Skip to content

PROGRESS.md

Latest commit

 

History

History
71 lines (48 loc) · 6.09 KB

PROGRESS.md

File metadata and controls

71 lines (48 loc) · 6.09 KB
Raw

Codechella Madrid 2026 — Progress Log

A running log of session-by-session work on the workshop materials. Newest entry on top.

2026-05-26 — Bias_ΔX restructure + Title X clinics event-study reveal

Slides/02-covariates.tex — Bias_ΔX section rebuilt

Distributed the dense slides for lower cognitive density. The old section crammed multiple ideas onto each frame (RCT analogy + DiD-no-randomization + arithmetic + formal product all on one slide). New layout breaks each idea onto its own slide so each frame breathes. The section is now 17 slides where it was 6.

Reframed the bias as a product. The formal statement now writes $$\text{Bias}{\Delta X} ;=; (\beta{\text{post}} - \beta_{\text{pre}}) ;\times; \big(\mathbb{E}[X|D{=}1] - \mathbb{E}[X|D{=}0]\big)$$ under the time-varying-coefficient model $Y(0)_{it} = \alpha_i + \gamma_t + \beta_t X_i + \varepsilon$. Either factor at zero kills the bias.

Imbalance is the root. Two-bullet slide makes the point that randomization zeroes the imbalance term (RCT) and conditioning on $X$ (DR / IPW / OR) does the same job. The two mechanisms — time-varying $X$ with constant $\beta$, vs. time-invariant $X$ with time-varying $\beta_t$ — collapse to the same factorization.

Slides/02-covariates.tex — Title X clinics reveal added (7 slides)

New worked example designed to demonstrate the violation of changing $\beta_t$ on an imbalanced $X$, with an event-study reveal as the pedagogical climax:

# Slide Purpose
1 A worked example of the violation Names the three ingredients: positive effect, $\beta_t$ shift at year of treatment, imbalanced $X$
2 What we built in Truth $+1.8$, urban slope shifts from $\sim!-0.1$ to $-2.5$ at year 6, 70/30 urban imbalance
3 The DGP Year-by-year replace trend = ... Stata block per Scott's exact specification
4 Run a DiD event study Naive event-study figure — flat pre-trends, ATT $\approx -3.34$ (sign-flipped from $+1.8$)
5 We can recover this with the methods today Same event study conditioned on urban — flat pre-trends, ATT $\approx +1.75$
6 One extra variable Brief code comparison; preview that DR / IPW / OR are coming next
7 Pre-trends don't protect us from imbalance Lesson: pre-trends are necessary but not sufficient

Slides/scripts/title_x_eventstudy.R — new simulation script

R script that recreates Scott's exact DGP (year-by-year trend block: pre-period urban decline $-0.15$ to $-0.10$, rural $+0.10$; post-period urban $-0.5$ to $-2.5$, rural $+0.10$). Runs did::att_gt twice — once with no covariate, once conditioning on urban — and aggregates via did::aggte to event-study form. Plots both event studies with the Madrid palette (terracotta for the wrong answer, navy for the truth) and saves to Slides/figures/.

Estimates produced:

  • No covariate: ATT $\approx -3.34$. Pre-trend coefficients $-0.09$, $+0.09$, $-0.18$, $+0.01$ (all CIs cover zero). Post: $-0.5$, $-1.4$, $-3.1$, $-4.8$, $-7.0$.
  • With urban: ATT $\approx +1.75$. Pre-trends still near zero. Post: $+0.5$, $+1.3$, $+1.6$, $+2.4$, $+3.1$.

Figures generated

  • Slides/figures/titlex_es_naive.pdf — event study with no covariate, sign flips
  • Slides/figures/titlex_es_urban.pdf — event study conditioning on urban, truth restored

Other

  • Copied mit_deck.pdf (1.5 MB) from claudecode33/ into Slides/ as reference material.

Pedagogical decisions worked through

  • Imbalance is the root, $X$-trends are the channel. Originally framed the bias as "X-specific trends in $Y(0)$"; Scott pushed back that imbalance is the deeper engine because if treated and control had been equally urban, the urban-specific decline would wash out across groups. Reformulated the bias as the explicit product so either factor at zero kills it.
  • The Title X simulation has time-invariant $X$ ($\Delta X = 0$) — so the bias engine is the time-varying $\beta_t$, not $\Delta X$. Updated the formalization slide accordingly: bias = $(\beta_{\text{post}}-\beta_{\text{pre}})\cdot(\mathbb{E}[X|D{=}1]-\mathbb{E}[X|D{=}0])$ unifies both the time-varying-X and time-varying-$\beta$ cases.
  • DGP designed so pre-trends pass but post-period sign-flips. The pre-period urban-rural slope gap is small ($\sim!0.1$ vs $\sim!0.1$ in magnitude). The post-period gap grows fast ($-0.5$ to $-2.5$ vs $+0.1$). Pre-trend coefficients in the event study sit near zero with CIs covering zero; post-trend coefficients hockey-stick down.
  • Don't introduce csdid by name in this section. The estimators (DR, IPW, OR) and their Stata wrappers come later in the deck — these slides only show "a DiD event study" and "the methods today."

Commit landed

365d463 — "02-covariates: Title X clinics event-study reveal + add mit_deck reference" (9 files changed, 594 insertions, 266 deletions)

2026-05-25 — Madrid-2026 deck series migration

Foundational rebuild that converted the 2025 Glasgow-era materials into the Madrid-2026 deck series.

  • Codechella Madrid 2026: migrate from 2025 decks to Glasgow-2026 series — full deck inventory rebuilt, 2025 decks archived under Slides/archived_2025/
  • Madrid theme: factor into madrid_theme.sty, apply to all 7 decks — palette switched from Glasgow green to Madrid (terracotta #C44536, twilight navy #2C3E5C, peach #FBE5D0, gold #C9982A, cream #FAF6EE); \codechellatitle{Title}{Subtitle}{Author}{Venue} macro added with Gran Vía banner (codechella_2026.jpg)
  • README: switch banner to codechella_2026.jpg, fix deck list, add Registration row — schedule table updated
  • 01-basics: restructure Medicaid arc as building case + honest pre-trends — §10 reorganized with framework-emphasis slides (Bite / Falsification / Event-study+Main-result / Mechanism each get a "five pieces" recall with current piece highlighted); honest pre-trends caveat added per Rambachan-Roth 2024 sensitivity
  • Fix Day-N labels: workshop is 4 days not 5; drop morning/afternoon — corrected all "Day 2 morning/afternoon" framings; workshop runs Mon–Thu

Going forward: each working session adds a new dated entry on top.