POPH90111 · Genetic Epidemiology
Genetic Screening
Screening is the end of the genetic-epidemiology pipeline: once we understand a genetic role, have discovered the variants and characterised their risk, the practical question is whether testing asymptomatic people does more good than harm. The course splits this into disease screening (find latent disease or its precursors) and genetic screening (find risk-raising variants before disease), judges any programme against the WHO Wilson–Jungner checklist — whose hardest principle is that a test only helps if knowing the result changes the outcome (“we can test” ≠ “we should screen”) — and then quantifies the benefit with the signature numbers NNT (carriers treated per case prevented) and NNS (people screened per case prevented). It evaluates the test itself with sensitivity, specificity and PPV/NPV, drives home that PPV depends on prevalence (a great test is useless at tiny prevalence), and summarises discrimination across thresholds with the ROC curve and AUC. The recurring lesson: screening must be targeted to high-prior-risk groups, because that lifts both PPV and the efficiency NNS.
What this chapter covers
- 018.1 Disease vs genetic screening; the Wilson–Jungner principles (WHO, 1968)
- 02The ‘can test’ ≠ ‘should screen’ gap (principles 7–9 are where genetic screens fail)
- 038.2 NNT and NNS — the course’s signature calculation (ARR → NNT → NNS)
- 04Why targeting beats population screening (NNS dominated by carrier frequency)
- 058.3 Sensitivity, specificity, PPV/NPV and the prevalence-dependence of PPV
- 068.4 The ROC curve & AUC (sensitivity vs 1−specificity; AUC as discrimination)
- 078.5 Critically appraising a screening programme
Worked example: NNT and NNS for a BRCA screen
- +1(a) Absolute risk reduction. ARR = carrier risk × proportion removed = 0.4 × 0.5 = 0.2 (use the absolute reduction, not the relative 50%).
- +1(a) NNT among carriers. NNT = 1 / ARR = 1 / 0.2 = 5 — treat 5 carriers to prevent one breast cancer.
- +2(b) NNS in the general population. Carrier frequency ≈ 1/150 ≈ 0.0067, so NNS = NNT / carrier freq = 5 / 0.0067 ≈ 746 women screened per cancer prevented.
- +1(c) Targeted NNS. With carrier frequency ≈ 0.25, NNS = 5 / 0.25 = 20 — about 37× more efficient.
- +1The lesson. NNS is dominated by carrier frequency, not by how good the treatment is, so raising the prior probability of carriage by targeting high-family-history groups is what makes a genetic screen worthwhile.
Key terms
- Wilson–Jungner principles
- The WHO (1968) 10-point checklist for whether a screen is worth running, grouped as condition (important problem, recognisable latent stage, understood natural history), test (suitable, acceptable, agreed policy on who is positive), treatment/action (an effective intervention exists, facilities available) and programme (cost balanced, a continuing process). For genetics, principles 7–9 — is there an effective action, and is the benefit worth the cost — are where most candidate screens fail.
- NNT (number needed to treat)
- The number of carriers you must treat to prevent one case: NNT = 1 / ARR, where the absolute risk reduction ARR = carrier risk × the proportion of risk removed by treatment. Using the BRCA example, ARR = 0.4 × 0.5 = 0.2 gives NNT = 5. Use the absolute, not the relative, reduction.
- NNS (number needed to screen)
- The number of people you must screen to prevent one case: NNS = NNT / carrier frequency. It folds the treatment benefit together with how rare the variant is, so it is dominated by carrier frequency — a wonderful treatment (low NNT) is wasted if you must screen thousands to find one carrier. Targeting high-prior-risk groups raises the carrier frequency and collapses the NNS (746 → 20 in the BRCA example).
- PPV and its prevalence-dependence
- Positive predictive value = TP/(TP+FP) = the probability of disease given a positive test. Unlike sensitivity and specificity (intrinsic to the test), PPV depends on who you test: as prevalence falls, the false-positive term (1−Spec)(1−Prev) dominates and PPV falls. A test with sensitivity 0.90 and specificity 0.95 has PPV ~8% at 0.5% prevalence but ~82% at 20% — the statistical reason genetic screening must be targeted.
- ROC curve and AUC
- The receiver-operating-characteristic curve plots sensitivity (true-positive rate) against 1−specificity (false-positive rate) as the test threshold sweeps; the AUC condenses it into one discrimination number = the probability a random case scores higher than a random control (0.5 = chance, 1.0 = perfect). AUC measures discrimination, not usefulness — it says nothing about PPV at a given prevalence or whether acting on the result helps.
Genetic Screening FAQ
What does ‘we can test ≠ we should screen’ mean?
It is the single biggest screening trap. The existence of an accurate test is necessary but not sufficient: a screen only helps if knowing the result reduces disease, disability or death and the benefits outweigh the harms (false positives, anxiety, insurance and legal consequences, variants of unknown significance, overdiagnosis). In the Wilson–Jungner checklist, principles 7–9 — is there an effective intervention, are facilities available, is the cost balanced — are where most candidate genetic screens fail, even when a valid test exists.
Why does NNS, not NNT, decide whether a genetic screen is worthwhile?
Because NNS = NNT / carrier frequency folds in how rare the variant is, and it is dominated by that carrier frequency rather than by how good the treatment is. In the BRCA example NNT is just 5, but at a general-population carrier frequency of 1 in 150 the NNS is about 746 — you must sift ~150 people to find each carrier. Target a high-family-history group (carrier frequency ~0.25) and the NNS falls to 20, about 37× more efficient. That is the quantitative case for targeting over population-wide screening.
Why can a great test still be useless for population screening?
Because PPV depends on prevalence. Sensitivity and specificity can both be excellent, but at tiny prevalence the false-positive term dominates and PPV approaches zero — most positives are false alarms. A test with sensitivity 0.90 and specificity 0.95 yields a PPV of only ~8% at 0.5% prevalence. This is the exact statistical reason genetic screening must be targeted to high-prior-risk groups: targeting raises the prevalence/carrier frequency and lifts PPV into a useful range.
What does the AUC tell you — and what does it not?
The AUC is the probability that a random case scores higher than a random control: 0.5 is chance (the ROC diagonal), 1.0 is perfect separation, and a CAD polygenic score sits around 0.81. It measures how well the test discriminates cases from controls. It does not tell you the PPV at a given prevalence or whether acting on the result helps. Two common slips: the ROC x-axis is 1−specificity (not specificity), and moving the cut-off trades sensitivity against specificity — you only improve both with a genuinely better test.
Exam move
The chapter rewards appraisal backed by numbers, so build the model answer to “should we screen for X?”: state the condition against Wilson–Jungner 1–3, quote test sensitivity/specificity then compute PPV at the target prevalence (where most marks sit), name the effective action and compute NNT (among carriers) and NNS (population vs targeted), weigh harms and ethics, and conclude targeted vs population vs not-at-all with the deciding number. Drill the signature chain ARR → NNT = 1/ARR → NNS = NNT/carrier frequency on the BRCA numbers, and internalise the two traps: ‘we can test’ ≠ ‘we should screen’, and a great test is useless at tiny prevalence because PPV collapses. Read the ROC carefully — x-axis is 1−specificity, AUC is discrimination not usefulness.
