Problem set 11A — Applications and ethics
Covers Ch 57–58.
Attempt before revealing. Part 12 looks like the argumentative part of the course and mostly is not. Nearly every disagreement in it turns on a number somebody declined to compute: what a positive is worth at a realistic prevalence, how many people a database exposes through relatives who never joined it, how far a actually moves one person. Six of these eight problems are arithmetic; the two that are not are built to stop you substituting a verdict for an argument. What the set will expose, if you read Part 12 rather than calculating it, is the habit of accepting a percentage — 99.7% sensitive, 2% , 3× relative risk — without asking what it is conditional on. Several answers are the opposite of the intuition, problems 3, 4 and 7 most sharply.
Conventions fixed for this set, so nothing turns on a guess. False-positive rate means 1 − specificity throughout. Every rate, , threshold, detection rate and cost supplied below is given data for the problem — the chapters supply the machinery and the direction of the effect, not these particular values. Where a table is labelled supplied, use it as stated rather than any figure you may know from elsewhere.
1. A positive screen, and what it is worth
A prenatal cell-free DNA screening programme runs in a cohort of younger mothers where the birth prevalence of 21 is 1 in 1,500. The assay is the one from Ch 57 §3: sensitivity 99.7%, false-positive rate 0.04%.
(a) A patient positive. Answer her. Work in a cohort of 150,000 pregnancies. (b) A different patient screens negative. Answer her too, and name the asymmetry. (c) The laboratory can fund exactly one improvement: sensitivity to a perfect 100%, or the false-positive rate halved to 0.02%. Compute the new PPV under each and say which to buy. (d) A classmate reaches for the chapter's shortcut PPV ≈ prevalence × sensitivity / FPR, gets 1.66, and reports a PPV of 166%. What went wrong, and state the condition under which the shortcut is safe.
Solution
(a) Build the 2×2 table. Never reason about a screen without one.
trisomy 21 euploid total
screen positive 99.70 59.96 159.66
screen negative 0.30 149,840.04 149,840.34
---------------------------------------------------------
total 100.00 149,900.00 150,000.00
- affected = 150,000 / 1,500 = 100; unaffected = 149,900
- true positives = 100 × 0.997 = 99.7
- false positives = 149,900 × 0.0004 = 59.96
PPV = 99.7 / 159.66 = 62.4%
So: this result raises the probability the fetus has trisomy 21 from about 0.07% to about 62%. It is a large change and it is not a diagnosis. Nearly two positives in five are wrong. The next step is a diagnostic test — CVS or amniocentesis — not a decision.
Sanity check you should do every time. Ch 57 §3 computes 71.4% for the same assay at a prevalence of 1 in 1,000. Our prevalence is lower, so our PPV must be lower, and 62.4% < 71.4%. If your answer had come out above 71%, the arithmetic was wrong, not the intuition.
(b) NPV = 149,840.04 / 149,840.34 = 99.9998%, i.e. about 1 false reassurance in 500,000.
The asymmetry is the definition of a screen. The same laboratory, on the same run, returns a negative that is nearly conclusive and a positive that is barely more likely than not. That is not a defect and it is not marketing: it follows from the false positives being drawn from a pool 1,499 times larger than the true positives. A negative cfDNA screen is genuinely informative; a positive cfDNA screen is a reason to do a different test.
(c) Recompute each in turn, holding everything else fixed.
| Change | True positives | False positives | PPV |
|---|---|---|---|
| (none) | 99.70 | 59.96 | 62.4% |
| sensitivity → 100% | 100.00 | 59.96 | 62.5% |
| FPR → 0.02% | 99.70 | 29.98 | 76.9% |
Perfect sensitivity buys 0.07 percentage points. Halving the false-positive rate buys 14.4 points — more than 200 times as much, for a change in specificity from 99.96% to 99.98%, which sounds like nothing.
Buy the specificity. And notice why the trade-off is invisible when quoted as specificity: two hundredths of a percentage point of specificity is half the false positives. Specificity is the wrong unit to think in; the false-positive count is the right one, because that is the number that lands in the denominator of the PPV.
(d) The shortcut drops the true positives out of the denominator:
exact PPV = (π × sens) / (π × sens + (1 − π) × FPR), and the shortcut is what remains when π × sens is negligible against (1 − π) × FPR.
Here it is not negligible — it is the majority, because PPV = 62%. The shortcut asserts that true positives are a small minority of positives and then divides by a number that is too small, so it produces a "probability" above 1, which is the error announcing itself.
The condition: the shortcut is safe when π × sens is far below FPR, equivalently when π is far below FPR/sens = 0.0004/0.997 = 1 in 2,492. Our prevalence, 1 in 1,500, is 1.66 times larger than that — hence the 1.66.
Where it is valid, it is excellent. On the chapter's 22q11.2 example (π = 2.5 × 10^-4, sens = 0.978, FPR = 0.0076): π × sens = 2.4 × 10^-4 against (1 − π) × FPR = 7.6 × 10^-3, a ratio of 0.032, and the shortcut gives 3.22% against an exact 3.12%.
Generalise. The shortcut fails exactly when the screen is working well, and works exactly when the screen is returning mostly false positives. That is a useful diagnostic in itself: if the shortcut and the exact answer agree, your test is dominated by false positives.
2. Newborn screening: which error costs more ★
A genomic newborn screening programme of the kind described in Ch 57 §4 includes gene G for a condition that is treatable in childhood. 1,000,000 newborns. Supplied:
| Quantity | Value |
|---|---|
| Newborns carrying a variant on the panel's pathogenic list for G | 1 in 5,000 |
| of those variants, from families ascertained because a child was affected | 0.90 |
| Penetrance of those variants in an unselected population (the truth, unknown to the programme) | 0.10 |
| Sequencing call accuracy | assume perfect — every variant carrier is flagged, nobody else is |
| The panel's variant list captures every disease-causing variant in G | assume true |
The biochemical predecessor — a dried blood spot on tandem mass spectrometry, which measures the disease process rather than the — has sensitivity 0.98 and a false-positive rate of 1 × 10^-5 for the same condition.
(a) How many babies does the genomic screen , and what does the programme implicitly claim? What is the true PPV? (b) Compute the PPV of the biochemical screen. Be careful which prevalence you use. (c) Let C_FP be the harm of one false positive and C_FN the harm of one missed case. At what ratio C_FN / C_FP do the two screens break even? (d) Ch 57 §4 says replacing biochemistry with sequencing "breaks that logic in a specific way". Which logic, which way, and what single piece of evidence would fix it?
Solution
(a) Flagged = 1,000,000 / 5,000 = 200 babies.
The programme's literature quotes penetrance 0.90, so it implicitly claims PPV = 90% — 180 of the 200 will develop the condition.
The truth: 200 × 0.10 = 20 cases. True PPV = 20/200 = 10%.
180 healthy children have been labelled, nine for every one helped. And note where the error lives. Not in the assay — the sequencing is stipulated perfect. Not in the variant list — it is stipulated complete. The error is entirely in a penetrance estimate imported from families ascertained because someone was affected, which is the one number the assay cannot measure and the programme did not check.
Scale check. Is 1 in 5,000 plausible for one gene on a 500-gene panel? If every panel gene flagged at this rate the programme would flag 500/5,000 = 1 in 10 babies. The Generation Study expects about 1 in 100 (Ch 57 §4), so the average panel gene must flag around 1 in 50,000 and G is an unusually common one. Do this check whenever a per-item rate is offered for a many-item panel.
(b) The trap is the prevalence. The biochemical assay detects the disease, not the genotype, so its denominator is the 20 true cases — not the 180 non-penetrant carriers, who are biochemically normal by construction, and not the 200 flagged.
affected unaffected total
positive 19.60 10.00 29.60
negative 0.40 999,970.00 999,970.40
------------------------------------------------------------
total 20.00 999,980.00 1,000,000.00
- true positives = 20 × 0.98 = 19.6
- false positives = 999,980 × 1 × 10^-5 = 10.0
PPV = 19.6 / 29.6 = 66.2%, against the genomic screen's 10%.
The wrong path, which is worth walking. Using 180 as the case count (the penetrance the programme believes) gives affected = 180, TP = 176.4, FP = 999,820 × 10^-5 = 10.0, PPV = 94.6% — and it is nonsense, because those 180 do not have the disease and a biochemical assay would not find one. Choosing the denominator is the whole problem here, and it is not a detail of the arithmetic: the two screens are being asked different questions, and only one of them is about a sick child.
(c) Compare the two programmes on the same cohort.
| True positives | False positives | False negatives | |
|---|---|---|---|
| Genomic | 20.0 | 180.0 | 0.0 |
| Biochemical | 19.6 | 10.0 | 0.4 |
| Genomic minus biochemical | +0.4 | +170.0 | −0.4 |
Genomic sequencing buys 0.4 additional true cases and pays 170 additional false positives. Break-even:
0.4 × C_FN = 170 × C_FP → C_FN / C_FP = 425
So the genomic screen is the better buy only if one missed treatable case is worse than 425 false positives. Whether it is depends entirely on what a false positive actually costs — a repeat blood spot and a fortnight of parental fear, or a childhood of surveillance appointments, an insurance record, and a family that treats a well child as ill. The chapter's phrase is that the Generation Study reports a suspicion, not a diagnosis, and the whole weight of the design sits on how quickly and how completely a suspicion can be retired.
(d) The logic is Wilson–Jungner: screen only for conditions that are serious, detectable presymptomatically, and treatable, with earlier treatment better. Every clause is about the disease.
The way it breaks: biochemistry measures the disease process; sequencing measures a genotype. A raised metabolite in a well baby is the disease, early. A pathogenic variant in a well baby is a probability whose value nobody has measured in well babies — because every published penetrance estimate came from families ascertained because someone was affected. That is what turned a 90% claim into a 10% reality in (a), and it is the difference between the two screens' PPVs in (b).
The evidence that would fix it: penetrance estimated in an unselected population — a large population biobank, genotype-first, counting how many carriers of the listed variants actually develop the condition without having been recruited through an affected relative. Nothing else substitutes: not a better assay, not a longer gene list, not a bigger cohort of families.
Generalise. Every time a test moves upstream of the disease it is measuring — genotype instead of metabolite, screen instead of symptom — the prior falls and the PPV falls with it. Problem 1 is the same sentence in a different clinic.
3. Designing a carrier-screening panel
A laboratory is designing a pan-ethnic carrier screening panel. Supplied inclusion rule: include a gene if its is at least 1 in 200 in at least one of the reference populations against which the laboratory estimates ancestry. Carrier frequencies:
| Gene | Reference population A | Reference population B |
|---|---|---|
| G1 | 1 in 25 | 1 in 90 |
| G2 | 1 in 150 | 1 in 400 |
| G3 | 1 in 1,200 | 1 in 60 |
| G4 | 1 in 900 | 1 in 1,100 |
(a) Which genes go on the panel, and via which population? (b) A couple whose genomes both resemble reference population B are considering a pregnancy. Before any testing, what is their risk of a child affected by the G3 condition? (c) Both negative for G3. The panel's variant list detects 90% of the pathogenic present in G3 in this population. What is their residual risk, and by what factor did the test reduce it? Predict the factor before you compute it. (d) The laboratory's competitor sells an ancestry-targeted panel: G3 is offered only to people who self-report an ancestry associated with population B. Name the two ways this fails, and state what a "negative" on a three-variant BRCA panel has in common with it.
Solution
(a) Threshold 1 in 200 = 0.005. Convert everything to a decimal before comparing — comparing denominators invites a sign error.
| Gene | A | B | Verdict |
|---|---|---|---|
| G1 | 0.0400 | 0.0111 | include — clears in both |
| G2 | 0.00667 | 0.0025 | include — clears via A only |
| G3 | 0.00083 | 0.01667 | include — clears via B only |
| G4 | 0.00111 | 0.00091 | exclude — clears in neither |
G2 and G3 are the point of the exercise. Each clears the threshold in exactly one reference population and would be missing from a panel built around the other. A rule of the form "common enough anywhere" is what makes a single panel serve everyone, and it is why the pan-ethnic design does not need to know anybody's self-reported ancestry to work.
(b) Both parents must be carriers, and both must transmit:
risk = (1/60) × (1/60) × 1/4 = 1/14,400 = 6.94 × 10^-5, about 1 in 14,400.
(This is the Ch 26 random-couple calculation you did in problem set 07 problem 2(c), reused here as a prior.)
(c) First the per-person posterior. Count, per 60,000 people from population B:
carriers 1,000
variant on the list 900 -> screen POSITIVE
variant not on the list 100 -> screen NEGATIVE
non-carriers 59,000 -> screen NEGATIVE
-----------------------------------------------------
screen negatives 59,100 of which 100 are carriers
P(carrier | negative) = 100 / 59,100 = 0.001692 = 1 in 591, down from 1 in 60.
Algebraically, with carrier frequency c and detection rate DR:
P(carrier | negative) = c(1 − DR) / (1 − c·DR) = (0.016667 × 0.10) / (1 − 0.015) = 0.001692 ✓
Residual couple risk = 0.001692^2 × 1/4 = 7.16 × 10^-7, about 1 in 1,400,000.
Reduction factor = (1/14,400) / (1/1,397,000) = 97-fold.
The prediction most people make is 10-fold, reasoning that the test misses 10% of alleles. It is wrong by an order of magnitude, and the reason is structural: both parents must be carriers, so the couple's risk is the square of a per-person probability, and squaring a 9.85-fold reduction gives 97.
Exactly: the per-person reduction is c / [c(1 − DR)/(1 − c·DR)] = (1 − c·DR)/(1 − DR) = 0.985/0.10 = 9.85, and 9.85^2 = 97.02. As DR → 1 the per-person factor grows like 1/(1 − DR) and the couple factor like 1/(1 − DR)^2, which is why the last few percentage points of detection rate are worth so much more than the first few.
(d) Two failures, both named in Ch 57 §4.
- Self-reported ancestry is a poor proxy for which founder variants someone carries. The thing that determines carrier status is which segments of genome the person actually has; the thing the intake form records is a social self-description. They correlate imperfectly, so the targeting mis-assigns people in both directions — and mis-assignment here means a test that was never offered.
- It fails completely for people of admixed background. There is no box to tick that maps onto a genome with segments from several reference populations, so the rule has no defined behaviour for a growing fraction of the population.
What a three-variant BRCA panel has in common: of a gene is not coverage of a condition (Ch 57 §11). A test genotyping three founder variants out of thousands of known pathogenic variants in BRCA1 and BRCA2 returns a "negative" whose detection rate is low and unstated — which is precisely part (c) run with a DR near zero, where the reduction factor is near 1 and the negative is close to uninformative. In both cases the failure is that the report says negative and the patient hears not a carrier.
Generalise. A negative screening result is a likelihood ratio applied to a prior, exactly like a positive one. You cannot interpret it without the detection rate, and the detection rate is a property of the variant list, not of the gene.
4. Who your relatives sign up for ★
A national genealogy service holds 600,000 profiles in a population of 40 million. Use the exposure model from Ch 58 §1:
P(findable) = 1 − (1 − c)^N
where c is the fraction of the population in the database and N is an effective number of relatives close enough to leave a detectable IBD segment. Use the chapter's deliberately conservative N = 45 unless told otherwise.
(a) What fraction of the population is findable, how many people is that, and how many of them are not customers? (b) The service doubles to 1,200,000 profiles. Recompute. Does exposure double? How many additional non-customers did each additional customer expose? (c) What would make 95% of the population findable, and how many profiles is that? (d) Ch 58 §1 is blunt that no single N reproduces Erlich et al.'s published results: 60% at ~0.9% coverage needs N ≈ 107, and
99% at 2% coverage needs N ≈ 228. Recompute (a) at both. Given that the answer moves that much,
what does the model still establish, and what may you not claim with it?
Solution
(a) c = 600,000 / 40,000,000 = 0.015.
P = 1 - (1 - 0.015)^45
= 1 - exp(45 x ln(0.985))
= 1 - exp(45 x (-0.0151136))
= 1 - exp(-0.680114)
= 1 - 0.506559
= 0.4934 -> 49.3%
People findable = 0.4934 × 40,000,000 = 19,740,000. Of those, 600,000 are customers, so 19,140,000 are not.
About 32 non-consenting people are exposed per consenting customer. That ratio, not the percentage, is the sentence to carry out of this problem.
(b) c = 0.03. P = 1 − 0.97^45 = 1 − 0.253938 = 0.7461 → 74.6%.
Findable = 29,840,000; non-customers among them = 28,640,000.
Exposure does not double. The naive linear answer is 2 × 49.3% = 98.6%, and the truth is 74.6% — because c sits in the exponent, so the function saturates. Every additional customer is increasingly likely to expose people already exposed by somebody else's cousin.
Additional non-customers exposed = 28,640,000 − 19,140,000 = 9,500,000, from 600,000 additional customers = 15.8 each, against 31.9 each on average for the first 600,000. The marginal customer exposes half as many people as the average one, and that number keeps falling — which is exactly why the harm is front-loaded and why a small database is not a safe one.
(c) Solve 1 − (1 − c)^45 = 0.95:
(1 - c)^45 = 0.05
1 - c = 0.05^(1/45) = exp(ln(0.05)/45) = exp(-2.99573/45) = exp(-0.0665718) = 0.935596
c = 0.0644 -> 6.44%
Profiles = 0.0644 × 40,000,000 = 2,576,000 — about 4.3 times the current database. To index 95% of a country you need one person in every 1/0.0644 = 15.5.
(d) Same coverage, three values of N:
| N | P(findable) at c = 1.5% | Non-customers exposed |
|---|---|---|
| 45 | 49.3% | 19.1 million |
| 107 | 80.2% | 31.5 million |
| 228 | 96.8% | 38.1 million |
The answer spans 49% to 97%. The model does not predict.
Check the two anchors before trusting either N. At N = 107 and c = 0.9% the model gives 62.0%, against Erlich's reported "nearly 60%". At N = 228 and c = 2% it gives 99.0%, against the paper's projected ">99%". Each N reproduces the endpoint it was fitted to and neither reproduces the other, which is the chapter's point: the functional form is wrong, not merely the parameter. Relatives are not independent draws into a database, and detectability decays with relationship, so no single N can hold.
What survives. The shape, and it survives at every N:
- Coverage is in the exponent, so exposure saturates fast. At every N in the table, a few per cent coverage exposes a large majority.
- The exposed vastly outnumber the enrolled, at every N. At the smallest, most conservative N, 600,000 people's decision exposes 19 million others — a ratio of 32 to 1. Larger N makes that worse, never better. N = 45 is a floor, not an estimate.
- Therefore the conclusion the chapter draws does not depend on the parameter at all: the unit of decision (a person who signs up) and the unit of exposure (their extended family) are different objects, so consent does not have the right shape for this, and no improvement to the consent form changes that.
What you may not claim. Any specific percentage as a prediction. Not "60% of this population is exposed", not "we would need 2.6 million profiles" as an engineering target. Write it as a bound with the N stated, or do not write a number.
Generalise. This is the failure mode of every model with a free parameter back-solved from one endpoint: it will reproduce that endpoint and nothing else. The way to use such a model honestly is to check whether your conclusion is monotone in the parameter over its whole plausible range. Here it is, which is why the argument holds and the prediction does not.
5. Reading a CPIC-style recommendation
Ch 57 §5 says a diplotype maps to an activity score and the score maps to a class. Here are the three supplied tables that make that concrete for CYP2D6 and codeine.
activity values
| Allele | Function | Activity |
|---|---|---|
| *1 | normal (reference) | 1.0 |
| *2 | normal | 1.0 |
| *4 | no function (splice-disrupting) | 0 |
| *10 | decreased | 0.25 |
| *41 | decreased | 0.5 |
A duplication is written xN and contributes N copies of that allele's activity.
Activity score to phenotype
| Activity score | Phenotype |
|---|---|
| 0 | Poor metaboliser |
| greater than 0, below 1.25 | Intermediate metaboliser |
| 1.25 to 2.25 inclusive | Normal metaboliser |
| above 2.25 | Ultrarapid metaboliser |
Codeine recommendation
| Phenotype | Recommendation |
|---|---|
| Ultrarapid | Avoid codeine. Risk of respiratory depression. Use a non-codeine, non-tramadol analgesic |
| Normal | Use the standard age- or weight-appropriate dose |
| Intermediate | Use the label dose; if analgesia is inadequate, switch to a non-codeine, non-tramadol analgesic |
| Poor | Avoid codeine. Lack of efficacy. Use a non-codeine, non-tramadol analgesic |
Five patients are prescribed codeine after surgery.
| Patient | Diplotype | Other |
|---|---|---|
| A | *1/*4 | — |
| B | *4/*4 | — |
| C | *1/*1x2 | — |
| D | *4/*41 | — |
| E | *1/*1 | started on a strong CYP2D6 inhibitor last week |
(a) Give each patient's activity score, phenotype and recommendation. (b) Two patients receive the same instruction for opposite reasons. Identify them, state both reasons, and say why a clinician needs to know which is which. (c) What does the recommendation table, keyed on , get wrong about patient E? Name the phenomenon. (d) A short-read pipeline that calls SNVs and small only is used to generate these diplotypes. Which patient is misclassified, into what, and what would the clinical consequence be?
Solution
(a) Add the allele activities; look up the band; read off the action.
| Patient | Activity score | Phenotype | Recommendation |
|---|---|---|---|
| A | 1.0 + 0 = 1.0 | Intermediate (0 < 1.0 < 1.25) | Label dose; switch if inadequate |
| B | 0 + 0 = 0 | Poor | Avoid codeine |
| C | 1.0 + (1.0 × 2) = 3.0 | Ultrarapid (3.0 > 2.25) | Avoid codeine |
| D | 0 + 0.5 = 0.5 | Intermediate | Label dose; switch if inadequate |
| E | 1.0 + 1.0 = 2.0 | Normal on genotype | Standard dose — and this is wrong, see (c) |
Watch the CYP2D6 *1x2 notation: *1/*1x2 is three functional copies, not two. Scoring it as 2.0 puts the patient in the Normal band and hands them a drug that could kill them.
Watch the band edges too. A score of exactly 1.25 is Normal; 1.24 is Intermediate. A score of exactly 2.25 is Normal; 2.26 is Ultrarapid. Real diplotypes land on them — *1/*10 scores 1.0 + 0.25 = exactly 1.25, so it is Normal and not Intermediate — and guessing an edge is how a lookup table becomes a coin flip. Do not interpolate a published table; read it.
(b) B and C both get "avoid codeine", for opposite reasons.
- B (poor metaboliser): codeine is a prodrug and CYP2D6 O-demethylates it to morphine. B makes almost no morphine, so codeine does nothing. The failure is no analgesia.
- C (ultrarapid, gene duplication): C makes morphine fast enough to reach concentrations far above the intended exposure. The failure is respiratory depression — the mechanism behind the deaths in children after tonsillectomy and in breastfed infants of ultrarapid mothers that drove the contraindications.
Why the clinician must know which: the substitute differs, and so does the monitoring. B needs an analgesic that does not depend on CYP2D6 activation, and B's earlier reports that "codeine doesn't work" were true rather than drug-seeking. C needs the same substitution and must not be given tramadol either, and anyone else in C's family may share the duplication. Same instruction, different clinical world.
This is the shape Ch 57 §5 insists on: variation in the same enzyme produces opposite failures depending on whether the enzyme activates the drug or clears it. Contrast DPYD with 5-fluorouracil, where the enzyme clears an active drug, so means accumulation and toxicity — the opposite direction from CYP2D6 loss of function on codeine.
(c) The table gives E "standard dose" because E's genotype is *1/*1, activity 2.0, Normal. But E is on a strong CYP2D6 inhibitor, so E's enzyme is inhibited and E behaves as a poor metaboliser: little morphine, no analgesia, and a clinician who concludes the genotype was wrong.
The phenomenon is . The chapter's formulation is the one to keep: genotype is not phenotype. A pharmacogenomic result is a statement about capacity, not about current activity, and the drug list on the same chart can override it. A decision-support system that fires on the diplotype alone and ignores concurrent medication will get E wrong every time.
(d) Patient C. A SNV/indel-only caller cannot see a whole-gene duplication: it observes *1 alleles and no variants, and returns *1/*1, activity 2.0, Normal metaboliser.
Consequence: an ultrarapid metaboliser is issued a standard dose of codeine — the exact scenario that killed children post-tonsillectomy. The mis-call is silent, because the report looks completely normal.
This is why Ch 57 §5 singles CYP2D6 out as genuinely hard to genotype: whole-gene deletions, duplications up to a dozen copies, and hybrid genes with the neighbouring CYP2D7. It needs copy-number-aware calling and, for the hybrids, long reads (Ch 46). Note the asymmetry in the errors a SNV-only pipeline makes here: it can only ever under-call activity variation toward the reference, so it turns Ultrarapid into Normal — a false reassurance — rather than the other way round.
6. What GINA does not cover
Six situations, all in the United States unless stated.
| # | Situation |
|---|---|
| 1 | A health insurer raises a premium after obtaining a predictive BRCA1 result |
| 2 | A life insurer asks for, and underwrites on, the same result |
| 3 | A long-term-care insurer declines an applicant who has a predictive Huntington disease result |
| 4 | An employer with 9 staff withdraws a job offer after learning the applicant's father has Huntington disease |
| 5 | An employer with 400 staff dismisses an employee who has developed symptomatic Huntington disease |
| 6 | A direct-to-consumer genomics company transfers its customer database to an acquirer |
(a) For each, say whether Title I, GINA Title II, both or neither applies, and if neither, name the regime that is actually operating. (b) The strongest empirical argument for letting insurers underwrite on genetic results is adverse selection. Supplied: among 10,000,000 adults, Huntington carriers are 1 in 10,000; 8% of adults hold long-term-care cover; carriers are five times as likely to hold it (Ch 58 §3); and a carrier's expected claim is 4 units against a non-carrier's 1. Compute the carriers' share of the insurer's book, their share of its claims, and the premium loading a ban imposes on everyone else. (c) Documented adverse decisions are far rarer than fear of them. Why is that not an argument that protection is unnecessary? Name the loop. (d) Recompute the loading in (b) if predictive results became routine and 20% of adults held one. What does the comparison say about the durability of the "adverse selection is small" finding?
Solution
(a)
| # | Covered by | What is actually operating |
|---|---|---|
| 1 | GINA Title I | Health insurers may not use genetic information in eligibility or premiums, and may not require a test. This is the case GINA was written for |
| 2 | Neither | Life insurance is outside GINA entirely. In the UK the Code on Genetic Testing and Insurance (2018) bars asking, except Huntington disease on life cover above £500,000; Canada criminalised the demand in 2017; Australia's ban commences October 2026. In most of the US, nothing |
| 3 | Neither | Long-term-care insurance is outside GINA — and it is the cover most directly implicated by a late-onset neurodegenerative variant, which is what makes this the sharpest gap in the statute. Some US states extend protection; most do not |
| 4 | Neither | Title II does not apply at all to employers with fewer than 15 employees, so the headcount decides the case however the information is classified. Note also Ch 58 §3's separate observation that much of the discrimination people report was triggered by family history, which no genetic-privacy statute regulates — so this is the route least reached by law even at a large employer |
| 5 | Neither | GINA protects genetic information, not people whose disease has manifested. That is ADA territory, with different tests and different remedies |
| 6 | Neither | Not GINA, and not HIPAA — a DTC genomics company is not a covered entity, so the governing document is a terms-of-service agreement. The 2025 bankruptcy showed what that is worth: a privacy policy is a contract a company can amend and a court can transfer. State genetic-privacy statutes are a growing patchwork (Utah's, effective May 2021, was first) and none of them reaches a company incorporated elsewhere |
The pattern: GINA covers one insurer and one employer, and every other row falls to a regime that either does not exist or was not designed for this.
(b) Build the book.
carriers = 10,000,000 x 1/10,000 = 1,000
non-carriers = 10,000,000 - 1,000 = 9,999,000
carrier uptake = 5 x 8% = 40%
carrier policies = 1,000 x 0.40 = 400
non-carrier policies = 9,999,000 x 0.08 = 799,920
---------------------------------------------------------
total policies = 800,320
Carriers' share of the book = 400 / 800,320 = 0.050%, i.e. 1 in 2,001 policyholders — against 1 in 10,000 of the population, so a 5-fold enrichment, exactly as advertised.
Claims: 400 × 4 + 799,920 × 1 = 1,600 + 799,920 = 801,520 units over 800,320 policies.
- Carriers' share of claims = 1,600 / 801,520 = 0.20%
- Claims per policy = 801,520 / 800,320 = 1.0015, so the premium loading is 0.15%
Both halves of the chapter's sentence are now numbers. In relative terms the selection is real and large: a 5-fold enrichment in the book and a 4-fold claim cost. In aggregate market terms it is negligible: fifteen basis points on the premium. And the reason is in the arithmetic rather than in anybody's opinion — carriers are a vanishing fraction of applicants, so a large multiple of a tiny number is still a tiny number.
(c) Because the low observed rate is partly produced by the thing protection would prevent. People decline testing precisely to keep a result off the record: they forgo surveillance that would extend their lives, and they decline research participation, degrading the evidence base for everyone. A discrimination event that never happened because the test was never taken does not appear in any count of discrimination events.
The loop, stated plainly: the aggregate adverse selection in (b) is small because few people have tested; few people have tested because the results are underwritable; and the smallness of the measured selection is then offered as the reason not to change that. Each step is true and the circle is vicious. You cannot use the 0.15% as evidence about a world in which testing is normal, because the 0.15% is a measurement of a world in which it is not.
(d) Let f be the fraction of adults holding an adverse predictive result, with the same 5-fold uptake and 4-fold claim cost. The loading is
loading(f) = f × 5 × 3 / (f × 5 + 1 − f)
| f | Premium loading |
|---|---|
| 1 in 10,000 (today) | 0.15% |
| 2% | 27.8% |
| 20% | 166.7% |
Check the last one directly: policies = 2,000,000 × 0.40 + 8,000,000 × 0.08 = 800,000 + 640,000 = 1,440,000; claims = 800,000 × 4 + 640,000 = 3,840,000; per policy = 2.667, a loading of 166.7%. ✓
What the comparison says. "Adverse selection is empirically small" is not a fact about genetics; it is a fact about f, and f is the one quantity the whole field is trying to increase. The finding is not wrong, it is conditional on the current rarity of predictive testing, and it inverts under exactly the conditions everyone is working toward. Anybody who quotes the small measured selection in defence of a permanent policy is extrapolating a measurement outside the range it was taken in.
And underneath all of it sits a question no dataset settles: whether insurance pools risk or sorts it. A society that thinks pooling is the point restricts underwriting on unchosen characteristics; one that thinks sorting is the point does not. Both are coherent. The arithmetic tells you the price of each; it does not tell you which to buy.
7. What a polygenic score tells one person, and what it does not ★★
Part A — the clinic. A patient is told: your for this disease is in the top 2% of the distribution, which carries 3 times the population-average lifetime risk. The disease has a lifetime prevalence of 10%.
(a) Convert that into the three numbers the patient should be given, and the one number the clinic did not mention. (b) A colleague proposes offering a preventive intervention to everyone in the top 2%. What does (a) say about the ceiling on that programme's benefit? (c) The patient's genome most resembles African-ancestry reference populations; the score was trained in European-ancestry biobanks. Which quantity in (a) becomes untrustworthy first, and by what mechanisms? Is this a statement about the patient?
Part B — the clinic that sells embryos. A couple has 5 viable embryos from one IVF cycle and is offered selection on a polygenic score for adult height. Supplied:
| Quantity | Value |
|---|---|
| Standard deviation of the trait in the population | 7 cm |
| Correlation between the score and the trait, estimated between families | 0.50 |
| Standard deviation of the score among full siblings, relative to its population SD | 1 / sqrt(2) ≈ 0.707 |
| E[maximum of n independent standard normal draws] | n=1: 0.00 · n=2: 0.56 · n=3: 0.85 · n=4: 1.03 · n=5: 1.16 · n=10: 1.54 |
| Correlation between score and trait within a sibship | 0.30 |
Embryo scores may be treated as independent draws about the parental mean (independent meioses).
(d) Compute the expected height gain from picking the top-scoring of the 5, first using the between-family correlation and then the within-family one. How much would 10 embryos buy? Then list which of Ch 58 §8's four objections your calculation has and has not accounted for.
Solution
(a) Three numbers to give her, one she was not given.
- Absolute lifetime risk: 3 × 10% = 30%. A relative risk without a baseline is not information.
- Absolute increase: 20 percentage points (30% against the 10% population average).
- 70% of people in the top 2% never develop the disease. The single most likely outcome for this patient, on this score, is that nothing happens.
And the number the clinic did not mention — what share of all cases the top 2% actually contains:
cases arising in the top 2% = 0.02 x 0.30 = 0.0060
cases in the whole population = 0.1000
share = 0.006/0.10 = 6%
94% of all cases occur outside the top 2%. Both facts are true simultaneously: this group is at genuinely elevated risk, and it is nearly irrelevant to where the disease actually comes from. Consistency check: the other 98% carry 0.094 of the risk mass over 0.98 of the people, an average risk of 9.59% — so the top 2% is 30/9.59 = 3.13 times the risk of everyone else, slightly more than the 3× quoted against the population average, which includes them.
(b) The programme's ceiling is 6% of cases, because that is all the top 2% contains. Even a perfectly effective intervention, perfectly delivered, with perfect adherence, prevents at most six cases in a hundred while touching 2% of the population.
This is the general shape of risk stratification on a common disease and it is not a defect of this particular score: most cases arise in the large middle of the distribution, where each individual's risk is unremarkable. A screening programme aimed at an extreme tail has a small ceiling by construction, and the comparison a health system needs is against a population-wide measure that moves everyone's risk a little. Nothing in (a) argues the score is uninformative for this patient; it argues that "informative for an individual" and "useful as a policy instrument" are different properties, and the second does not follow from the first.
(c) The percentile goes first — before the relative risk, before the absolute risk. "Top 2%" is a statement about where the patient falls in a distribution, and the distribution that was measured is the training cohort's. The mechanisms, all from Ch 58 §6:
- The found , not causal variants, and LD between tag and causal variant differs between populations, so the tag stops tagging (Ch 29).
- frequencies differ, so each variant explains a different share of variance, and the score's own variance — the denominator of the percentile — is different.
- The effect estimates absorb stratification and indirect genetic effects that do not transfer.
- Environments and their interactions differ.
A score developed in European-ancestry samples can lose most of its predictive value in African-ancestry individuals. So the honest report is not "top 2%, 30% risk"; it is that the score's calibration in this patient's ancestry is unestablished, and both the percentile and the risk attached to it are unreliable — with the direction of the error unknown, which is worse than a known bias.
Is this a statement about the patient? No. It is a statement about who was sampled. failure is caused by biased training data and by LD structure, and the asymmetry reverses: train the score in an African-ancestry cohort of equivalent size and it works there and transfers poorly the other way. This is the same lesson as Ch 57's cattle-versus-humans contrast — identical estimator, different study design — and it is fixed by sampling differently, not by any claim about biology.
(d) Work in units of the score's population SD, then convert.
Step 1 — how far above the sibling mean is the best of 5? The embryos' scores are independent draws about the parental mean with SD 0.707 population SDs. The best of 5 standard normal draws sits 1.16 SDs above the mean, so:
best-of-5 advantage = 1.16 × 0.707 = 0.820 population score SDs
Step 2 — convert score to trait. With a standardised score S and trait Y correlated at r, E[Y | S = s] = r·s.
Between-family r = 0.50: expected gain = 0.50 × 0.820 = 0.410 trait SDs = 0.410 × 7 cm = 2.9 cm
That is the marketable number, and it matches the chapter's "on the order of a few centimetres of height".
Step 3 — use the right correlation. The weights were estimated between families, and the selection happens within one. With r = 0.30:
expected gain = 0.30 × 0.820 = 0.246 trait SDs = 1.7 cm — 60% of the headline figure.
Step 4 — buy more embryos? Best of 10 sits 1.54 × 0.707 = 1.089 SDs above the mean, giving 0.544 trait SDs = 3.8 cm at r = 0.50. Doubling the number of embryos from 5 to 10 buys 0.9 cm. The order statistic grows roughly like the square root of the logarithm of n; there is no number of embryos that turns this into a large effect, which is the quantitative content of "the variance available is within-family".
Which objections the calculation covers. Of the four in Ch 58 §8:
| Objection | In the calculation? |
|---|---|
| 1. The variance available is within-family | Yes — the 1/sqrt(2) factor and the order statistic are exactly this |
| 2. Weights estimated between families, applied within one | Yes — Step 3, the drop from 0.50 to 0.30 |
| 3. — selecting on one score moves every genetically correlated trait | No. Nothing above touches it. The 1.7 cm is a gain in one dimension of an unknown-dimensional move |
| 4. Portability — every problem in Part A applies to every embryo whose ancestry differs from the training cohort's | No. The r values are stipulated; in a mismatched embryo they are unknown and lower |
So the calculation is an upper bound under favourable assumptions, not an estimate. The honest summary for the couple: about 1.7 cm in expectation under assumptions chosen to flatter the method, against a population SD of 7 cm — a quarter of a standard deviation — with two of the four known objections not yet priced in at all, on an uncertain causal model.
Generalise, and note what the argument is not. Every step here came from statistics, not from an ethical premise. That matters for problem 8: PGT-P is weak for reasons that could in principle change with better data — sibling-design GWAS, calibrated within-family weights — whereas the objections to, say, heritable editing are ones better technique does not answer. Knowing which kind of objection you are making is the difference between an argument and a position.
8. The strongest version of a position you rejected
Ch 58 states repeatedly that where a disagreement is genuine it gives both sides their strongest form. This problem asks you to do it, and it is not an invitation to opine: each part has a required structure and a checkable answer.
Below are three claims. Assume you accept all three — they are the positions Part 12 will most likely have left you holding.
| Claim | |
|---|---|
| I | Forensic investigative genetic genealogy should require judicial authorisation, because it searches non-consenting third parties |
| II | Insurers should be barred from underwriting on predictive genetic test results |
| III | Embryo selection on should not be offered clinically |
(a) For each claim, write the strongest opposing case in three or four sentences. It must be one a competent advocate would actually make, must use the chapter's own facts and numbers, and must not be a position the chapter has already shown to be defective. (b) For each, name the single premise the opposing case must defeat for the original claim to fall. Not a list — one premise each. (c) For each, state what evidence, if it existed, would move you. If the honest answer is "none, because the disagreement is not empirical", say so and say why. (d) One of these three is an empirical disagreement wearing ethical clothes, and two are value disagreements wearing empirical clothes. Which is which, and what does the misclassification cost in practice?
Solution
(a) Three steelmen. Note that a steelman is not a summary of what opponents say; it is the best available version, which often nobody has bothered to state.
Against I (require judicial authorisation for FIGG). The exposure is a property of IBD, not of police procedure: at the levels of problem 4 a large fraction of the population — on the larger fitted values of N, the overwhelming majority — is already indexed through relatives whether or not any warrant is ever issued. A warrant requirement therefore does not un-expose a single person — it rations access to an exposure that already exists, and rations it against the one use with an unambiguous public benefit, since hundreds of cold cases have been closed and wrongly convicted people exonerated. Meanwhile every conventional investigative technique — canvassing, phone records, informants — also implicates non-consenting third parties without their notice, so singling out this one requires an argument that genetic information is different in kind, not merely in scale, and that argument has to be made rather than assumed.
Against II (bar genetic underwriting). Voluntary insurance is priced on symmetric information; if applicants may know their and insurers may not ask, high-risk applicants buy more cover at pooled prices and low-risk applicants buy less. This is not hypothetical: Huntington carriers are up to five times as likely as the general population to hold long-term-care cover, and Oster and colleagues argue that even a modest expansion of genetic information could threaten that market's viability — which is precisely the coverage the statutory gap leaves unprotected. A ban is therefore not neutral fairness but an unstated, unlegislated transfer from low-risk to high-risk purchasers, and problem 6(d) shows the transfer is small only for as long as testing stays rare, which is not a state anybody is defending.
Against III (do not offer PGT-P). The expected gain computed in problem 7(d) is small but it is not zero, and "small" is not a category that ordinarily licenses prohibition — PGT-A is sold routinely on evidence Ch 57 §4 calls equivocal for live birth per cycle started, and reproductive autonomy is a strong and widely shared commitment. The objections in Ch 58 §8 are objections to the magnitude and reliability of the effect, which is an argument for mandatory disclosure of the magnitude — 1.7 cm, two objections unpriced — rather than for withholding the choice. A rule that prospective parents may not buy a service because a professional society judges the expected benefit too small is a paternalism that would not survive being stated in any other area of reproductive medicine.
(b) The single load-bearing premise in each case.
| Claim | The premise the opposing case must defeat |
|---|---|
| I | That a search which imposes a cost on identifiable non-consenting people requires authorisation even when the exposure enabling it already exists. If exposure alone extinguishes the interest, claim I falls |
| II | That insurance ought to pool risk on unchosen characteristics rather than sort it. Every actuarial fact in the world is compatible with claim II if this premise holds, and claim II falls the moment it does not |
| III | That an intervention with small, poorly characterised expected benefit and unquantified correlated effects should not be sold as a clinical service. If disclosure is a sufficient remedy for uncertainty, claim III falls |
The discipline here is the point: if you cannot name one premise, you have not understood your own position well enough to defend it, and you will find yourself arguing about the numbers when the numbers were never what you disagreed about.
(c) What would move you.
I — mostly no. Empirical work could narrow the edges: how many cases are actually closed, how often the technique produces a wrong lead, how unevenly the exposure falls across a population — Ch 57 §8 says it is uneven, and says the related technique of familial searching inherits offender databases' over-representation of some communities. But the core is a question about whether a person retains an interest in information they cannot control, and no dataset answers it. Note the trap in the opposing case: "the exposure already exists" is a fact about the world that arrived through a sequence of unreviewed product decisions (Ch 58 §2), so treating it as a given is treating the outcome of the disputed process as the premise of the argument.
II — the empirical half is settleable and has been settled; the disagreement is not. How large adverse selection is, is measurable, and problem 6 measures it: 0.15% today, 28% at f = 2%, 167% at f = 20%. What no measurement touches is whether a market that sorts on unchosen characteristics is the right institution. The chapter is explicit that the same unresolved disagreement sits under every debate about pre-existing conditions, and that is the tell: a dispute that recurs identically across unrelated policy domains is not waiting on data from any one of them.
III — yes, genuinely. This one is empirical almost all the way down. Within-sibship at scale would pin the within-family correlation that problem 7 had to be handed. Long-term outcome data on children selected this way would price . Ancestry-diverse training cohorts would fix . If those data came back favourable the calculation changes and so should the recommendation — and if they came back unfavourable the case closes. Nothing about claim III rests on a value premise that data cannot reach, which is exactly why the chapter says PGT-P is weak "for reasons that follow from the statistics, not from any ethical premise".
(d) III is the empirical disagreement wearing ethical clothes. I and II are value disagreements wearing empirical ones.
What the misclassification costs, in both directions:
- Treating a value disagreement as empirical (I and II) produces an endless demand for more studies as a substitute for making a decision, and it lets whichever side currently controls the data-collection agenda set the default by delay. It also produces bad-faith argument: if the real disagreement is pooling versus sorting, then arguing about the size of adverse selection is arguing about something neither party would change their mind over, which is how a debate runs for twenty years without moving.
- Treating an empirical disagreement as ethical (III) is the more expensive error, because it is self-sealing. Framing PGT-P as a values question means the studies that would settle it are never prioritised, the professional-society position hardens into a norm rather than a reading of the evidence, and the market — which Ch 58 §8 notes exists anyway — fills the space with claims nobody is measuring. A recommendation that is really about effect size should be stated as being about effect size, so that it can be revisited when the effect size is known.
Generalise. Before entering any argument in this chapter, ask which kind it is. If new data would change your position, say what data. If none would, say that, and argue about the premise instead. Almost all of the heat in this field comes from people making value arguments in the vocabulary of evidence, and both sides usually do it at once.
Where you went wrong
| Error pattern | What to re-read |
|---|---|
| Read a sensitivity or specificity as the probability the result is right | Problem 1(a) — accuracy is conditional on disease status, PPV on the result |
| Bought sensitivity when specificity was the | Problem 1(c) — 0.07 points against 14.4 |
| Applied PPV ≈ prevalence × sensitivity / FPR outside its range | Problem 1(d) — it needs prevalence far below FPR/sens, and announces failure by exceeding 1 |
| Used the from ascertained families as the PPV of a population | Problem 2(a) — 90% claimed, 10% true |
| Used the -positive count as the disease prevalence | Problem 2(b) — the biochemical screen's denominator is cases, not carriers |
| Assumed more sensitive screening is better screening without pricing the false positives | Problem 2(c) — 0.4 extra cases for 170 extra false positives |
| Compared carrier frequencies as denominators rather than decimals | Problem 3(a) |
| Predicted a 10-fold risk reduction from a 90% detection rate | Problem 3(c) — the couple's risk is a square, so it is 97-fold |
| Heard "negative" as "not a carrier" | Problem 3(d) — a negative is a likelihood ratio and needs the detection rate |
| Expected exposure to scale linearly with database size | Problem 4(b) — is in the exponent; the marginal customer exposes half as many as the average one |
| Quoted a findability percentage as a prediction | Problem 4(d) — 49% to 97% on the same coverage; N = 45 is a floor |
| Scored *1/*1x2 as two functional copies | Problem 5(a) — it is three, activity 3.0, Ultrarapid |
| Read "avoid codeine" as one clinical situation | Problem 5(b) — no analgesia and respiratory depression are opposite failures |
| Took a genotype-keyed recommendation as a | Problem 5(c) — |
| Assumed covers genetic discrimination | Problem 6(a) — not life, disability or long-term-care cover, not employers under 15, not once disease has manifested, not a DTC company |
| Cited "adverse selection is small" as a permanent finding | Problem 6(d) — 0.15% at f = 1 in 10,000, 167% at f = 20% |
| Reported a relative risk without a baseline | Problem 7(a) — 3× means 30%, and 70% of that group stay well |
| Assumed a high-risk tail is where the cases are | Problem 7(b) — the top 2% holds 6% of cases |
| Trusted a percentile in an ancestry the score was not trained in | Problem 7(c) — the percentile fails before the risk does |
| Used a between-family effect estimate for a within-family choice | Problem 7(d) — 2.9 cm becomes 1.7 cm, and two objections are still unpriced |
| Answered an ethical prompt with a verdict instead of the opposing argument | Problem 8(a)–(b) — name the premise, not the conclusion |
| Demanded more data for a disagreement no data can settle | Problem 8(d) — and neglected the data for the one it could |