Appendix — Model Deliverable — PS 7.5 (Identification Memo + Threats Table)

Problem set: book/weeks/week-07/ps7.5.md (PS 7.5, Week 7). Chapter: Ch 7.5 — The Identification Memo.

PS 7.5 has no numeric answer key: it is a project deliverable, and what is graded is the design, not a finding. So this appendix entry is not a worked solution but a model deliverable — a complete, A-grade identification memo for one cast project, written exactly as a strong student would submit it, followed by instructor grading notes that key each section to the research-design rubric and call out the moves that earn (and lose) points. Read the exemplar the way Ch 7.5 §6 walked through Maya’s HMDA memo: not to copy, but to see the standard.

The chapter’s worked memo was Maya’s staggered fair-lending DiD. To give you a second fully-worked design rather than a paraphrase of the chapter, this exemplar is Priya’s climate-insurance shock — the disasters-and-insurance-costs DiD she scoped in Ch 7.1 (Worked Example B), built from NOAA storm events, FEMA disaster declarations, and FRED cost indices. It is a non-staggered (or lightly staggered) county-level DiD, so its threat menu overlaps Maya’s but is not identical, which is exactly the point: the menu of Ch 7.5 §2 is design-specific, and you pull your design’s row off it. No empirical magnitudes appear anywhere below, because the memo is filed before the first-look regression is run — the freeze of Ch 7.5’s “Your Turn.”

THE MODEL DELIVERABLE — `memo.md`

1. The specification (CONVENTIONS §4 form)

Outcome: the annual log average homeowner-insurance cost in county i, year t — log_premium, built from the FRED/state-filing homeowner premium series merged to county, deflated to constant dollars. (Logged so the coefficient reads as an approximate percent change; the level series is right-skewed.)
Treatment / key regressor: post_disaster = D_{it} = \mathbf{1}\{t \ge G_i\}, an indicator that county i has experienced its first billion-dollar FEMA major-disaster declaration on or before year t, where G_i is the county’s first-declaration year (counties with no qualifying declaration in the window are never-treated controls, G_i = \infty). The coefficient \beta on D_{it} is the one number I read off to answer H1.
Controls: county-year median household income and housing-unit count (from the ACS), entering as covariate-conditional parallel-trends controls — both shift the level of insurable exposure and could trend differently across counties for reasons unrelated to disasters. I deliberately do not control for the post-disaster rebuilding rate or the local construction-cost index, because higher rebuilding cost is a channel through which a disaster raises premiums (the over-controlling trap of Ch 7.3 §2 / Mentor 4): conditioning on it would regress away part of the very effect I am estimating.
Fixed effects: county fixed effects (so the comparison is within a county, sweeping out fixed exposure — coastal, wildfire-belt, floodplain) and year fixed effects (so a nationwide reinsurance-cost shock that hits every county in a given year is absorbed and not mistaken for a treatment effect). The comparison is therefore within-county, net of common-year shocks.
Clustering: by county. Treatment turns on at the county level and a county’s premiums are serially correlated across years, so the plausible within-cluster error correlation lives at the county. (I will also report state-clustered SEs as a robustness check, since regulatory rate-setting is a state-level process — see the table.)
Sample: all U.S. counties with continuous premium coverage over the window 2010–2022, dropping counties with a population below 10,000 (premium series too noisy / often imputed) and dropping the two hurricane mega-years’ directly-flooded counties only from a robustness cut, not the main sample. Expected ~2,800 counties × 13 years. Vintage pinned: NOAA Billion-Dollar Disasters table and FEMA OpenFEMA Disaster Declarations snapshot dated and recorded in the data card.
Identifying assumption (one sentence): absent the disaster, treated and never-/not-yet-treated counties’ homeowner-insurance costs would have followed parallel trends. (Expanded in §2.)

2. The identifying-assumption paragraph

Our estimate of the effect of a billion-dollar FEMA disaster declaration on a county’s average homeowner-insurance cost is credible as long as, absent the disaster, the insurance costs of treated counties and of never- and not-yet-treated counties would have evolved in parallel — a parallel-trends assumption we defend three ways: with an event study whose pre-treatment leads are flat and statistically indistinguishable from zero, plotted with confidence bands tight enough to rule out an economically meaningful pre-existing divergence; with a clean-control estimator that compares each treated cohort only against never- and not-yet-treated counties, never against already-hit counties whose still-rising premiums would contaminate the comparison under dynamic effects (Ch 4.2); and with the institutional fact that the precise timing of a billion-dollar disaster is meteorological, not chosen by the county, so adoption timing is plausibly unrelated to a county’s latent premium trajectory. We report a single overall ATT and an event-time profile, with county-clustered inference, a state-clustered SE as a robustness check, and a placebo test that randomly reassigns disaster years to untreated counties. The verb is credible, not causal: parallel trends in the post-period is a counterfactual no test can reach, so we have argued it as hard as the institutions allow and we claim a conditional belief, not a proof.

3. The threats-and-responses table

Threat	Why it’s plausible	What we do about it	Residual concern
Differential pre-trend `[testable]` (parallel trends already failing before the disaster)	Disaster-prone counties — coast, floodplain, wildfire belt — may already be on a steeper premium-growth path as insurers price rising climate risk, independent of any single declaration.	Estimate the event study; normalize at k=-1; report and plot the leads with confidence bands. Reassuring result: leads flat and insignificant, bands tight enough to exclude a meaningful slope.	A flat pre-trend cannot confirm the post-period counterfactual; it only fails to refute it (Ch 4.1), and low power can make a noisy series look falsely flat.
Compound / coincident shock `[arguable]` (a disaster co-occurs with a state regulatory or reinsurance change)	A major disaster often triggers a contemporaneous state response — a rate-freeze, an insurer-of-last-resort expansion, a building-code update — so the premium move could be the bundled policy, not the disaster itself.	Document each treated state’s contemporaneous insurance-regulation changes from NAIC and state-filing records; argue case by case that no co-timed policy plausibly drives the premium in the same direction and window. State FE absorb time-invariant regulatory regimes; year FE absorb national reinsurance shocks.	No statistical test detects a perfectly co-timed state policy; this rests on the completeness of the institutional review, which is necessarily incomplete.
Anticipation `[arguable]` (insurers reprice before the official declaration date)	Insurers watch forecasts and prior near-misses; a county repeatedly threatened may see premiums rise before the disaster that crosses the billion-dollar threshold, contaminating the “pre” period.	Partly testable: inspect the leads in the one to two years just before G_i and, as a robustness cut, drop the year before declaration and re-estimate. The institutional argument: the exact declaration year is meteorological and unforecastable at the pricing horizon.	Anticipation that began before any observed lead is invisible to the leads; the claim ultimately rests on the timing-is-meteorological argument, not a test.
Forbidden comparison / negative weights `[testable]` (lightly staggered timing using already-hit counties as controls)	Declarations arrive in different years, and the premium effect plausibly builds as insurers re-rate over several renewal cycles — the exact recipe for the staggered-DiD bias (Ch 4.2).	Design choice: use a clean-control group-time estimator (Callaway–Sant’Anna) against never- and not-yet-treated counties only; run a Goodman–Bacon decomposition to show how much weight a naive TWFE put on forbidden comparisons.	The clean estimator needs a sufficient never-treated pool; if climate risk becomes near-universal late in the window, late cohorts are thinly identified.
Outcome measurement / coverage gaps `[testable]` (the premium series is imputed or missing for small counties)	Homeowner-premium data are assembled from state filings of uneven quality; sparse small-county cells are sometimes imputed, and imputation could correlate with disaster status (a Week-1 measurement-error problem).	Restrict the main sample to counties with continuous, non-imputed coverage; show the result is stable when the population floor is raised, and report how many county-years are imputed by treatment status.	Conditioning on coverage may itself select counties; the surviving sample is non-imputed but not a random subset of all counties.
Spillover / SUTVA violation `[arguable]` (a disaster raises premiums in control counties too)	Reinsurance and insurer capital are pooled regionally, so a large disaster can push up premiums in nearby un-hit counties — contaminating the controls and biasing the estimated effect toward zero.	Argue from market structure which counties share a reinsurance pool; as a robustness cut, drop control counties adjacent to (or in the same state as) a treated county and re-estimate; year FE absorb the purely national component.	If spillovers are diffuse and national, no “drop the neighbors” cut removes them, and the estimate is a lower bound on the true effect — stated as such.

Ordering note: row 1 is the parallel-trends question — the hardest, led not buried. Rows 2–3 are the most dangerous arguable threats (compound shock, anticipation). Rows 4–6 descend to the design-mechanics and data-specific threats.

4. Honesty audit

(a) Most dangerous threat. The compound / coincident shock (row 2), because it is arguable, not testable, and could in principle account for the entire effect: if every billion-dollar disaster is reliably followed by a state rate action in the same year, my \beta is the policy, not the disaster. My two strongest sentences: (i) the direction matters — many post-disaster regulatory responses (rate freezes, last-resort pools) suppress premiums, so a co-timed policy would bias me toward zero, making a positive estimate conservative; (ii) state and year fixed effects already absorb time-invariant regulatory regimes and national reinsurance shocks, so the residual confounder must be a county-or-state-specific, disaster-timed, premium-raising policy, a narrow target I can hunt in NAIC records. Honestly: this narrows the threat but does not kill it, so the verb stays credible and the residual cell stays full.

(b) Verb audit. I wrote credible, not causal, and I defend that: the load-bearing assumption (post-period parallel trends) is untestable, so a flat pre-trend earns a conditional belief, not a proof. There is no ungameable cutoff or randomized instrument here that would license causal. Every claim in the memo uses “consistent with” and “we cannot reject,” never “proves.”

(c) Empty-cell test. No residual cell is blank. No cell reads “this could be the entire effect” outright, but row 2’s residual comes closest, which is why it leads the audit. I am proceeding (not redesigning) because the fixed-effects structure plus the direction-of-bias argument bound the compound-shock threat to a conservative one — but I will state it as an explicit limitation in the Week-8 paper, not hide it.

INSTRUCTOR GRADING NOTES (keyed to the research-design rubric)

These notes show why this deliverable earns an A and where a weaker submission would lose points. The rubric criteria mirror Assessment 4 / Assessment 7 and the PS 7.5 part allocation.

Part 1 — Spec discipline (15 pts) → Excellent

All seven CONVENTIONS §4 slots are named to the level of a variable name (log_premium, post_disaster =D_{it}, the ACS controls, county+year FE, county clustering, the 2010–2022 sample with a stated row count and a pinned data vintage). Two A-grade moves: (i) the student names the control they deliberately omit — the rebuilding-cost channel — and says why (over-controlling on a mediator, Ch 7.3 §2), which is exactly the discipline the spec slot is meant to force; (ii) the clustering choice is justified by where the errors correlate (serial correlation within county) rather than asserted, and the alternative (state) is previewed as a robustness check rather than chosen post hoc to manufacture significance. Where points are lost on a typical submission: “Outcome: insurance cost” with no variable definition; controls listed without the omit-the-mediator reasoning; clustering stated but unjustified; no row count or vintage. Each is a slot scored down. A spec that has silently drifted from the PAP is the most serious Part-1 failure and is flagged as a contract violation regardless of polish.

Part 2 — Identification reasoning (25 pts) → Excellent

The sentence fills all three blanks correctly. (a) The effect is named (“the effect of a billion-dollar FEMA disaster declaration on a county’s average homeowner-insurance cost”), not the coefficient. (b) The threat is specific — post-disaster parallel trends — never “endogeneity”; this alone separates a passing memo from a failing one and is the single most common place weak memos collapse. (c) The third blank does real work: three distinct defenses (flat tight leads, clean-control estimator, meteorological-timing institutional fact), not a restatement of the assumption. (d) The verb is credible, and — crucially — the student defends the verb choice explicitly (“parallel trends in the post-period is a counterfactual no test can reach”). Where points are lost: “credible as long as there is no endogeneity” → automatic zero on (b), because no threat is named (PS 7.5 Part 2b states this). “Defended by parallel trends” → near-zero on (c), because it restates rather than defends. Writing causal for an untestable DiD → loses (d) and signals the overclaiming Ch 7.5 §1 calls the tell of an amateur. The exemplar avoids all three.

Part 3 — Threats-and-responses table (45 pts) → Excellent

(a) Coverage (12 pts). The table carries the full DiD menu of Ch 7.5 §2 — differential pre-trend, forbidden comparison, anticipation — plus data-specific threats the menu cannot supply (measurement/coverage gaps, the SUTVA/spillover problem that is acute for regionally-pooled insurance). Six rows; a DiD memo missing the parallel-trends row would be capped at half on this part. (b) Plausibility (6 pts). Every “why plausible” cell tells a concrete, setting-specific story (coastal counties already on a steeper risk-pricing path; reinsurance pooled regionally so disasters leak into neighbors). No strawmen. (c) Testable vs. arguable (15 pts) — the heaviest single criterion. Every row is tagged [testable] or [arguable], and — the part weak students get wrong — the column-3 response matches the tag. Pre-trend, forbidden comparison, and coverage are tagged testable and answered with diagnostics (event-study leads, Goodman–Bacon decomposition, an imputation count). Compound shock, anticipation, and diffuse spillover are tagged arguable and answered with institutional argument (NAIC records, the timing-is-meteorological claim, reinsurance-pool structure), with column 4 admitting “no test can detect this.” The student commits neither of the two fatal errors of Ch 7.5 §4: no test is run against an untestable threat (e.g., no claim that a balance table “proves” no compound shock), and no testable threat is merely argued away. (d) No empty residuals (8 pts). Every column-4 cell is substantive — including the hardest admissions (“a flat pre-trend cannot confirm the counterfactual”; “the estimate is a lower bound” under diffuse spillover). This is the column a mentor reads first, and it is full and honest. (e) Danger order (4 pts). Row 1 is parallel trends, the first-ninety-seconds question; the dangerous arguable threats follow; mechanics and data threats descend last. The student even annotates the ordering logic. Where points are lost: the classic failure is a table where every column-3 entry is “we control for it” (no test/argument distinction, no tags) and column 4 is all “none.” That table scores below half on (c) and zero on (d), even if it looks complete.

Part 4 — Honesty audit (15 pts) → Excellent

The student (a) names the most dangerous threat correctly as the arguable compound-shock row — not the showy testable one — and gives a genuine direction-of-bias argument that bounds rather than dismisses it; (b) defends the credible verb against the evidence; (c) passes the empty-cell test and, most tellingly, decides explicitly to proceed with a stated limitation rather than soften the language — which is exactly the Ch 7.5 §5 disposition: change the design or state the residual plainly, never launder the threat away. A submission that wrote “no concerns remain” in (c) would be marked down, per the chapter’s rule that a clean residual column is a less credible paper.

One-line grading heuristic

The two highest-signal questions for any PS 7.5 submission: (1) Does the identifying-assumption sentence name a specific threat with a real defense and a calibrated verb (Part 2b–d)? (2) Is every table row correctly sorted into testable-with-a-statistic vs. arguable-with-an-argument, with a non-empty residual (Part 3c–d)? A student who clears both has the Week-7 mindset — they keep “identified under a defended-but-untestable assumption” cleanly apart from “proven causal,” and they lead with their hardest question instead of burying it. The rest is execution. This exemplar clears both cleanly, which is what an A looks like.

End of model deliverable for PS 7.5. The exemplar is Priya’s climate-insurance DiD; the chapter’s worked memo (Ch 7.5 §6) is Maya’s HMDA fair-lending DiD — read both to see the menu of Ch 7.5 §2 instantiated on two different designs. Citation for the disaster-insurance framing and data sources: NOAA National Centers for Environmental Information, Billion-Dollar Weather and Climate Disasters* (data card pinned); FEMA OpenFEMA Disaster Declarations Summaries. The staggered-DiD machinery invoked in row 4 is Callaway and Sant’Anna (2021) and Goodman-Bacon (2021), as built in Ch 4.2. No empirical magnitudes are reported, by design: the memo is filed before the first-look regression of nb7.5.*

# Model Deliverable — PS 7.5 (Identification Memo + Threats Table) **Problem set:** `book/weeks/week-07/ps7.5.md` (PS 7.5, Week 7). **Chapter:** Ch 7.5 — The Identification Memo. PS 7.5 has no numeric answer key: it is a project deliverable, and what is graded is the *design*, not a finding. So this appendix entry is not a worked solution but a **model deliverable** — a complete, A-grade identification memo for one cast project, written exactly as a strong student would submit it, followed by **instructor grading notes** that key each section to the research-design rubric and call out the moves that earn (and lose) points. Read the exemplar the way Ch 7.5 §6 walked through Maya's HMDA memo: not to copy, but to see the standard. The chapter's worked memo was Maya's staggered fair-lending DiD. To give you a *second* fully-worked design rather than a paraphrase of the chapter, this exemplar is **Priya's climate-insurance shock** — the disasters-and-insurance-costs DiD she scoped in Ch 7.1 (Worked Example B), built from NOAA storm events, FEMA disaster declarations, and FRED cost indices. It is a non-staggered (or lightly staggered) county-level DiD, so its threat menu overlaps Maya's but is not identical, which is exactly the point: the menu of Ch 7.5 §2 is design-specific, and you pull *your* design's row off it. No empirical magnitudes appear anywhere below, because the memo is filed *before* the first-look regression is run — the freeze of Ch 7.5's "Your Turn." --- ## THE MODEL DELIVERABLE — `memo.md` ### 1. The specification (CONVENTIONS §4 form) - **Outcome:** the annual log average homeowner-insurance cost in county $i$, year $t$ — `log_premium`, built from the FRED/state-filing homeowner premium series merged to county, deflated to constant dollars. (Logged so the coefficient reads as an approximate percent change; the level series is right-skewed.) - **Treatment / key regressor:** `post_disaster` $= D_{it} = \mathbf{1}\{t \ge G_i\}$, an indicator that county $i$ has experienced its first billion-dollar FEMA major-disaster declaration on or before year $t$, where $G_i$ is the county's first-declaration year (counties with no qualifying declaration in the window are never-treated controls, $G_i = \infty$). The coefficient $\beta$ on $D_{it}$ is the one number I read off to answer H1. - **Controls:** county-year median household income and housing-unit count (from the ACS), entering as covariate-conditional parallel-trends controls — both shift the *level* of insurable exposure and could trend differently across counties for reasons unrelated to disasters. I deliberately do **not** control for the post-disaster rebuilding rate or the local construction-cost index, because higher rebuilding cost is a *channel* through which a disaster raises premiums (the over-controlling trap of Ch 7.3 §2 / Mentor 4): conditioning on it would regress away part of the very effect I am estimating. - **Fixed effects:** county fixed effects (so the comparison is *within* a county, sweeping out fixed exposure — coastal, wildfire-belt, floodplain) and year fixed effects (so a nationwide reinsurance-cost shock that hits every county in a given year is absorbed and not mistaken for a treatment effect). The comparison is therefore within-county, net of common-year shocks. - **Clustering:** by county. Treatment turns on at the county level and a county's premiums are serially correlated across years, so the plausible within-cluster error correlation lives at the county. (I will also report state-clustered SEs as a robustness check, since regulatory rate-setting is a state-level process — see the table.) - **Sample:** all U.S. counties with continuous premium coverage over the window 2010–2022, dropping counties with a population below 10,000 (premium series too noisy / often imputed) and dropping the two hurricane mega-years' directly-flooded counties only from a robustness cut, not the main sample. Expected ~2,800 counties × 13 years. Vintage pinned: NOAA Billion-Dollar Disasters table and FEMA OpenFEMA Disaster Declarations snapshot dated and recorded in the data card. - **Identifying assumption (one sentence):** absent the disaster, treated and never-/not-yet-treated counties' homeowner-insurance costs would have followed parallel trends. (Expanded in §2.) ### 2. The identifying-assumption paragraph *Our estimate of the effect of a billion-dollar FEMA disaster declaration on a county's average homeowner-insurance cost is **credible** as long as, absent the disaster, the insurance costs of treated counties and of never- and not-yet-treated counties would have evolved in parallel — a parallel-trends assumption we defend three ways: with an event study whose pre-treatment leads are flat and statistically indistinguishable from zero, plotted with confidence bands tight enough to rule out an economically meaningful pre-existing divergence; with a clean-control estimator that compares each treated cohort only against never- and not-yet-treated counties, never against already-hit counties whose still-rising premiums would contaminate the comparison under dynamic effects (Ch 4.2); and with the institutional fact that the precise timing of a billion-dollar disaster is meteorological, not chosen by the county, so adoption timing is plausibly unrelated to a county's latent premium trajectory. We report a single overall ATT and an event-time profile, with county-clustered inference, a state-clustered SE as a robustness check, and a placebo test that randomly reassigns disaster years to untreated counties.* The verb is **credible, not causal**: parallel trends in the post-period is a counterfactual no test can reach, so we have argued it as hard as the institutions allow and we claim a conditional belief, not a proof. ### 3. The threats-and-responses table | Threat | Why it's plausible | What we do about it | Residual concern | |---|---|---|---| | **Differential pre-trend** `[testable]` (parallel trends already failing before the disaster) | Disaster-prone counties — coast, floodplain, wildfire belt — may already be on a *steeper* premium-growth path as insurers price rising climate risk, independent of any single declaration. | Estimate the event study; normalize at $k=-1$; report and plot the leads with confidence bands. Reassuring result: leads flat and insignificant, bands tight enough to exclude a meaningful slope. | A flat pre-trend cannot *confirm* the post-period counterfactual; it only fails to refute it (Ch 4.1), and low power can make a noisy series look falsely flat. | | **Compound / coincident shock** `[arguable]` (a disaster co-occurs with a state regulatory or reinsurance change) | A major disaster often triggers a *contemporaneous* state response — a rate-freeze, an insurer-of-last-resort expansion, a building-code update — so the premium move could be the bundled policy, not the disaster itself. | Document each treated state's contemporaneous insurance-regulation changes from NAIC and state-filing records; argue case by case that no co-timed policy plausibly drives the premium *in the same direction and window*. State FE absorb time-invariant regulatory regimes; year FE absorb national reinsurance shocks. | No statistical test detects a perfectly co-timed state policy; this rests on the completeness of the institutional review, which is necessarily incomplete. | | **Anticipation** `[arguable]` (insurers reprice before the official declaration date) | Insurers watch forecasts and prior near-misses; a county repeatedly threatened may see premiums rise *before* the disaster that crosses the billion-dollar threshold, contaminating the "pre" period. | Partly testable: inspect the leads in the one to two years just before $G_i$ and, as a robustness cut, drop the year before declaration and re-estimate. The institutional argument: the *exact* declaration year is meteorological and unforecastable at the pricing horizon. | Anticipation that began before any observed lead is invisible to the leads; the claim ultimately rests on the timing-is-meteorological argument, not a test. | | **Forbidden comparison / negative weights** `[testable]` (lightly staggered timing using already-hit counties as controls) | Declarations arrive in different years, and the premium effect plausibly *builds* as insurers re-rate over several renewal cycles — the exact recipe for the staggered-DiD bias (Ch 4.2). | Design choice: use a clean-control group-time estimator (Callaway–Sant'Anna) against never- and not-yet-treated counties only; run a Goodman–Bacon decomposition to show how much weight a naive TWFE put on forbidden comparisons. | The clean estimator needs a sufficient never-treated pool; if climate risk becomes near-universal late in the window, late cohorts are thinly identified. | | **Outcome measurement / coverage gaps** `[testable]` (the premium series is imputed or missing for small counties) | Homeowner-premium data are assembled from state filings of uneven quality; sparse small-county cells are sometimes imputed, and imputation could correlate with disaster status (a Week-1 measurement-error problem). | Restrict the main sample to counties with continuous, non-imputed coverage; show the result is stable when the population floor is raised, and report how many county-years are imputed by treatment status. | Conditioning on coverage may itself select counties; the surviving sample is non-imputed but not a random subset of all counties. | | **Spillover / SUTVA violation** `[arguable]` (a disaster raises premiums in *control* counties too) | Reinsurance and insurer capital are pooled regionally, so a large disaster can push up premiums in nearby un-hit counties — contaminating the controls and biasing the estimated effect *toward zero*. | Argue from market structure which counties share a reinsurance pool; as a robustness cut, drop control counties adjacent to (or in the same state as) a treated county and re-estimate; year FE absorb the purely national component. | If spillovers are diffuse and national, no "drop the neighbors" cut removes them, and the estimate is a lower bound on the true effect — stated as such. | *Ordering note:* row 1 is the parallel-trends question — the hardest, led not buried. Rows 2–3 are the most dangerous *arguable* threats (compound shock, anticipation). Rows 4–6 descend to the design-mechanics and data-specific threats. ### 4. Honesty audit **(a) Most dangerous threat.** The **compound / coincident shock** (row 2), because it is *arguable*, not testable, and could in principle account for the entire effect: if every billion-dollar disaster is reliably followed by a state rate action in the same year, my $\beta$ is the policy, not the disaster. My two strongest sentences: (i) the *direction* matters — many post-disaster regulatory responses (rate freezes, last-resort pools) *suppress* premiums, so a co-timed policy would bias me *toward zero*, making a positive estimate conservative; (ii) state and year fixed effects already absorb time-invariant regulatory regimes and national reinsurance shocks, so the residual confounder must be a *county-or-state-specific, disaster-timed, premium-raising* policy, a narrow target I can hunt in NAIC records. Honestly: this narrows the threat but does not kill it, so the verb stays `credible` and the residual cell stays full. **(b) Verb audit.** I wrote **credible**, not causal, and I defend that: the load-bearing assumption (post-period parallel trends) is untestable, so a flat pre-trend earns a conditional belief, not a proof. There is no ungameable cutoff or randomized instrument here that would license `causal`. Every claim in the memo uses "consistent with" and "we cannot reject," never "proves." **(c) Empty-cell test.** No residual cell is blank. No cell reads "this could be the entire effect" outright, but row 2's residual comes closest, which is why it leads the audit. I am proceeding (not redesigning) because the fixed-effects structure plus the direction-of-bias argument bound the compound-shock threat to a conservative one — but I will state it as an explicit limitation in the Week-8 paper, not hide it. --- ## INSTRUCTOR GRADING NOTES (keyed to the research-design rubric) These notes show *why* this deliverable earns an A and where a weaker submission would lose points. The rubric criteria mirror Assessment 4 / Assessment 7 and the PS 7.5 part allocation. ### Part 1 — Spec discipline (15 pts) → **Excellent** All seven CONVENTIONS §4 slots are named to the level of a variable name (`log_premium`, `post_disaster` $=D_{it}$, the ACS controls, county+year FE, county clustering, the 2010–2022 sample with a stated row count and a pinned data vintage). Two A-grade moves: (i) the student **names the control they deliberately omit** — the rebuilding-cost channel — and says *why* (over-controlling on a mediator, Ch 7.3 §2), which is exactly the discipline the spec slot is meant to force; (ii) the clustering choice is *justified by where the errors correlate* (serial correlation within county) rather than asserted, and the alternative (state) is previewed as a robustness check rather than chosen post hoc to manufacture significance. *Where points are lost on a typical submission:* "Outcome: insurance cost" with no variable definition; controls listed without the omit-the-mediator reasoning; clustering stated but unjustified; no row count or vintage. Each is a slot scored down. A spec that has silently **drifted from the PAP** is the most serious Part-1 failure and is flagged as a contract violation regardless of polish. ### Part 2 — Identification reasoning (25 pts) → **Excellent** The sentence fills all three blanks correctly. **(a)** The effect is named ("the effect of a billion-dollar FEMA disaster declaration on a county's average homeowner-insurance cost"), not the coefficient. **(b)** The threat is *specific* — post-disaster parallel trends — never "endogeneity"; this alone separates a passing memo from a failing one and is the single most common place weak memos collapse. **(c)** The third blank does real work: three distinct defenses (flat tight leads, clean-control estimator, meteorological-timing institutional fact), not a restatement of the assumption. **(d)** The verb is `credible`, and — crucially — the student *defends the verb choice explicitly* ("parallel trends in the post-period is a counterfactual no test can reach"). *Where points are lost:* "credible as long as there is no endogeneity" → automatic zero on (b), because no threat is named (PS 7.5 Part 2b states this). "Defended by parallel trends" → near-zero on (c), because it restates rather than defends. Writing `causal` for an untestable DiD → loses (d) and signals the overclaiming Ch 7.5 §1 calls the tell of an amateur. The exemplar avoids all three. ### Part 3 — Threats-and-responses table (45 pts) → **Excellent** **(a) Coverage (12 pts).** The table carries the full DiD menu of Ch 7.5 §2 — differential pre-trend, forbidden comparison, anticipation — **plus** data-specific threats the menu cannot supply (measurement/coverage gaps, the SUTVA/spillover problem that is acute for regionally-pooled insurance). Six rows; a DiD memo missing the parallel-trends row would be capped at half on this part. **(b) Plausibility (6 pts).** Every "why plausible" cell tells a concrete, *setting-specific* story (coastal counties already on a steeper risk-pricing path; reinsurance pooled regionally so disasters leak into neighbors). No strawmen. **(c) Testable vs. arguable (15 pts) — the heaviest single criterion.** Every row is tagged `[testable]` or `[arguable]`, and — the part weak students get wrong — the column-3 response *matches the tag*. Pre-trend, forbidden comparison, and coverage are tagged testable and answered with diagnostics (event-study leads, Goodman–Bacon decomposition, an imputation count). Compound shock, anticipation, and diffuse spillover are tagged arguable and answered with *institutional argument* (NAIC records, the timing-is-meteorological claim, reinsurance-pool structure), with column 4 admitting "no test can detect this." The student commits **neither** of the two fatal errors of Ch 7.5 §4: no test is run against an untestable threat (e.g., no claim that a balance table "proves" no compound shock), and no testable threat is merely argued away. **(d) No empty residuals (8 pts).** Every column-4 cell is substantive — including the hardest admissions ("a flat pre-trend cannot confirm the counterfactual"; "the estimate is a lower bound" under diffuse spillover). This is the column a mentor reads first, and it is full and honest. **(e) Danger order (4 pts).** Row 1 is parallel trends, the first-ninety-seconds question; the dangerous arguable threats follow; mechanics and data threats descend last. The student even annotates the ordering logic. *Where points are lost:* the classic failure is a table where every column-3 entry is "we control for it" (no test/argument distinction, no tags) and column 4 is all "none." That table scores below half on (c) and zero on (d), even if it looks complete. ### Part 4 — Honesty audit (15 pts) → **Excellent** The student (a) names the most dangerous threat correctly as the *arguable* compound-shock row — not the showy testable one — and gives a genuine direction-of-bias argument that *bounds* rather than dismisses it; (b) defends the `credible` verb against the evidence; (c) passes the empty-cell test and, most tellingly, decides explicitly to *proceed with a stated limitation* rather than soften the language — which is exactly the Ch 7.5 §5 disposition: change the design or state the residual plainly, never launder the threat away. A submission that wrote "no concerns remain" in (c) would be marked *down*, per the chapter's rule that a clean residual column is a less credible paper. ### One-line grading heuristic The two highest-signal questions for any PS 7.5 submission: **(1)** Does the identifying-assumption sentence name a *specific* threat with a *real* defense and a *calibrated* verb (Part 2b–d)? **(2)** Is *every* table row correctly sorted into testable-with-a-statistic vs. arguable-with-an-argument, with a non-empty residual (Part 3c–d)? A student who clears both has the Week-7 mindset — they keep "identified under a defended-but-untestable assumption" cleanly apart from "proven causal," and they lead with their hardest question instead of burying it. The rest is execution. This exemplar clears both cleanly, which is what an A looks like. --- *End of model deliverable for PS 7.5. The exemplar is Priya's climate-insurance DiD; the chapter's worked memo (Ch 7.5 §6) is Maya's HMDA fair-lending DiD — read both to see the menu of Ch 7.5 §2 instantiated on two different designs. Citation for the disaster-insurance framing and data sources: NOAA National Centers for Environmental Information, *Billion-Dollar Weather and Climate Disasters* (data card pinned); FEMA OpenFEMA *Disaster Declarations Summaries*. The staggered-DiD machinery invoked in row 4 is Callaway and Sant'Anna (2021) and Goodman-Bacon (2021), as built in Ch 4.2. No empirical magnitudes are reported, by design: the memo is filed before the first-look regression of `nb7.5`.*

THE MODEL DELIVERABLE — memo.md