Numbers Do Not Lie, But Framing Can Mislead

Numbers Do Not Lie, But Framing Can Mislead

When people hear, "The average annual income in Japan is about 4.78 million yen," many imagine the standard income they might earn if they worked full-time in the future. Yet what that number actually means is usually narrower than imagined, and at the same time broader than imagined.

As of May 2026, the latest National Tax Agency Survey of Private Sector Salaries available is the 2024 survey. In that survey, the average salary of private-sector salary earners who worked throughout the year was 4.78 million yen. By gender, it was 5.87 million yen for men and 3.33 million yen for women. For regular employees such as full-time regular staff, it was 5.45 million yen; for non-regular employees, 2.06 million yen. Broken down further, regular male employees averaged 6.09 million yen, regular female employees 4.30 million yen, non-regular male employees 2.71 million yen, and non-regular female employees 1.74 million yen.

At this point, it is already clear that the single sentence "the average annual income is 4.78 million yen" is not enough. Within the same survey, we have 4.78 million, 5.87 million, 3.33 million, 5.45 million, 2.06 million, 6.09 million, 4.30 million, 2.71 million, and 1.74 million yen. These numbers create completely different impressions. None of them is false. But the impression changes greatly depending on which number is placed in front.

The first thing to examine is the population. A population is the range of people or organizations that a statistic covers. The National Tax Agency survey covers people who receive salaries from private-sector establishments, and it includes not only company employees but also corporate officers, part-time workers, and temporary workers. Public servants are not included. There is also the condition of having worked throughout the year. Salaries include bonuses, but do not include non-taxable commuting allowances and similar payments.

In other words, a number described as "the average annual income of Japanese people" may actually be "the total salary paid to private-sector salary earners who meet certain conditions, divided by the number of those people." That is a large difference. The figure a university graduate entering a private company as a regular full-time employee in an urban area wants to know is not the same answer as an average that also includes people working a few days a week, part-time workers, corporate officers receiving executive compensation, and non-regular workers.

That is why the topic of average income is not merely about money. It is an entrance to the ability to read data.

The Average Does Not Necessarily Represent the Typical Person

An average has strong persuasive power because it represents a whole group with one number. But an average is not a number that shows the "ordinary person." It is only the result of adding all values and dividing by the number of people.

Suppose 10 people earn 2.5 million yen, 2.8 million yen, 3.0 million yen, 3.2 million yen, 3.4 million yen, 3.6 million yen, 3.8 million yen, 4.0 million yen, 4.2 million yen, and 20 million yen. The average for these 10 people is 5.05 million yen. But 9 of the 10 do not reach 5.05 million yen. The average is correct, but it is hard to say that it represents the lived reality of this group.

In distributions such as annual income, where a small number of high earners stretch the upper tail, the mean is often higher than the median. The median is the value that comes in the middle when everyone is arranged from lowest to highest. If we also look at the mode, the income band with the largest number of people, a different picture appears again.

In the National Tax Agency's 2024 data, the largest salary bracket overall was more than 3 million yen and up to 4 million yen, with 8.26 million people, or 16.1% of the total. The next largest was more than 4 million yen and up to 5 million yen, with 7.87 million people, or 15.3%. Among men, the largest group was more than 4 million yen and up to 5 million yen, with 4.93 million people, or 16.9%. Among women, the largest group was more than 2 million yen and up to 3 million yen, with 4.21 million people, or 19.0%.

When these numbers are placed side by side, the impression created by the headline "average salary: 4.78 million yen" changes. The average may be 4.78 million yen, but the largest group is in the 3-million-yen range. The peak differs between men and women. It differs between regular and non-regular employment. It differs by industry. The average salary in electricity, gas, heat supply, and water was 8.32 million yen; in finance and insurance, 7.02 million yen; in information and communications, 6.60 million yen. By contrast, accommodation and food services averaged 2.79 million yen. Even within the same box called "private-sector salary earners," the inside of the box is not uniform.

When reading numbers, it is necessary to ask not only "what does this number represent?" but also "what has been mixed together?"

Official Statistics Usually State Their Definitions. The Problem Is That Headlines Remove Them

The important point here is that government and public statistics are not necessarily made carelessly. On the contrary, official statistics usually describe their target population, exclusions, and definitions in detail. The problem is that these definitions are stripped away when the statistics become news headlines, social media posts, sales materials, or advertising copy.

The National Tax Agency statistics explain which establishments are covered, what salary means, who counts as a person who worked throughout the year, and what an "otsuran" taxpayer means. An otsuran taxpayer is a person who receives salary from multiple payers or otherwise falls under a different treatment from ordinary year-end tax adjustment. When these definitions are read, the true identity of the number begins to appear.

But the words that circulate through society are short: "average annual income: 4.78 million yen." That sentence walks around on its own. This is where deception is born. The number itself is not the lie. The conditions are removed from the number.

This structure is not limited to annual income. When reading the news, what matters is not only what is being reported. It is also what is not being reported. The same is true when looking at graphs. We must look not only at what is displayed but also at what is not displayed. We must read not only the numbers in the table but also the conditions pushed outside the table.

University Advancement Rates Also Change With Wording

The phrase "now almost everyone goes to university" looks a little different when checked against statistics. According to the Ministry of Education, Culture, Sports, Science and Technology's 2025 School Basic Survey, the advancement rate to higher education institutions, including universities, junior colleges, and specialized training colleges, was 85.4%. That is very high. But the undergraduate advancement rate for four-year universities was 58.6%. The share of high school graduates who entered employment after graduation was 13.8%, a record-low level.

So "few people start working immediately after high school" is largely correct. But "more than 90% go to university" is not correct. Does "university" mean universities only, or does it include junior colleges and specialized training colleges? Does it include only current-year graduates, or also previous-year graduates? Here again, the meaning changes unless we check the population and definitions.

This difference is also important when reading average income. Are we looking only at wages for university graduates, or also including high school graduates? Are we looking only at regular employees, or also including non-regular workers? Are we looking only at full-time workers, or also including short-time workers? If the question changes, the answer changes.

In the Ministry of Health, Labour and Welfare's 2025 Basic Survey on Wage Structure, the scheduled monthly cash earnings of general workers were 340,600 yen. For men, the figure was 373,400 yen; for women, 285,900 yen. With men's wages set at 100, women's wage level was 76.6. By educational background, high school graduates earned 297,200 yen, specialized training college graduates 313,700 yen, technical college and junior college graduates 321,200 yen, university graduates 396,300 yen, and graduate school graduates 517,400 yen.

However, the wage in this survey is not annual income. It is the scheduled cash earnings for June 2025 and differs from annual pay that includes overtime allowances and bonuses. In addition, the main tabulation covers private establishments with 10 or more regular workers. Here too, the numbers are correct, but they become misleading if we ask the wrong question.

The Minimum Breakdown Needed to Read Annual Income

When reading annual income, at least five distinctions are necessary.

First is employment status. Is the worker regular or non-regular? In the 2024 private-sector salary survey, regular employment averaged 5.45 million yen and non-regular employment averaged 2.06 million yen, a difference of 3.39 million yen. If we only look at the 4.78 million yen average, this gap disappears.

Second is gender. The gap between 5.87 million yen for men and 3.33 million yen for women is not simply the difference between people of the same age doing the same job. Many factors overlap: years of service, occupation, management ratio, maternity and childcare leave, shorter working hours, and the share of non-regular employment. Still, the resulting gap exists in reality. Numbers do not explain the cause of discrimination in one stroke, but they also do not hide that a difference exists.

Third is age. Annual income is not the same from age 22 to 65. In the Ministry of Health, Labour and Welfare's 2025 survey, men's scheduled monthly cash earnings were near their peak at 445,600 yen for ages 55 to 59. For women, the peak was 305,700 yen for ages 45 to 49 and 55 to 59, and the rise with age was smaller than for men. If ages are mixed together, both the reality of younger workers and that of middle-aged and older workers become harder to see.

Fourth is industry. If information and communications, finance, electricity and gas, and accommodation and food services are mixed into one average, the number moves away from the reality felt on the ground. In accommodation and food services, some tabulations show that more than 60% of workers fall at or below 3 million yen in annual income. A cross-industry average dilutes differences in industrial structure.

Fifth is values other than the average. Median, mode, top 10%, bottom 10%, gender, employment status, age, and region all add dimensionality that the average alone cannot show. Reading numbers does not mean believing one value. It means viewing the same phenomenon from multiple angles.

How Dangerous Is "Four Times More Dangerous"?

A classic case in which numerical impressions change dramatically is the difference between relative risk and absolute risk.

The National Cancer Center explains that people who smoke have a lung cancer risk about 4.4 times higher for men and 2.8 times higher for women than people who do not smoke. Passive smoking also raises lung cancer risk by about 20% to 30%. This is a serious health risk and should not be treated lightly.

However, when people hear "four times," danger can expand too much in their minds. Four times means a multiplier applied to the underlying probability of developing lung cancer. It does not mean that most people who smoke will die of lung cancer.

In Japan in 2024, total deaths were 1,605,378. Deaths from malignant neoplasms, that is, cancer, were 384,111, accounting for 23.9% of all deaths. Deaths from lung cancer were 75,569. That is about 4.7% of all deaths and about 19.7% of cancer deaths.

Seen this way, the expression "smoking makes lung cancer risk four times higher" becomes more precise. Smoking is clearly dangerous. But if only the phrase "four times" is extracted, the absolute number of people and the position of lung cancer within all deaths disappear. Conversely, it is also wrong to say that lung cancer is not a big issue because lung cancer deaths are about 4.7% of all deaths. 75,569 people is an extremely large number, and if quitting smoking can reduce that risk, it has major public health significance.

The lesson is not to jump at only one side of the number. Relative risk shows the size of change. Absolute risk shows the real-world scale. One alone is not enough.

Psychologists Daniel Kahneman and Amos Tversky showed in the 1970s that human beings are poor at intuitively handling probability and base rates. People react strongly to phrases like "four times." At the same time, they easily overlook "how many out of how many" and "what was the original probability?" This is less a matter of carelessness than a habit of human cognition.

Fertility Rate Is Well Defined, Yet Still Misunderstood

Some numbers have relatively clear definitions. One example is the total fertility rate. Japan's total fertility rate in 2024 was 1.15, and the number of births was 686,173.

However, the total fertility rate is not "the average number of children actually born over the lifetimes of women currently alive." It is the number of children a woman would have if she experienced, from ages 15 to 49, the age-specific fertility rates observed in that year. In other words, it differs from completed fertility.

This number is important for reading population dynamics. If we think simply about the next generation being born from two parents, then if a level far below roughly 2 continues over the long term, the population tends to shrink. But even when looking at a fertility rate of 1.15, we cannot grasp the overall social picture unless we also consider marriage rates, age at first birth, economic conditions, childcare environments, regional differences, immigration, mortality, and other factors.

Even a well-defined number cannot explain the world by itself. A number is a map, not the terrain itself.

"No. 1 in Satisfaction" Means No. 1 in What?

Corporate advertising often uses phrases such as "No. 1 in customer satisfaction," "No. 1 market share in the industry," or "the service people most want to recommend." These are also good teaching materials for learning how to read numbers.

In 2024, Japan's Consumer Affairs Agency conducted a survey on No. 1 claims and high-rating percentage claims. The survey collected 368 claims from web advertisements and other sources, conducted a survey of 1,000 consumers, and interviewed advertisers. Of the collected claims, 275 were No. 1 claims and 93 were high-rating percentage claims. Among No. 1 claims, satisfaction-related claims stood out.

The advertiser interviews are especially interesting. According to the Consumer Affairs Agency materials, among the 15 advertisers interviewed, only one specifically understood what questions were asked, which comparison targets were selected, and which web pages were shown in the survey. The materials also pointed out problems with so-called image surveys, in which people with no actual usage experience answer based only on impressions of websites.

When we hear "No. 1," it can sound like an absolute champion. In reality, the result changes depending on how many comparison targets were included, which region was surveyed, whether respondents were actual users or non-users, how the question was written, and when the survey was conducted.

The Consumer Affairs Agency says that a No. 1 claim with a reasonable basis requires selection of comparison products, selection of respondents, fairness of survey method, and correspondence between the claim and the survey result. Put another way, if these four elements are vague, that No. 1 claim is quite risky.

There is a point here that looks like a joke but is actually important. There are many No. 1 companies in the world. But No. 1 originally means first place under one criterion. In English, one can say "one of them," but "one of number one" is a strange idea. When dozens of No. 1 claims line up, it probably does not mean "first place overall." It probably means "first place somewhere under narrowed conditions."

When looking at advertising, we need to search for the conditions before thinking, "That is impressive." Look for small print stating the survey period, survey organization, survey target, target service, comparison range, and number of respondents. That is where the real body of the claim is.

Cumulative Counts Are Not the Number of Human Beings

Another important numerical trick is the cumulative count.

If a cram school advertises "100 admissions to difficult universities," the meaning changes depending on whether this is the number of individuals or the cumulative number of acceptances. If one excellent student is accepted by the University of Tokyo, Waseda University, Keio University, Sophia University, and Meiji University, the cumulative acceptance count is 5. But the actual number of accepted students is 1. A cumulative count is not a lie. But if it is read as the number of individuals, it becomes misleading.

The same is true of YouTube views. One hundred million views does not mean one hundred million people watched. The same person may watch three, four, or ten times. Someone may leave partway through and play it again. View count means "times," not "people."

Store visitors, event attendance, app downloads, and website page views have similar structures. Downloads are not users. Page views are not visitors. Registrations are not continuing users. Followers are not buyers.

When reading numbers, we must look at the unit. Is it people, times, cases, households, contracts, accounts? If the unit changes, the world changes.

Sales, Profit, Annual Revenue, and Share Do Not Mean the Same Strength

Company numbers contain the same trap. "Annual revenue of 100 million yen" sounds large. But some businesses leave almost no profit. Conversely, annual revenue of 10 million yen with a 70% profit margin leaves 7 million yen. Sales of 10 billion yen at a 1% profit margin produce 100 million yen in profit. Sales of 1 billion yen at a 20% profit margin produce 200 million yen in profit.

Of course, corporate strength is not determined by profit amount alone. Market size, growth rate, debt, fixed costs, inventory risk, customer concentration, people, technology, brand, and regulatory environment must also be examined. Still, at minimum, sales alone do not justify the conclusion that "a large company is safe."

Market share is the same. Is it domestic share or global share? Is it value-based or volume-based? Is it shipments or actual sales? Is it only a particular category, or does it include adjacent markets? Are foreign companies included or excluded? Is it last year or the most recent quarter?

The phrase "No. 1 market share in the industry" can become true if the industry is sliced differently. A company that ranks third in a large market can be first in a subdivided category. This is not necessarily improper. But if listeners understand it as first place in the overall large market, it invites misunderstanding.

"Same as Rent" in Home Buying Is Often Not the Same

The way numbers are shown also affects major life decisions. A common expression in real estate advertising is "monthly payments comparable to your current rent."

If the monthly mortgage payment is the same as rent, buying seems like the better choice. But owning a home brings property tax, fire insurance, earthquake insurance, repair costs, management fees, reserve funds for repairs, and equipment replacement costs. For a detached house, owners must prepare for exterior walls, roof, water heater, and plumbing repairs themselves. For a condominium, there are risks related to the finances of the management association and future increases in reserve funds.

Interest rate rises, job changes, divorce, care for parents, children's education, local population decline, disaster risk, resale price, brokerage fees, and registration costs also matter. Rent and ownership cannot be compared by monthly payment alone.

"90,000 yen per month" is easy to see. But how much will be paid over 35 years? What happens if the interest rate rises by 1 percentage point? What happens if repair costs every 10 years are included? How should the cost of reduced mobility be evaluated? The core of the decision lies outside the displayed number.

The Magic of Time: Monthly, Daily, First Month Free

Units of time also change human judgment.

A service that sounds expensive at 120,000 yen per year becomes easier to accept when described as 10,000 yen per month. Even 10,000 yen per month feels lighter when described as about 333 yen per day. Conversely, even 333 yen per day becomes about 1.21 million yen over 10 years. Daily pricing lowers psychological resistance. Annual pricing shows the total.

"First month free" works the same way. Even if the first month is free, charges begin from the second month onward unless the user cancels. The essence of a subscription is not the unit price but the duration. 980 yen per month looks cheap, but five unused services cost 4,900 yen per month and 58,800 yen per year.

These numbers are not lies. Rather, they are often calculated precisely. The problem is that dividing the burden into short time units makes the long-term burden harder to see.

Numbers About AI Eliminating Jobs Also Have Populations

AI also produces strong numerical headlines: "AI will take jobs," "programmers will lose value," "white-collar work will disappear." These statements contain part of reality. But here too, what we need to examine is which jobs, which tasks, which parts of those tasks, and over what time frame are changing.

The World Economic Forum's 2025 Future of Jobs Report projected that by 2030, structural change equivalent to 22% of current jobs would occur, 170 million new roles would be created, 92 million roles would be displaced, and the net result would be a gain of 78 million jobs. It also projected that 39% of key skills required of workers would change by 2030.

The International Monetary Fund analyzed that about 40% of global employment could be affected by AI, rising to about 60% in advanced economies. But "affected" does not mean "everything disappears." Some jobs become more productive through AI, while demand for others declines. Some jobs are complemented; others are substituted.

The important point is not to see work only by occupational title. The occupation called programmer contains requirements definition, design, implementation, testing, operations, customer communication, incident response, security, documentation, and team coordination. Even if AI increases the speed of writing code, the work of deciding what a company should build, understanding business operations, aligning stakeholders, and ensuring quality and accountability remains.

We must also look at the population of demand. The amount of system development that companies have historically ordered is not necessarily the total amount of what companies actually wanted to do. Because of constraints in budget, time, people, and technology, companies may have selected only part of what they wanted. If development costs fall and speed rises, projects that were previously abandoned may become executable.

For example, suppose a company has 100 improvement ideas but, because of budget, executes only the single most important one. If AI makes development cheaper and faster, the company may move to the second, third, or tenth idea. Even if the unit price per project falls, total demand may increase, so the total amount of work does not automatically decline. Of course, this will not happen in every occupation. There will be areas where unit prices fall, areas where work disappears, areas where specialization rises, and areas where entry-level workers have a harder time entering.

The Edo-period hikyaku courier is an easy analogy. In the past, delivering a letter over a long distance was itself major labor. People ran, connected post towns, and spent time. Today, no single specialist receives enough money to support a household just to deliver one letter from Tokyo to Osaka. Postal services, railways, roads, logistics, and communications developed.

Did the disappearance of hikyaku work eliminate information transmission as a whole? No. Postal services, parcel delivery, newspapers, telephones, the internet, cloud infrastructure, data centers, smartphones, e-commerce, and video streaming emerged. The value per letter fell, but the volume of information flowing through society exploded.

AI may have the same structure. A lower unit price and the disappearance of work are not the same thing. Unit prices can fall, volume can increase, and roles can change. Therefore, AI employment debates need more than the headline "how many jobs will disappear?" They require decomposition: which tasks will be substituted, which tasks will be complemented, which demand will newly surface, and which skills will be valued more highly.

Looking Outside the Number Is Not Suspicion, But Understanding

After all of this, numbers may look dangerous. But the truth is the opposite. Read carefully, numbers supplement human intuition.

The twentieth-century statistician John Tukey emphasized the importance of exploratory data analysis in the 1970s. Data is not only something that imposes conclusions; it is also a tool for finding questions. In the nineteenth century, Florence Nightingale visualized causes of death in the Crimean War and argued for sanitary improvements in hospitals. Numbers and diagrams can deceive people, but they can also make previously invisible problems visible to society.

Darrell Huff's 1954 book How to Lie with Statistics has been read for so long not because statistics are bad, but because people can receive the wrong impression from correct numbers unless they understand how statistics are presented.

If Pocho Lab addresses this theme, the conclusion is not "do not believe numbers." It is "read numbers to the end."

If it is an average, look at the median and the distribution. If it is a rate, look at the denominator. If it is a risk, look at both relative and absolute values. If it is a ranking, look at the comparison targets. If it is market share, look at how the market was defined. If it is a cumulative count, look at the number of actual people. If it is a monthly fee, look at the total. If it is an AI forecast, look not at job titles but at tasks.

When reading news or advertisements, look not only at what is reported but also at what is not reported. When looking at graphs or tables, look not only at which values are displayed but also at which values are not displayed. Consider whether information was deliberately hidden, omitted because of limited space, or removed because it was too technical.

This attitude is not the same as doubting everything. There is no need to become paranoid. We do not need to interrogate friends in everyday conversation about the population behind every casual statement. But when buying a home, choosing work, investing, judging health risks, choosing a school, or believing a company's advertisement, it is worth looking outside the number.

Practical Questions for Reading Numbers

Finally, here are questions to use when you encounter a number.

Is the number a mean, a median, or a mode?

Who is in the population? Regular employees only, or non-regular workers too? Students included? Elderly people included? Public servants included? Actual users only, or non-users too?

What is the denominator? A survey of 1,000 people? Administrative statistics covering 100,000 people? An analysis of several hundred web advertisements?

What is the period? Annual income for 2024? Monthly wages for June 2025? A survey from the most recent month? A cumulative total over 10 years?

What is the unit? People, times, cases, households, contracts, accounts?

What is the comparison target? The previous year, the pre-COVID period, the world average, competitors in the same industry, or only a specific category?

Is it nominal or real? Even if wages rise, living conditions can worsen if prices rise more.

How does the picture change if the denominator changes? Lung cancer deaths mean different things when viewed among all deaths, among cancer deaths, or among smokers and non-smokers.

What is not displayed? Are taxes, insurance premiums, repair costs, fees, churn rates, turnover rates, failures, non-respondents, excluded companies, or unexamined products missing?

Having these questions alone changes how numbers appear.

Seeing the Truth Requires a Little Patience

Numerical magic does not always come from flashy lies. Often, it comes from omission. Employment status disappears from average annual income. Absolute numbers disappear from health risk. Comparison conditions disappear from No. 1 claims. Actual people disappear from cumulative admissions counts. Maintenance costs disappear from "same as rent." New demand disappears from AI unemployment narratives.

Human beings prefer short phrases. We want to feel that we understand with one number. But society usually cannot be explained by one number. That is why seeing the truth requires a little patience.

A number is an entrance. We must not look only at the entrance and assume we have seen the whole building. Behind the average annual income of 4.78 million yen are gender, employment status, age, industry, education, region, working hours, bonuses, non-regular employment ratios, childrearing, caregiving, and industrial structure. Behind the fourfold smoking risk are relative risk, absolute risk, mortality statistics, smoking cessation effects, and public health. Behind No. 1 claims are survey design, comparison targets, advertising budgets, respondents' experience, and question wording.

Look at what is displayed. Look at what is not displayed. Look at what is emphasized and what is pushed into small print.

This is not only a technique for avoiding deception. It is also a technique for presenting numbers honestly when explaining things to others. In sales, planning, education, research, or any other field, people who handle numbers have responsibility. Hiding information we do not want to show may bring temporary advantage. Over the long term, it destroys trust.

The ability to read numbers is a basic literacy of the modern age. It is necessary for junior high school students, high school students, university students, and adults. If we only bathe in the fast conclusions of short videos, our brains lose practice in reading conditions. We need the ability to follow premises like reading a mystery novel, search for contradictions, pick up overlooked information, and compare multiple hypotheses.

Reading data is not merely a statistical technique. It is an intellectual habit of questioning stories, constructing other stories, and moving closer to reality.

---

References

• National Tax Agency, "2024 Survey of Private Sector Salaries"
• National Tax Agency, "Overview of the Survey of Private Sector Salaries"
• Ministry of Education, Culture, Sports, Science and Technology, "Final Figures for the 2025 School Basic Survey"
• Ministry of Health, Labour and Welfare, "2025 Basic Survey on Wage Structure"
• Ministry of Health, Labour and Welfare, "Vital Statistics of Japan 2024, Final Data"
• National Cancer Center Japan, "Smoking and Lung Cancer Risk"
• National Cancer Center Japan, "Passive Smoking Raises Lung Cancer Risk in Japanese People by About 1.3 Times"
• Consumer Affairs Agency, "Survey Report on No. 1 Claims"
• World Economic Forum, "The Future of Jobs Report 2025"
• International Monetary Fund, "AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity"
• Daniel Kahneman and Amos Tversky, "Judgment under Uncertainty: Heuristics and Biases," Science, 1974.
• John W. Tukey, Exploratory Data Analysis, 1977.
• Darrell Huff, How to Lie with Statistics, 1954.