Step 3 · 11–13%

Biostatistics, Epidemiology & Literature

Study designs, test math, bias, and reading a drug ad or an abstract for what it really shows.

Diagnostic tests & screening accuracy

Build the 2x2 table, then decide which number answers the question: a fixed property of the test, or this patient's chance of disease after the result.

Rules to know

Sensitivity → true positives ÷ everyone with the disease, a ÷ (a + c); a very sensitive test that comes back negative helps rule disease out
Specificity → true negatives ÷ everyone without the disease, d ÷ (b + d); a very specific test that comes back positive helps rule disease in
Positive predictive value → true positives ÷ all positive results, a ÷ (a + b); the chance someone with a positive result truly has the disease
Negative predictive value → true negatives ÷ all negative results, d ÷ (c + d); the chance someone with a negative result is truly free of disease
Same test used where the disease is rarer → PPV falls and NPV rises; sensitivity and specificity stay the same
Sensitivity 90%, specificity 90%, prevalence 1% → per 1,000 tested: 9 true and 99 false positives, PPV about 8%; most positives are false
Positive screening result in a low-risk person → confirm with a more specific test before labeling or treating
Cutoff moved so more people test positive → sensitivity and NPV rise; specificity and PPV fall
Cutoff moved so fewer people test positive → specificity and PPV rise; sensitivity and NPV fall
Positive likelihood ratio → sensitivity ÷ (1 − specificity); values above 10 raise the probability of disease a great deal
Negative likelihood ratio → (1 − sensitivity) ÷ specificity; values below 0.1 lower the probability of disease a great deal
Turning a likelihood ratio into post-test probability → pretest odds × LR = post-test odds; then probability = odds ÷ (1 + odds)
Pretest probability 20%, positive result with LR 8 → odds 0.25 × 8 = 2, so post-test probability is about 67%
Likelihood ratio near 1 → the result barely moves the probability; the test is not worth ordering for that question
Low pretest probability plus a negative highly sensitive test → disease is effectively excluded; stop the workup
High pretest probability with a negative result → still too likely to dismiss; pursue a better test or treat
ROC curve → sensitivity plotted against 1 − specificity across every cutoff; the cutoff nearest the top-left corner balances the two
Area under the ROC curve → 0.5 is no better than a coin flip, 1.0 is perfect; a larger area means better discrimination
Two tests in series, called positive only if both are positive → higher specificity and PPV, lower sensitivity; used to confirm
Two tests in parallel, called positive if either is positive → higher sensitivity and NPV, lower specificity; used when a miss is dangerous
Repeated results agree closely but sit away from the true value → precise (reliable) but not accurate (valid); suspect systematic error
Results average out to the true value but scatter widely → accurate but imprecise; random error, reduced by repeating or enlarging the sample
Kappa statistic → agreement between observers beyond what chance predicts; measures reliability, not whether the reading is correct
Only test-positives received the reference standard → verification bias; sensitivity is overstated and specificity understated
Test studied only in very sick patients and healthy volunteers → spectrum bias; accuracy will look worse in an ordinary clinic
Probability that two independent events both occur → multiply them (0.1 × 0.3 = 0.03); for either of two mutually exclusive events, add them
Decision tree → multiply each branch's probability by its outcome value, sum per option, and choose the option with the best expected value

Traps that cost points

Saying specificity drops when the test moves to a lower-prevalence clinic→ Sensitivity and specificity belong to the test and do not change with prevalence; only the predictive values shift.
Quoting sensitivity when a patient asks what a positive result means for them→ That question is answered by PPV, which depends on how common the disease is in people like this patient.
Starting treatment on one positive result from a sensitive screening test→ Sensitive tests are tuned to catch cases and so produce false positives; confirm with a specific test first.
Multiplying the likelihood ratio by the pretest probability directly→ Likelihood ratios act on odds. Convert probability to odds, multiply, then convert back; the shortcut fails badly at moderate probabilities.
Running tests in parallel to confirm a diagnosis→ Parallel testing maximizes sensitivity and suits ruling out; confirmation needs series testing, which raises specificity.
Calling a tightly reproducible test accurate→ Reproducibility is precision. Accuracy means closeness to the truth; a miscalibrated device can be precise and consistently wrong.
Ordering a test whose result would not change management→ If the post-test probability stays on the same side of the treatment threshold whatever the result, skip the test.

Risk, effect size & number needed to treat

In the Brain

Lay out the 2x2 table by exposure and outcome, use the measure the design allows, and turn every relative effect into an absolute one.

Rules to know

Relative risk from a cohort or trial → risk in exposed ÷ risk in unexposed: [a ÷ (a + b)] ÷ [c ÷ (c + d)]
Odds ratio from a case-control study → (a × d) ÷ (b × c), the odds of exposure among cases over the odds among controls
Outcome rare (under about 10%) → odds ratio approximates relative risk; with a common outcome the OR sits further from 1 than the RR
Ratio measure equal to 1 → no association; below 1 the exposure is protective, above 1 it is harmful
Absolute risk reduction → control event rate minus treated event rate; 10% minus 6% gives an ARR of 4%
Relative risk reduction → ARR ÷ control event rate, or 1 − RR; 4% ÷ 10% gives an RRR of 40%
Number needed to treat → 1 ÷ ARR written as a decimal; 1 ÷ 0.04 = 25, always rounded up to a whole patient
Number needed to harm → 1 ÷ absolute increase in the adverse effect; bleeding 3% vs 1% gives an NNH of 50
Trial with 8% events on drug vs 12% on placebo → RR 0.67, RRR 33%, ARR 4%, NNT 25
Same 25% relative reduction at baseline risk 20% vs 2% → ARR 5% (NNT 20) vs 0.5% (NNT 200); high-risk patients gain most
Stating an NNT → always attach the time frame (25 over 3 years); NNTs from trials of different length are not directly comparable
Attributable risk (risk difference) → risk in exposed minus risk in unexposed; the excess cases among the exposed caused by the exposure
Attributable risk percent → (RR − 1) ÷ RR; the share of disease in exposed people that the exposure explains
Population attributable risk → risk in the whole population minus risk in the unexposed; grows with how common the exposure is
Common exposure with modest RR vs rare exposure with high RR → the common one often causes more cases overall (larger PAR%)
Hazard ratio → compares event rates over time between groups in a time-to-event analysis; 0.75 means a 25% lower rate throughout follow-up
KM (product-limit) survival curve → stepwise share of patients still event-free over time; each drop is an event, tick marks are censored patients
Median survival on a KM survival curve → read the time at which the curve crosses 50% event-free
Two survival curves that cross → hazards are not proportional; a single hazard ratio misrepresents the effect
Confidence interval for RR, OR or HR includes 1 → not statistically significant
Confidence interval for an absolute difference includes 0 → not statistically significant
Unequal follow-up times → incidence rate = new cases ÷ person-time at risk (for example per 1,000 person-years)
Probability to odds → odds = p ÷ (1 − p); 20% probability is odds of 0.25, 50% probability is odds of 1
Relative risk requested from a case-control study → cannot be calculated because investigators fixed the number of cases; report the odds ratio

Traps that cost points

Judging benefit from the relative risk reduction alone→ A large RRR can mean a tiny ARR when baseline risk is low. Convert to ARR and NNT before advising a patient.
Computing NNT as 1 ÷ 4 when the ARR is 4%→ Use the decimal: 1 ÷ 0.04 = 25. Using the whole number gives an impossible NNT below 1.
Rounding an NNT of 24.3 down to 24→ Round NNT and NNH up to the next whole patient (25) so the benefit is never overstated.
Reading an odds ratio of 3 as triple the risk when the outcome is common→ With a frequent outcome the OR exaggerates the RR; describe it as odds or convert using the baseline risk.
Calling a hazard ratio of 0.8 with CI 0.6 to 1.05 a 20% reduction→ The interval crosses 1, so the result is not statistically significant; describe it as no clear difference.
Comparing NNTs from a 1-year trial and a 5-year trial head to head→ NNT depends on follow-up length; compare absolute risks at matched time points instead.
Subtracting relative risks to get attributable risk→ Attributable risk is a difference of absolute risks (incidence in exposed minus unexposed), not a difference of ratios.

Reading drug ads & abstracts

In the Brain

Check who was studied, against what, on which outcome and for how long; then accept only the conclusion the numbers actually support.

Rules to know

Any abstract or drug ad → first name the population, intervention, comparator, outcome and follow-up time before reading the claims
Your patient differs from those enrolled (older, kidney disease, pregnant) → results may not apply; check the inclusion and exclusion lists
New drug compared only with placebo → shows it beats nothing; says nothing about whether it beats current standard therapy
Comparator given at a low dose or an outdated drug → comparison tilted toward the new drug; superiority is not shown
Ad headline gives only a relative reduction such as 'halves the risk' → find both groups' event rates, then compute ARR and NNT
Harm reported as doubled risk, 0.1% to 0.2% → absolute increase 0.1%, NNH 1,000; present benefit and harm in absolute terms
Surrogate endpoint improved (LDL, A1c, bone density, tumor size) → does not prove fewer deaths, events or fractures; some drugs improved markers yet worsened outcomes
Composite endpoint such as death, MI or hospitalization → find which component drove the result; usually the softest, most frequent one
Benefit claimed in a subgroup of a negative trial → hypothesis-generating unless prespecified and backed by a significant interaction test
Primary endpoint not significant but a secondary endpoint is → the trial is negative; secondary findings are exploratory
Noninferiority trial result → the new drug is not worse by more than the preset margin; it does not show the drug is better
Wide noninferiority margin → a meaningfully worse drug can still pass; judge whether the margin is clinically acceptable
Short follow-up for a drug taken for years → long-term benefit and delayed harms are unknown
Trial too small to see rare harms → zero events in n patients still allows a true rate up to about 3 ÷ n (rule of three)
Heavy or unequal dropout between arms (over about 20%) → results less trustworthy; check how missing data were handled
Unblinded trial with a subjective outcome (pain, symptom scores) → expectation bias likely; hard outcomes like death resist it
Run-in period that removed nonadherent or intolerant patients → tolerability and adherence look better than they will in practice
Ad describes the drug as 'well tolerated' → compare discontinuation rates for adverse effects in each arm
Graph with a truncated or stretched y-axis → small differences look large; read the actual numbers
Observational data used for a causal claim in an ad → association only; confounding has not been excluded
Efficacy in a trial vs effectiveness in practice → expect smaller benefit with real-world adherence, comorbidity and monitoring
Dose or patient group in the ad differs from what was studied → the data do not support that use
Industry funding or authors with conflicts of interest → does not void the data; look harder at comparator, endpoints and what was left out
Primary outcome switched after enrollment began, or trial never registered → suspect selective outcome reporting
Nonsignificant result with a wide confidence interval → inconclusive, not proof of no effect; the trial may be underpowered
Asked which conclusion the data support → choose the statement limited to the population, comparator, outcome and time frame actually studied

Traps that cost points

Recommending the new drug because it beat placebo→ The patient's real alternative is current standard care, not placebo. Without a head-to-head trial you cannot call it better.
Concluding the new drug is superior after a noninferiority trial→ Noninferiority only rules out being worse than the margin; superiority needs its own prespecified test.
Accepting a better lab marker as proof of patient benefit→ Ask for clinical outcomes such as death, stroke, fracture or symptoms; marker gains have often failed to translate.
Applying results to a patient the trial excluded→ Check the exclusion criteria; frail older adults, kidney disease and pregnancy are often left out, so benefit and harm are unknown.
Trusting a subgroup benefit from a trial that failed overall→ Unplanned subgroups throw up chance findings; treat them as ideas for a new trial, not as practice changers.
Dismissing a study solely because a drug company funded it→ Funding is a reason to scrutinize design and reporting, not to reject the data outright.
Reading no significant difference as proof the drugs are equal→ Only an equivalence or noninferiority design, or a narrow CI hugging zero, supports that conclusion.
Quoting the relative risk reduction to a patient choosing therapy→ Give absolute numbers: out of 100 people like you, how many benefit and how many are harmed over the stated time.

Study designs

In the Brain

Match the question to the design, and know the one thing each design cannot show.

Rules to know

Testing whether a treatment works → randomized controlled trial; randomization balances known and unknown confounders
Rare disease, searching for past exposures → case-control study: start with cases and controls, look back, report an odds ratio
Rare exposure, or incidence needed → cohort study: start with exposed and unexposed, follow for outcomes, report relative risk
Prospective vs retrospective cohort → both begin from exposure; prospective follows forward from now, retrospective uses past records where outcomes already happened
One survey at a single point in time → cross-sectional study; gives prevalence but cannot show which came first
Rates compared across countries or regions → ecological study; group-level links cannot be applied to individuals (ecological fallacy)
Each participant gets both treatments in turn with a washout → crossover trial; patients serve as their own controls
Crossover design chosen → needs a stable chronic condition and a treatment whose effect wears off; carryover is the main threat
One or a few unusual patients described → case report or case series; no comparison group, so it only raises hypotheses
All studies on a question found by a prespecified search and appraisal → systematic review; pooling their numbers is meta-analysis
Meta-analysis → more power and precision, but only as sound as the included studies; heterogeneity weakens the pooled answer
Forest plot → each square is one study with its CI line; the diamond is the pooled estimate, significant only if it misses the no-effect line
Asymmetric funnel plot with small negative studies missing → publication bias
Intervention delivered to whole schools, clinics or towns → cluster or community trial; randomize the groups, not individuals
New drug cheaper, safer or easier, not expected to work better → noninferiority trial with a prespecified margin
Equivalence trial → shows the difference lies within a margin in both directions
Exposure is harmful (smoking, a toxin) → randomizing it is unethical; use cohort or case-control designs
Outcome takes decades to appear → case-control is fastest and cheapest; a prospective cohort would take too long
Cases and controls both drawn from an existing cohort → nested case-control; efficient use of stored samples and data
Interviews or focus groups exploring experiences → qualitative study; produces themes, not effect sizes
Judging causation → temporality is essential; strength, dose-response, consistency, plausibility and reversibility add support
Exposure may be a result of the outcome (sick people stop drinking) → reverse causation; cross-sectional data cannot separate the two
Dependent vs independent variable → the outcome measured is dependent; the exposure or intervention is independent
Treatment questions, strongest evidence first → meta-analysis of good RCTs, single RCT, cohort, case-control, case series, expert opinion
Rates compared before and after a policy change → pre-post or interrupted time series; background trends can mimic an effect

Traps that cost points

Picking a cohort study for a very rare disease→ You would need enormous numbers to collect cases; a case-control design starts from the cases.
Labeling any study that uses old records a case-control study→ Design depends on the starting point: grouped by exposure is a cohort, even if retrospective; grouped by outcome is case-control.
Drawing conclusions about individuals from country-level data→ Ecological links may not hold for people; confirm with individual-level studies.
Using a crossover trial for an acute, curable illness→ Once cured there is nothing left to cross over; crossover suits stable chronic conditions with short-acting treatment.
Treating a case series as proof that a therapy works→ Without a control group, natural history, regression to the mean and placebo effects explain much of the improvement.
Assuming a meta-analysis always outranks one large rigorous trial→ Pooling biased or heterogeneous small studies gives a precise but wrong answer; a large well-run RCT may be more reliable.
Calling a cross-sectional association causal→ Exposure and outcome are measured together, so the timing is unknown and reverse causation cannot be excluded.

Bias, confounding & validity

In the Brain

Name the bias by where the error enters (who got in, how data were gathered, who left, what else differs), then name the design fix.

Rules to know

Sample unlike the target population (volunteers, referral-center patients) → selection bias; limits generalizability or distorts comparisons
Cases recall past exposures more thoroughly than controls → recall bias; use records or exposure data gathered before the outcome
Outcome assessor knows the treatment assignment → observer (ascertainment) bias; blind the assessors
Participants behave differently because they know they are being studied → Hawthorne effect
Sicker patients drop out more from one arm → attrition bias; report the losses and analyze by intention to treat
Screening finds disease earlier but death comes at the same time → lead-time bias; survival from diagnosis looks longer, so compare mortality
Screening picks up slow-growing tumors more often than fast ones → length-time bias; screen-detected cases look more benign
Screening finds disease that would never have caused symptoms → overdiagnosis; adds treatment harm without benefit
Third factor tied to both exposure and outcome → confounding (smoking makes coffee look like a cause of lung cancer)
Controlling confounding → randomization at the design stage; restriction, matching, stratification or multivariable adjustment in observational studies
Stratum-specific results agree with each other but differ from the crude result → that variable was a confounder
Effect differs across strata (works in women, not men) → effect modification; report each stratum, do not adjust it away
Sicker patients receive the drug more often in observational data → confounding by indication; the drug looks harmful
Allocation concealment → staff enrolling patients cannot foresee the next assignment; prevents selection at enrollment, distinct from blinding
Double blinding → patients and assessors unaware of assignment; limits placebo effects and observer bias
Intention-to-treat analysis → everyone analyzed in the group assigned, whatever their adherence; preserves randomization, gives a conservative estimate
Per-protocol analysis → only those who followed their assigned treatment; can break randomization, but is reported alongside ITT in noninferiority trials
Same measurement error in both groups (nondifferential misclassification) → usually pulls the estimate toward no effect
Working populations compared with the general public → healthy worker effect; occupational hazards look smaller
Survey responders differ from nonresponders → nonresponse bias
Controls matched to cases on age and sex → matched factors can no longer be studied as risk factors
Extreme first measurement moves toward average on retest → regression to the mean; recheck before crediting an intervention
Improvement in the placebo arm → placebo effect; the reason a blinded placebo-controlled comparison is needed
Internal vs external validity → internal: results are right for those studied; external: they apply to other people and settings
Hospital patients used as controls whose illnesses relate to the exposure → hospital-control (Berkson) bias
Clinicians search harder for disease in exposed patients → detection (surveillance) bias; apply the same workup to every group

Traps that cost points

Enlarging the sample to remove bias→ Bigger samples shrink random error only; systematic error stays and needs a design fix.
Adjusting away an effect modifier→ Effect modification is a real difference between groups; report results separately for each stratum.
Equating allocation concealment with blinding→ Concealment protects who enters which arm; blinding protects what happens after assignment. A trial can have one without the other.
Claiming a screening test works because 5-year survival rose→ Earlier diagnosis lengthens measured survival without delaying death; ask for lower disease-specific mortality in a randomized trial.
Analyzing only patients who took the drug as directed→ This breaks randomization and inflates benefit; the primary analysis should be intention to treat.
Blaming recall bias in a prospective cohort→ Exposure was recorded before the outcome, so memory cannot differ by outcome; recall bias belongs to case-control and retrospective designs.
Crediting the new treatment for all improvement in patients enrolled at their worst→ Extreme values drift back toward average anyway; a randomized control group separates that drift from the treatment effect.

Hypothesis testing & choosing the statistical test

In the Brain

Know what a p-value and a confidence interval actually say, what drives power, and which test fits the type of data.

Rules to know

p-value → chance of a result at least this extreme if the null hypothesis were true; not the chance the null is true
p below the preset alpha (usually 0.05) → reject the null hypothesis of no difference
Type I error (alpha) → finding a difference that is not real; a false positive
Type II error (beta) → missing a difference that is real; a false negative, usually from too few patients
Power → 1 − beta, the chance of detecting a true effect of the stated size; trials usually aim for 80 to 90%
Power goes up with → more patients, a larger true effect, less variability in the outcome, and a higher alpha
Sample size calculation → set alpha, power, expected variability and the smallest effect worth detecting before enrolling
95% confidence interval → range of values compatible with the data; narrows as the sample grows
95% CI excludes the no-effect value → p below 0.05; if it includes that value, p is above 0.05
Confidence intervals of two group means do not overlap → difference is significant; modest overlap can still be significant
Many comparisons at alpha 0.05 → 20 independent tests give about a 64% chance of at least one false positive; Bonferroni divides alpha by the number of tests
Statistically significant but tiny difference (2 mm Hg) → large samples make trivial effects significant; judge clinical importance separately
Comparing means of two groups → t-test; paired t-test when the same patients are measured before and after
Comparing means of three or more groups → analysis of variance (ANOVA)
Comparing proportions across categories → chi-square test; Fisher exact test when expected counts are small
Skewed or ranked data that break t-test assumptions → nonparametric tests (Mann-Whitney, Wilcoxon signed-rank, Kruskal-Wallis)
Association between two continuous variables → Pearson correlation; Spearman for ranked or skewed data
Correlation coefficient r → runs from −1 to +1; sign gives direction, size gives strength, r² is the share of variation explained
Strong correlation between two variables → does not prove causation; a confounder or reverse causation may explain it
Predicting an outcome while adjusting for several factors → multiple linear regression for continuous outcomes, logistic regression for yes/no outcomes
Comparing time-to-event curves → log-rank test; Cox proportional hazards regression adjusts for covariates and yields a hazard ratio
Normal distribution → mean, median and mode coincide; about 68% within 1 SD, 95% within 2 SD, 99.7% within 3 SD
Right (positive) skew, long tail toward high values → mean above median above mode (hospital stay, income)
Left (negative) skew, long tail toward low values → mean below median below mode
Skewed data or outliers → summarize with median and interquartile range rather than mean and SD
Standard deviation vs standard error → SD describes spread of individual values; SE = SD ÷ √n describes precision of the mean
Nominal, ordinal and continuous data → unordered categories, ordered categories, measured numbers; the data type picks the test
Analysis planned before data collection vs after looking → prespecified results carry their stated alpha; post hoc findings need confirmation

Traps that cost points

Reading p = 0.03 as a 3% chance the drug does nothing→ The p-value assumes the null is true; it gives the chance of data this extreme, not the chance any hypothesis is right.
Calling p = 0.08 proof of no effect→ Failing to reject the null does not prove it; look at the confidence interval and whether the study had enough power.
Using a t-test to compare the percentage of patients cured→ Proportions are categorical outcomes; use chi-square, or Fisher exact for small counts.
Running a separate t-test for every pair among four groups→ Multiple t-tests inflate false positives; use ANOVA, then a corrected post hoc comparison.
Reporting the mean length of stay for heavily skewed data→ A few very long stays drag the mean up; the median with interquartile range describes the typical patient.
Confusing standard error with standard deviation→ SE shrinks as more patients are added and describes the mean's precision; SD describes how individuals vary and does not shrink.
Assuming overlapping confidence intervals rule out significance→ Intervals can overlap modestly while the difference is still significant; test the difference directly.

Population health & prevention

In the Brain

Know what moves incidence vs prevalence, compare populations fairly with adjusted rates, and place each intervention on the prevention ladder.

Rules to know

Incidence → new cases over a period ÷ people at risk; the risk of developing disease
Prevalence → all existing cases at one time ÷ total population; roughly incidence × average duration
Treatment prolongs life without curing → prevalence rises while incidence is unchanged
Disease becomes rapidly fatal or rapidly curable → prevalence falls while incidence is unchanged
New exposure, wider case definition or better detection → incidence rises, and prevalence follows
Mortality rate vs case-fatality rate → deaths ÷ whole population per period vs deaths ÷ people who have the disease
Comparing death rates of populations with different age mixes → use age-adjusted (standardized) rates, not crude rates
Direct standardization → apply each population's age-specific rates to one standard population and compare the totals
Standardized mortality ratio → observed ÷ expected deaths (indirect standardization); above 1 means more deaths than expected
Infant mortality rate → deaths before age 1 per 1,000 live births; the neonatal rate covers the first 28 days
Maternal mortality ratio → deaths from pregnancy-related causes during pregnancy or within 42 days after, per 100,000 live births
Primary prevention → stops disease before it starts: vaccines, smoking cessation, seat belts, statins for high-risk people
Secondary prevention → finds disease early before symptoms: mammography, cervical cytology, blood pressure checks
Tertiary prevention → limits harm from established disease: cardiac rehab, diabetic foot care, stroke rehabilitation
Quaternary prevention → protects patients from overtesting and overtreatment
Disease worth screening for → important, with a detectable presymptomatic phase, an acceptable accurate test, and treatment that works better when started early
Proof a screening program works → lower disease-specific or all-cause mortality in randomized trials, not longer survival after diagnosis
Herd immunity → enough immune people block chains of transmission, protecting those who cannot be vaccinated
Herd immunity threshold → about 1 − 1/R0; measles, with R0 of 12 to 18, needs roughly 92 to 95% immunity
Basic reproduction number R0 → average new infections from one case in a fully susceptible group; above 1 spreads, below 1 fades
Years of potential life lost → weights deaths at younger ages; injuries and overdoses rank high
Quality-adjusted life year → a year of life weighted by health quality from 0 to 1; cost per QALY compares interventions
Disability-adjusted life year → years lost to early death plus years lived with disability; a measure of disease burden
Population-wide policy (tobacco taxes, smoke-free laws, sugary drink taxes) → shifts the whole community's risk; often reaches more people than office counseling
Unmet social needs found at a visit (food, housing, transport) → screen routinely and link to community resources as part of the plan
Health disparity → worse outcomes in a disadvantaged group; examine access, quality of care, bias and structural causes
Population pyramid narrowing at the base and widening at the top → aging population with rising chronic disease and caregiving needs

Traps that cost points

Thinking a better treatment lowers prevalence→ If it keeps people alive without curing them, more people live with the disease, so prevalence rises.
Comparing crude death rates between a retirement community and a college town→ Age drives death rates; compare age-adjusted rates or a standardized mortality ratio.
Calling mammography primary prevention→ Screening finds existing disease early, which is secondary prevention; primary prevention stops disease from starting.
Using case-fatality and mortality rate interchangeably→ Case-fatality uses only people with the disease as the denominator; mortality uses the whole population.
Screening for a disease that has no better outcome when treated early→ Early detection helps only if earlier treatment changes the outcome; otherwise it adds anxiety, cost and overdiagnosis.
Assuming a small dip in vaccine coverage is harmless→ For highly contagious infections the threshold is high; coverage just below it lets outbreaks take off.
Treating disparities as purely individual choices→ Outcomes track access, neighborhood, income and bias; address barriers in the plan and connect patients to resources.

Outbreaks & disease surveillance

In the Brain

Confirm the outbreak, define a case, read the epidemic curve, find the vehicle with attack rates, and use the right public-health measure.

Rules to know

More cases than the usual baseline in a place and time → first verify the diagnosis and confirm the excess is real
Steps of an outbreak investigation → verify, define cases, find cases, describe by time, place and person, form and test hypotheses, control, communicate
Case definition → clinical criteria limited by person, place and time; may be graded confirmed, probable or possible
Point-source epidemic curve → a single sharp peak within one incubation period, as after a shared meal
Continuous common-source curve → cases plateau for as long as the exposure lasts, as with a contaminated water supply
Propagated epidemic curve → successive, growing waves about one incubation period apart; person-to-person spread
Attack rate → ill people ÷ people exposed during the outbreak period
Finding the food vehicle → compare attack rates in eaters vs non-eaters of each item; the biggest gap, with most cases among eaters, points to it
Secondary attack rate → new cases among contacts ÷ susceptible contacts; measures person-to-person spread
Vaccine effectiveness → (attack rate unvaccinated − attack rate vaccinated) ÷ attack rate unvaccinated; 20% vs 2% gives 90%
Hypothesis about the source after descriptive data → test it analytically, often with a case-control study of exposures
Passive surveillance → clinicians and labs report notifiable diseases on their own; cheap but undercounts
Active surveillance → health department reaches out to providers and labs to find cases; used in outbreaks and elimination efforts
Sentinel surveillance → chosen sites report all cases of a condition (influenza-like illness) to track trends
Syndromic surveillance → prediagnostic data such as ED complaints or pharmacy sales to spot outbreaks early
Isolation vs quarantine → isolation separates people who are ill; quarantine restricts exposed, well people for an incubation period
Contact tracing → identify, notify and monitor contacts, offering vaccine or prophylaxis when available
Ring vaccination → vaccinate contacts and contacts of contacts around each case to halt spread
Case count jumps after a new test or wider definition → possible surveillance artifact; check before declaring an outbreak
Endemic vs epidemic vs pandemic → usual baseline level vs excess over expected in an area vs epidemic spread across countries or continents
Unusual agent, or severe illness clustering in healthy people → think bioterrorism (also disease out of season or place); notify public health at once
Urgently notifiable condition suspected (measles, meningococcal disease, botulism) → call the health department on suspicion, not after confirmation
Health advisory issued by the health department → follow its testing, reporting and patient-notification guidance in your practice

Traps that cost points

Declaring an outbreak after a lab adopts a more sensitive test→ Changes in detection can raise counts without more disease; compare against the case definition and the baseline first.
Calling the confinement of exposed but well contacts isolation→ Isolation is for the sick; separating exposed, healthy people through the incubation period is quarantine.
Reading growing waves of cases as a single shared exposure→ Repeated peaks one incubation period apart mean person-to-person spread; control must target transmission chains.
Blaming the food with the highest attack rate among eaters alone→ Compare with non-eaters; a dish nearly everyone ate can show a high rate without being the source.
Waiting for culture confirmation before calling public health about suspected measles→ Some conditions require notification on suspicion so contacts can be protected within the window for prophylaxis.
Using the whole population as the attack-rate denominator→ Divide by the people actually exposed during the outbreak period, or the rate is diluted.

Research ethics & drug development

In the Brain

Know what each trial phase can show, when a placebo is ethical, and what consent, review boards and monitoring committees protect.

Rules to know

Phase I trial → small group, often healthy volunteers; safety, dose range and pharmacokinetics
Phase II trial → a few hundred patients with the disease; early efficacy, dose finding and short-term harms
Phase III trial → large randomized comparison with placebo or standard care; the basis for approval
Phase IV → postmarketing surveillance for rare, long-term and real-world adverse effects
Any research involving human participants → institutional review board approval before enrollment
Consent for research → purpose, procedures, risks, benefits, alternatives, confidentiality, and the right to refuse or withdraw without losing care
Child enrolled in research → parental permission plus the child's own assent when developmentally able
Adult research participant who lacks capacity → a legally authorized representative may consent where the review board allows; extra protections apply
Placebo control → acceptable only when no proven therapy exists or withholding it risks no serious or lasting harm
Equipoise → genuine uncertainty about which arm is better; needed to randomize ethically
Data safety monitoring board → independent group reviews unblinded interim data and can stop a trial for benefit, harm or futility
Trial stopped early for benefit → effect size tends to be overestimated; interpret with caution
Serious unexpected adverse event during a trial → report promptly to the review board and sponsor
Participant asks to leave a study → allowed at any time without penalty or any change in usual care
Physician enrolling their own patients → disclose the dual role and any payment; a refusal cannot affect the patient's care
Payment to research participants → reasonable compensation for time and burden; not so large it pressures people into risk
Health records used for research → requires patient authorization, a review-board waiver, or properly de-identified data
Trial registration → design and primary outcome recorded in a public registry before the first patient enrolls
Off-label prescribing → legal and common when evidence supports it; explain it to the patient and document the reasoning
Serious illness, no alternative and no trial available → expanded access (compassionate use) to an investigational drug

Traps that cost points

Using a placebo arm when an effective treatment exists for a serious disease→ Compare against standard care; withholding proven therapy that prevents death or lasting harm is unethical.
Skipping consent because the study is low risk→ Consent is required unless the review board formally waives it under strict criteria.
Expecting phase III trials to reveal rare side effects→ A few thousand patients cannot detect a 1-in-10,000 harm; postmarketing surveillance finds those.
Implying a patient who declines a trial will lose their doctor→ Participation is voluntary; usual care continues unchanged after a refusal or withdrawal.
Treating all off-label prescribing as illegal or experimental→ Prescribing off label is permitted and often standard; manufacturer promotion of off-label uses is what is restricted.
Letting investigators view unblinded interim results to decide on stopping→ An independent monitoring board reviews interim data so the trial team stays blinded.