The Field Guide

How to read science news

Short, plain explainers for the words and warning signs that decide how much a study is really worth. One page each, so you can tell a real breakthrough from a hopeful headline.

01 Clinical trial phases Phase 1 tests whether a drug is safe, phase 2 tests whether it works, and phase 3 confirms both in a large group before approval.
02 Animal model translation A result in mice is an early clue, not a cure, because most treatments that work in animals never work in people.
03 Effect size Effect size measures how big a difference is, which is the question that decides whether a real result actually matters in your life.
04 Relative and absolute risk A "50% risk cut" is relative; the absolute change can be tiny, so always ask what the risk went from and to in real numbers.
05 Preprint and peer review A preprint is a study posted before independent experts have vetted it, while peer review is that vetting step, so preprints deserve extra caution.
06 Correlation and causation Two things happening together does not prove one causes the other, so "linked to" headlines rarely establish cause and effect.
07 Reproducibility and replicability A finding earns trust when other scientists can repeat it, so one study is a data point, not a settled conclusion.
08 P-value A p-value measures how surprising the data would be if there were no real effect, not the probability that a finding is true.
09 Confidence interval A confidence interval is the plausible range around a result, which tells you how precise, or shaky, that single reported number really is.
10 Randomized and observational studies A randomized trial assigns people to groups to test cause, while an observational study only watches, which is why trials give stronger evidence.
11 Meta-analysis A meta-analysis pools many separate studies into one combined estimate, which usually beats any single study, as long as the studies it pools are sound.
12 Survivorship bias Survivorship bias is drawing conclusions from the things that made it through while ignoring the ones that did not, which quietly warps the picture.
13 Regression to the mean Regression to the mean is the tendency for an unusually high or low measurement to be closer to average the next time, which can make a treatment look like it worked when nothing did.
14 Nutritional epidemiology Diet headlines reverse so often because most rest on observational studies with tiny effects and countless confounders, so the same food gets linked to both harm and benefit.
15 Number needed to treat The number needed to treat is how many people have to take a treatment for one of them to benefit, which is often the plainest measure of how much a drug really does.
16 Publication bias Studies that find a positive result are more likely to be published than those that find nothing, so the literature you can see overstates how real and how big an effect is.
17 Funding effect (industry sponsorship bias) Studies funded by the company that sells the product tend to come out more favorably, a pattern strong enough that who paid for the research is part of how you should read it.
18 Lead-time bias Detecting a disease earlier automatically lengthens 'survival since diagnosis' even when it does not delay death by a single day, which makes screening look more effective than it is.
19 Healthy-user bias People who take a preventive treatment or follow health advice are, on average, already healthier and more careful, so their better outcomes can be mistaken for the effect of the treatment.
20 Composite endpoint A composite endpoint bundles several separate outcomes into one, which makes a trial smaller and faster but can blur together events of very different importance.
21 Base rate A base rate is how common something already is before any test result comes in, and ignoring it is the main reason an accurate test can still mislead.
22 Confounding Confounding happens when a third factor is tied to both the exposure and the outcome, so a real statistical link can appear where the exposure itself causes nothing.
23 Simpson's paradox Simpson's paradox is what happens when a trend appears in several separate groups of data but reverses, or disappears, once those groups are combined into one.
24 Ecological fallacy The ecological fallacy is assuming that a pattern true of a group, such as a country or a state, also holds for the individuals inside that group.
25 Immortal time bias Immortal time bias inflates a treatment's apparent benefit when part of the follow-up period could not, by design, have counted against it.
26 Recall bias Recall bias is the systematic difference in how accurately or completely people remember past events, and it can distort a study that relies on memory.
27 Attrition bias Attrition bias happens when the people who drop out of a study are systematically different from those who stay, skewing the results left behind.
28 Statistical power Statistical power is the probability that a study will detect a real effect if one truly exists, and a study with low power can miss it entirely.
29 Multiple comparisons problem The multiple comparisons problem is that running enough separate statistical tests will turn up some significant results by chance alone, even in pure noise.
30 Preregistration Pre-registration means a study's hypothesis, methods, and analysis plan are filed publicly before any data are collected, closing off room to change the question afterward.
31 Intention-to-treat analysis Intention to treat means every participant is counted in the group they were originally randomly assigned to, even if they dropped out or switched treatments.
32 Blinding Blinding keeps participants, caregivers, or the people measuring outcomes from knowing who received which treatment, so belief cannot quietly skew the result.
33 Placebo effect A placebo can measurably ease symptoms even when patients are told exactly what it is, and its mirror image, the nocebo effect, can produce real side effects from an inert pill.
34 Hazard ratio A hazard ratio compares how fast an event is happening in one group versus another at each moment in a trial, and it assumes that comparison holds steady for the whole study, which real trials do not always do.
35 Natural experiment A natural experiment uses a real-world circumstance that sorts people into different conditions as if at random, letting researchers compare outcomes without running a trial themselves.
36 Sensitivity and specificity Sensitivity and specificity describe how well a test tells sick from healthy, but neither one tells you the chance a positive result is real, which also depends on how common the condition is.
37 Difference-in-differences Difference-in-differences compares the change over time in a group exposed to a policy against the change in a similar unexposed group, canceling out trends both shared anyway.
38 Overdiagnosis Overdiagnosis is finding a real disease, often through screening, that would never have caused symptoms or death, so the diagnosis itself becomes the only harm the person experiences.
39 Regression discontinuity design Regression discontinuity compares people just above and just below an arbitrary cutoff, such as a legal age or a test-score threshold, treating them as close to randomly assigned to each side.
40 Surrogate endpoint A surrogate endpoint is a stand-in measure, like a lab value or scan result, used in place of a real outcome such as death, and moving the stand-in does not guarantee the real outcome moves with it.
41 Mendelian randomization Mendelian randomization uses a genetic variant, assigned randomly at conception, as a stand-in for a real experiment, to test whether an association between a trait and a disease is actually causal.
42 Levels of evidence Levels-of-evidence systems rank study designs by how much they can be trusted, from systematic reviews of randomized trials at the top to case reports and mechanism-based reasoning at the bottom.
43 Odds ratio An odds ratio and a risk ratio answer different questions and diverge most when the outcome is common, so reading a large odds ratio as if it were a risk ratio can overstate an effect.
44 Digital object identifier (DOI) Most headlines are built on a real, findable paper, and locating it usually takes only a few minutes once you know to check the press release for a DOI or journal name, then search PubMed and PubMed Central.
45 Margin of error (polls) An election poll is an estimate, with a margin of error around each candidate's share, and the gap between two candidates is roughly twice as uncertain as either share alone.

Every quoted line is verbatim from the source linked on its page. The explanations are ours, written plainly, so you can read a study for what it is worth without a statistics degree.

The best filter is a daily habit.