Week 5 Assessment — Reading the Frontier I

This is the end-of-week assessment for Week 5. Weeks 1–4 handed you a toolkit — OLS and its assumptions, robust and clustered standard errors, omitted-variable bias, instrumental variables, difference-in-differences, regression discontinuity — and taught you to run it. This week taught you something harder: to read it in someone else’s hands. Every chapter applied one fixed procedure, the seven-part Reader’s-Guide anatomy from the reading pack, to one famous paper — Fama–French (1992) and (1993), Jegadeesh–Titman (1993), Petersen (2009), Bertrand–Duflo–Mullainathan (2004). The assessment asks you to do it once more, alone, on a paper you have never seen.

The assessment has three parts. Part A is the centerpiece: you take an unseen empirical-finance paper that uses methods from Weeks 1–4 and produce your own four-page Reader’s Guide on it, applying the same seven boxes — research question, identification strategy, data, table-by-table reading order, what’s clever, what’s vulnerable, and three replication exercises. Part B is a short conceptual section, four quick items that test the reading reflexes in isolation: spot the headline table, separate identification from estimation, flag a vulnerable claim, name a robustness check. Part C is the analytic rubric with explicit point totals, followed by an instructor answer key and a worked example of a strong versus a weak “identifying assumption in one sentence.”

Total: 100 points. Part A = 70, Part B = 20, presentation and intellectual honesty woven through both = 10. This is take-home: budget an evening to read the paper the professional way (tables first, prose last) and a second sitting to write. The Guide is the seed of your Week 7–8 capstone, so treat it as a rehearsal for defending tables you will one day build yourself.

Throughout, hold every author to CONVENTIONS §4: every specification gets named explicitly — outcome · treatment/key regressor · controls · fixed effects · clustering · sample · identifying assumption in one sentence — with no hand-wavy appeals to “controlling for endogeneity.” The skill is the same one we hold ourselves to, pointed outward.

Part A — Write a four-page Reader’s Guide on an unseen paper (70 points)

A.0 The paper

Your instructor will hand you a paper, or you may choose one from the candidate types below and clear it with your mentor. The paper must (i) be empirical — it estimates something from data, it does not merely theorize; (ii) use a design you met in Weeks 1–4 — a portfolio sort, a panel regression with fixed effects and clustered standard errors, an instrumental-variables design, a difference-in-differences, or a regression discontinuity; and (iii) be one you have not read before this assessment. A short paper (15–35 pages) with a clear headline table is ideal. Do not pick a pure theory paper, a literature survey, or a structural-estimation paper whose machinery is past Week 4 — you cannot grade an identification strategy you were never taught.

Six candidate paper types (pick a type, then find a real paper of that type with your mentor — we deliberately do not hand you a citation, because finding the paper is part of the exercise):

A difference-in-differences on a financial regulation. A rule that switched on in some places or for some firms and not others — a state-level lending law, a disclosure mandate, a change to deposit-insurance limits, a short-sale ban applied to some stocks. Outcome: lending, prices, default, risk-taking. The identifying assumption will be parallel trends; the chief threat, selection into who got the rule and when.
An instrumental-variables paper in corporate finance or banking. Something instruments an endogenous regressor — leverage instrumented by tax-code variation, credit supply instrumented by a bank’s pre-existing exposure to a shock, financial development instrumented by legal origin. The identifying assumption is the exclusion restriction; the chief threat, that the instrument affects the outcome through a second channel.
A regression discontinuity at an index threshold. The Russell 1000/2000 reconstitution cutoff, a credit-score approval line, an S&P 500 addition rule, a covenant trigger, an analyst-coverage threshold. Outcome: ownership, voting, returns, financing. Identifying assumption: continuity of potential outcomes at the cutoff; chief threat, sorting/manipulation or a compound treatment that the same cutoff also triggers.
A factor / cross-sectional asset-pricing paper. A paper proposing a new return predictor or anomaly via portfolio sorts and Fama–MacBeth or time-series factor regressions — a Week 5 paper’s close cousin. “Identification” lives in the portfolio construction and the risk adjustment; the chief threat is that the premium is compensation for an omitted risk factor, or a data-mining artifact, or driven by tiny illiquid stocks.
A panel fixed-effects study with clustered standard errors. Firm-year or state-year panels on, say, climate risk and insurance pricing, BNPL adoption and household delinquency, or fair- lending outcomes. Identification is “within-unit, after absorbing fixed effects”; the chief threats are time-varying confounders the fixed effects do not absorb, and standard errors clustered at the wrong level (the Petersen/BDM disease).
An event study around a financial shock or announcement. Abnormal returns around an earnings surprise, a regulatory announcement, an index reconstitution, a crypto-protocol change. Identification rests on the event being unanticipated and the estimation window being clean; the chief threat is confounding events in the window and cross-sectional correlation in abnormal returns.

A.1 The seven-part Reader’s Guide (60 points)

Write roughly four pages — about 1,200–1,800 words — organized into the seven numbered boxes, in order. Read the paper the professional way before you write a word: abstract once, then straight to the tables and figures, find the headline result and read its sign / magnitude / t-statistic / sample size off the table itself, then the identification and data sections, then robustness, and only last the introduction. The boxes:

(1) Research question (5 pts). One sentence, plain English, ending in a question mark. Name the outcome and the key right-hand-side variable explicitly. “Does X predict / cause Y?” — not “this paper studies Y.”

(2) Identification strategy (12 pts). Name the design (sort, panel FE, IV, DiD, RD), then — the single most important sentence in the Guide — state the identifying assumption in one sentence, in the form “this estimate equals the true effect as long as ___.” Then name the chief threat to that assumption: the single most plausible thing that, if true, makes that sentence false, and say which way it would bias the estimate. Distinguish, in a line, identification (the assumption that lets the number mean what you want) from estimation (the regression machinery that produces the number).

(3) Data (8 pts). Unit of observation (firm-month? person? state-year?), sample period, source (CRSP, Compustat, EDGAR, a survey, an exchange feed), and number of observations. List the sample filters — “drops financial firms,” “requires two years of prior data,” “winsorizes at 1%” — because filters are decisions and decisions are levers. Flag any look-ahead or survivorship trap you suspect, even if the paper does not mention it.

(4) Table-by-table reading order (10 pts). Identify the headline table — the one that, if it broke, would sink the paper — and name the one number in it that carries the result. Then give the order in which a reader should look at the tables to absorb the argument fastest, labeling each as setup (summary stats, an IV first stage), headline, or defense (robustness, placebo, alternative specs), with one line on what each is for. Write the headline specification in the CONVENTIONS §4 format: outcome · key regressor · controls · fixed effects · clustering · sample · identifying assumption.

(5) What’s clever (5 pts). Name the one genuinely good idea — a construction, a comparison, a dataset nobody had assembled, a way of turning an awkward problem into a clean one. One paragraph.

(6) What’s vulnerable (8 pts). State the single most threatening alternative explanation for the headline result, and the specific evidence that would settle it for or against. This must be consistent with the chief threat you named in box (2), and it must be a real soft spot, not a generic “correlation isn’t causation.”

(7) Three replication exercises (12 pts). Three concrete things a student could do on real data on a laptop. Exactly one of each kind: (a) reproduce a specific number from the headline table — name the table, the cell, and the data you would pull; (b) stress an assumption — re-run with different controls, a different sample window, a different standard-error flavor (cluster differently, add HAC), or a placebo; say what result would break the paper; (c) extend the idea to a new setting — a new asset class, a more recent decade, a different country or market. Each must be feasible with data the camp can actually reach (CRSP, Compustat, EDGAR, public APIs); “obtain proprietary hedge-fund records” is not a feasible exercise.

A.2 The two load-bearing sentences (10 points, scored inside A but called out separately)

Two sentences in your Guide carry disproportionate weight, and we grade them on their own:

The identifying assumption, in one sentence (box 2). In the exact form “this estimate equals the true effect as long as ___,” naming the specific condition for this paper’s design — not the word “endogeneity,” not “no confounders” in the abstract, but the concrete bet this paper is making (parallel trends in the post period; the instrument affecting Y only through the endogenous regressor; potential outcomes continuous at the cutoff; the sort not loading on an omitted priced risk). (5 pts)
The chief threat, named (box 2 / box 6). The single most plausible violation of that assumption, named concretely for this paper, with the direction of the resulting bias. (5 pts)

A Guide that nails these two sentences and is merely competent elsewhere outscores a polished one that fudges them. This is the whole point of Week 5.

Part B — Conceptual reading drills (20 points)

Four short items. Answer each in two to five sentences. These test the reflexes from the reading pack in isolation; you may answer them about the paper you chose for Part A, or your instructor may attach a one-page table excerpt to answer them against. Cite specific cells, columns, or rows.

B1. Identify the headline table (5 pts). Of all the tables in your paper, which is the headline — the one that, if it broke, would sink the paper? Name it, name the single number in it that carries the result, and say in one sentence how you can tell it apart from a setup table (summary statistics, a first stage) and a defense table (robustness, placebo). State the headline number’s sign, magnitude, and t-statistic (compute the t-stat yourself as \hat\beta/\text{SE} if the table reports the standard error).

B2. Distinguish identification from estimation (5 pts). Pick the headline specification and separate its two halves. What is the estimation — the mechanical regression that produces the number (e.g., “OLS of default on debt-to-income with lender and year fixed effects, SEs clustered by lender”)? What is the identification — the assumption under which that number equals the causal or structural quantity the authors want (e.g., “within a lender and year, debt-to-income is as good as randomly assigned with respect to unobserved borrower quality”)? Explain why a paper can have flawless estimation and still be wrong, but not the reverse.

B3. Spot a vulnerable claim (5 pts). Quote one sentence from your paper’s prose — the abstract, intro, or conclusion — that claims more than the corresponding table delivers: a causal verb where the design only supports correlation, a “robust across all specifications” where one column wobbles, an economic-importance claim resting on a statistically-significant-but-tiny coefficient, or a t-stat that looks inflated because the standard errors are not clustered at the level the key variable varies. Name which kind of overreach it is, and what the table actually supports.

B4. Name a robustness check (5 pts). Name the one robustness check the authors ran that most reassured you, and say what alternative explanation it was designed to rule out and whether it succeeds. Then name one robustness check they should have run but did not — the one a skeptical seminar audience would ask for first — and say what it would test. (Examples of the genre: a placebo on a fake event date, re-clustering standard errors, dropping the largest firms, a McCrary density test at an RD cutoff, a different formation/holding window for a sort, a pre-trend event study for a DiD.)

Part C — Analytic rubric (point allocations explicit)

Each row is scored at one of four levels. The Part-A rows distribute the 70 Part-A points across the seven boxes and the two load-bearing sentences; Part B is scored by its four items (5 points each); the presentation-and-honesty row spans the whole assessment.

Criterion	Excellent	Proficient	Developing	Missing	Points
Identifying assumption + chief threat (boxes 2, 6; the two load-bearing sentences A.2)	The identifying assumption is stated in one clean sentence in the “equals the true effect as long as ___” form, naming the concrete condition for this design (parallel trends in the post period / exclusion / continuity at the cutoff / sort not loading on an omitted risk); the chief threat is named specifically and its bias direction given; identification is cleanly separated from estimation; box 6’s vulnerability is the same threat, with the settling evidence named.	Assumption and threat both correct and specific, but bias direction missing, or box 6 only loosely tied to box 2, or the identification-vs-estimation line is thin.	Assumption stated but generic (“controls for endogeneity,” “no confounders”) or attached to the wrong design; threat named but vague; bias direction absent or wrong.	Assumption absent, the word “endogeneity” doing all the work, or threat not named.	22
Table-reading accuracy (box 4, A1; plus B1, B2)	Headline table correctly identified and distinguished from setup/defense; the one carrying number read correctly (sign, magnitude, t-stat, N) off the table, not the prose; reading order sensible; the §4 spec line complete (outcome · key regressor · controls · FE · clustering · sample · assumption); B1/B2 correct, with identification cleanly split from estimation.	Headline correct and number read correctly, but the §4 spec line drops one element (e.g., clustering or sample), or the setup/defense labeling has one slip.	Headline plausible but the carrying number misread or taken from the prose; spec line mostly missing; confuses a setup or defense table for the headline.	Wrong headline table, or numbers invented / not located in any table.	18
Quality & feasibility of the three replications (box 7)	Three exercises, one of each kind (reproduce a named cell / stress an assumption / extend to a new setting); each is concrete, names the data source the camp can actually reach, and the stress exercise says what result would break the paper; the reproduce exercise names table, cell, and the data pull.	Three exercises of the right three kinds and feasible, but one is vague on the data or the stress test does not say what would break the paper.	Exercises present but two are the same kind, or one requires data the camp cannot reach, or they restate the paper rather than test it.	Fewer than three, or generic (“run more regressions”), or infeasible.	14
The other four boxes (1 research question, 3 data, 5 clever; plus B3, B4)	Question is one sentence as a question with outcome and key regressor named; data box gives unit / period / source / N and lists filters and flags a look-ahead or survivorship trap; “clever” names the one real good idea; B3 quotes a genuine overreach and classifies it; B4 names a reassuring check and a missing one.	All four boxes solid; one item thin (e.g., filters listed but no trap flagged, or B4 names a reassuring check but a weak “missing” one).	Boxes present but the question is a statement not a question, data box missing N or source, “clever” is generic; B3/B4 superficial.	One or more boxes missing or wrong.	16
Presentation, spec-discipline, and honest limitations	Clean prose in the book’s voice; specs in §4 format; reads tables before prose and says so; explicitly flags which assumptions are untestable and what would be needed to defend them; keeps “identified under a defended-but-untestable assumption” cleanly separate from “proven causal”; admits where the paper — or the student’s own reading — is uncertain rather than overclaiming.	One labeling or stylistic lapse; honesty present but a touch thin.	Several lapses; treats a passed diagnostic as proof of an untestable assumption; overclaims causality.	Unreadable, or asserts the paper is “proven” with no caveats.	10

Total: 100 points. (Part A = 70: the identifying-assumption/threat row 22 + table-reading 18 + replications 14 + the-other-four-boxes 16 = 70. Part B’s 20 points are folded into the table-reading row [B1, B2 = 10] and the other-four-boxes row [B3, B4 = 10]. The presentation/honesty row adds 10. The rubric is normalized so the maximum awarded is 100.)

A note on the spirit of the honesty row. Week 5 rewards the reader who can say, out loud and unprompted, which assumption the paper’s whole result rests on and why no test in the paper can prove it, and who refuses to launder “the authors ran a placebo and it passed” into “the design is proven valid.” A Guide that writes “the parallel-trends assumption is untestable in the post period; the flat pre-trends are reassuring but not proof, and here is the sensitivity check I would run” outscores a Guide that presents a clean pre-trend plot as a verdict. Equally, a Guide that admits it could not tell, from the paper alone, whether the standard errors were clustered correctly — and says what it would need to check — outscores one that confidently asserts a t-stat it never interrogated. Knowing what kind of claim the paper is making, and what kind you are making about it, is the entire point.

Instructor guidance and answer key

This is an open-ended writing task, so the key is a set of standards and exemplars rather than a single correct answer. The two highest-signal things to grade are (1) the identifying-assumption sentence and the named threat — does the student state the concrete bet this paper makes, in the “as long as” form, and name the specific, direction-signed violation — and (2) table-reading fidelity — did the student find the real headline table and read its carrying number off the table rather than parroting the abstract? A student who does these two things well has the Week-5 mindset regardless of polish; the rest is execution.

Model-answer sketch for Part B

B1 (headline table). A strong answer points to a results table (the main regression or the main portfolio-sort spread), not Table 1 (which is almost always summary statistics — a setup table) and not a robustness appendix table (defense). The tell: the headline table’s dependent variable is the paper’s outcome and its key row is the treatment/key regressor; a setup table has no coefficient of interest; a defense table re-runs the headline under a perturbation. Strong answers read the carrying number’s sign, size, t-stat (computing t=\hat\beta/\text{SE} where only the SE is printed, per the reading pack’s table glossary), and N. Common error: naming the first stage of an IV as the headline — it is setup; the headline is the second-stage (or reduced-form) effect on the outcome.

B2 (identification vs. estimation). Full credit separates the two cleanly. Estimation is mechanical and checkable from the table: which estimator, which controls, which fixed effects, which clustering — anyone with the data can reproduce it. Identification is the untestable bridge from the estimate to the quantity of interest: the assumption (parallel trends / exclusion / continuity / correct factor model) under which the coefficient equals the effect. The required insight: a paper can have flawless estimation and false identification (the regression is computed perfectly, but the assumption that makes the coefficient causal is wrong — and no amount of estimation care fixes a dead identifying assumption), whereas the reverse — true identification undone purely by an estimation slip — is usually a fixable bug, not a fatal flaw. Reward the student who says estimation errors are correctable and identification failures are often fatal.

B3 (vulnerable claim). Reward a specific quoted sentence and a correct classification of the overreach. The four genres to recognize: (i) a causal verb on a correlational design (“X raises Y” when the design is a cross-sectional panel with no exogenous variation); (ii) “robust across all specifications” when the coefficient visibly shrinks or loses significance in the most demanding column; (iii) economic-importance inflation — a three-star coefficient that is economically trivial (Week 2’s “stars are not size”); (iv) inflated t-stats from standard errors not clustered at the level the key variable varies (the Petersen/BDM lesson — un-clustered SEs on persistent panel data are too small). A weak answer quotes a sentence that is actually well-supported, or calls something an overreach without saying what the table does support.

B4 (robustness). A strong answer names a real check from the paper and the specific alternative it rules out (a placebo event date rules out “anything that happened at that time”; a McCrary test rules out sorting at an RD cutoff; re-clustering rules out understated SEs; dropping the largest firms rules out “one mega-cap drives the result”). Then it names a missing check that a referee would demand first — and it should be the check that bears on the chief threat from box 2. A weak answer names a cosmetic robustness check (e.g., “added one more control”) and a missing check unrelated to the actual vulnerability.

Worked example: a strong vs. a weak “identifying assumption in one sentence”

Take a candidate of type 1, a difference-in-differences on a state lending regulation — Maya’s territory. Suppose the headline table shows that after some states capped payday-loan APRs, household delinquency in those states fell relative to non-capping states.

Weak (do not reward): > “The paper uses a difference-in-differences design to control for endogeneity and isolate the causal > effect of the cap, since fixed effects account for confounding factors.”

This fails on every count the reading pack and CONVENTIONS §4 warn about. It names no concrete condition; “control for endogeneity” is exactly the banned hand-wave; “fixed effects account for confounding” is false comfort (fixed effects absorb time-invariant confounders, not the time-varying ones that threaten parallel trends); and it is not in the “as long as” form, so it states no falsifiable bet. There is no threat named and no bias direction.

Strong (full credit): > “This DiD estimate equals the true effect of the APR cap on delinquency as long as the > capping and non-capping states would have followed parallel delinquency trends absent the cap — the > chief threat being that states chose to cap precisely when their own delinquency was already turning > (reverse causation in the timing), which would bias the estimated effect away from zero and make > the cap look more protective than it was.”

This names the design’s concrete bet (parallel trends in the post period), states it as a falsifiable “as long as” condition, names a specific and plausible threat for this setting (policy timing responding to the outcome’s own trajectory — not a generic “omitted variable”), and gives the direction of the bias. A student who then adds, in box 6, “I would check the pre-period event-study leads and run a sensitivity analysis for how large a differential trend would overturn the result” has connected the assumption, the threat, and a replication into one coherent reading — the Week-5 target.

A second quick contrast, for an IV paper (type 2): weak — “the instrument is valid because it is exogenous”; strong — “the 2SLS estimate equals the causal effect of leverage on investment as long as the tax-code change shifts a firm’s leverage but has no direct effect on its investment except through leverage — the chief threat being that the same tax change also altered the after-tax return to investment directly, violating exclusion and biasing the estimate upward if the direct channel and the leverage channel push the same way.” The pattern is identical across designs: name the design’s own assumption concretely, in the “as long as” form, then the specific threat and its direction. That is the sentence the entire week was built to teach.

# Week 5 Assessment — Reading the Frontier I This is the end-of-week assessment for Week 5. Weeks 1–4 handed you a toolkit — OLS and its assumptions, robust and clustered standard errors, omitted-variable bias, instrumental variables, difference-in-differences, regression discontinuity — and taught you to *run* it. This week taught you something harder: to *read* it in someone else's hands. Every chapter applied one fixed procedure, the seven-part Reader's-Guide anatomy from the reading pack, to one famous paper — Fama–French (1992) and (1993), Jegadeesh–Titman (1993), Petersen (2009), Bertrand–Duflo–Mullainathan (2004). The assessment asks you to do it once more, alone, on a paper you have never seen. The assessment has three parts. **Part A** is the centerpiece: you take an unseen empirical-finance paper that uses methods from Weeks 1–4 and produce your own four-page Reader's Guide on it, applying the same seven boxes — research question, identification strategy, data, table-by-table reading order, what's clever, what's vulnerable, and three replication exercises. **Part B** is a short conceptual section, four quick items that test the reading reflexes in isolation: spot the headline table, separate identification from estimation, flag a vulnerable claim, name a robustness check. **Part C** is the analytic rubric with explicit point totals, followed by an instructor answer key and a worked example of a strong versus a weak "identifying assumption in one sentence." **Total: 100 points.** Part A = 70, Part B = 20, presentation and intellectual honesty woven through both = 10. This is take-home: budget an evening to read the paper the professional way (tables first, prose last) and a second sitting to write. The Guide is the seed of your Week 7–8 capstone, so treat it as a rehearsal for defending tables you will one day build yourself. Throughout, hold every author to CONVENTIONS §4: every specification gets named explicitly — **outcome · treatment/key regressor · controls · fixed effects · clustering · sample · identifying assumption in one sentence** — with no hand-wavy appeals to "controlling for endogeneity." The skill is the same one we hold *ourselves* to, pointed outward. --- ## Part A — Write a four-page Reader's Guide on an unseen paper (70 points) ### A.0 The paper Your instructor will hand you a paper, or you may choose one from the candidate **types** below and clear it with your mentor. The paper must (i) be empirical — it estimates something from data, it does not merely theorize; (ii) use a design you met in Weeks 1–4 — a portfolio sort, a panel regression with fixed effects and clustered standard errors, an instrumental-variables design, a difference-in-differences, or a regression discontinuity; and (iii) be one you have **not** read before this assessment. A short paper (15–35 pages) with a clear headline table is ideal. Do not pick a pure theory paper, a literature survey, or a structural-estimation paper whose machinery is past Week 4 — you cannot grade an identification strategy you were never taught. **Six candidate paper types** (pick a *type*, then find a real paper of that type with your mentor — we deliberately do not hand you a citation, because finding the paper is part of the exercise): 1. **A difference-in-differences on a financial regulation.** A rule that switched on in some places or for some firms and not others — a state-level lending law, a disclosure mandate, a change to deposit-insurance limits, a short-sale ban applied to some stocks. Outcome: lending, prices, default, risk-taking. The identifying assumption will be parallel trends; the chief threat, selection into who got the rule and when. 2. **An instrumental-variables paper in corporate finance or banking.** Something instruments an endogenous regressor — leverage instrumented by tax-code variation, credit supply instrumented by a bank's pre-existing exposure to a shock, financial development instrumented by legal origin. The identifying assumption is the exclusion restriction; the chief threat, that the instrument affects the outcome through a second channel. 3. **A regression discontinuity at an index threshold.** The Russell 1000/2000 reconstitution cutoff, a credit-score approval line, an S&P 500 addition rule, a covenant trigger, an analyst-coverage threshold. Outcome: ownership, voting, returns, financing. Identifying assumption: continuity of potential outcomes at the cutoff; chief threat, sorting/manipulation or a compound treatment that the same cutoff also triggers. 4. **A factor / cross-sectional asset-pricing paper.** A paper proposing a new return predictor or anomaly via portfolio sorts and Fama–MacBeth or time-series factor regressions — a Week 5 paper's close cousin. "Identification" lives in the portfolio construction and the risk adjustment; the chief threat is that the premium is compensation for an omitted risk factor, or a data-mining artifact, or driven by tiny illiquid stocks. 5. **A panel fixed-effects study with clustered standard errors.** Firm-year or state-year panels on, say, climate risk and insurance pricing, BNPL adoption and household delinquency, or fair- lending outcomes. Identification is "within-unit, after absorbing fixed effects"; the chief threats are time-varying confounders the fixed effects do not absorb, and standard errors clustered at the wrong level (the Petersen/BDM disease). 6. **An event study around a financial shock or announcement.** Abnormal returns around an earnings surprise, a regulatory announcement, an index reconstitution, a crypto-protocol change. Identification rests on the event being unanticipated and the estimation window being clean; the chief threat is confounding events in the window and cross-sectional correlation in abnormal returns. ### A.1 The seven-part Reader's Guide (60 points) Write roughly four pages — about 1,200–1,800 words — organized into the seven numbered boxes, in order. Read the paper the professional way before you write a word: abstract once, then straight to the **tables and figures**, find the headline result and read its sign / magnitude / t-statistic / sample size *off the table itself*, then the identification and data sections, then robustness, and only last the introduction. The boxes: **(1) Research question (5 pts).** One sentence, plain English, ending in a question mark. Name the **outcome** and the **key right-hand-side variable** explicitly. "Does X predict / cause Y?" — not "this paper studies Y." **(2) Identification strategy (12 pts).** Name the design (sort, panel FE, IV, DiD, RD), then — the single most important sentence in the Guide — **state the identifying assumption in one sentence**, in the form "this estimate equals the true effect *as long as* ___." Then **name the chief threat to that assumption**: the single most plausible thing that, if true, makes that sentence false, and say which way it would bias the estimate. Distinguish, in a line, *identification* (the assumption that lets the number mean what you want) from *estimation* (the regression machinery that produces the number). **(3) Data (8 pts).** Unit of observation (firm-month? person? state-year?), sample period, source (CRSP, Compustat, EDGAR, a survey, an exchange feed), and number of observations. List the sample filters — "drops financial firms," "requires two years of prior data," "winsorizes at 1%" — because filters are decisions and decisions are levers. Flag any look-ahead or survivorship trap you suspect, even if the paper does not mention it. **(4) Table-by-table reading order (10 pts).** Identify the **headline table** — the one that, if it broke, would sink the paper — and name the one number in it that carries the result. Then give the order in which a reader should look at the tables to absorb the argument fastest, labeling each as *setup* (summary stats, an IV first stage), *headline*, or *defense* (robustness, placebo, alternative specs), with one line on what each is *for*. Write the headline specification in the CONVENTIONS §4 format: outcome · key regressor · controls · fixed effects · clustering · sample · identifying assumption. **(5) What's clever (5 pts).** Name the one genuinely good idea — a construction, a comparison, a dataset nobody had assembled, a way of turning an awkward problem into a clean one. One paragraph. **(6) What's vulnerable (8 pts).** State the *single most threatening* alternative explanation for the headline result, and the specific evidence that would settle it for or against. This must be consistent with the chief threat you named in box (2), and it must be a real soft spot, not a generic "correlation isn't causation." **(7) Three replication exercises (12 pts).** Three concrete things a student could *do* on real data on a laptop. Exactly one of each kind: (a) **reproduce** a specific number from the headline table — name the table, the cell, and the data you would pull; (b) **stress** an assumption — re-run with different controls, a different sample window, a different standard-error flavor (cluster differently, add HAC), or a placebo; say what result would *break* the paper; (c) **extend** the idea to a new setting — a new asset class, a more recent decade, a different country or market. Each must be feasible with data the camp can actually reach (CRSP, Compustat, EDGAR, public APIs); "obtain proprietary hedge-fund records" is not a feasible exercise. ### A.2 The two load-bearing sentences (10 points, scored inside A but called out separately) Two sentences in your Guide carry disproportionate weight, and we grade them on their own: - **The identifying assumption, in one sentence** (box 2). In the exact form "this estimate equals the true effect *as long as* ___," naming the specific condition for *this* paper's design — not the word "endogeneity," not "no confounders" in the abstract, but the concrete bet this paper is making (parallel trends in the *post* period; the instrument affecting Y only through the endogenous regressor; potential outcomes continuous at the cutoff; the sort not loading on an omitted priced risk). **(5 pts)** - **The chief threat, named** (box 2 / box 6). The single most plausible violation of that assumption, named concretely for this paper, with the *direction* of the resulting bias. **(5 pts)** A Guide that nails these two sentences and is merely competent elsewhere outscores a polished one that fudges them. This is the whole point of Week 5. --- ## Part B — Conceptual reading drills (20 points) Four short items. Answer each in two to five sentences. These test the reflexes from the reading pack in isolation; you may answer them about the paper you chose for Part A, or your instructor may attach a one-page table excerpt to answer them against. Cite specific cells, columns, or rows. **B1. Identify the headline table (5 pts).** Of all the tables in your paper, which is the **headline** — the one that, if it broke, would sink the paper? Name it, name the single number in it that carries the result, and say in one sentence how you can tell it apart from a *setup* table (summary statistics, a first stage) and a *defense* table (robustness, placebo). State the headline number's sign, magnitude, and t-statistic (compute the t-stat yourself as $\hat\beta/\text{SE}$ if the table reports the standard error). **B2. Distinguish identification from estimation (5 pts).** Pick the headline specification and separate its two halves. What is the **estimation** — the mechanical regression that produces the number (e.g., "OLS of default on debt-to-income with lender and year fixed effects, SEs clustered by lender")? What is the **identification** — the assumption under which that number equals the causal or structural quantity the authors want (e.g., "within a lender and year, debt-to-income is as good as randomly assigned with respect to unobserved borrower quality")? Explain why a paper can have flawless estimation and still be wrong, but not the reverse. **B3. Spot a vulnerable claim (5 pts).** Quote one sentence from your paper's *prose* — the abstract, intro, or conclusion — that claims **more than the corresponding table delivers**: a causal verb where the design only supports correlation, a "robust across all specifications" where one column wobbles, an economic-importance claim resting on a statistically-significant-but-tiny coefficient, or a t-stat that looks inflated because the standard errors are not clustered at the level the key variable varies. Name which kind of overreach it is, and what the table actually supports. **B4. Name a robustness check (5 pts).** Name the **one robustness check the authors ran that most reassured you**, and say what alternative explanation it was designed to rule out and whether it succeeds. Then name **one robustness check they should have run but did not** — the one a skeptical seminar audience would ask for first — and say what it would test. (Examples of the genre: a placebo on a fake event date, re-clustering standard errors, dropping the largest firms, a McCrary density test at an RD cutoff, a different formation/holding window for a sort, a pre-trend event study for a DiD.) --- ## Part C — Analytic rubric (point allocations explicit) Each row is scored at one of four levels. The Part-A rows distribute the 70 Part-A points across the seven boxes and the two load-bearing sentences; Part B is scored by its four items (5 points each); the presentation-and-honesty row spans the whole assessment. | Criterion | Excellent | Proficient | Developing | Missing | Points | |---|---|---|---|---|---| | **Identifying assumption + chief threat (boxes 2, 6; the two load-bearing sentences A.2)** | The identifying assumption is stated in one clean sentence in the "equals the true effect *as long as* ___" form, naming the *concrete* condition for this design (parallel trends in the post period / exclusion / continuity at the cutoff / sort not loading on an omitted risk); the chief threat is named specifically and its *bias direction* given; identification is cleanly separated from estimation; box 6's vulnerability is the same threat, with the settling evidence named. | Assumption and threat both correct and specific, but bias direction missing, or box 6 only loosely tied to box 2, or the identification-vs-estimation line is thin. | Assumption stated but generic ("controls for endogeneity," "no confounders") or attached to the wrong design; threat named but vague; bias direction absent or wrong. | Assumption absent, the word "endogeneity" doing all the work, or threat not named. | 22 | | **Table-reading accuracy (box 4, A1; plus B1, B2)** | Headline table correctly identified and distinguished from setup/defense; the one carrying number read correctly (sign, magnitude, t-stat, $N$) off the table, not the prose; reading order sensible; the §4 spec line complete (outcome · key regressor · controls · FE · clustering · sample · assumption); B1/B2 correct, with identification cleanly split from estimation. | Headline correct and number read correctly, but the §4 spec line drops one element (e.g., clustering or sample), or the setup/defense labeling has one slip. | Headline plausible but the carrying number misread or taken from the prose; spec line mostly missing; confuses a setup or defense table for the headline. | Wrong headline table, or numbers invented / not located in any table. | 18 | | **Quality & feasibility of the three replications (box 7)** | Three exercises, one of *each* kind (reproduce a named cell / stress an assumption / extend to a new setting); each is concrete, names the data source the camp can actually reach, and the stress exercise says what result would *break* the paper; the reproduce exercise names table, cell, and the data pull. | Three exercises of the right three kinds and feasible, but one is vague on the data or the stress test does not say what would break the paper. | Exercises present but two are the same kind, or one requires data the camp cannot reach, or they restate the paper rather than test it. | Fewer than three, or generic ("run more regressions"), or infeasible. | 14 | | **The other four boxes (1 research question, 3 data, 5 clever; plus B3, B4)** | Question is one sentence as a question with outcome and key regressor named; data box gives unit / period / source / $N$ and lists filters and flags a look-ahead or survivorship trap; "clever" names the one real good idea; B3 quotes a genuine overreach and classifies it; B4 names a reassuring check *and* a missing one. | All four boxes solid; one item thin (e.g., filters listed but no trap flagged, or B4 names a reassuring check but a weak "missing" one). | Boxes present but the question is a statement not a question, data box missing $N$ or source, "clever" is generic; B3/B4 superficial. | One or more boxes missing or wrong. | 16 | | **Presentation, spec-discipline, and honest limitations** | Clean prose in the book's voice; specs in §4 format; reads tables before prose and says so; explicitly flags which assumptions are *untestable* and what would be needed to defend them; keeps "identified under a defended-but-untestable assumption" cleanly separate from "proven causal"; admits where the paper — or the student's own reading — is uncertain rather than overclaiming. | One labeling or stylistic lapse; honesty present but a touch thin. | Several lapses; treats a passed diagnostic as proof of an untestable assumption; overclaims causality. | Unreadable, or asserts the paper is "proven" with no caveats. | 10 | **Total: 100 points.** (Part A = 70: the identifying-assumption/threat row 22 + table-reading 18 + replications 14 + the-other-four-boxes 16 = 70. Part B's 20 points are folded into the table-reading row [B1, B2 = 10] and the other-four-boxes row [B3, B4 = 10]. The presentation/honesty row adds 10. The rubric is normalized so the maximum awarded is 100.) A note on the spirit of the honesty row. Week 5 rewards the reader who can say, out loud and unprompted, *which assumption the paper's whole result rests on and why no test in the paper can prove it*, and who refuses to launder "the authors ran a placebo and it passed" into "the design is proven valid." A Guide that writes "the parallel-trends assumption is untestable in the post period; the flat pre-trends are reassuring but not proof, and here is the sensitivity check I would run" outscores a Guide that presents a clean pre-trend plot as a verdict. Equally, a Guide that *admits* it could not tell, from the paper alone, whether the standard errors were clustered correctly — and says what it would need to check — outscores one that confidently asserts a t-stat it never interrogated. Knowing *what kind of claim* the paper is making, and what kind you are making about it, is the entire point. --- ## Instructor guidance and answer key This is an open-ended writing task, so the key is a set of *standards* and exemplars rather than a single correct answer. The two highest-signal things to grade are (1) **the identifying-assumption sentence and the named threat** — does the student state the concrete bet this paper makes, in the "as long as" form, and name the specific, direction-signed violation — and (2) **table-reading fidelity** — did the student find the real headline table and read its carrying number off the table rather than parroting the abstract? A student who does these two things well has the Week-5 mindset regardless of polish; the rest is execution. ### Model-answer sketch for Part B **B1 (headline table).** A strong answer points to a *results* table (the main regression or the main portfolio-sort spread), not Table 1 (which is almost always summary statistics — a *setup* table) and not a robustness appendix table (*defense*). The tell: the headline table's dependent variable is the paper's outcome and its key row is the treatment/key regressor; a setup table has no coefficient of interest; a defense table re-runs the headline under a perturbation. Strong answers read the carrying number's sign, size, t-stat (computing $t=\hat\beta/\text{SE}$ where only the SE is printed, per the reading pack's table glossary), and $N$. Common error: naming the first stage of an IV as the headline — it is *setup*; the headline is the second-stage (or reduced-form) effect on the outcome. **B2 (identification vs. estimation).** Full credit separates the two cleanly. *Estimation* is mechanical and checkable from the table: which estimator, which controls, which fixed effects, which clustering — anyone with the data can reproduce it. *Identification* is the untestable bridge from the estimate to the quantity of interest: the assumption (parallel trends / exclusion / continuity / correct factor model) under which the coefficient equals the effect. The required insight: a paper can have *flawless estimation and false identification* (the regression is computed perfectly, but the assumption that makes the coefficient causal is wrong — and *no amount of estimation care fixes a dead identifying assumption*), whereas the reverse — true identification undone purely by an estimation slip — is usually a fixable bug, not a fatal flaw. Reward the student who says estimation errors are correctable and identification failures are often fatal. **B3 (vulnerable claim).** Reward a *specific quoted sentence* and a correct classification of the overreach. The four genres to recognize: (i) a **causal verb on a correlational design** ("X *raises* Y" when the design is a cross-sectional panel with no exogenous variation); (ii) **"robust across all specifications"** when the coefficient visibly shrinks or loses significance in the most demanding column; (iii) **economic-importance inflation** — a three-star coefficient that is economically trivial (Week 2's "stars are not size"); (iv) **inflated t-stats** from standard errors not clustered at the level the key variable varies (the Petersen/BDM lesson — un-clustered SEs on persistent panel data are too small). A weak answer quotes a sentence that is actually well-supported, or calls something an overreach without saying what the table *does* support. **B4 (robustness).** A strong answer names a *real* check from the paper and the *specific* alternative it rules out (a placebo event date rules out "anything that happened at that time"; a McCrary test rules out sorting at an RD cutoff; re-clustering rules out understated SEs; dropping the largest firms rules out "one mega-cap drives the result"). Then it names a *missing* check that a referee would demand first — and it should be the check that bears on the chief threat from box 2. A weak answer names a cosmetic robustness check (e.g., "added one more control") and a missing check unrelated to the actual vulnerability. ### Worked example: a strong vs. a weak "identifying assumption in one sentence" Take a candidate of **type 1**, a difference-in-differences on a state lending regulation — Maya's territory. Suppose the headline table shows that after some states capped payday-loan APRs, household delinquency in those states fell relative to non-capping states. **Weak (do not reward):** > "The paper uses a difference-in-differences design to control for endogeneity and isolate the causal > effect of the cap, since fixed effects account for confounding factors." This fails on every count the reading pack and CONVENTIONS §4 warn about. It names no concrete condition; "control for endogeneity" is exactly the banned hand-wave; "fixed effects account for confounding" is false comfort (fixed effects absorb *time-invariant* confounders, not the *time-varying* ones that threaten parallel trends); and it is not in the "as long as" form, so it states no falsifiable bet. There is no threat named and no bias direction. **Strong (full credit):** > "This DiD estimate equals the true effect of the APR cap on delinquency *as long as* the > capping and non-capping states would have followed parallel delinquency trends absent the cap — the > chief threat being that states chose to cap precisely when their own delinquency was already turning > (reverse causation in the timing), which would bias the estimated effect *away from zero* and make > the cap look more protective than it was." This names the design's concrete bet (parallel trends in the *post* period), states it as a falsifiable "as long as" condition, names a *specific* and plausible threat for *this* setting (policy timing responding to the outcome's own trajectory — not a generic "omitted variable"), and gives the *direction* of the bias. A student who then adds, in box 6, "I would check the pre-period event-study leads and run a sensitivity analysis for how large a differential trend would overturn the result" has connected the assumption, the threat, and a replication into one coherent reading — the Week-5 target. A second quick contrast, for an **IV** paper (type 2): weak — "the instrument is valid because it is exogenous"; strong — "the 2SLS estimate equals the causal effect of leverage on investment *as long as* the tax-code change shifts a firm's leverage but has no direct effect on its investment except through leverage — the chief threat being that the same tax change also altered the after-tax return to investment directly, violating exclusion and biasing the estimate *upward* if the direct channel and the leverage channel push the same way." The pattern is identical across designs: name the design's own assumption concretely, in the "as long as" form, then the specific threat and its direction. That is the sentence the entire week was built to teach.