Lessons

Every article carries a short lesson: a piece of how science works that you can use on the next headline. They are collected here.

Short lesson

What a calorie-matched control is for

Many drugs change how much an animal eats. Appetite suppressants do it on purpose; others do it through nausea or illness. That creates a trap, because eating less has powerful effects of its own. In rodents, cutting food by a fifth or more often lowers blood sugar, trims fat and lengthens life; in one large mouse study, the more the cut, the longer the life.

So when a drug that reduces eating makes animals healthier, there are two stories. The drug may act directly on the body, or it may work simply because the animals ate less.

A calorie-matched control separates them. Researchers measure how much the treated animals eat, then give a second group without the drug exactly that amount. If both groups do equally well, the food explains the benefit. If the drug group does better, something extra is going on. Details matter: a matched group fed one meal a day also goes hungry between meals, which is its own difference.

From: Ozempic's drug gave old female mice about three extra months

Short lesson

Two kinds of error bar

Every measurement has two kinds of uncertainty. Statistical error is scatter: repeat the measurement and you get slightly different answers because of noise or limited samples. It shrinks predictably as you collect more data, roughly with the square root of the sample size. Four times the data halves it.

Systematic error is bias: a miscalibrated instrument, a contaminating signal, a simplifying assumption in the analysis. It pushes every repeat the same way, so more data doesn't shrink it, and it can hide inside a tidy-looking result.

A careful paper reports both, often as "± statistical ± systematic". When only one error bar appears, ask which kind it is. One guard against a subtler bias is blinding: analysts hide the answer from themselves until every choice is fixed, so they can't nudge the analysis toward the result they expect.

From: Across 750 million light-years, galaxy groups fall together as Newton predicts

Short lesson

When the disease comes first

Studies that measure a habit and then follow people for years face a trap called reverse causation. Many diseases start quietly, years before anyone diagnoses them. Early heart failure can make people breathless at night, early dementia can scramble sleep, and early Parkinson's can change how people move. So a habit measured at the start may already be a symptom of the disease the study later "finds".

The usual check is a washout: set aside diagnoses in the first months or years of follow-up, when hidden disease is most likely, and see whether the link survives. If it weakens a lot, some of it was probably disease showing up early. If it survives, that helps but doesn't settle the question, because some diseases take longer than any washout to surface.

From: Less REM on a wrist monitor preceded 83 diseases. Cause or warning?

Short lesson

How do you find a 5,500-year-old infection?

Most infections leave no mark on the skeleton. Plague kills too fast to scar bone, so for most of history a prehistoric plague death was invisible. Ancient DNA changed that. When a bacterium is in the bloodstream at the moment of death, traces of its DNA can end up sealed inside the dense tissue in the roots of the teeth. Researchers grind up a little of it, sequence everything, and match the fragments against a library of known pathogens.

There is a catch. A "hit" is good evidence the microbe was present, but a blank is not proof it was absent: DNA decays unevenly, and many genuinely infected people leave too little behind to detect. So a detection rate — 39% here — is a floor, not a true infection rate, and can't be compared cleanly with a rate measured a different way, in different tissue, at another site.

From: Plague was already striking Siberian families 5,500 years ago

Short lesson

Misses, false alarms, and why "most accurate" depends on what you count

Any yes-or-no test can go wrong in two ways. It can miss someone who has the condition, or raise a false alarm in someone who doesn't. Sensitivity is the share of true cases a test catches; specificity is the share of unaffected people it correctly clears.

The two pull against each other. Make a test flag more people and it catches more cases, but it also accuses more of the healthy. Make it stricter and the reverse happens.

Researchers often boil a test down to one summary score. A common one, the area under the curve, works out for a simple yes-or-no rule to the average of sensitivity and specificity, so it treats a miss and a false alarm as equally bad. Whether they are is a judgment, not a statistic. A missed cancer and an unneeded biopsy don't cost the same. Two tests can swap places in a ranking depending on which kind of mistake you care about more.

From: BMI missed half the adults a body-fat scan flagged. Waists caught more.

Short lesson

Why you can't pick studies by their results

A meta-analysis pools many studies so that their small, noisy results add up to one clearer answer. It only works if the studies are chosen by a rule fixed before anyone looks at the results: this dose, this kind of patient, this outcome.

Suppose instead you keep only the studies that found a clear benefit. Every study mixes the true effect with luck. Some came out high by chance, some low. Keeping the winners keeps the lucky ones, so the pooled effect comes out bigger than the truth, whatever the truth is. Throw out the studies that found nothing and you would reach the same "clear" answer for a sugar pill.

That's why careful reviewers write down their rules in advance, often in a public registry, and why readers should ask how the studies were chosen. A review that picks its studies after seeing the results can still point somewhere interesting. It just can't tell you how big the effect is.

From: A daily gram of vitamin C made colds a little milder

Short lesson

Energy density: why a bowl of soup fills you up

Energy density is the number of calories in each gram of food. Water has none. Fat has about 9 per gram, while carbohydrate and protein have about 4. So foods full of water and fibre, like vegetables, fruit, soups and cooked beans, carry few calories for their weight. Oil, cheese, nuts and pastries carry many.

This matters because people tend to eat a fairly steady weight of food from day to day, whatever it contains. In controlled feeding experiments, when researchers quietly lowered the calories per gram of meals, people ate about the same amount of food, felt as full, and ended up with fewer calories.

Researchers can calculate energy density with or without drinks, and the choice matters: counting water, coffee and diet soda pulls the number down a lot. So energy density figures can only be compared between studies that calculated them the same way.

From: On a low-fat vegan diet, plates stayed as heavy but calories fell

Short lesson

Why "ranked first" doesn't mean "works best"

A treatment ranking answers a narrow question: if you had to bet on which option comes out on top, which would you pick? A common score for this, called SUCRA, turns each treatment's estimated effect and its uncertainty into a single number between 0 and 100%.

The trouble is that a ranking hides how big the differences are and how sure anyone is about them. A treatment tested in one tiny trial with a lucky result can top the table, because its estimate is large even though its range runs from harm to huge benefit. Two treatments can sit first and fifth while differing by less than anyone would feel.

So the useful questions come after the ranking: how much better, in points on the pain scale? Is that more than patients notice? Does the range rule out no difference? And how many people's results is the estimate built on?

From: A 14-person trial put knee braces top of an arthritis ranking

Short lesson

Before-and-after change versus the difference between groups

A trial with a comparison group produces two kinds of numbers. One is the change within a group: how much the people on treatment A improved from where they started. The other is the difference between groups: how much more A's group improved than B's.

Only the second tells you what the treatment itself did. The change within a group mixes the treatment's effect with everything else that happened to those people over the same months: weight loss, a new routine, the extra attention of being in a study, even measurements drifting back toward normal. If the comparison group improved almost as much, most of the change wasn't down to the treatment.

So when a headline gives one group's before-and-after figure, the next question is what happened to the comparison group. A result that sounds large can shrink to a modest difference, or vanish, once you subtract what the comparison group did anyway.

From: Losing the same weight, keto dieters shed more liver fat

Short lesson

Using genes as a natural experiment, and where it can go wrong

People with high levels of something in their blood often differ in other ways: diet, weight, illness. Mendelian randomization tries to sidestep that. Genes are dealt out at conception, before any of those habits, so if variants that raise a substance also go with worse health, the substance itself is a suspect. It works like a lottery that nature ran for us.

The method rests on one big assumption: the variants affect health only through the substance being studied. Many genes do several jobs. A variant that raises a blood chemical might also raise cholesterol or blood sugar, and those could be what shortens life. Geneticists call this pleiotropy.

Researchers check for it with extra methods that tolerate some misbehaving variants. When those methods agree with the main one, the result is sturdier. When they don't, the main result may be carried by a few variants doing something else.

From: Genes for higher blood tyrosine went with shorter lives. Supplements weren't studied.

Short lesson

Big risk, small share

A risk can be large for a few people and small for everyone else, and epidemiologists keep the two apart. Relative risk asks how much more often something happens in an exposed group than in an unexposed one. "Three times as often" is a big relative risk.

The attributable fraction asks a different question: of all the cases in a population, what share would disappear if the exposure disappeared? That depends on the relative risk and on how common the exposure is. A tripled risk carried by 1 in 3,000 people accounts for a tiny share of cases. A small rise in risk carried by half the population can account for many.

The attributable fraction assumes the exposure really causes the outcome, and so does reading a relative risk as the risk an exposure adds. If part of the link has another explanation, the true share is smaller still. Families facing a diagnosis need the first number. Planners deciding where to spend on prevention need the second. Headlines tend to give only one.

From: One in ten babies born with diagnosed infections was later diagnosed autistic

Short lesson

Why the reason for taking something can masquerade as its effect

People don't take pills at random. They take them because of something: a symptom, a diagnosis, a doctor's advice. When a study compares takers with non-takers, that "something" comes along for the ride. Epidemiologists call it confounding by indication.

Take a joint supplement. Its users are more likely to have painful arthritis, move less, use painkillers and visit doctors more often. Any of those could be tied to memory decline or to a diagnosis being recorded sooner.

Researchers push back in a few ways. They adjust for the condition itself. They compare users with people who have the same condition but chose a different remedy. And they check "comparator" exposures, recorded the same way, that shouldn't matter: if multivitamin users show the same pattern, the pattern probably belongs to the kind of person who reports supplements, not to the pill.

From: Glucosamine dulled memory in Alzheimer's mice; the evidence in people is thinner

Short lesson

How a rock remembers where it was born

The atoms in the Solar System were forged in many different stars before the Sun existed, and they never mixed perfectly. So the proportions of an element's isotopes (versions of an atom with different numbers of neutrons) differ very slightly from place to place, by parts in ten thousand. These are nucleosynthetic anomalies: fingerprints of stellar ancestry, not of anything that happened to the rock later.

Because a region of the young disc kept its fingerprint, a meteorite's isotopes say roughly where its parent body formed. Mixing is simple arithmetic: blend two sources and the result falls on a straight line between them on a plot of one isotope ratio against another. If a planet lies on that line, it could be a mix of the two ends. If it lies beyond the end of the line, no blend of those two sources can make it.

From: A reanalysis of ten isotope systems points to Earth's inner Solar System origins

Short lesson

Why a trial bets on one main outcome

A large trial measures dozens of things: deaths, hospital stays, symptoms, side effects. If researchers could pick the best-looking one afterwards, chance alone would often hand them a "significant" result. So before starting, a trial names one primary outcome and calculates how many patients it needs to detect the effect it expects.

The result comes with a range, the confidence interval: the effects compatible with the data. If that range includes "no difference," the trial hasn't shown the drug works, even if the average looks better. It hasn't shown the drug fails, either.

When the primary outcome misses, everything else becomes exploratory: worth reporting, useful for planning the next trial, but not proof. That's why a trial can find fewer hospital stays on one measure and still be reported, correctly, as not showing a benefit.

From: The oldest heart failure drug missed its goal in a 1,001-patient trial

Short lesson

Why more cancer cases doesn't have to mean higher cancer risk

Cancer is mostly a disease of age: the rate for people in their seventies is many times the rate for people in their thirties. So the number of cases in a country depends heavily on how many older people live there.

That's why epidemiologists report two kinds of figure. A count, or a crude rate (cases divided by the whole population), tells a health system how many patients to expect. An age-standardised rate asks a different question: what would the rate be if every country, or every year, had the same age mix? It strips out the effect of ageing so that the risk itself can be compared.

The two can move in opposite directions. A country whose population is ageing can see its case count climb year after year while the risk for a 60-year-old stays flat or even falls. When a headline says cases are rising, the first question is which of the two it means.

From: The 67% cancer jump assumes risk at every age stays put

Short lesson

What a "biological age" score measures

Many body measurements drift with age across a population. Blood pressure tends to rise, the blood protein albumin tends to fall, and kidney markers change. A biological-age score takes a set of these readings and uses formulas fitted to a large reference group to ask: what age is typical for someone with these readings? The Klemera-Doubal method is one common way to build such a score; the versions used in this study were fitted to US national health surveys. If your readings look like a typical 60-year-old's, your score will be close to 60.

In large populations, people whose score runs ahead of their real age tend to get sick and die sooner, on average. That's why researchers find these scores useful.

But a score is only as steady as its ingredients. Cholesterol, blood sugar and blood pressure can shift within days of a diet change, a new medicine or a stressful week. A score that falls after a month shows that those readings changed. Whether ageing itself slowed is a separate question, and only long follow-up can answer it.

From: Did a month of new menus make older Australians biologically younger?

Short lesson

Why a result found by slicing the data needs a fresh test

A trial compares a treatment with a control across everyone. When that comparison finds nothing clear, it's tempting to look inside: men and women, young and old, heavy and light users. Researchers call these subgroups.

The trouble is chance. Every extra slice is another roll of the dice, and at the usual cut-off for "significant", about 1 in 20 comparisons of pure noise will look like a real effect. With enough slices, something almost always turns up, and it usually comes with a sensible-sounding reason.

That's why careful researchers name their subgroups before they see the data, test whether the groups truly differ, and treat a subgroup found afterwards as a lead, not a result. The lead becomes a finding when a new study names that subgroup in advance and sees the same thing again.

From: Air purifiers sped a brain test only after researchers sorted by age

Short lesson

How to read a curve with a shaded band

Many studies ask whether more of something, like cups of coffee, goes with more or less of something else. Instead of splitting people into groups, researchers often fit a smooth curve through everyone's data. The curve is the model's best guess at each dose.

The shaded band around it matters more. It shows the range of values the data are compatible with, usually a 95% confidence interval. Where the band still covers the "no difference" line, the study can't tell the curve apart from a flat line at that dose.

Two more things to check. First, how many people sit at each end: curves are often steepest where data are thinnest, because a few people pull them around. Second, whether the dose of interest was chosen before or after looking. A threshold spotted on a finished chart is a question for the next study, not an answer from this one.

From: Five-cup coffee warning for older women's bones rests on a faint curve

Short lesson

When "no difference" is a real answer

A study that finds no difference hasn't necessarily shown that nothing happens. A small study can miss a real effect simply because it was too small to see it. So the useful question about a null result is: how big an effect could still be hiding?

The answer is in the range around the result, often called a confidence interval. It's the span of effects that fit the data reasonably well. A wide range that runs from "large benefit" to "some harm" means the study couldn't tell. A narrow range centred on zero means it could.

The last step is deciding how big a benefit would matter. If the whole range sits below that size, the study has answered the question: whatever effect there is, it's too small to be worth having. That's a positive finding about a negative result, and it needs lots of data, which is why pooling many trials helps.

From: Across 46 trials, vitamin D pills didn't cut fractures or falls