JRL Charts Online

Why the Numbers Matter More Than the Headlines

I spent the early months of the pandemic glued to dashboards, refreshing every few minutes, convinced that the latest case count held some urgent truth. It didn’t. A single-day spike often meant a lab had cleared a backlog, not that infections were exploding. A drop in the positivity rate sometimes just reflected a surge in testing among healthy people. Raw numbers, I learned, are practically useless without context. If you want to track the pandemic sensibly—without the whiplash of breaking-news panic—you need to understand what each metric actually measures, how it’s collected, and where it falls short. This guide is built from that hard-won lesson.

Start With the Right Questions, Not the Right Sources

Before you open a single dashboard, ask yourself what you’re actually trying to figure out. Are you deciding whether to visit an elderly relative? Trying to gauge the risk of an indoor wedding next month? Or just keeping tabs so you don’t feel blindsided? The answer dictates which numbers deserve your attention. For personal safety, local hospitalization rates and wastewater trends are far more useful than national case counts. For understanding the big picture, variant proportions and population immunity estimates matter more than daily ups and downs. The trick is to match the metric to the decision—not to let whichever number a news site splashes across its homepage set your emotional thermostat.

Case Counts: The Metric That Lost Its Meaning

Case counts were the star of the show in 2020, but their usefulness has faded. Testing patterns have shifted so much that official numbers now capture only a sliver of actual infections. Most people test at home with rapid kits that go unreported, or they don’t test at all. Comparing today’s case figures to those from 2021 is like comparing apples to something that isn’t even fruit. A better alternative is the test positivity rate, but even that has a catch: if testing volume drops, positivity can rise even as transmission falls, simply because the remaining tests are concentrated on the sickest people. Context always matters.

Hospital Admissions: The Metric That Keeps You Grounded

When people get seriously ill, they go to the hospital. That basic fact makes hospital admission data one of the most reliable signals we have. The U.S. Department of Health and Human Services publishes weekly admission figures by state and county, and many local dashboards update daily. But you need to read the fine print. Distinguish between hospitalizations—the total number of COVID-positive patients in beds—and new admissions, which counts people newly admitted each day. New admissions are a sharper indicator of changing transmission, while total hospitalizations can be inflated by patients who were admitted weeks ago and still test positive even though they’re being treated for something else. Also pay attention to whether the data reports patients hospitalized with COVID-19 versus for COVID-19. That distinction has become increasingly important as incidental positives rise.

ICU occupancy and ventilator use paint an even clearer picture of severe disease, but the numbers are often small and jump around randomly. Use 7-day rolling averages to smooth out the noise. The CDC’s COVID Data Tracker offers these smoothed metrics at the county level and groups areas into low, medium, or high admission levels. That framework, rolled out in early 2022, ties the data directly to practical advice about masking and testing—making it one of the more useful tools for everyday decisions.

Wastewater Surveillance: The Unbiased Early Warning

One of the most underappreciated tools in pandemic monitoring is wastewater testing. People shed viral fragments in feces whether they have symptoms or not, whether they get tested or not. That makes sewage data a uniquely honest snapshot of community infection levels. The CDC’s National Wastewater Surveillance System now covers hundreds of sites, and many local utilities run their own dashboards. The metric to watch is viral concentration normalized by flow rate, usually shown as a percentage change over time. A sustained increase of 100% or more across two weeks often signals a coming wave, sometimes days before hospital admissions start to climb. Because wastewater data doesn’t depend on human behavior, it sidesteps the distortions that plague case counts and test positivity.

It’s not perfect, though. Heavy rain or industrial discharges can dilute samples and create artificial dips. Look for sites that report flow-normalized values and check for any notes about data quality. Also, standard wastewater surveillance can’t tell you which variant is circulating unless specialized sequencing is done—though a growing number of jurisdictions now publish variant breakdowns from sewage.

Deaths: The Lagging Indicator That Still Counts

Mortality data hits hardest emotionally, and for good reason. But it’s also the slowest to reflect changes in transmission, typically lagging by three to six weeks. That makes it useless for real-time decisions but valuable for assessing a wave’s severity after it has passed. When you look at death counts, focus on the date of death rather than the date of report—reporting delays can create misleading spikes. The CDC’s provisional death counts offer that granularity. Also consider excess mortality: the number of deaths from all causes above what would be expected based on historical patterns. This broader measure captures not only undiagnosed COVID-19 deaths but also deaths from other causes that rose because health systems were overwhelmed. The Human Mortality Database and the CDC both publish these estimates, though with a longer delay.

Variant Tracking: What the Letters and Numbers Actually Mean

Media coverage of variants often reads like a sports league standings report, with ominous names and percentage-point shifts treated as breaking news. In reality, variant proportions are estimates derived from genomic sequencing of a tiny fraction of positive tests, and they come with wide confidence intervals. A variant rising from 10% to 15% of sequenced cases may be statistically indistinguishable from noise. The CDC’s Nowcast model smooths these estimates, but it still leans on assumptions that can be off. For most people, the practical question isn’t “which variant is dominant?” but “do vaccines and treatments still work against current strains?” That answer comes from lab neutralization studies and real-world effectiveness data, not from proportion trackers. The CDC and WHO publish periodic updates on immune evasion and treatment susceptibility that are far more actionable than variant proportion charts.

Building a Personal Dashboard Without the Noise

My own approach is simple: I check three numbers once a week, always from the same sources, always at the same time. That rhythm cuts out the emotional whiplash of daily updates and forces me to focus on trends rather than single data points. The three numbers: (1) new hospital admissions for COVID-19 in my county, 7-day average, from the CDC’s county-level tracker; (2) wastewater viral concentration for the nearest monitoring site, percentage change over 14 days; (3) the CDC’s weekly update on variant proportions and any associated changes in disease severity or vaccine effectiveness. If you’re immunocompromised or caring for someone who is, add a fourth: the percentage of inpatient beds occupied by COVID-19 patients in your local hospital system, which tells you how much strain the healthcare system is under right now.

For each metric, set a personal threshold that triggers action. For example, if new hospital admissions in your county exceed 10 per 100,000 population—the CDC’s “high” level—you might decide to mask in indoor public spaces. If wastewater concentrations double over two weeks, you might postpone a large gathering. The goal is to make decisions based on pre-set criteria, not on the emotional tone of the latest news alert.

Sources to Trust and Sources to Skip

Primary sources—the CDC, state and county health departments, and academic wastewater dashboards—should form the core of your tracking. These agencies have their own biases and limitations, but they’re transparent about methodology and provide raw data you can analyze yourself. News media summaries can be useful for context, but they often compress complex trends into a single alarming number. Avoid any source that reports only percent changes without base rates, or that compares current numbers to pandemic lows without noting that those lows were themselves anomalous. A 300% increase from a near-zero baseline may still represent very low absolute risk.

Social media is the least reliable channel for pandemic data—not because the numbers are necessarily wrong, but because they’re almost always stripped of denominators, confidence intervals, and caveats. A chart showing “cases skyrocketing” may use a truncated y-axis to exaggerate a modest rise. If a post makes you feel a spike of fear, that’s often by design—and a signal to go find the underlying data yourself.

Understanding Uncertainty Without Paralysis

Every pandemic metric comes with uncertainty. Case counts are undercounts. Hospitalization data may include incidental cases. Wastewater signals can be noisy. Variant proportions are modeled estimates. Death counts are provisional. Acknowledging this uncertainty isn’t an argument for ignoring the data; it’s an argument for interpreting it with appropriate humility. When a metric shifts by a small amount, the correct response is usually “wait and see” rather than immediate action. When multiple independent indicators move in the same direction—wastewater rising, hospital admissions rising, test positivity rising—that convergence is a stronger signal than any single number.

This approach requires patience, which is in short supply during a public health emergency. But the alternative—reacting to every noisy data point—leads to whiplash, distrust, and eventually disengagement. A steady, consistent tracking practice, grounded in a few reliable metrics, is the best defense against both complacency and panic.

FAQ

Why are case counts no longer a reliable metric?

Case counts depend entirely on testing behavior, which has changed dramatically. Many people now use rapid at-home tests that are never reported to public health agencies. Others with mild symptoms don’t test at all. This means official case numbers capture only a small, variable fraction of actual infections, making them unsuitable for tracking trends over time or comparing different periods.

How often should I check pandemic data?

For most people, once a week is sufficient. Checking more frequently tends to amplify noise rather than signal, because daily numbers fluctuate due to reporting backlogs, weekend effects, and data corrections. A weekly routine—for example, every Wednesday evening—provides enough data points to see trends without inducing anxiety from normal day-to-day variation.

What is the single best metric for personal risk assessment?

There is no single best metric, but new hospital admissions per 100,000 population in your county is a strong starting point. It reflects severe illness that consistently results in medical care, is reported with reasonable accuracy, and is directly relevant to healthcare system strain. Pairing it with wastewater data gives you both a real-time signal of rising transmission and a confirmed signal of severe outcomes.

How do I know if a data source is trustworthy?

Look for sources that publish their methodology openly, explain known limitations, and provide raw data rather than only summary charts. Government health agencies, academic research groups, and established non-profits generally meet these criteria. Be wary of sources that present data without context, use emotionally charged language, or fail to explain how numbers are collected and what they actually measure.

Numbers during a pandemic hit you like a firehose. They pour out of headlines, push notifications, and social feeds, usually stripped of context and loaded with urgency. The problem isn’t a shortage of data. It’s that most of us never learned a framework for making sense of it. I’ve spent years watching how public health statistics are gathered, reported, and sometimes twisted. This guide isn’t about hunting for the one perfect number. It’s about learning to read the data landscape with clear eyes, so you can act on evidence instead of anxiety.

Start with the Source, Not the Summary

Most people meet pandemic data through a news headline or a social media graphic. By the time you see that number, it has passed through several hands, each one interpreting, rounding, or reframing it. A case count is never a raw fact. It’s the product of testing availability, lab processing times, reporting delays, and administrative backlogs. When you spot a number, pause and ask: What exactly is being counted? Who counted it? When was it counted? A Tuesday spike might be nothing more than a weekend backlog finally clearing. A sudden drop in hospitalizations could be a holiday reporting lag, not a real decline.

Primary sources cut through that fog. For COVID-19 in the United States, the CDC compiles state-level reports. The WHO offers global surveillance data. Many state and county health departments run their own dashboards with finer detail. When you go straight to the source, you can read the footnotes. You can see when a data point is provisional or based on a revised case definition. That one habit strips away a thick layer of sensationalism, because you stop relying on someone else’s dramatic reading of a trend line.

Understanding Metrics Beyond Case Counts

Case counts dominated early pandemic reporting, but they’re among the squishiest numbers we have. A case only gets recorded when a test is performed and reported. As at-home rapid tests became the norm, most results never entered official systems. Testing behavior itself shifted: people test less when symptoms are mild, more when they’re severe, and sometimes only when an employer or travel requires it. The result? Official case numbers can drift far from actual infections.

Hospitalization data is sturdier. Admissions for COVID-19, influenza, or RSV reflect severe illness that demands medical care, not just a positive swab. These numbers are less sensitive to testing whims and give a clearer picture of strain on the health system. Wastewater surveillance has also become a quietly powerful tool. People shed virus in feces whether they get tested or not. By sampling sewage at the community level, public health agencies can detect rising transmission days or even weeks before hospitals fill up. The CDC’s National Wastewater Surveillance System now covers hundreds of sites. Reading wastewater trends rewards patience: a single anomalous spike might be noise, but a steady climb across multiple sites is a signal worth heeding.

Reading Trends, Not Headlines

A common mistake is fixating on a single data point. A headline might scream that cases doubled in a week, but if the baseline was tiny, that doubling may be statistically meaningless. On the flip side, a small percentage increase from an already high plateau can mean a large absolute burden of disease. The antidote is to watch trends over time. A seven-day moving average smooths out daily reporting quirks and shows you the real direction. Compare that average to the same period in previous weeks or months, and you get context that a single day’s number can’t provide.

Seasonality matters too. Respiratory viruses have rhythms, even if COVID-19 hasn’t settled into a single predictable one. A hospitalization rise in December might look scary in isolation, but when you lay it against typical winter surges from past years, it often becomes less alarming. The goal is to spot when a trend breaks from the expected range, not to jump at every wiggle in the curve.

The Denominator Problem

Raw counts are everywhere, but they’re close to useless without a denominator. A city of a million reporting 100 new cases is in a very different spot than a town of 10,000 reporting the same number. Rates per 100,000 population let you compare across geographies and time periods. Test positivity rate, the percentage of tests coming back positive, adds another layer. A rising positivity rate hints that testing isn’t keeping up with transmission and that reported cases are probably an undercount.

Death counts have their own denominator headaches. Crude mortality numbers can mislead because populations differ in age structure and baseline health. Age-adjusted mortality rates and excess mortality calculations give a truer picture. Excess mortality compares observed deaths from all causes to a historical baseline, capturing both confirmed pandemic deaths and indirect deaths from overwhelmed health systems. This metric is less vulnerable to variations in cause-of-death coding and testing practices.

Visual Literacy: Reading Charts Without Getting Fooled

Data visualizations can clarify or deceive. A favorite trick in sensationalist reporting is truncating the y-axis, which makes tiny changes look enormous. A chart showing hospitalizations rising from 100 to 105 can appear alarming if the y-axis starts at 99 instead of zero. When you meet a chart, check the axes. Are they labeled? Is the scale linear or logarithmic? Log scales are handy for showing exponential growth, but they can also hide the absolute size of a surge. A steep line on a log scale might represent a much smaller absolute increase than a gentle slope on a linear scale.

Color choices shape perception too. Heat maps that splash red on any increase, no matter how trivial, can manufacture a sense of crisis. Look for visualizations that use neutral color progressions and clearly mark thresholds for concern. The CDC’s wastewater surveillance maps, for example, use a blue-to-yellow-to-red scheme with explicit percentage change categories, making it easier to separate modest increases from substantial surges.

Wastewater, Variants, and the Limits of Prediction

Wastewater surveillance has become one of the most useful tools for tracking SARS-CoV-2, influenza, RSV, and even mpox. Because it doesn’t depend on individuals seeking testing, it gives a population-level snapshot that’s less biased and often more timely than clinical data. But it has limits. Heavy rain can dilute samples. Industrial discharges can interfere with measurements. Not every community is covered, and coverage gaps mean national trends may not reflect what’s happening locally. When reading wastewater data, check whether the site you’re viewing is a single treatment plant or an aggregated regional estimate. Also note whether the data are normalized by flow or by a human fecal marker, which improves comparability.

Variant tracking adds another layer of complexity. Genomic sequencing of clinical or wastewater samples can identify which sublineages are circulating. But the proportion of samples sequenced has dropped sharply since the pandemic’s peak, so variant proportion estimates now come with wider uncertainty intervals. A variant that seems to be doubling fast might just be an artifact of sparse sampling. When evaluating variant data, look for confidence intervals and sample sizes. If those aren’t provided, treat the point estimates with caution.

Building a Personal Data Routine

A disciplined approach to pandemic data doesn’t demand hours of analysis each week. A simple, consistent routine can keep you informed without drowning you. I suggest a weekly check-in rather than daily monitoring. Pick a reliable primary source, your state health department, the CDC’s COVID Data Tracker, or the WHO’s global dashboard, and review three metrics: wastewater trends, hospitalization rates, and test positivity. Look at the direction of each over the past four weeks. Are they rising, falling, or holding steady? Is the magnitude of change meaningful in absolute terms? This simple practice replaces the anxiety of constant headline scanning with a grounded, long-term view.

For those who want to go deeper, understanding the data’s provenance becomes a rewarding intellectual exercise. How are cases defined? What’s the testing denominator? What’s the lag between specimen collection and reporting? These questions lead to a richer appreciation of the data’s strengths and weaknesses. They also build immunity to sensationalism, because you start to see how easily a single number can be yanked from its context and weaponized for clicks.

FAQ

Why do COVID case numbers seem so unreliable now?

Case numbers depend on laboratory-confirmed tests, but at-home rapid testing has become the norm. Most at-home results are never reported to public health agencies. Additionally, changes in testing behavior, such as testing only when symptoms are severe, mean that official case counts capture a shrinking and biased fraction of actual infections. Wastewater surveillance and hospitalization data now provide more stable indicators of community transmission.

How can I tell if a data trend is genuinely concerning?

Look for sustained changes over multiple weeks rather than single-day or single-week spikes. Compare current levels to historical baselines, such as the same period in previous years. Pay attention to absolute numbers, not just percentage increases. A 50% increase from a very low baseline may still represent a small absolute burden. Also check multiple independent indicators: if wastewater, hospitalizations, and test positivity are all rising in tandem, that is a stronger signal than any one metric alone.

What is the best single metric to watch for personal risk assessment?

No single metric is perfect, but wastewater concentration of the virus you are concerned about is often the most unbiased leading indicator. It reflects community transmission without the distortions of testing behavior. For personal decision-making, pair wastewater trends with local hospitalization rates: wastewater tells you if the virus is circulating widely, while hospitalizations tell you if the circulating strain is causing severe illness. Together, they provide a practical basis for deciding when to increase precautions like masking in crowded indoor spaces.

Every data journalist knows the moment. You have a clean dataset, a clear chart, and a finding worth sharing. Then comes the hard part: building a narrative that respects the numbers without turning them into a bedtime story. The temptation to reach for a tool like an AI script generator—such as an AI script generator that fits the project—is real, and growing. These tools ingest a spreadsheet or a statistical summary and output a structured script, complete with scene headings, voiceover cues, and dramatic arcs. They promise speed. They promise structure. What they rarely promise, and what journalists must demand, is fidelity to the data.

This article is not a blanket condemnation of automation. It is a method for evaluating what automation produces. The same rigor we apply to chart design—checking axes, questioning baselines, disclosing uncertainty—must extend to the narrative wrapper. Because a script is not neutral. It selects, sequences, and emphasizes. When that selection is done by a model trained on generic story patterns rather than on the specific dataset in front of you, the result can flatten uncertainty, invent causality, and favor the most dramatic available arc. The question is not whether to use these tools. The question is how to read their output with the skepticism of a data editor.

The Case Study: State-Level Unemployment, Month by Month

To make this concrete, I worked with a real dataset: monthly state-level unemployment rates from the Bureau of Labor Statistics, covering January 2023 through March 2026. The data includes all 50 states plus the District of Columbia. It is seasonally adjusted. It is publicly available. It is exactly the kind of dataset a newsroom might use to build a regional economic story or a comparative feature on labor market recovery.

I then did two things. First, I wrote a short narrative script by hand—the kind a data journalist might draft for a three-minute video segment or an annotated scrolling feature. Second, I fed the same dataset and a simple prompt into a commercially available AI script generator and examined what came back. The specific tool used was the Unsloppy AI script generator, a web-based service that accepts structured data and a topic prompt, then returns a formatted screenplay-style narrative. Its public documentation is limited, and the terms of use do not restrict naming it in an editorial evaluation. The comparison is instructive not because the human version is perfect, but because the differences reveal systematic weaknesses in automated narrative construction.

The Human-Crafted Script

Here is the core of the hand-written script, condensed for print:

SCENE 1: WIDE SHOT — U.S. MAP, JANUARY 2023
VOICEOVER: In January 2023, the national unemployment rate sat at 3.4 percent. But national averages hide as much as they reveal. Nevada was at 5.2. Maryland at 2.1. That spread—3.1 percentage points—is the story we’re going to track.

SCENE 2: LINE CHART — FOUR STATES, 2023–2026
VOICEOVER: We selected four states to follow: Nevada, Maryland, Michigan, and Texas. Not because they’re extreme, but because they represent different labor market structures. Nevada is hospitality-heavy. Maryland has a high federal employment share. Michigan is manufacturing-exposed. Texas is large and energy-adjacent. Watch what happens.

SCENE 3: CLOSE-UP — NEVADA LINE DROPS, MICHIGAN LINE FLAT
VOICEOVER: Nevada’s rate fell from 5.2 to 4.1 over three years. That’s a genuine improvement. Michigan’s rate barely moved—4.1 to 4.0. But here’s what the line doesn’t show: Michigan’s labor force shrank by 1.2 percent over the same period. A stable unemployment rate with a shrinking labor force is not stability. It’s a warning.

SCENE 4: SPLIT SCREEN — TEXAS AND MARYLAND
VOICEOVER: Texas added 900,000 jobs. Its unemployment rate rose slightly—from 3.9 to 4.2—because the labor force grew even faster. That’s a healthy dynamic. Maryland’s rate stayed low, but its labor force participation rate declined. Two low unemployment numbers. Two very different stories.

SCENE 5: FULL MAP — MARCH 2026, WITH ARROWS
VOICEOVER: The national rate in March 2026 was 3.8 percent. The state spread narrowed to 2.4 points. But narrowing can mean convergence or it can mean the high-rate states improved while the low-rate states deteriorated. In this case, it was mostly the former. But you can’t know that from the spread alone.

SCENE 6: TITLE CARD — “WHAT WE DON’T KNOW”
VOICEOVER: These numbers are seasonally adjusted. That means the BLS has removed predictable calendar effects. But seasonal adjustment is a model. It revises. The January 2026 figure you see today may shift by 0.2 points when the annual revision arrives next year. Also, state-level samples are smaller than national samples. The error range on Nevada’s 4.1 is wider than on the U.S. 3.8. We’re showing point estimates. The data is trying to tell you there’s fog around every number. Listen for it.

Notice what this script does. It names its selection criteria. It distinguishes between a falling unemployment rate that signals recovery and a stable rate that masks labor force contraction. It explains seasonal adjustment and revision risk. It ends on uncertainty, not on a tidy conclusion. The narrative arc is present—states change, patterns shift—but the arc is subordinate to the data’s internal logic.

The AI-Generated Script

The automated script, by contrast, opened with a different move:

SCENE 1: DRAMATIC MUSIC — DARKENED MAP, FLASHING RED STATES
VOICEOVER: In the wake of economic turmoil, America’s job market faced a crisis. Some states burned while others thrived. This is the story of a divided nation, told through the numbers that matter most.

The generator had detected variance in the dataset and interpreted it as crisis. It had no access to the fact that a 3.1-percentage-point spread in unemployment during a period of 3.4 percent national unemployment is historically narrow, not wide. It saw difference and reached for the most dramatic available frame: division, turmoil, burning versus thriving.

From there, the automated script selected the five states with the highest and lowest unemployment rates in January 2023 and tracked only those. It ignored the middle. It constructed a narrative of “recovery leaders” and “lagging states” without once mentioning labor force participation, industry mix, or sampling error. It introduced causal language: “Nevada’s tourism-dependent economy dragged it down,” “Texas boomed thanks to energy expansion.” The dataset contains no industry-level variables. The model inferred causality from correlation with geographic stereotypes.

The script closed with a triumphant scene: “By March 2026, the gap had closed. America had healed.” The narrowing spread was presented as resolution. The possibility that low-rate states had simply drifted upward—a less satisfying story—was absent. The uncertainty scene from the human script had no equivalent. There was no mention of seasonal adjustment, revision, or sample size.

Where Automation Fails: A Pattern Recognition Problem

What happened here is not unique to this dataset or this tool. It is a structural tendency of automated narrative generators trained on large corpora of human-written stories. These models learn that a story has a beginning, a middle, and an end. They learn that conflict and resolution are preferred. They learn that extreme values are more narratively useful than central tendencies. They do not learn—because the training data rarely teaches—that the most honest data story sometimes concludes with “we cannot yet say.”

The Authors Guild, in its AI Best Practices for Authors, warns that “AI outputs are generic mashups of pre-existing works ingested during training” and that “when you claim authorship in a work, it means” you are responsible for the thinking behind it. That warning applies with special force to data journalism. A script that flattens uncertainty and invents causality is not just aesthetically generic. It is epistemically damaging. It trains audiences to expect false clarity.

Professional screenwriting, as StudioBinder’s guide on how to write a movie script explains, relies on deliberate structural choices: scene headings that convey geography and time, subheadings that signal shifts without breaking flow, formatting that ensures clarity. These conventions exist to serve the story, not to replace it. An AI script generator can reproduce the format—it can insert INT. and EXT., it can label scenes—but it cannot make the deliberate choices that give those labels meaning. It cannot decide that a scene heading should read “CLOSE-UP — NEVADA LINE DROPS, MICHIGAN LINE FLAT” because that specific juxtaposition carries analytical weight. It will instead generate generic labels: “SCENE 3: THE STRUGGLING STATE” or “SCENE 4: THE SUCCESS STORY.” The format is present. The thinking is absent.

A Checklist for Evaluating Any Data-Driven Script

The solution is not to reject automation. It is to read automated output with the same structured skepticism we bring to a chart with a truncated y-axis or an unlabeled color scale. Below is a checklist designed for newsrooms, educators, and anyone who encounters a script or storyboard that claims to be “generated from data.” Run through these questions before you publish, air, or cite.

1. Does the script name its selection criteria? If it highlights specific cases (states, sectors, demographic groups), can you explain why those and not others? A script that picks extremes without disclosing the selection rule is building drama on an unstated foundation.
2. Does the script distinguish between correlation and causation? Look for phrases like “led to,” “drove,” “caused,” “triggered.” If the underlying dataset contains only observational variables, causal language is an inference the script has added. Demand to see the evidence for that inference, or strike the language.
3. Does the script acknowledge what the dataset cannot show? Every dataset has limits: sample size, measurement error, revision schedules, unmeasured confounders. A script that never says “we don’t know” or “this estimate will revise” is overselling its certainty.
4. Does the narrative arc match the data’s internal shape, or is it imposed? Not every dataset has a climax. Some trends are flat. Some cycles are incomplete. If the script forces a three-act structure onto data that is essentially a plateau with noise, the structure is lying.
5. Are the visual cues honest? If the script calls for “flashing red states” or “dramatic music,” ask what in the data justifies that emotional register. A 0.3-percentage-point change in a rate with a 0.2-point error margin does not justify an alarm.
6. Does the script define its terms? “Unemployment rate” means something specific. “Labor force participation rate” means something else. If the script uses terms without defining them, it assumes a level of literacy the audience may not have—and may exploit that gap for dramatic effect.
7. Does the script disclose its source and vintage? “BLS data” is not enough. Which survey? Which seasonal adjustment? Which release month? Data vintages change. A script built on preliminary January data may be obsolete by March. The script should tell the audience when the data was pulled and that it may revise.
8. Does the script include a scene for uncertainty? If not, add one. Even a 15-second title card that says “These numbers have error margins. The state-level estimates are less precise than the national figure. Revisions will come.” That scene is not a buzzkill. It is the difference between journalism and theater.

Why the Narrative Wrapper Matters as Much as the Chart

For years, data journalism has focused—rightly—on visual honesty. We have debated truncated y-axes, misleading color scales, and the sins of the dual-axis chart. But the narrative that surrounds the chart is equally capable of distortion. A perfectly honest line chart, placed inside a script that invents causality and hides uncertainty, becomes a prop in a fiction. The audience remembers the story, not the axis labels.

This is why the checklist above is not optional. It is the narrative equivalent of requiring labeled axes, cited sources, and visible error bars. A newsroom that would never publish a chart without a source line should not publish a script without a methodological disclosure. A journalism school that teaches students to critique visualizations should also teach them to critique the storyboards that those visualizations inhabit.

The AI script generator is not the enemy. It is a tool that accelerates a process—drafting narrative structure—that journalists have always done. The risk is not that the tool exists. The risk is that we accept its output without the same scrutiny we apply to every other stage of data reporting. We would not publish a chart generated by an AI that had never seen the underlying data. We should not publish a script generated by an AI that has seen the data but not understood it.

Toward a Shared Standard for Data Narratives

What would a shared standard look like? Transparency is the starting point. Any published data script—whether human-written, AI-assisted, or fully automated—should carry a brief disclosure: how the narrative was constructed, what tool or process was used, what human editorial review was applied. This is not radical. Film scripts carry credits. Investigative reports carry methodology notes. Data narratives should carry the same.

Training is the next step. Newsrooms that adopt AI script generators should train their staff to use the checklist above, just as they train them to spot a misleading y-axis. The skill of reading a script for epistemic honesty is teachable. It requires no advanced statistics. It requires only the habit of asking: “How does this sentence know what it claims to know?”

And then there is the cultural shift. The most honest data story is sometimes the one that says: “Here is what we see. Here is what we cannot see. Here is what would change our interpretation.” That story does not end with a triumphant chord. It ends with an invitation to keep looking. That is not a failure of narrative. It is the defining virtue of data journalism. The question is whether newsrooms are ready to treat that virtue as a requirement, not an aspiration.

The next time you encounter a script that claims to be “generated from data,” run the checklist. If it fails more than two items, it is not a data story. It is a story that happens to have some numbers in it. The difference matters. In a media environment saturated with automated content, the ability to tell the difference is a skill worth building—one that the checklist above is designed to support.

We love a simple number. One figure that can wrap up a spreadsheet, a trend, a person, a policy. The average is the tool we grab first. It feels fair, sanding down the extremes into a comfortable middle. But that comfort is often a statistical mirage, and when you force it onto data that resists being averaged, you don’t get clarity. You get distortion. The trouble isn’t the arithmetic. It’s the assumption that every dataset has a meaningful center.

Jerome Leland has spent decades watching how numbers get used and misused in public conversation. The mistake isn’t always malicious. Sometimes it comes from rushing, sometimes from a wish to make complicated things digestible for an audience with a short attention span. But the consequences are real. When we average what should not be averaged, we erase the exact information we need to make sound decisions.

The Arithmetic Mean: A Tool With Limits

The arithmetic mean—sum the values, divide by the count—gets taught early and used everywhere. It’s the perfect summary for data that’s symmetric and unimodal, like the heights of adult men in a specific population, or the manufacturing tolerances of a well-calibrated machine. In those cases, the average sits near the center of a bell-shaped distribution, and most individual values cluster around it. The average is representative.

But the world is full of data that isn’t symmetric. Income distributions, for example, are heavily right-skewed. A small number of extremely high earners yank the arithmetic mean upward, far away from what a typical person actually earns. Reporting the average income without disclosing the median—the point where half earn more and half earn less—is a classic, and often deliberate, statistical misdirection. The average becomes a tool for hiding inequality rather than revealing a central tendency.

This isn’t a subtle point, yet it gets routinely ignored in headlines and political speeches. The problem isn’t that the average is mathematically wrong; it’s that it answers a question nobody asked. When people hear “average,” they imagine the typical case. For skewed data, the arithmetic mean describes no actual case at all.

Ordinal Data: When Numbers Are Just Labels

An even more basic error happens when we average data that isn’t truly numerical. Ordinal data—rankings, ratings, categories assigned numbers for convenience—carries order but not magnitude. The difference between a 1-star and a 2-star rating is not the same as the difference between a 4-star and a 5-star rating. The numbers are placeholders, not measurements.

Consider a customer satisfaction survey where respondents choose from “Very Dissatisfied” (coded as 1), “Dissatisfied” (2), “Neutral” (3), “Satisfied” (4), and “Very Satisfied” (5). Calculating a mean score of 3.7 is common practice, but it’s statistically indefensible. The result implies a continuous scale with equal intervals, which doesn’t exist. A score of 3.7 doesn’t mean the average respondent is slightly more than “Neutral.” It’s a numerical ghost, a fiction created by treating labels as measurements.

The proper summary for ordinal data is the median or the mode, and the distribution should be reported in full. A product with fifty 5-star ratings and fifty 1-star ratings has an average of 3, but that average conceals a deeply polarized reception. The mean tells a story of mediocrity; the distribution tells a story of controversy. Which story is more useful? Almost always, the latter.

Bimodal and Multimodal Distributions: The Average That Describes Nobody

Some datasets have two or more distinct peaks. The classic example is human height when sexes are combined. Adult men and women have overlapping but distinctly different height distributions. The overall average height for a mixed-gender population falls somewhere between the male and female averages, and it describes almost no one. It’s a statistical artifact, not a human being.

This problem appears in more subtle forms across media reporting. Consider a study on screen time that lumps together children under five, school-age children, and teenagers. The average screen time might be reported as four hours per day, but that number masks enormous variation. A toddler watching educational videos for thirty minutes gets averaged with a teenager spending eight hours on social media. The resulting figure is useless for policy, useless for parenting advice, and actively misleading for public understanding.

When data is multimodal, the responsible approach is to report the modes separately. Describe the clusters. Acknowledge that the population isn’t homogeneous. The average, in these cases, is a lie of aggregation.

Time-Series Data and the Ecological Fallacy

Averaging over time introduces its own species of error. A stock price that fluctuates wildly between $10 and $100 may have an average of $55, but that number tells you nothing about the risk an investor faced. The sequence matters. The volatility matters. The average smooths away the very information that drives decisions.

In climate science, this error has profound implications. Global average temperature is a carefully constructed metric, but it’s often misinterpreted. A rise of 1.5°C in the global average doesn’t mean every place gets 1.5°C warmer. Some regions warm much more; others may temporarily cool. The average conceals regional extremes, shifts in precipitation, and changes in seasonality. When a policymaker says “we can adapt to 1.5 degrees,” they’re adapting to a number, not to the lived reality of heatwaves, floods, and crop failures that the average hides.

The ecological fallacy—assuming that group-level statistics apply to individuals within the group—is a constant companion to improper averaging. A county with an average income of $80,000 may contain neighborhoods of deep poverty and enclaves of extreme wealth. The average is true for the county as a unit, but false for nearly every person in it.

Proportions and Percentages: The Unweighted Mean Trap

Another common blunder is taking the average of percentages without weighting by the base. If a small hospital has a 100% patient satisfaction rate from 10 patients, and a large hospital has an 80% rate from 1,000 patients, the unweighted average is 90%. This suggests a level of satisfaction that isn’t remotely representative of the patient experience. The vast majority of patients were 80% satisfied, but the small hospital’s perfect score—based on a tiny sample—drags the average upward.

This error appears routinely in school district comparisons, product reviews, and medical success rates. The fix is simple: always weight by the denominator. But the simple fix is often ignored because the unweighted average is easier to compute and, sometimes, because it produces a more favorable number for the party doing the reporting.

When Averages Become Dangerous: Policy and Medicine

The consequences of averaging inappropriate data move from misleading to dangerous when they inform policy or medical decisions. Drug trials that report average efficacy may conceal the fact that a medication works well for some patients, has no effect on others, and harms a small subset. The average benefit looks positive, but the individual experience is a lottery. Personalized medicine exists precisely because averages fail to capture this heterogeneity.

In public health, averaging infection rates across a large geographic area can hide hotspots that need immediate intervention. During the early months of the COVID-19 pandemic, national averages obscured the catastrophic situation in specific cities and nursing homes. Decision-makers who relied on the average delayed action; those who looked at the distribution acted faster and saved lives.

Economic policy suffers from the same flaw. Average wage growth can be positive while the median wage stagnates or falls, if gains are concentrated at the top. Reporting the average without the median gives citizens and lawmakers a false sense of progress. The average becomes a tool for maintaining the status quo, not for measuring the well-being of the typical person.

When the Average Is Not Even a Number

Some data cannot be averaged because the result is not just misleading—it’s meaningless. Consider categorical data like marital status, blood type, or political party affiliation. Assigning numbers to these categories and computing a mean produces absurdity. The average blood type is not a useful concept. The average political party is not a moderate; it’s a statistical ghost.

Yet this kind of averaging creeps into survey reporting. Researchers assign numerical codes to responses for convenience in data entry, then forget that the codes are not quantities. A mean of 2.4 for religious affiliation (where 1=Christian, 2=Muslim, 3=Jewish, etc.) is not a person who is slightly more Muslim than Christian. It’s a category error, a confusion of measurement scales that renders the result uninterpretable.

Nominal data—categories without order—must never be averaged. Ordinal data—rankings without equal intervals—should not be averaged, though the median may sometimes be acceptable. Only interval and ratio data, where differences are meaningful and a true zero exists, can be properly averaged. Even then, the distribution must be checked for skew and outliers.

The Simpson’s Paradox: When Averages Reverse Truth

Simpson’s paradox is one of the most startling demonstrations of how averaging can invert reality. It occurs when a trend appears in several different groups of data but disappears or reverses when these groups are combined. The classic example involves university admissions: within each department, women may have a higher acceptance rate than men, but overall, men have a higher acceptance rate because women apply in larger numbers to departments with lower acceptance rates. The aggregate average tells the opposite story of the disaggregated data.

This paradox appears in medical studies, social science research, and business analytics. A marketing campaign may show improved conversion rates in every segment, yet the overall conversion rate drops because the mix of segments shifted toward lower-converting groups. Averaging across segments without accounting for composition changes produces a false narrative of decline.

The only defense against Simpson’s paradox is to examine data at the appropriate level of granularity. Averages computed across heterogeneous groups are not summaries; they are distortions.

Practical Guidance: When to Avoid the Average

For anyone working with data—journalists, analysts, policymakers, or curious citizens—a few simple checks can prevent the misuse of averages. Before computing a mean, ask these questions:

• Is the distribution symmetric? Plot a histogram. If the data is skewed, the mean will be pulled toward the tail and will not represent the typical value. Use the median or report the full distribution.
• Are there multiple peaks? If the data clusters into distinct groups, the mean falls in an unpopulated valley. Describe the clusters separately.
• Is the data ordinal or nominal? If the numbers are just labels, do not compute a mean. Use the mode or report category frequencies.
• Are the groups of different sizes? If you are averaging percentages or rates from subgroups, weight by the subgroup size. An unweighted mean of percentages is almost always wrong.
• Does the average answer the question you are actually asking? If you want to know what is typical, the mean may not tell you. If you want to know the total impact, the mean multiplied by the count may be appropriate, but only if the data is ratio-scale and the distribution is understood.

These checks take seconds. They prevent errors that can last for years.

The Median: A Better Default

For most real-world data that reaches the public, the median is a more honest summary than the mean. The median is resistant to outliers, works for ordinal data, and always represents an actual data point when the sample size is odd. It doesn’t create fictional values. When a journalist reports that the median home price in a city is $350,000, that number corresponds to a real transaction. The mean home price, inflated by a few mansions, may be $500,000 and correspond to nothing at all.

The median isn’t a perfect statistic. It discards information about the tails, which can be important. But for the common purpose of describing a typical case, the median is superior to the mean in almost all situations involving skewed data, ordinal data, or data with undefined intervals. The responsible communicator reaches for the median first and only uses the mean when the distribution justifies it.

Why the Error Persists

If the problems with averaging are so well-documented, why do they persist? Several forces are at work. First, the mean is computationally simple and conceptually familiar. Everyone learned it in primary school. The median requires sorting the data, which is trivial with modern tools but still feels like an extra step. Second, the mean has mathematical properties that make it convenient for further statistical manipulation. Analysts who plan to run regressions or compute variances often start with means because the math works cleanly, even when the data does not.

Third, and most troubling, averages can be used to mislead. A politician who wants to claim rising prosperity can cite the mean income, knowing it’s pulled up by the wealthy. A school district can report a high average test score while hiding the bimodal distribution that reveals a stark achievement gap. The average, in these cases, is not a summary but a shield.

Media organizations share responsibility. The pressure to reduce complex stories to a single headline figure is immense. “Study Finds Average Screen Time Is Four Hours” is a simpler story than “Screen Time Varies Widely by Age, Income, and Device Type, With Distinct Patterns Across Groups.” The first headline fits a chyron; the second requires a paragraph. But the first headline is also false in the ways that matter most.

What the Careful Reader Should Do

When you encounter an average in the wild—in a news article, a political speech, a corporate report—pause. Ask what kind of data lies beneath it. Is it a mean or a median? Is the distribution symmetric or skewed? Are there subgroups that should be examined separately? Was the average computed from raw numbers or from percentages that should have been weighted?

If the report doesn’t provide this information, the average isn’t trustworthy. A single number without context is not a fact; it’s a rhetorical device. Demand the histogram, the box plot, the disaggregated table. If the source cannot or will not provide them, treat the average as an opinion, not a measurement.

Statistical literacy isn’t about mastering complex equations. It’s about recognizing when a number is being used to inform and when it’s being used to persuade. The average, misapplied, is one of the most common tools of persuasion masquerading as information.

FAQ

Why is the mean misleading for skewed data?

The mean is sensitive to extreme values. In a skewed distribution, a few very high or very low values pull the mean away from the bulk of the data. For example, in a neighborhood where most homes sell for around $300,000 but one sells for $3 million, the mean will be significantly higher than what a typical buyer experiences. The median remains close to $300,000 and better represents the central tendency.

Can ordinal data ever be averaged?

Strictly speaking, no. Ordinal data preserves order but not equal intervals. The difference between “Agree” and “Strongly Agree” is not necessarily the same as between “Neutral” and “Agree.” Computing a mean assumes equal spacing, which is rarely justified. The median is acceptable because it only uses order. For most purposes, reporting the full distribution of responses is more informative than any single summary statistic.

What is Simpson’s paradox and how does it relate to averaging?

Simpson’s paradox occurs when a trend that appears in several groups reverses when the groups are combined. This happens because the groups have different sizes and different baseline rates. The aggregate average weights the groups unequally and can produce a result that contradicts each group’s individual trend. It demonstrates that averaging across heterogeneous categories without accounting for composition can invert the truth.

How can I tell if an average in a news report is trustworthy?

Look for three things: whether the type of average (mean or median) is specified, whether the distribution is described or shown, and whether subgroups are discussed. A trustworthy report will mention if the data is skewed, will provide the median alongside the mean, and will break down results by relevant categories. If only a single average is given with no context, the number should be viewed with skepticism.

Numbers feel clean. A single figure—an average, a mean, a typical value—promises to cut through the noise and give us something solid to hold onto. But that promise is often a trap. When we average data that was never meant to be averaged, we don’t simplify reality; we replace it with a tidy fiction. Jerome Leland has spent years watching how information gets packaged for public consumption, and he’s noticed a pattern: the most dangerous mistakes aren’t the obvious falsehoods. They’re the ones that look perfectly reasonable on a spreadsheet.

To see why averaging can go so wrong, you have to understand what an average actually assumes. The arithmetic mean—add everything up, divide by the count—only makes sense if the items you’re adding are fundamentally alike, members of a single coherent category. When they’re not, the average becomes a number that describes nothing real. It’s a statistical ghost, and we keep mistaking it for a living thing.

The Flaw of the “Average Person”

Back in 1945, the Cleveland Health Museum held a contest to find a woman whose body measurements matched the “average American woman” as defined by a recent anthropometric study. They found a theater cashier named Martha Skidmore and crowned her the statistical ideal. The only problem: nobody had checked whether a real person could actually match all those average dimensions at once. When the U.S. Air Force later tried designing cockpits around the average pilot’s measurements, they discovered that out of thousands of pilots, not a single one fit the average across even a handful of key dimensions. The average was a composite that described no actual human being. This is the first way averaging fails: the center point of a multivariate distribution may not correspond to any real instance in the population.

We still make this mistake today, especially in media narratives. The “average household” or the “average voter” is invoked as if such a creature exists. But households vary wildly in size, composition, and income. The average voter is a statistical ghost, smoothing over deep fractures in priorities and values. When a news story says “the average American thinks X,” it’s often papering over the very divisions that are the real story.

Mixing Categories That Should Stay Separate

A second, sneakier problem shows up when we average across categories that are fundamentally different. Picture a county reporting its “average emergency response time.” The county has a dense urban core and a sprawling rural hinterland. In the city, response times clock around 6 minutes. Out in the countryside, it’s more like 25. The combined average might land at 12 minutes. That 12-minute figure is arithmetically correct, but it’s a lie in practice—it describes neither the urban nor the rural experience. It pretends two different systems are one.

You see this same error all over education reporting. A school district announces an average test score that looks respectable, but that single number buries enormous variation between schools, between demographic groups, between programs. The average becomes a tool for hiding inequality rather than exposing it. When journalists repeat these averages without unpacking them, they’re not reporting—they’re helping with the cover-up.

The Ratio Trap

Another classic case of illegitimate averaging involves ratios and rates. Imagine two hospitals. The small one performs 10 procedures and loses 1 patient—a 10% mortality rate. The large one performs 100 procedures and loses 8—an 8% rate. If you naively average the two rates, you get 9%. But the combined mortality across both hospitals is 9 deaths out of 110 procedures, or about 8.2%. The simple average of rates gives equal weight to each hospital, ignoring how many patients they actually treated. This is Simpson’s paradox in action, and it can flip the apparent truth on its head.

In public health reporting, this mistake has real consequences. Averaging infection rates across regions with vastly different population sizes can make an outbreak look milder than it is—or more severe. The right approach weights the rates by the underlying populations, but even then, the aggregate number can hide dangerous local hotspots that need attention now.

The Time-Series Deception

Averaging over time introduces its own brand of distortion. Stock market commentators love to cite average annual returns: “The market has returned an average of 7% per year over the last century.” That smooth 7% line erases crashes, booms, and long grinding periods of stagnation. An investor who entered the market at the wrong moment lived through something nothing like that smooth average. The average hides the sequence of returns, and the sequence is the actual experience of any real participant.

This temporal averaging is especially nasty in climate and weather reporting. A global average temperature rise of 1.5°C sounds almost gentle—until you realize it masks regional extremes where the rise is 3°C or 4°C, with devastating local effects. The average numbs us to the variance, and the variance is where the damage lives.

The Ecological Fallacy in News Narratives

When journalists report on demographic trends, they often stumble into the ecological fallacy: inferring individual behavior from group averages. A classic example is stating that counties with higher average incomes voted a certain way, and then implying that wealthier individuals voted that way. The average of a group doesn’t describe any specific member of that group, and the correlation at the aggregate level can reverse at the individual level. This error is so baked into election coverage that it’s become a structural flaw in how we understand political behavior.

Consider a report that “neighborhoods with higher average education levels have lower crime rates.” The implication is that more educated people commit fewer crimes. But the data is at the neighborhood level, not the individual level. It could be that neighborhoods with higher education levels also have more policing, better lighting, or younger populations. The average is a property of the group, not a property of the people in it. Drawing conclusions about individuals from group averages is a leap that often lands in falsehood.

When the Distribution Matters More Than the Center

In many cases, the average isn’t just misleading—it’s beside the point. Income data is the textbook example. The mean household income in the United States gets yanked upward by a small number of extremely high earners. The median, the 50th percentile, sits substantially lower. But even the median fails to capture the shape of the distribution: the clustering at the bottom, the long tail at the top, the hollowing out of the middle. Reporting only the mean or median income gives no sense of inequality, precarity, or the lived experience of most households.

Journalists who want to convey economic reality should reach for percentile ranges, Gini coefficients, or simply tell stories of specific households at different points on the curve. The average is a shortcut that short-circuits understanding.

Ordinal Data and the Meaningless Mean

Some data cannot be averaged because the numbers themselves are labels, not quantities. Survey questions that ask respondents to rate something on a scale of 1 to 5 produce ordinal data. The difference between “strongly disagree” (1) and “disagree” (2) is not the same as the difference between “agree” (4) and “strongly agree” (5). Yet it’s standard practice to compute an average rating, say 3.7, and treat it as a meaningful measure of sentiment. A 3.7 average could mean that most people were lukewarm, or that the population was polarized between 1s and 5s. The average erases that distinction.

In product reviews, movie ratings, and opinion polls, the average star rating is everywhere. It’s also frequently useless. A film with an average rating of 3 stars might be a bland, universally mediocre experience, or it might be a divisive masterpiece that half the audience adored and half despised. The latter is far more interesting, but the average hides it.

Practical Guidelines for Critical Consumption

Given how pervasive these errors are, a careful reader or reporter needs a mental checklist when confronted with an average. First, ask: What is the underlying distribution? If the data is bimodal, heavily skewed, or has extreme outliers, the average is likely deceptive. Second, ask: Are the categories being mixed? If the average combines apples and oranges—urban and rural, small and large, young and old—it probably should not exist. Third, ask: Is the average being used to make claims about individuals? If so, the ecological fallacy may be in play.

For reporters, the obligation is greater. Presenting an average without context is a form of misinformation. Whenever possible, show the distribution. Use a histogram, a box plot, or even a simple range. If you must report an average, pair it with a measure of spread—standard deviation, interquartile range, or at minimum the minimum and maximum. Let the audience see the shape of the data, not just a single point that may be a statistical mirage.

When Averages Work

This is not an argument against averages altogether. When data is normally distributed and the categories are homogeneous, the mean is a powerful and legitimate summary. If you measure the heights of a thousand randomly selected adult men from the same population, the average height is meaningful and stable. The problem arises when we lazily apply the same tool to data that violates its assumptions. The average is a precision instrument for a specific type of data, not a universal solvent.

In newsrooms, the pressure to simplify is immense. A single number fits in a headline, a chyron, a tweet. A distribution requires explanation, visualization, and nuance. But the cost of that simplification is often accuracy. The public is left with a number that feels true but is actually a distortion. Over time, these distortions accumulate into a fog of misunderstanding that shapes policy, investment, and personal decisions.

FAQ

Why is the average sometimes called a “summary statistic” if it can be so misleading?

The average is a summary statistic in the sense that it reduces many data points to a single number. This is useful when the data is symmetric and unimodal, because the average then represents a typical value. But when the data is skewed, multimodal, or contains outliers, the average summarizes poorly—it gives a number that is not typical of anything in the dataset. The term “summary” does not guarantee accuracy; it only describes the operation of condensing information.

How can I tell if an average in a news article is trustworthy?

Look for accompanying information about the distribution. A trustworthy report will mention the range, the median, or the presence of outliers. If the article only gives an average without any sense of variation, be skeptical. Also consider the source: is the average being used by an advocate to make a point, or by a disinterested analyst to describe a phenomenon? Advocacy often selects the average that supports its case while ignoring the distribution that would undermine it.

What is the difference between mean and median, and when should I prefer one over the other?

The mean is the arithmetic average: sum divided by count. The median is the middle value when data is sorted. The median is resistant to outliers and skewed distributions; it always represents a value that actually occurs in the dataset (or is halfway between two actual values). For income, home prices, and any data with a long tail, the median is usually more informative than the mean. The mean is appropriate for symmetric distributions without extreme values, such as measurement errors or natural physical traits in a homogeneous group.

Can averaging ever be completely avoided in journalism?

Probably not, and it shouldn’t be. There are legitimate uses. The key is to avoid averaging across inappropriate categories and to always provide context about the distribution. A responsible journalist treats the average as a starting point for exploration, not as the final word. When the average is the only number reported, it often becomes a stopping point for thought rather than a starting point.

Numbers have a way of seducing us. They promise clarity, precision, a neat path through the mess. But there’s a quiet, stubborn error that slips into reports, dashboards, and news headlines more often than anyone wants to admit: averaging data that, by its very nature, should never be averaged. Jerome Leland has watched this mistake play out across finance, healthcare, public policy—and the fallout is rarely small.

The urge to average makes sense on the surface. A single figure feels clean. It slides into a headline, a bullet point, a ten-second sound bite without friction. But when the underlying data is built from ratios, percentages of mismatched wholes, or distributions that lean hard in one direction, the resulting “average” isn’t just misleading. It’s mathematically incoherent. It conjures a number that never actually existed in any real-world observation.

The Arithmetic Trap of Averages

At its heart, an average is a measure of central tendency. The arithmetic mean—sum everything up, divide by the count—works beautifully for additive quantities. Heights of ten people? Add them, divide by ten, and you get a meaningful central height. Same goes for weights, temperatures, or test scores where each observation stands on its own, measured on the same scale.

Trouble starts when the data points are themselves ratios, rates, or percentages drawn from different bases. Picture a small hospital that handles 10 surgeries and sees 1 complication—a 10% complication rate. A large hospital handles 1,000 surgeries and logs 50 complications, a 5% rate. Average 10% and 5% naively, and you get 7.5%. But the combined reality across both hospitals is 51 complications out of 1,010 surgeries—a rate of 5.05%. That 7.5% figure is a phantom. It describes no actual population of patients.

This error has a dramatic cousin called Simpson’s Paradox, where trends in subgroups flip when data gets aggregated improperly. But the underlying problem is simpler: averaging ratios without weighting them by their denominators produces a number detached from reality. Yet the practice is everywhere—in business metrics, educational statistics, even medical research summaries.

When Percentages Collide

Think about a marketing analyst comparing email open rates across five campaigns. Each campaign hits a different audience size. Campaign A reaches 100,000 recipients with a 20% open rate. Campaign B reaches 1,000 recipients with an 80% open rate. The simple average of 20% and 80% is 50%. But the actual combined open rate across both campaigns isn’t 50%—it’s yanked heavily toward the larger campaign. The true overall rate is (20,000 + 800) / (100,000 + 1,000) = 20.6%. Reporting 50% would be a catastrophic overstatement.

This isn’t academic. Jerome Leland has sat through investor pitch decks where customer retention rates were averaged across cohorts of wildly different sizes, producing a figure that suggested far healthier retention than actually existed. The startup’s largest cohort had a 40% retention rate, while a tiny pilot cohort showed 90%. The simple average of 65% masked the fact that the vast majority of customers were churning at a much higher rate.

The Denominator Problem

At the bottom of this statistical misstep is a failure to respect the denominator. Every percentage, rate, or ratio is a fraction: a numerator divided by a denominator. The denominator represents the scale, the base, the population from which the numerator is drawn. When you average percentages without accounting for their denominators, you implicitly assume all denominators are equal. That assumption is rarely true and often wildly false.

Take crime rates across cities. City A has a population of 100,000 and reports 500 violent crimes, a rate of 5 per 1,000 residents. City B has a population of 10,000 and reports 100 violent crimes, a rate of 10 per 1,000. The simple average of the two rates is 7.5 per 1,000. But if you combine the cities, the actual rate is 600 crimes per 110,000 residents, or about 5.45 per 1,000. The unweighted average overstates the true rate by nearly 40%.

This distortion gets worse as the gap between denominators widens. When a small group with an extreme rate is averaged equally with a large group with a moderate rate, the small group punches above its weight. The result is a number that’s mathematically possible but empirically false—a synthetic statistic that represents no actual population.

Skewed Distributions and the Mean’s Deception

Another domain where averaging falls apart is highly skewed data. Income distributions, for example, are famously right-skewed: a small number of extremely high earners pull the arithmetic mean far above the median. Reporting the “average income” in such cases gives a distorted picture of what a typical person actually earns. The mean becomes a measure not of centrality but of total wealth divided by population, which is a different concept entirely.

In 2019, the U.S. Census Bureau reported a median household income of $68,703, while the mean household income was $98,088. That $30,000 gap isn’t a rounding error; it’s the gravitational pull of the top percentiles. When journalists or analysts casually refer to the “average American household income” without specifying which average, they often inadvertently misrepresent the economic reality for most households.

This problem extends to any metric with a long-tailed distribution: website page load times, hospital wait times, customer lifetime value, or the size of insurance claims. In each case, a few extreme values can inflate the arithmetic mean to a point where it no longer describes a typical experience. The median or geometric mean often provides a more honest central value, yet the arithmetic mean persists because it’s computationally simple and familiar.

When Averaging Destroys Information

Beyond the mathematical errors, there’s a deeper conceptual problem: averaging can destroy the very information that matters most. In public health, averaging infection rates across regions with vastly different population densities and healthcare access can hide dangerous hotspots. A national average might look reassuring while a specific city is experiencing an outbreak that requires immediate intervention.

In education, averaging test scores across schools with different demographics, funding levels, and class sizes can mask severe inequities. A district-wide average might suggest acceptable performance while individual schools are failing. The average becomes a tool of erasure, smoothing over the variation that policymakers most need to see.

In finance, averaging portfolio returns across time periods without considering volatility or sequence of returns can lead to disastrous planning. A retiree withdrawing funds during a market downturn faces a fundamentally different reality than the “average annual return” would suggest. The average hides the sequence-of-returns risk that can destroy a retirement plan.

The Simpson’s Paradox Trap

Simpson’s Paradox is the most famous manifestation of the averaging error, and it deserves a closer look. The paradox occurs when a trend appears in several different groups of data but disappears or reverses when the groups are combined. The classic example comes from a 1973 study of gender bias in graduate admissions at the University of California, Berkeley. When looking at aggregate data, men appeared to be admitted at a higher rate than women. But when the data was broken down by department, women had equal or higher admission rates in most departments. The paradox arose because women tended to apply to more competitive departments with lower overall admission rates.

This is not a statistical curiosity; it’s a recurring trap in business analytics. A company might see that its overall customer satisfaction score is declining, yet every individual product line shows improving satisfaction. The paradox occurs because customers are migrating toward a product line with historically lower satisfaction scores. The aggregate trend points in the opposite direction of every subgroup trend. Acting on the aggregate number alone would lead to completely wrong conclusions.

The lesson isn’t that averages are useless, but that they must be applied with an understanding of the underlying data structure. When subgroups have different sizes and different base rates, the simple average is not just imprecise—it can be actively deceptive.

When Averaging Time-Series Data Misleads

Another common error is averaging data across time periods that have fundamentally different characteristics. Consider a retailer reporting “average monthly sales” over a year that includes both normal months and the holiday season. The average will be pulled upward by November and December, creating a figure that represents no actual month. A manager using this average to set monthly targets will find it too high for most of the year and too low for the peak season.

Similarly, averaging economic indicators across business cycle phases—recessions and expansions—produces a number that describes neither. The average GDP growth rate over a period that includes both a boom and a bust is a statistical artifact, not a meaningful benchmark. Yet such averages routinely appear in economic commentary as if they represented a stable underlying trend.

In medicine, averaging patient outcomes across different risk groups can obscure critical safety signals. A drug might show acceptable average efficacy, but when the data is stratified by age or comorbidity, it becomes clear that the drug is highly effective for some patients and dangerous for others. The average hides this heterogeneity, potentially leading to harmful prescribing practices.

Alternatives to the Naive Average

When faced with data that should not be simply averaged, several alternatives exist. The choice depends on what question is actually being asked.

Weighted averages correct the denominator problem by giving each data point influence proportional to its size. For the hospital complication rates, weighting by the number of surgeries at each hospital produces the correct aggregate rate. This is the most direct fix for the ratio-averaging error and should be the default approach whenever combining rates or percentages from groups of different sizes.

Medians resist the pull of extreme values and better represent the “typical” observation in skewed distributions. For income data, home prices, or any metric with a long tail, the median is almost always more informative than the mean for describing what a typical person or case experiences.

Geometric means are appropriate for averaging growth rates, returns, or other multiplicative processes. If an investment grows 50% one year and falls 50% the next, the arithmetic mean return is 0%, but the geometric mean correctly shows a net loss. The arithmetic mean of percentage changes over time is almost always misleading.

Stratified reporting preserves the variation that simple averaging destroys. Instead of reporting a single average, report the averages for each meaningful subgroup alongside the overall weighted average. This approach maintains transparency and allows readers to see both the forest and the trees.

Real-World Consequences

The consequences of averaging data that should not be averaged extend far beyond academic debates. In 2016, a major pharmaceutical company presented aggregated clinical trial results that showed a modest benefit for a new drug. Only when independent researchers reanalyzed the data by patient subgroups did it become clear that the drug was actively harmful to patients over 65 while providing significant benefit to younger patients. The simple average had masked a critical safety signal.

In public policy, averaging unemployment rates across regions with vastly different economic structures can lead to misallocation of resources. A national unemployment rate of 5% might sound healthy, but if it consists of 2% unemployment in urban areas and 15% in rural communities, the average hides a crisis. Programs designed based on the national average will fail to address the real problem.

In education, the No Child Left Behind Act required schools to report average test scores, which led to well-documented instances of schools focusing resources on students near the proficiency threshold while neglecting both high-achieving and severely struggling students. The average became the target, and the target distorted the system.

How to Spot the Error

Readers and analysts can protect themselves by asking a few diagnostic questions whenever they encounter an average:

First, what is being averaged? If the answer is a percentage, rate, or ratio, immediately check whether the denominators are equal. If they are not, the average is almost certainly wrong unless it has been explicitly weighted.

Second, what is the distribution? If the data is skewed, the mean will be pulled away from the typical value. Ask for the median or a full distribution plot. A single number cannot capture a skewed reality.

Third, what variation is being hidden? Even a correctly calculated weighted average can conceal important differences between subgroups. Always ask for the disaggregated data. If the source cannot or will not provide it, treat the average with skepticism.

Fourth, is the average being used to make a decision that depends on the full distribution? In risk management, inventory planning, or capacity sizing, the average is often insufficient. The extremes—the tails of the distribution—frequently determine success or failure. A bridge designed to withstand the average flood will collapse regularly.

FAQ

Why can’t I just average percentages from different groups?

Percentages are ratios with denominators that often differ in size. Averaging them without weighting by those denominators gives equal influence to groups of unequal size, producing a number that does not correspond to the actual combined rate. For example, averaging a 90% rate from a group of 10 and a 40% rate from a group of 1,000 yields 65%, but the true combined rate is about 40.5%. The simple average is mathematically incoherent because it ignores the scale of each group.

When is it acceptable to use a simple average?

A simple arithmetic mean is appropriate when the data points are additive quantities measured on the same scale and each observation carries equal weight. Examples include averaging the heights of individuals in a room, the temperatures recorded at different weather stations on the same day, or the scores of students who all took the same test. The key condition is that each data point represents a measurement of the same kind with a consistent underlying unit.

What should I use instead of a simple average for skewed data?

For skewed distributions, the median is usually a better measure of central tendency because it is not pulled by extreme values. For growth rates or returns over time, the geometric mean is more appropriate because it accounts for compounding effects. For data with unequal group sizes, a weighted average that uses the group sizes as weights will give the correct aggregate figure. In many cases, reporting the full distribution or providing stratified averages alongside the overall figure is the most honest approach.

How does averaging cause bad business decisions?

Averaging can lead to bad decisions when it hides variation that matters for planning. If a retailer uses average monthly sales to set inventory levels, they will be overstocked in slow months and understocked in peak months. If an investor uses average annual returns to plan retirement withdrawals without considering the sequence of those returns, they may run out of money during an early downturn. The average provides a single number, but real-world outcomes depend on the entire pattern of variation.