I’m always careful to explain why these guidelines work. I encourage my students to root their decisions in a nuanced consideration of context, not in a simplistic algorithm. When you fully understand why good rules of thumb work well in general, you can identify specific situations when they don’t apply. In other words, you can break the rules when the situation demands it.

Good teachers help people think at the conceptual level, navigating nuance, not merely at the procedural level. We humans are capable of thinking that is more sophisticated than blind obedience to algorithms. Procedural knowledge (“If A happens, then do X; if B happens, then do Y; else do Z.”) exhibits little if any understanding. Conceptual knowledge, on the other hand, allows us to master context, the realm of kings. If you want to become an expert in data visualization (or an any other field), avoid teachers, books, and courses that say “Do it this way” without explaining why. Don’t settle for being a lackey when you can become a king.

Data Sensmaking, Science, and Atheism

January 13th, 2020

I’m an atheist. Despite the stigma that most Americans still attach to atheism, I embrace it without reservation. My perspective as an atheist is tightly coupled with my perspectives as an advocate of science and a data sensemaking professional. Atheism, science, and data sensemaking all embrace evidence and reason as the basis for understanding. All three shun beliefs that are not based on evidence and reason as the enemy of understanding.

I wasn’t always an atheist. Like most Americans born in the 1950s, I was raised as a Christian—in my case, a version of fundamentalist Protestantism known as Pentecostalism. Not satisfied with a nominal commitment to religion, I was that weird kid who carried his Bible with him to high school every day. While still a teenager, I felt called by God into the ministry and pursued that calling as my initial profession. Despite a genuine commitment, however, I sometimes felt a bit uneasy about my faith. From time to time, I was faced with facts and a sense of morality that conflicted with my faith. These conflicts became increasingly difficult to ignore. In my mid-20s, after many dark nights of the soul, I pulled out of the ministry and gradually abandoned my faith altogether while searching for a new foundation to build my life upon. Eventually, science became that foundation. The transition was painful, but also exciting. I went on to study religion in graduate school from an academic perspective (comparative religion, psychology of religion, sociology of religion, history of religion, etc.) because I wanted to better understand the powerful role of religion in people’s lives and in the world at large.

After leaving the ministry, I didn’t embrace atheism immediately. I first spent a few years exploring liberal expressions of religion (e.g., Unitarianism, the Society of Friends, and even Reformed Judaism), hoping to find a like-minded community, but they all had something in common that never felt right to me. That something was faith. As I increasingly embraced science as the best path to understanding, I increasingly recognized faith as a problem. Faith delivers ready-made answers based on authority—end of story—but science encourages open-ended curiosity, continuous self-correction, and discovery.

For many years, I called myself an agnostic. Without reservation, I could say, “I don’t know if a god exists.” Since no one really knows, in a sense everyone is an agnostic, whether they admit it or not. At the time, I didn’t think of myself as an atheist because I misunderstood the term. I thought that an atheist was someone who claimed to know for sure that no gods exist. That isn’t the case. Agnosticism is an epistemological expression—it’s concerned with knowledge, or more precisely, with the lack of knowledge: “I don’t know.” Atheism, on the other hand, is an expression of belief, or more precisely, the absence of belief. An atheist says, “I don’t believe that any gods exist.” One can also embrace a slightly firmer version of atheism that declares “I believe that no gods exist.” Either way, atheism does not claim “I know for sure that no gods exist.” Agnosticism and atheism represent the same epistemological perspective: “I don’t know if any gods exist.” Atheism just goes one step further by extending a lack of knowledge to the realm of belief.

Science resists certainty; it deals in probabilities. Based on the available evidence, something is either likely or unlikely to a statistically calculated degree. I lack belief in gods because I’m not aware of any evidence for their existence. During my years as a Christian, I accepted the Christian god’s existence as a matter of faith. At the time, I made this leap to make sense of the world, but I no longer need faith in a god to make sense of the world or my role in it. If evidence for a god’s existence ever emerges, I’ll reconsider my position.

It bears mentioning that, just as everyone is in a sense an agnostic, whether they realize it or not, everyone is also an atheist. Even if you’re a religious fundamentalist, as I was, you’re also an atheist. This is because, while you believe in your god, you don’t believe in other gods. In other words, in respect to most of the gods that people believe in—all but your own—you’re an atheist. In this respect, you and I are a lot alike. We only differ in that I include one more god on my list of those that I don’t believe in.

Religions codify faith-based beliefs. They declare what is true about the world, about humans, about our role in the world, and, of course, about the role of supernatural beings. They do so without evidence. Faith discourages curiosity and the search for truth. As Richard Dawkins wrote, “One of the truly bad effects of religion is that it teaches us that it is a virtue to be satisfied with not understanding” (The God Delusion, 2008, page 126). As a data sensemaking professional, my commitment to reason and evidence as the basis for understanding puts me at odds with faith.

We can thank the late Harvard evolutionary biologist Stephen J. Gould for the conceptual basis on which many scientists and data sensemakers who are also religious reconcile these conflicting perspectives. Gould proposed that science and religion occupy two “Nonoverlapping Magisteria.” I admire Gould’s work greatly. He was a marvelous scientist who did a great deal to popularize science, but I find this awkward construction intolerable. According to Gould, science has its domain, religion has its domain, and the two don’t overlap. Furthermore, these two domains should respect one another and consistently stick to their own distinct areas of expertise. As explained by Adam Neiblum in his book Unexceptional when describing Gould’s position:

Each magisteria has its own epistemic foundation, each fulfilling a different role in human needs and affairs. Science, epistemologically founded on empirical observation, evidence, data and reason, necessarily deals with facts about the world, while religion, epistemologically founded on personal revelation and faith, deals with values and morality, which have nothing to do with matters of fact about the world. (Unexceptional: Darwin, Atheism and Humanity, Adam Neiblum, 2017, p. 166)

Since the emergence of modern science, it and religion have always co-existed uncomfortably. I suspect that Gould wanted to make science more palatable for religious folks—the majority of Americans—so he separated science and religion into exclusive, non-competing realms.

According to a Pew Research Center survey of scientists (specifically members of the American Association for the Advancement of Science), only 33% believe in a god and over 40% identify themselves as atheists or agnostics (“Scientists and Belief,” Pew Research Center, November 5, 2009, This is extraordinary in light of the following statistics, also reported by Pew:

The vast majority of Americans (90%) believe in some kind of higher power, with 56% professing faith in God as described in the Bible and another 33% saying they believe in another type of higher power or spiritual force. Only one-in-ten Americans say they don’t believe in God or a higher power of any kind. (“Key Findings about Americans’ Belief in God,” Pew Research Center, April 5, 2018,

The correlation between scientific work and atheism, while extraordinary, is not surprising. Pursuit of science is not necessarily responsible for a lack of theistic belief, but my own exposure to science definitely influenced my departure from theism, and to a great degree.

There is a fundamental problem with Gould’s concept of Nonoverlapping Magisteria: it isn’t scientific. Science is definitely concerned with religion’s claims that the world was created by a god and that supernatural entities (gods, angels, spirits, demons, leprechauns, dead people, etc.) continue to intervene in the world’s affairs. Religion is definitely agitated by the fact that more and more of its territory is being reduced by scientific discoveries. This conflict cannot be defined out of existence. To do so defies the tenets and methods of science.

I reject the notion that morality is the rightful and exclusive domain of religion. Morality does not require religion. To say that it does makes morality an obligation that’s imposed on us by an external authority rather than a personal choice. I am no less moral as an atheist than I was as a Christian. Actually, I am more moral, for my behavior is based entirely on a personal sense of good behavior, never on a belief that I must behave in certain ways because a god demands it. When I was religious, my morality was governed, at least in part, by fear. You don’t dare piss off a god.

As religions develop, they codify morality in various ways, but they don’t create it. Morality began to evolve in social animals before the emergence of Homo sapiens. Certain ways of behaving towards others naturally evolved as moral instincts in all social animals, not just humans. Altruism, justice, and fairness are exhibited quite naturally in our species and in several others as well.

If you’re a scientist, or similarly, if you’re a data sensemaking professional, and you’re also religious, you must come to grips with the conflict that exists between these perspectives. You must divide your life, as Gould proposed, into two distinct realms. You can’t allow your willingness to accept things on faith to influence your work. Professionally, you must always go where the evidence leads you. If you do this successfully in your work, it may become increasingly difficult to do otherwise in your personal life.

Despite the stigma about atheism that still persists, an increasing number of people embrace it as a reasonable position. This is especially true among the young. They are less militant about it than my generation, however. Unlike my generation, many of them haven’t needed to claw their way out of religion. Atheism simply makes sense to them and has from an early age.

As with almost everything that I write about in this blog, this article was prompted by a particular event. Not long ago I was approached by the business school of a nearby religiously affiliated college to help them put together a data analytics program, and potentially, to also teach in the program, so I reviewed their website to find out just how religious they were. Despite the fact that members of the denomination that founded and runs this college are often quite liberal and known for their work as social activists, I found that this college is quite fundamentalist in its statement of faith. Here it is, word for word:

The Trinity
We believe in one eternal God, the source and goal of life, who exists as three persons in the Trinity: the Father, the Son, and the Holy Spirit. In love and joy, God creates and sustains the universe, including humanity, male and female, who are made in God’s image.
God the Father
We believe in God the Father Almighty, whose love is the foundation of salvation and righteous judgment, and who calls us into covenant relationship with God and with one another.
God the Son
We believe in Jesus Christ, the Word, who is fully God and fully human. He came to show us God and perfect humanity, and, through his life, death, and resurrection, to reconcile us to God. He is now actively present with us as Savior, Teacher, Lord, Healer, and Friend.
God the Holy Spirit
We believe in the Holy Spirit, who breathed God’s message into the prophets and apostles, opens our eyes to God’s Truth in Jesus Christ, empowers us for holy living, and carries on in us the work of salvation.
We believe that salvation comes through Jesus Christ alone, to whom we must respond with repentance, faith, and obedience. Through Christ we come into a right relationship with God, our sins are forgiven, and we receive eternal life.
The Bible
We believe that God inspired the Bible and has given it to us as the uniquely authoritative, written guide for Christian living and thinking. As illumined by the Holy Spirit, the Scriptures are true and reliable. They point us to God, guide our lives, and nurture us toward spiritual maturity.
The Christian Life
We believe that God has called us to be and to make disciples of Jesus Christ and to be God’s agents of love and reconciliation in the world. In keeping with the teaching of Jesus, we work to oppose violence and war, and we seek peace and justice in human relationships and social structures.
The Church
We believe in the church as the people of God, composed of all who believe in Jesus Christ, who support and equip each other through worship, teaching, and accountability, who model God’s loving community, and who proclaim the gospel to the world.
Christian Worship
We believe Christ is present as we gather in his name, seeking to worship in spirit and in truth. All believers are joined in the one body of Christ, are baptized by the Spirit, and live in Christ’s abiding presence. Christian baptism and communion are spiritual realities, and, as Christians from many faith traditions, we celebrate these in different ways.
The Future
We believe in the personal return of Jesus Christ, in the resurrection of the dead, in God’s judgment of all persons with perfect justice and mercy, and in eternal reward and punishment. Ultimately, Christ’s kingdom will be victorious over all evil, and the faithful will reign.

Wow. This is an incredible statement of faith. I mean this quite literally: it isn’t credible. Not a shred of verifiable evidence exists for any of these assertions, but faculty members at this college must affirm these articles of faith in writing. Obviously, I can’t make this affirmation. I pointed this problem out to my contact at the college and she suggested a way to get around it. I didn’t want to offend her, but I had to make it clear that I could not affiliate myself with a faith-based religious organization, even if they allowed it. What I didn’t say was that I found the college’s statement of faith frightening. It brought back vivid memories of the faith-based beliefs that I fought hard to abandon when I was young.

This encounter left me wondering how people of faith manage to avoid the pitfalls of this orientation when making sense of data or doing science. The cognitive dissonance must be exhausting. I have no desire to offend anyone who navigates this tension, but I am genuinely concerned that it affects the work. Faith primes us to accept certain things as true, without question, regardless of the evidence. This is never a good approach to data sensemaking or science. The magisteria definitely overlap. The conflict is real. If you’re religious and also a scientist or a data sensemaker, you must navigate these conflicting perspectives with care. I, for one, couldn’t do it.