# Chapter 15 · Complete draft solution notes

**Constructed examples only. Technical and author review pending.** These answers distinguish calculation from causal identification.

1. One possible protocol assigns eligible cases at opening to form version 2 plus specified training or version 1, measures resolution within seven days and reopening by day fourteen, and targets the average assignment effect in the eight-week intake population. Requiring use changes the strategy, compliance problem and estimand.
2. After assignment to the new process, $Y_i(1)$ is observed and $Y_i(0)$ is missing. A different case's observed $Y_j(0)$ can stand in only under a design and assumptions making the cases comparable for the target question.
3. The rates are 0.60 and 0.65, so the risk difference is 0.05, or five percentage points. The independent-case normal standard error is $\sqrt{0.6(0.4)/1000+0.65(0.35)/1000}=0.02162175$, in proportion units. This assumes case-level independence and complete measurement.
4. The point estimate is 50 additional timely outcomes per 1,000 assigned cases under comparable conditions. Full rollout could change staffing, case mix, compliance or spillovers.
5. Compliance can depend on severity, staff availability or motivation. Restricting to compliers after assignment can select different kinds of cases in the two arms and discard the original random comparison.
6. New-arm successes among all assigned are between 585 and 685, giving 58.5–68.5%. Old-arm successes are between 570 and 620, giving 57–62%. The contrast is therefore between −3.5 and 11.5 points with no assumptions about the missing outcomes. Better follow-up, justified information about missingness or defensible sensitivity assumptions can narrow it.
7. A possible diagram has baseline severity pointing to both adoption and resolution, adoption pointing to referral and resolution, and referral pointing to resolution. Pre-treatment severity is a candidate adjustment. Referral may mediate the total effect. If unmeasured difficulty also causes referral and outcome, conditioning on referral can add collider bias. The arrows require subject-matter justification.
8. Routine rate difference: 90−80=+10 points. Complex: 40−20=+20. Equal-mix new is 65%, old 50%, contrast +15. Causal interpretation requires comparability of potential outcomes within relevant pre-treatment strata, overlap, consistent intervention and outcome measurement, plus attention to other common causes.
9. $(75-60)-(80-70)=+5$ percentage points. Changes in staffing, client mix, measurement, outages or interference between sites can produce differential shocks. Equal baseline levels are not the main assumption; the missing untreated trend is.
10. A broad above–below comparison confounds the intervention with the score itself and anything associated with being far from the threshold. A credible local design compares near-cutoff units under continuity and no precise manipulation assumptions; it estimates a local effect.
11. Winning changes both the tutoring invitation and the free textbook. The lottery may affect outcomes through the book even without tutoring, contradicting the proposed exclusion restriction for tutoring alone.
12. A defensible paragraph reports 650/1000 versus 600/1000, five points, an illustrative 0.76–9.24 point independent-case interval, assignment to the package and seven-day resolution, and the need to audit allocation, missingness, reopening and scaling. Do not recommend rollout if serious reopening harm appears or the assignment/outcome records cannot be trusted. It must not claim that the form alone caused the difference.
13. The old month resolves $80(0.8)+20(0.2)=68$ of 100, or 68%. The new month resolves $20(0.9)+80(0.4)=50$ of 100, or 50%: a crude change of −18 points. Under equal routine/complex weights, old is 50% and new is 65%: a standardised contrast of +15 points. The crude comparison weights the different observed mixes; standardisation applies one common target mix. Neither alone establishes a causal effect.
14. An old-process case with estimated new-process probability 0.99 receives inverse weight $1/(1-0.99)=100$ in the stated average-effect scheme. A result dominated by that case is sensitive to its measurement, outcome, score-model specification and the near-absence of comparable old-process cases. The weight is not one hundred independent observations.
15. A defensible plan sets time zero at case opening; fixes eligibility, process versions and seven-day outcome before viewing outcomes; records only pre-opening severity/site/season and staffing covariates for adjustment; standardises to the intended eight-week eligible case mix; checks outcome completeness and overlap by group; and reports an unmeasured-staffing sensitivity analysis. Unrecorded supervisor triage affecting both process assignment and resolution could prevent a causal interpretation even if observed covariates balance.
