Biostatistics

• Definition: Biostatistics is the application of probability theory to biological and clinical data — quantifying how likely an observed difference reflects a true effect rather than random sampling variation. Epidemiology supplies the study designs; biostatistics supplies the inference.
• Why it matters clinically: Every diagnostic decision is a probability revision. Ordering a test does not produce truth; it moves a pre-test probability toward or away from disease according to the test's likelihood ratios. Mechanistically, this is why the same test performs differently in an ED chest-pain population and in an asymptomatic screening population — the test's operating characteristics are treated as fixed for exam purposes (though spectrum bias can shift them in real populations, since a test performs better in advanced than in early disease), while the prior probability changes dramatically.
• Where it appears in practice:
• Screening policy: the USPSTF grades screening recommendations A through D plus I (insufficient evidence), and those grades rest explicitly on tradeoffs between detection and false-positive harm — i.e., on sensitivity, specificity, and prevalence.
• Therapeutics: guideline bodies (ACC/AHA, ADA Standards of Care, GOLD) grade recommendations by class of recommendation and level of evidence, which are statistical judgments about effect size and precision.
• Literature appraisal: CONSORT (randomized trials) and STROBE (observational studies) reporting standards exist because incomplete statistical reporting distorts clinical practice.
• Exam epidemiology: Biostatistics, epidemiology, and population health form a recurring content category on both Step 1 and Step 2 CK blueprints, and the items cluster tightly around 2×2 tables, PPV/NPV under changing prevalence, NNT, and p-value/confidence-interval interpretation.
• The recurring trap: statistical significance and clinical significance are separate questions, and a study can deliver one without the other.

The 2×2 table (disease across the top, test result down the side)

• Sensitivity = TP / (TP + FN): among people with disease, the fraction the test catches. Read across the diseased column.
• Specificity = TN / (TN + FP): among people without disease, the fraction correctly called negative.
• Positive predictive value = TP / (TP + FP): given a positive result, the probability disease is truly present. Read across the test-positive row.
• Negative predictive value = TN / (TN + FN): given a negative result, the probability of no disease.
• Likelihood ratios: LR+ = sensitivity / (1 − specificity); LR− = (1 − sensitivity) / specificity. These are prevalence-independent like sensitivity/specificity, but unlike them they directly convert pre-test odds into post-test odds — which is why they are the most clinically usable numbers on the table.

Cutoff behavior

• Lowering the diagnostic threshold raises sensitivity and lowers specificity; raising it does the reverse. The ROC curve plots this tradeoff (sensitivity vs. 1 − specificity); area under the curve near 1.0 = excellent discrimination, 0.5 = coin flip.

Effect measures

• Relative risk = risk in exposed / risk in unexposed (cohort, RCT). Odds ratio = (a×d)/(b×c) (case-control).
• Absolute risk reduction = control event rate − treatment event rate. Relative risk reduction = ARR / control event rate. NNT = 1/ARR. RRR looks impressive when baseline risk is tiny; ARR does not lie.

Inference

• P-value: probability of data this extreme if the null were true. Per the American Statistical Association's statement on p-values, it is not the probability the hypothesis is true and should not be used alone to declare an effect real.
• Power (1 − β): rises with larger sample size, larger true effect, and higher α. Underpowered negative trials show absence of evidence, not evidence of absence.

• Build the table: 100 have disease, 9,900 do not.
• True positives = 0.95 × 100 = 95; false negatives = 5.
• False positives = 0.10 × 9,900 = 990; true negatives = 8,910.
• PPV = 95 / (95 + 990) ≈ 9%. Over 90% of positives are false.
• NPV = 8,910 / (8,910 + 5) ≈ 99.9%.
• Interpretation: the test is unchanged, but because healthy people vastly outnumber diseased ones, the false-positive pool swamps the true-positive pool. This is the exact arithmetic behind USPSTF debates over screening harms — downstream biopsies, imaging, and anxiety generated by mathematically inevitable false positives.
• The follow-up question examiners love: "How do you raise PPV?" PPV rises most steeply with increased specificity (fewer false positives) and with higher prevalence — so screen a higher-risk group or apply a more specific confirmatory test in sequence. Increasing sensitivity raises PPV only marginally (here, going from 95% to 99% sensitivity moves PPV from roughly 8.8% to 9.1%), and in practice sensitivity is bought by lowering the threshold, which sacrifices specificity and can actually lower PPV.

• RRR = (10 − 8)/10 = 20% — the number the press release quotes.
• ARR = 2 percentage points; NNT = 1/0.02 = 50 patients treated for 5 years to prevent one MI.
• Always state the time horizon with an NNT; an NNT without a duration is uninterpretable.

• Likelihood ratios are the right answer when a stem asks what to use across different populations, because they are prevalence-independent yet still translate a pre-test probability into a post-test probability. Rough anchors: LR+ above roughly 10 substantially raises post-test probability; LR− below roughly 0.1 substantially lowers it.
• Spectrum bias: sensitivity measured in a referral population with advanced disease overstates performance in primary care, where disease is early and subtle. A test validated in the wrong spectrum will underperform when deployed.
• Nonsignificance is not equivalence. An underpowered trial (high β) yields type II error; before answering "no difference," check sample size, event count, and interval width. Demonstrating equivalence or noninferiority requires a prespecified margin and a different design.
• Lead-time and length-time bias make screened patients appear to survive longer without any true mortality benefit — earlier diagnosis simply starts the clock sooner, and indolent tumors are preferentially caught. This is why the USPSTF anchors screening grades to disease-specific and all-cause mortality rather than post-diagnosis survival duration.
• Intention-to-treat preserves randomization and gives the more conservative, real-world estimate; per-protocol analysis inflates efficacy by discarding nonadherent patients, reintroducing the confounding randomization was meant to eliminate.
• Match the statistical test to the data: two means → t-test; three or more means → ANOVA; proportions/categorical data → chi-square (Fisher exact for small cells).
• Incidence versus prevalence: prevalence ≈ incidence × duration. A therapy that prolongs life without curing disease raises prevalence while leaving incidence untouched — a favorite distractor with antiretroviral and heart failure stems.
• Surrogate endpoints (LDL, HbA1c, ejection fraction) can move without any change in the outcome patients feel; ACC/AHA and ADA recommendations carry the most weight when anchored to hard clinical events.

• Sensitivity & Specificity are properties of the TEST; Positive/Negative Predictive Value depend on DISEASE PREVALENCE
• Type I error (α) = false positive; Type II error (β) = false negative; Power = 1 - β
• Relative Risk (RR) used in cohort studies; Odds Ratio (OR) used in case-control studies; OR approximates RR when disease is rare
• Number Needed to Treat (NNT) = 1/ARR; Number Needed to Harm (NNH) = 1/ARR for adverse event
• P-value < 0.05 is standard cutoff for statistical significance; confidence intervals that exclude 1 (OR/RR) or 0 (mean difference) are significant

1. 1.Confusing Sensitivity/Specificity with PPV/NPV

• Sensitivity & specificity are FIXED test properties (intrinsic)
• PPV/NPV CHANGE with disease prevalence (extrinsic) — high prevalence → higher PPV

1. 1.Misinterpreting P-value

• P-value is NOT probability that null hypothesis is true; it's the probability of observing this data (or more extreme) IF null is true
• P-value doesn't measure effect SIZE; large studies can have small p-values with clinically insignificant effects

1. 1.Wrong Association Measure for Study Design

• Use RR for prospective cohort/RCT (you follow people forward in time)
• Use OR for case-control (you start with disease outcome and look backward)

• Choosing the correct study design for your research question
• Calculating/interpreting sensitivity, specificity, NPV, PPV, RR, OR, NNT
• Assessing statistical significance (p < 0.05) AND clinical significance (effect size)
• Recognizing bias and confounding sources and how to mitigate them in study design or analysis