ClearTea Logo ClearTea

ClearTea,
the AI researcher

Free to start — no payment required

Compare PM2.5-attributable DALY rates between countries that adopted a national standard and those that never did, 1990–2025, allowing for staggered adoption, then write up the result in the manuscript.
204 countries and 7,344 country-years: 68 adopters and 136 never-adopters. I fit a staggered difference-in-differences model with event-study leads and lags.
Five years after adoption, adopters had an 11.8% lower age-standardised PM2.5-attributable DALY rate (95% UI 7.9–15.6) — an ATT of −108 per 100,000. I saved the ATT table and an event-study figure to the notebook.
Drafted the main-result paragraph in Results and tightened the Discussion. Both are tracked changes for your review.
Were adopters already on a different trend before they adopted?
No. The five pre-adoption coefficients are all within ±1.2 points of zero (joint p = 0.41), so adopters and never-adopters were on parallel trends before adoption. Want me to add the event-study figure to the manuscript as Figure 1?

Top scientists choose to

Research with ClearTea

CUHK
CAS
HKU
PKU
Tsinghua
Zhejiang
PUMC
CUHK
CAS
HKU
PKU
Tsinghua
Zhejiang
PUMC

Use cases

Use ClearTea to reliably do the heavy lifting — all with proof of work and decision-making.

Deep research review

  • – PubMed + OpenAlex indexed retrieval across millions of papers
  • – Reads the full paper during synthesis, not just abstracts
  • – Remembers your context to propose new research directions

Where should we start?

Or just type a message to get started

Draft with citations

Reproduce findings

  • – Hand the agent a paper plus its data
  • – Reproduces the result and explores variants
  • – Every step visible — code, output, decisions

Sandboxed notebook analysis

  • – Full .ipynb integration in a secure sandbox
  • – Iterative code execution from natural language
  • – Upload CSV, XLSX, SAS, SPSS, and more
Run all Restart Clear outputs
Python Ready
Markdown 1

PM2.5-attributable DALYs — adopters vs never-adopters, 1990–2025

Staggered difference-in-differences on the 204-country panel: 68 adopters of a national PM2.5 standard vs 136 never-adopters.

Cell 2 [1] 0.61s
1 import pandas as pd
2 from differences import ATTgt
3
4 panel = pd.read_csv("gbd_pm25_panel.csv")
5 panel = panel.merge(pd.read_csv("standards_adoption.csv"), on="location_id")
6 panel.shape
Output 
(7344, 14)
Cell 3 Modified [3] 0.48s
1 att_gt = ATTgt(data=panel.set_index(["location_id", "year"]), cohort_name="standard_year")
2 att_gt.fit(formula="np.log(dalys_asr) ~ log_gdp_pc + urban_pct",
3 control_group="never_treated", cluster_var="location_id")
− report_att(att_gt, event_time=5, scale=["pct"])
4 report_att(att_gt, event_time=5, scale=["pct", "per_100k"])
Output
estimandestimatelower 95%upper 95%p
ATT, +5 yrs (%)−11.8−15.6−7.9<0.001
ATT, +5 yrs (per 100k)−108−143−72<0.001
Cell 4 [4] 1.12s
1 es = att_gt.aggregate("event", min_e=-5, max_e=5)
2 ax = plot_event_study(es, ci=0.95)
3 ax.axhline(0, ls="--", c="grey")
4 ax.axvline(-0.5, ls=":", c="grey")
5 ax.set_xlabel("Years relative to adoption")
6 ax.set_ylabel("Change in DALY rate (%)")
Output
Event-study coefficients — flat before adoption, falling to about −12% five years after

Read with you

  • – Import PDFs and supporting docs into the workspace
  • – Inline summaries as you read
  • – On-demand translation in any major language

And more..

  • – Share files and workspaces with co-authors — and the agent
  • – Comments and tracked changes on every paragraph
  • – Export, file sharing, and original / drafted / accepted view modes

Collaborate like
coworkers

Learning, scratchpads, and team workflows that compound with your work.

Learning

ClearTea stores its learnings in CLEARTEA.doc

Every preference, hypothesis, and project detail is written to a context doc the agent reads before each task. You edit it like any other file — or let ClearTea do it automatically.

Where should we start?

Or just type a message to get started

Proof of work

Every investigation written down

Scratchpads capture searches, reasoning, and code as the agent works. Open the trail, audit each step, refine the prompt — research with a clear record of what the agent did and why.

Run all Restart Clear outputs
Python Ready
Markdown 1

Event-study pre-trends — PM2.5 standard adoption (leads −5…−1)

The headline result rests on a staggered difference-in-differences model that assumes adopters and never-adopters were on parallel DALY-rate trends before adoption. Before citing the result, verify that assumption — jointly, and within each adoption cohort.

Plan: (1) joint test on leads −5…−1, (2) the same test by adoption cohort, (3) pre-period event-study plot.

Cell 2 [1] 0.19s
1 # reuse the fitted staggered DiD model + panel from did_analysis.ipynb
2 es = att_gt.aggregate("dynamic", min_e=-5, max_e=5)
3 pre = es.loc[es.event_time < 0]
4 print(f"pre-period coefficients all within ±{pre.att.abs().max():.1f} pts, joint p = {es.pretrend_pvalue:.2f}")
5 print(f"cohorts: {n_adopters} adopters / {n_never} never-adopters n = {panel.shape[0]} country-years")
Output 
pre-period coefficients all within ±1.2 pts, joint p = 0.41
cohorts: 68 adopters / 136 never-adopters   n = 7344 country-years
Cell 3 [2] 0.07s
1 # break the pre-trends test down by adoption cohort
2 pre_by_cohort = att_gt.pretrend_test(by="cohort_bin")
3 pre_by_cohort[["max_abs_lead_pts", "p"]].round(2)
Output
adoption_cohortmax_abs_lead_ptsp
1995–20040.90.52
2005–20111.20.38
2012–20161.10.44
2017–20200.80.61
Cell 4 [3] 0.88s
1 # visual check: pre-period event-study coefficients with 95% UI
2 # flat around 0 before adoption == parallel pre-trends
3 ax = plot_event_study(pre, ci=0.95)
4 ax.axhline(0, ls="--", c="grey")
5 ax.set_title("Event-study pre-trends — leads −5…−1")
6 ax.set_xlabel("Years relative to adoption")
Output
Pre-period event-study coefficients, leads −5 to −1 — flat around zero with 95% UI whiskers
Markdown 5

Conclusion. The parallel-trends assumption holds. The pre-period coefficients for leads −5…−1 all sit within ±1.2 points of zero and the joint test is non-significant (p = 0.41); no adoption cohort shows a pre-trend on its own — 1995–2004 (0.52), 2005–2011 (0.38), 2012–2016 (0.44), 2017–2020 (0.61). The plotted leads scatter flat around zero with every 95% UI crossing it.

The headline 11.8% lower age-standardised DALY rate five years after adoption (95% UI 7.9–15.6; ATT −108 per 100,000) can be cited as estimated — no cohort-specific trends or re-weighting needed.

Follow-up. Pre-trends are flat overall; re-run this check within the pre-specified super-region subgroups before finalizing the subgroup figure.

Teamwork

Works alongside your collaborators

Drop a comment on a paragraph and the agent responds, edits, and resolves it like a co-author. Same review workflow your team already runs — humans and ClearTea share the workspace.

Frequently Asked Questions

Also available

Just need a certificate of editing?

For ¥699/month, upload your manuscript and get an official editing certificate in about 10 minutes — no workspace needed.

Get an editing certificate
See our pricing here

Research better with
ClearTea

Free to start — no payment required