Migraine Research Statistics Explained: A Plain-Language Glossary

Levels of evidence

Researchers rank study designs by how well they can show that a treatment truly works. Combined analyses of several good trials sit at the top, single randomized trials come next, and studies that only observe people or describe a few patients sit lower down. Expert opinion that isn't based on study data sits at the bottom.

A study lower on the pyramid can still be useful. It just can't answer the question as firmly.

Migraine example: Evidence that erenumab (Aimovig) prevents migraine attacks comes from several randomized trials and meta-analyses, the strongest types of evidence. Evidence for many supplements comes from small or uncontrolled studies, which is weaker.

Randomized controlled trial (RCT)

A study where people are randomly assigned to get either the treatment being tested or a comparison, usually a placebo. Because chance decides who gets what, the groups start out alike, so differences at the end are most likely caused by the treatment.

RCTs are the strongest single type of study for showing that a treatment works.

Migraine example: The trials that led to the approval of atogepant (Qulipta) randomly assigned people with episodic migraine to the drug or a placebo pill and compared their monthly migraine days after 12 weeks.

Crossover trial

A trial where each person tries both the treatment and the comparison, one after the other, usually with a break in between called a washout period. Each person acts as their own comparison, so crossover trials can work with fewer people.

Migraine example: Several small trials of timolol eye drops for migraine attacks had people use timolol drops for some months and placebo drops for others, then compared how their attacks responded.

Parallel-group trial

The most common trial design, where each person stays in one group, treatment or comparison, for the whole study. The groups run side by side and are compared at the end.

Migraine example: Most prevention trials of CGRP antibodies, such as galcanezumab (Emgality), are parallel-group trials: one group gets the drug for months and another gets placebo.

Pilot study

A small early study used to test whether a larger study is practical, such as whether people will sign up and stick with the treatment. Pilot studies are not designed to prove that a treatment works.

Migraine example: A 30-person pilot trial of a new diet for migraine might check whether people can follow the diet for 12 weeks before a larger trial is run.

Observational study

A study where researchers watch what happens without assigning any treatment. People make their own choices, and researchers record the results.

Observational studies can show that two things are linked, but they can't prove that one causes the other.

Migraine example: A study comparing migraine attacks in people who already drink coffee with people who don't is observational, because nobody was told how much coffee to drink.

Cohort study (prospective and retrospective)

An observational study that follows a group of people over time. A prospective cohort study enrolls people first and then follows them forward. A retrospective cohort study looks back at records that already exist, such as insurance claims or medical charts.

Migraine example: Following 3,000 people with migraine for two years to see whether pet owners have fewer migraine attacks is a prospective cohort study. Using insurance records to see whether people who took a migraine drug in early pregnancy had more miscarriages is a retrospective cohort study.

Case-control study

A study that starts with people who already have an outcome (the cases) and compares them with similar people who don't (the controls), looking back at what was different between them.

Migraine example: Researchers might compare people with migraine who had a stroke (cases) with people with migraine who didn't (controls) to see which migraine medications each group had taken.

Cross-sectional study

A study that measures everything at one point in time, like a snapshot. It can show how common something is and what tends to go together, but not what came first.

Migraine example: A survey asking 10,000 adults whether they have migraine and how much screen time they get each day is cross-sectional.

Case series and case report

A description of one patient (case report) or a group of patients (case series) who received a treatment, without a comparison group. These reports can raise new ideas but can't show that a treatment works better than placebo.

Migraine example: A clinic describing 16 patients who tried a timolol nasal spray for their migraine attacks is a case series.

Real-world study

A study of how a treatment performs in everyday medical practice rather than in a controlled trial. Real-world studies include a wider range of patients but usually have no placebo group.

Migraine example: Tracking how 479 patients at headache clinics did on onabotulinumtoxinA (Botox) for chronic migraine is a real-world study.

Registry and claims database studies

Studies that use large collections of data gathered for other reasons, such as insurance claims, electronic health records, or disease registries. They can include millions of people, but the data may be incomplete or coded inconsistently.

Migraine example: A study might use a national insurance database to compare stroke rates in people with migraine who did and didn't take triptans.

Mendelian randomization

A method that uses genetic differences people are born with as a kind of natural experiment. Because genes are assigned at conception, this can help separate cause from coincidence, but it relies on assumptions that are hard to check.

Migraine example: Researchers have used genes linked to higher blood pressure to ask whether blood pressure itself raises the risk of migraine.

Systematic review

A review that searches for every study on a question using a set method, then assesses and summarizes them. It is more complete and less biased than an expert simply picking the studies they know.

Migraine example: A systematic review of ginger for migraine would search several databases for every trial of ginger, not just the ones with positive results.

Meta-analysis

A method that combines the results of many studies into one overall estimate. Pooling studies gives a more precise answer than any single study, but the result is only as good as the studies that went into it.

Migraine example: Combining 22 studies of screen time and migraine into one pooled estimate of risk is a meta-analysis.

Network meta-analysis

A type of meta-analysis that compares several treatments at once, even ones that were never tested directly against each other, by linking them through shared comparisons such as placebo.

Migraine example: A network meta-analysis can compare ibuprofen with ubrogepant (Ubrelvy) and lasmiditan (Reyvow) using trials where each was tested against placebo.

Pooled analysis

Combining the raw data from several trials, usually run by the same company or group, and analyzing them together. It is similar to a meta-analysis but uses individual patient data.

Migraine example: A company might combine the data from all phase 3 trials of a CGRP antibody to see how well it works in people over 60.

Placebo

A treatment with no active drug, made to look like the real one. A placebo pill, injection, or spray lets researchers separate the effect of the drug from the effect of simply being treated.

Migraine example: In trials of rimegepant (Nurtec ODT), the placebo was a dissolving tablet that looked and tasted the same but had no drug in it.

Placebo response

Improvement in people who received a placebo. It happens because of expectations, extra attention, and the natural ups and downs of migraine.

In migraine trials, the placebo response is often large, which is why a treatment has to beat placebo, not just help some people.

Migraine example: In prevention trials, people on placebo often have 2 to 3 fewer migraine days a month by the end of the study. A drug has to do better than that.

Nocebo effect

Side effects or worsening that happen because a person expects them, not because of the drug. People on placebo in trials often report side effects.

Migraine example: In many migraine prevention trials, a noticeable share of people taking placebo report fatigue or nausea and some stop the study because of it.

Active comparator

A trial that compares a new treatment with an existing treatment instead of a placebo. It answers whether the new treatment is better or worse than what people already use.

Migraine example: A trial comparing erenumab (Aimovig) with topiramate (Topamax) used topiramate as the active comparator.

Blinding (single-blind, double-blind)

Keeping people from knowing who got the real treatment. In a single-blind study, the participants don't know. In a double-blind study, neither the participants nor the researchers assessing them know. Blinding stops expectations from shaping the results.

Migraine example: Blinding can be hard with drugs that have obvious effects. People taking topiramate (Topamax) may guess they are on the real drug because of tingling in their hands.

Open-label study

A study where everyone knows which treatment they are getting. Open-label studies are easier to run but more open to bias, because expectations can affect the results.

Migraine example: After many CGRP antibody trials ended, people could continue in an open-label extension, where everyone got the real drug and knew it.

Randomization and allocation concealment

Randomization means using chance, like a computer coin flip, to decide who gets which treatment. Allocation concealment means the people enrolling patients can't see or predict the next assignment, so they can't steer certain patients into certain groups.

Migraine example: In a trial of a new CGRP drug, a computer decides whether each new participant gets the drug or placebo, and the clinic staff can't see the list in advance.

Stratified randomization

Randomizing separately within subgroups so the treatment groups end up balanced on important factors.

Migraine example: A prevention trial might randomize people with episodic and chronic migraine separately so both groups get a similar mix.

Baseline and run-in period

Baseline is the starting point measured before treatment begins. Many migraine trials have a run-in period, often 4 weeks, where people keep a headache diary without treatment so researchers know how many migraine days they usually have.

Migraine example: If someone has 10 migraine days during the baseline month and 5 during treatment, their change from baseline is 5 fewer days.

Washout period

A break between treatments, or before a study starts, so the effects of a previous drug wear off and don't affect the results.

Migraine example: A trial might require people to stop their old migraine preventive for a month before starting the study drug.

Trial phases (phase 1, 2, 3, and 4)

Phase 1 trials test safety and dosing in a small number of people. Phase 2 trials test whether the drug seems to work and which dose is best. Phase 3 trials are large studies that confirm the drug works and is safe, and they are usually what regulators review for approval. Phase 4 studies happen after approval to watch for rare side effects and long-term effects.

Migraine example: Zavegepant (Zavzpret) nasal spray was approved after phase 3 trials showed it beat placebo for 2-hour pain freedom.

Superiority, non-inferiority, and equivalence trials

A superiority trial asks whether a treatment is better than the comparison. A non-inferiority trial asks whether a new treatment is not meaningfully worse than an existing one, within a pre-set margin. An equivalence trial asks whether two treatments work about the same.

Non-inferiority trials are used when a new treatment has other advantages, such as fewer side effects, lower cost, or easier dosing.

Migraine example: A trial could test whether a cheaper generic migraine preventive is non-inferior to a brand-name one.

Primary and secondary endpoints

The primary endpoint is the main outcome a study was designed to measure, chosen before the study starts. Secondary endpoints are additional outcomes. Results for the primary endpoint carry the most weight because researchers can't choose them after seeing the data.

Migraine example: In most prevention trials, the primary endpoint is the change in monthly migraine days. A secondary endpoint might be the 50% responder rate or a quality-of-life score.

Pre-specified and post hoc analyses

A pre-specified analysis is planned before the data are collected. A post hoc analysis is decided after researchers have seen the data. Post hoc findings are more likely to be chance results and are best treated as ideas to test in future studies.

Migraine example: If a trial finds by looking afterward that a drug seemed to work better in left-handed people, that is a post hoc finding that needs to be confirmed.

Subgroup analysis

Looking at results in a smaller part of the study, such as only women, only older adults, or only people with chronic migraine. Subgroups have fewer people, so their results are less certain, and differences between subgroups are often due to chance.

Migraine example: A trial might report how a preventive worked in people who had already failed two or more other preventives.

Null hypothesis

The starting assumption that there is no real effect or difference. A study tests whether the data are different enough from what the null hypothesis predicts to reject it.

Migraine example: In a trial of a new migraine drug, the null hypothesis is that the drug reduces migraine days no more than placebo does.

P-value

A number between 0 and 1 that shows how likely it is to see a result at least this strong if the treatment or factor being studied actually had no effect. A small p-value means the result would be unlikely to happen by chance alone.

By convention, a p-value below 0.05 (written P<0.05) is called statistically significant. That means there would be less than a 5% chance of seeing a difference this large if there were truly no effect. P<0.001 means less than a 0.1% chance.

A p-value does not tell you how big or how important an effect is. A tiny difference can have a very small p-value in a large study.

Migraine example: If a preventive cut monthly migraine days more than placebo with P=0.002, there would be only a 0.2% chance of seeing a difference that large if the drug did nothing.

Statistical significance

A label for results with a p-value below a chosen cutoff, usually 0.05. It means the result is unlikely to be due to chance. It doesn't mean the result is large or meaningful in everyday life. A result can be statistically significant but too small to notice.

Migraine example: In a study of 100,000 people, a difference of a fraction of a migraine day a month could be statistically significant even though no one would notice it.

Clinical significance

Whether a difference is big enough to matter to patients. One fewer migraine day a month might be statistically significant in a large trial, but whether it is clinically significant depends on the person and the cost or side effects of the treatment.

Migraine example: For someone with 20 migraine days a month, going down to 19 may not feel meaningful. Going down to 10 usually does.

Minimal clinically important difference (MCID)

The smallest change in a score that patients would notice as meaningful. Researchers use it to judge whether a statistically significant result actually matters.

Migraine example: For the HIT-6 headache impact questionnaire, a drop of several points is generally considered the smallest change people notice.

Confidence interval (95% CI)

A range of values that likely contains the true result. A 95% confidence interval means that if the study were repeated many times, about 95% of the intervals calculated would include the true value.

A narrow interval means the estimate is precise. A wide interval means there is more uncertainty. For ratios such as hazard ratios, if the interval includes 1, the result is not statistically significant.

Migraine example: A hazard ratio of 1.42 (95% CI, 1.12 to 1.80) for miscarriage with CGRP antibodies means the best estimate is a 42% higher risk, and the true increase is likely somewhere between 12% and 80%.

Type I and type II errors (false positives and false negatives)

A type I error is concluding that a treatment works when it doesn't (a false positive). A type II error is missing a treatment that really does work (a false negative). The 0.05 p-value cutoff limits type I errors. Larger studies reduce type II errors.

Migraine example: A small trial of a supplement that finds no benefit might be a type II error: the supplement could help, but the study was too small to tell.

Sample size

The number of people in a study, often written as n. Larger studies give more reliable results and can detect smaller differences. Very small studies can easily produce results that don't hold up.

Migraine example: A trial of 10 people that finds a dramatic benefit is much less convincing than a trial of 1,000 people that finds a modest one.

Statistical power

The chance that a study will detect a real effect if one exists. A study with too few people is underpowered and may miss a real effect, so a negative result from a small study doesn't prove a treatment doesn't work.

Migraine example: Most migraine prevention trials are designed with at least 80% power, meaning an 80% chance of detecting the expected difference from placebo.

Multiple comparisons

When a study tests many outcomes or groups, some will look significant just by chance. Researchers can adjust for this, for example with a Bonferroni correction, which uses a stricter cutoff for each test.

Migraine example: If a study tests 20 different foods as migraine triggers, about 1 could reach P<0.05 by chance alone even if none are real triggers.

One-sided and two-sided tests

A two-sided test checks for a difference in either direction, better or worse. A one-sided test checks in only one direction. Most medical studies use two-sided tests because a treatment could also turn out to be harmful.

Migraine example: A trial of a new migraine preventive almost always uses a two-sided test, so it would also detect if the drug made migraine worse.

Bayesian analysis and credible interval

A different approach to statistics that combines what was known before a study with the new data to estimate how likely different results are. A 95% credible interval is the Bayesian version of a confidence interval and is read as a 95% probability that the true value lies in that range.

Migraine example: Some network meta-analyses of migraine treatments use Bayesian methods and report credible intervals instead of confidence intervals.

Mean, median, and mode

The mean is the average: add up all the values and divide by how many there are. The median is the middle value when all values are lined up in order. The mode is the most common value.

The median is less affected by a few very high or very low values, which is why studies often report it for things like monthly headache days.

Migraine example: If 11 people have 2 to 8 headache days a month and one person has 28, the mean is pulled up to about 6.6 while the median stays at 4.5.

Standard deviation (SD)

A measure of how spread out the values are. A small standard deviation means most people had similar results. A large one means results varied a lot from person to person.

When values follow a bell-shaped curve, about 68% fall within 1 standard deviation of the mean and about 95% within 2.

Migraine example: If a group averages 8 migraine days a month with an SD of 2, most people have between 6 and 10 days.

Variance

Another measure of spread, equal to the standard deviation squared. It is mostly used in calculations rather than reported on its own.

Migraine example: A study of migraine days might report variance when describing how much attack frequency differs between people.

Standard error (SE)

A measure of how precise an estimate, such as an average, is likely to be. It gets smaller as the sample size grows. Standard errors are used to calculate confidence intervals.

Migraine example: A trial might report that a drug reduced monthly migraine days by 4.2 (SE 0.3), meaning the estimate is fairly precise.

Range and interquartile range (IQR)

The range is the lowest and highest value. The interquartile range is the span of the middle half of people, from the 25th to the 75th percentile. The IQR is often reported with the median.

Migraine example: A study might report a median of 7 headache days (IQR 4 to 12), meaning half of the people had between 4 and 12 days.

Percentile

The value below which a given percentage of people fall. The 50th percentile is the median.

Migraine example: If your migraine disability score is at the 90th percentile, 90% of people in the study had lower scores than you.

Normal distribution and skew

A normal distribution is the symmetric, bell-shaped pattern many measurements follow. A skewed distribution has a long tail on one side. Monthly headache days are often skewed, with many people having few days and a smaller number having many.

Migraine example: In most migraine studies, many people have a few headache days a month and fewer people have many, so the data are skewed to the right.

Incidence and prevalence

Prevalence is how many people have a condition at a given time. Incidence is how many new cases appear over a period of time.

Migraine example: About 12% to 15% of adults have migraine (prevalence). New cases are most common in the teens and twenties, especially in women (incidence).

Rate and person-years

A rate describes how often something happens over time. Person-years add up the time each person was followed, so someone followed for 2 years counts as 2 person-years. This lets studies compare groups followed for different lengths of time.

Migraine example: A study might report 3 strokes per 1,000 person-years in people with migraine with aura.

Absolute risk

The actual chance of something happening, such as 2 in 1,000 people. Absolute risk tells you how common something really is.

Migraine example: The absolute risk of stroke in young women with migraine with aura is small, even though the relative risk is higher than in women without migraine.

Relative risk (risk ratio)

The risk of an outcome in one group divided by the risk in another. A relative risk of 1 means no difference. Above 1 means higher risk, and below 1 means lower risk.

Relative risk can make a change sound large even when the absolute risk is still small. Doubling a 1 in 10,000 risk only raises it to 2 in 10,000.

Migraine example: A relative risk of 2.0 for stroke in women with migraine with aura means the risk is twice as high, but because stroke is rare at young ages, the absolute risk is still low.

Absolute risk reduction and relative risk reduction

Absolute risk reduction is the difference between two groups' risks, such as 20% minus 10% = 10 percentage points. Relative risk reduction is that difference as a share of the original risk, such as 10 divided by 20 = 50%. The same result can sound much larger as a relative reduction.

Migraine example: If 20% of people on placebo and 10% on a preventive have a medication overuse headache over a year, the absolute risk reduction is 10 percentage points and the relative risk reduction is 50%.

Odds ratio (OR)

A way of comparing how likely an outcome is in two groups, often used in case-control studies. It is read like a relative risk: 1 means no difference, above 1 means more likely, and below 1 means less likely.

Migraine example: An odds ratio of 1.5 for migraine in people with poor sleep means the odds of having migraine were 1.5 times higher in poor sleepers.

Hazard ratio (HR)

A comparison of how often an outcome happens over time in two groups. It is read the same way: 1 means no difference, 1.5 means about 50% more often, and 0.7 means about 30% less often.

Migraine example: In one study, women with migraine who took propranolol (Inderal) had an adjusted hazard ratio of 0.45 for stroke, meaning stroke happened at less than half the rate seen in women with migraine who didn't take it.

Survival curve (Kaplan-Meier curve)

A graph showing the share of people who have not yet had an event, such as a side effect or stopping a drug, over time. Curves that separate show that one group has events sooner.

Migraine example: A study of how long people stay on CGRP antibodies might show a curve of the share still taking the drug at 6, 12, and 24 months.

Number needed to treat (NNT)

How many people need to take a treatment for one extra person to benefit compared with placebo. A lower NNT means a more effective treatment. It is calculated as 100 divided by the absolute difference in percentage points.

Migraine example: If 30% of people are pain-free 2 hours after a migraine drug and 20% are pain-free after placebo, the difference is 10 points, so the NNT is 100 divided by 10 = 10.

Number needed to harm (NNH)

How many people need to take a treatment for one extra person to have a particular side effect compared with placebo. A higher NNH means the side effect is less common.

Migraine example: If 15% of people on a triptan feel chest tightness compared with 5% on placebo, the NNH is 100 divided by 10 = 10.

Effect size

How big a difference or relationship is, as opposed to whether it is statistically significant. The effect size tells you how much a treatment actually helps.

Migraine example: "The drug reduced monthly migraine days by 2.5 more than placebo" describes the effect size.

Mean difference and change from baseline

Change from baseline is how much a measure changed from the start of the study. The mean difference is the difference between two groups' average changes, which shows the treatment's effect beyond placebo.

Migraine example: If people on a preventive drop from 9 to 4 migraine days (a change of 5) and people on placebo drop from 9 to 6 (a change of 3), the mean difference is 2 days.

Least squares mean (LS mean)

An average that has been adjusted by a statistical model to account for differences between groups at the start. Trial results are often reported as LS mean changes.

Migraine example: A trial might report an LS mean reduction of 4.1 monthly migraine days with the drug versus 2.5 with placebo.

Standardized mean difference (SMD) and Cohen's d

A way to express a difference between groups in standard deviation units, so studies that used different scales can be combined. As a rough guide, 0.2 is a small effect, 0.5 medium, and 0.8 large.

Migraine example: A meta-analysis combining several different headache disability questionnaires might report results as an SMD.

Dose-response relationship

A pattern where more of something leads to a bigger effect. Finding a dose-response pattern makes it more likely that a relationship is real, though in observational studies it still doesn't prove cause and effect.

Migraine example: If people who spend more time with their pets each day see larger drops in migraine attacks, that is a dose-response pattern.

Correlation and correlation coefficient (r)

Correlation describes how strongly two measurements move together. The correlation coefficient, r, runs from -1 to 1. A value near 1 means they rise together, near -1 means one falls as the other rises, and near 0 means no straight-line relationship. Correlation alone doesn't show cause.

Migraine example: Monthly headache days and disability scores are usually strongly positively correlated: people with more headache days tend to report more disability.

Regression (linear and logistic)

A statistical method that estimates how an outcome changes with one or more factors. Linear regression is used for measurements like migraine days. Logistic regression is used for yes or no outcomes, like whether someone has migraine, and produces odds ratios.

Migraine example: A study might use regression to estimate how many extra migraine days are linked to each hour of lost sleep.

Cox regression (proportional hazards model)

A type of regression used when the outcome is the time until something happens. It produces hazard ratios.

Migraine example: A study comparing how quickly people stop different migraine preventives might use Cox regression and report hazard ratios.

Adjusted analysis and multivariable analysis

A calculation that accounts for other factors, such as age, sex, or other health conditions, that could affect the result. A multivariable analysis includes several of these factors at once. An adjusted result (for example, an adjusted hazard ratio or aHR) tries to isolate the effect of the factor being studied.

Adjustment can only account for factors that were measured.

Migraine example: A study of migraine and heart disease might adjust for smoking, blood pressure, and age, because those affect heart disease too.

Covariate

Any factor included in an analysis besides the main one being studied, usually to adjust for it.

Migraine example: Age, sex, and baseline migraine days are common covariates in migraine studies.

Propensity score matching

A method used in observational studies to compare similar people. Each person gets a score for how likely they were to receive a treatment based on their characteristics, and people with similar scores in each group are compared. It balances measured factors but not unmeasured ones.

Migraine example: To compare people who took a CGRP antibody with people who took topiramate (Topamax), researchers might match people with similar ages, migraine frequency, and other conditions.

Interaction (effect modification)

When the effect of one factor depends on another. Studies test for interaction to see whether a treatment works differently in different groups.

Migraine example: If a drug lowers stroke risk only in women and not in men, sex modifies the drug's effect.

Mediation analysis

A method that tries to estimate how much of an effect works through a particular pathway.

Migraine example: A study might ask how much of the link between dog ownership and fewer migraine days works through more exercise.

Sensitivity analysis

Repeating an analysis with different assumptions or methods to check whether the main result holds up. If the result stays similar, it is more trustworthy.

Migraine example: A meta-analysis might rerun its results after removing small or low-quality studies to see if the conclusion changes.

Prediction model, validation, and calibration

A prediction model uses information about a person to estimate their chance of an outcome. Validation tests the model on data it wasn't built from: internal validation uses part of the same data, and external validation uses a different group of people. Calibration describes how well predicted risks match what actually happens.

Migraine example: A model that predicts who will respond to a CGRP antibody from their headache history needs external validation in other clinics before it can be trusted.

Machine learning

Computer methods that find patterns in large amounts of data to make predictions. They can find complex patterns but can also pick up patterns that only exist in the data they were trained on.

Migraine example: Researchers have used machine learning on headache questionnaires and wearable sensor data to try to predict migraine attacks.

Forest plot

The standard graph in a meta-analysis. Each line shows one study's result and confidence interval, and a diamond at the bottom shows the combined result. Results on one side of the center line favor the treatment and on the other side favor the comparison.

Migraine example: A forest plot of acupuncture trials for migraine would show each trial's effect on migraine days and the pooled effect.

Heterogeneity and I-squared (I²)

Heterogeneity is how much the results of different studies disagree. I² estimates the share of that disagreement caused by real differences between studies rather than chance. Roughly, 25% is low, 50% moderate, and 75% high. High heterogeneity means the pooled result should be read with caution.

Migraine example: Studies of diet and migraine often have high heterogeneity because they test different diets in different groups of people.

Fixed-effect and random-effects models

Two ways to combine studies in a meta-analysis. A fixed-effect model assumes all studies are measuring the same true effect. A random-effects model assumes the true effect varies somewhat between studies, which is usually more realistic and gives wider confidence intervals.

Migraine example: Meta-analyses of migraine prevention trials usually use random-effects models because the trials differ in who they enrolled and how long they lasted.

Publication bias and funnel plots

Publication bias happens when studies with positive results are more likely to be published than studies with negative results, which can make a treatment look better than it is. A funnel plot is a graph used to look for it. Small studies should scatter evenly on both sides, so a missing corner suggests unpublished negative studies.

Migraine example: If only the small supplement trials with good results were published, a meta-analysis would overestimate how well the supplement works.

SUCRA and treatment rankings

In a network meta-analysis, SUCRA is a score from 0% to 100% that summarizes how likely a treatment is to rank among the best. Rankings can shift a lot with small differences in data, so they should be read alongside the actual effect sizes.

Migraine example: A network meta-analysis might rank several acute migraine treatments by SUCRA for 2-hour pain freedom.

Risk of bias assessment

A structured check of how well each study in a review was designed and run, such as whether randomization and blinding were done properly. Common tools include the Cochrane risk of bias tool.

Migraine example: A review of acupuncture for migraine might rate trials without sham acupuncture as high risk of bias because participants knew what they were getting.

GRADE (certainty of evidence)

A system for rating how confident we can be in a body of evidence: high, moderate, low, or very low. Evidence is downgraded for problems like bias, inconsistent results, imprecision, or indirectness, and can be upgraded for large effects or dose-response patterns.

Migraine example: A review might rate the evidence for a treatment as low certainty because the trials were small and not well blinded.

Sensitivity and specificity

Sensitivity is how good a test is at catching people who have a condition. Specificity is how good it is at correctly clearing people who don't. A test with high sensitivity rarely misses cases. A test with high specificity rarely gives false alarms.

Migraine example: A 3-question migraine screening questionnaire with 81% sensitivity correctly flags about 81 of every 100 people who have migraine.

False positive and false negative

A false positive is a test that says someone has a condition when they don't. A false negative is a test that misses someone who does.

Migraine example: A wearable that warns of a migraine attack that never comes has given a false positive.

Positive and negative predictive value

Positive predictive value is the chance that a person with a positive test actually has the condition. Negative predictive value is the chance that a person with a negative test really doesn't. Both depend on how common the condition is in the group being tested.

Migraine example: A migraine screening questionnaire will have a higher positive predictive value in a headache clinic, where most people have migraine, than in the general public.

ROC curve and AUC

An ROC curve shows how well a test or model separates people with and without a condition across different cutoffs. The area under the curve (AUC) summarizes it: 0.5 is no better than a coin flip, and 1.0 is perfect. Values above 0.8 are generally considered good.

Migraine example: A model predicting who will respond to CGRP antibodies with an AUC of 0.83 separates likely responders from non-responders fairly well.

Likelihood ratio

How much a test result changes the odds that someone has a condition. A likelihood ratio well above 1 makes the condition more likely, and one well below 1 makes it less likely.

Migraine example: If a headache feature has a likelihood ratio of 3 for migraine, having that feature multiplies the odds that the headache is migraine by about 3.

Episodic and chronic migraine

Trials usually group people by how often they have headaches. Chronic migraine means headache on 15 or more days a month for more than 3 months, with migraine features on at least 8 of those days. Episodic migraine means fewer headache days than that. High-frequency episodic migraine usually means about 8 to 14 migraine days a month.

Migraine example: Someone with 10 migraine days a month has high-frequency episodic migraine. Someone with headache on 18 days a month, 9 of them migraine, has chronic migraine.

Monthly migraine days (MMD) and monthly headache days (MHD)

The number of days in a month with a migraine or with any headache. Prevention trials usually measure how much a treatment lowers this number compared with placebo. Some studies also count moderate-to-severe headache days.

Migraine example: A trial might report that a preventive lowered monthly migraine days by 4.3 compared with 2.5 on placebo.

Acute medication use days

The number of days a month someone takes medication to treat a migraine attack. Lower numbers are good and also lower the risk of medication overuse headache.

Migraine example: A preventive that lowers acute medication days from 10 to 4 a month also lowers the risk of medication overuse headache.

Responder rate (50% responder rate)

The percentage of people whose migraine days dropped by at least a set amount, usually half. Some studies also report 30%, 75%, or 100% responder rates. A 30% responder rate is often used for chronic migraine.

Migraine example: A 50% responder rate of 40% means 4 in 10 people had their monthly migraine days cut at least in half.

Pain scales (4-point, 0 to 10, and VAS)

Trials of acute treatments usually rate pain on a 4-point scale: 0 none, 1 mild, 2 moderate, 3 severe. Other studies use a 0 to 10 number rating scale or a visual analog scale (VAS), where people mark a point on a line.

Migraine example: A trial might count a migraine attack as treated successfully if pain drops from 3 (severe) to 0 (none) within 2 hours.

2-hour pain freedom

The standard outcome for treatments taken during a migraine attack: the percentage of people who are completely pain-free 2 hours after taking the medication.

Migraine example: In their main trials, ubrogepant (Ubrelvy) and rimegepant (Nurtec ODT) each left about 20% of people pain-free at 2 hours, compared with about 11% to 14% on placebo.

2-hour pain relief

The percentage of people whose pain went from moderate or severe to mild or none 2 hours after treatment. It is a lower bar than pain freedom, so the percentages are higher.

Migraine example: A drug might give 2-hour pain relief to 60% of people but 2-hour pain freedom to only 25%.

Most bothersome symptom (MBS)

The symptom besides pain that each person finds hardest during a migraine attack, such as nausea, light sensitivity, or sound sensitivity. Newer trials measure whether this symptom is gone at 2 hours.

Migraine example: If light sensitivity is your most bothersome symptom, a trial would check whether it was gone 2 hours after you took the study drug.

Sustained pain freedom

Being pain-free at 2 hours and staying pain-free, without rescue medication, for 24 or 48 hours.

Migraine example: A person who is pain-free 2 hours after a triptan but has the headache return the next morning had pain freedom but not 24-hour sustained pain freedom.

Rescue medication

A second medication taken when the first treatment doesn't work well enough. Trials track how often people need rescue medication as a sign of how well the study drug worked.

Migraine example: If someone takes ibuprofen 2 hours after a study drug because the pain hasn't improved, the ibuprofen is a rescue medication.

Disability and quality-of-life scales (MIDAS, HIT-6, MSQ)

Questionnaires that measure how much migraine affects daily life. MIDAS counts days of missed or reduced activity over 3 months, with scores of 21 or more meaning severe disability. HIT-6 scores range from 36 to 78, with 60 or higher meaning severe impact. The Migraine-Specific Quality of Life Questionnaire (MSQ) scores from 0 to 100, with higher scores meaning better quality of life.

Migraine example: A trial might report that a preventive improved MSQ scores by 15 points compared with 8 points on placebo.

Self-report and headache diaries

Information that participants record themselves, such as pain scores, migraine days, or sleep quality. Most migraine outcomes are self-reported, often in a daily paper or electronic diary.

Migraine example: Whether someone had a migraine attack on a given day usually comes from their own diary entry, not from a test.

Adverse event and treatment-emergent adverse event

An adverse event is any health problem that happens during a study, whether or not it was caused by the treatment. A treatment-emergent adverse event is one that started or got worse after treatment began. Comparing rates with the placebo group shows which ones are likely caused by the drug.

Migraine example: Constipation was reported more often with erenumab (Aimovig) than with placebo, which suggests it is a true side effect.

Serious adverse event (SAE)

An adverse event that leads to hospitalization, disability, a birth defect, a life-threatening situation, or death. Serious does not necessarily mean the drug caused it.

Migraine example: A participant hospitalized for a broken leg during a migraine trial would count as a serious adverse event, even though the drug didn't cause it.

Discontinuation rate

The share of people who stopped a treatment during a study, often reported separately for stopping because of side effects. It is a practical measure of how tolerable a drug is.

Migraine example: In topiramate (Topamax) trials, a noticeably larger share of people stopped because of side effects than in placebo groups.

Persistence and adherence

Persistence is how long people continue a treatment. Adherence is how consistently they take it as prescribed. Both are often lower in real life than in trials.

Migraine example: A real-world study might find that 60% of people were still taking a CGRP antibody after 12 months.

Efficacy and effectiveness

Efficacy is how well a treatment works under ideal trial conditions. Effectiveness is how well it works in everyday practice, where people miss doses, have other conditions, and take other medications.

Migraine example: A preventive may cut migraine days by 4 in a trial (efficacy) but by less in everyday practice if people miss doses (effectiveness).

Intention-to-treat and per-protocol analysis

An intention-to-treat analysis counts everyone in the group they were originally assigned to, even if they stopped the treatment. A per-protocol analysis counts only people who completed the study as planned. Intention-to-treat keeps the benefits of randomization and gives a more realistic picture.

A modified intention-to-treat analysis counts everyone who was randomized and took at least one dose or had at least one follow-up measurement.

Migraine example: If 100 people are assigned to a preventive and 20 stop early because of side effects, an intention-to-treat analysis still includes all 100.

Dropout, loss to follow-up, and missing data

Dropout is when people leave a study early. Loss to follow-up is when researchers lose contact with them. If many people leave, especially for reasons related to the treatment, the results can be biased. Studies handle missing data with methods such as imputation, which estimates the missing values, or last observation carried forward, which uses each person's last recorded value.

Migraine example: If many people on a supplement quit a migraine study because it wasn't helping, the ones who stayed may make the supplement look better than it is.

Association vs causation

An association means two things tend to happen together. Causation means one actually causes the other. Observational studies can show associations, but only well-designed experiments such as randomized trials can show causation directly.

Migraine example: People with migraine are more likely to have depression. That association alone doesn't tell us whether one causes the other or whether something else causes both.

Confounding

When a third factor affects both of the things being studied and creates a link that isn't direct. Researchers try to adjust for known confounders, but unmeasured ones can still affect the result.

Migraine example: If people with more free time are both more likely to own a dog and more likely to exercise, free time is a possible confounder in a study of dogs and migraine.

Reverse causation

When the outcome actually causes the exposure, rather than the other way around.

Migraine example: If people give up a pet because their migraine got worse, a study might wrongly conclude that giving up the pet made their migraine worse.

Regression to the mean

When something is unusually high or low, it tends to move back toward its usual level on its own. People often join migraine studies or start a new treatment during a bad stretch, so some improvement would happen even without treatment. This is one reason trials need a placebo group.

Migraine example: Someone who starts a supplement during their worst month of the year will probably have a better next month whether or not the supplement works.

Bias (selection, recall, and reporting)

Bias is any systematic error that pushes results in one direction. Selection bias happens when the people studied aren't representative. Recall bias happens when people remember past events inaccurately, often because of their current condition. Reporting bias happens when only some results are reported.

Migraine example: People with frequent migraine attacks may remember past triggers differently than people without migraine, which can create recall bias in trigger studies.

Healthy user bias

People who take a treatment or adopt a habit consistently often have other healthy habits too, which can make the treatment or habit look more beneficial than it is.

Migraine example: People who reliably take a daily supplement may also sleep more regularly and exercise more.

Immortal time bias

A mistake in some observational studies where one group gets credit for time during which the outcome couldn't have happened, making a treatment look better than it is.

Migraine example: If a study counts people as CGRP antibody users from the date of their migraine diagnosis, even though they started the drug a year later, the drug gets credit for that first year.

Generalizability

Whether the results of a study apply to people outside it. A study done in one country, age group, or clinic may not apply equally to everyone.

Migraine example: Many migraine trials exclude people over 65 and people with heart disease, so their results may not apply fully to older adults.

Funding and conflicts of interest

Who paid for a study and whether the authors have financial ties to a company. Industry-funded trials can still be high quality, but readers should know the funding source.

Migraine example: Most large trials of CGRP drugs were funded by the companies that make them. That is normal for drug approval, and readers can check the funding statement in each paper.