Method
How do we know?
Every number on this site comes from one kind of model: a Poisson regression of daily emergency department visits on which temperature interval the day fell into, with thirty days of lags, fitted to the universe of Mexican public hospitals from 2008 to 2021. This page shows what the model holds fixed, and what happens to the answer when those choices are changed.
The design in five parts
- What is counted
- Emergency department visits per municipality, age group and day. Counts, not rates, with population entering as a weight.
- What is compared
- A day in one temperature interval against a day between 20 and 25 °C in the same municipality, same age group, same month of the same year.
- What is absorbed
- Municipality by age by year-month, and municipality by age by weekday. That removes any difference between places, any local seasonal pattern, and any day-of-week rhythm.
- What else is controlled
- Precipitation and its square, and relative humidity, each with the same thirty lags as temperature.
- How uncertainty is computed
- Standard errors clustered at the municipality level, so days within a municipality are allowed to be correlated with each other.
Change the fixed effects and see what moves
estimate, with the 95% interval beneath
| Group | ≤10 °C | 10–15 °C | 15–20 °C | 20–25 °C | 25–30 °C | >30 °C |
|---|---|---|---|---|---|---|
| 1 | −5.5%−7.0% to −4.1% | −3.8%−4.5% to −3.1% | −2.1%−2.5% to −1.6% | reference | +2.0%+1.5% to +2.5% | +3.0%+2.1% to +3.9% |
| 2 | −5.6%−6.9% to −4.3% | −3.7%−4.4% to −3.1% | −2.0%−2.4% to −1.6% | reference | +2.2%+1.7% to +2.6% | +3.2%+2.4% to +4.0% |
| 3 | −10.0%−11.1% to −8.8% | −5.9%−6.5% to −5.3% | −3.1%−3.5% to −2.7% | reference | +2.7%+2.2% to +3.2% | +4.1%+3.3% to +5.0% |
| 4 | −9.4%−10.6% to −8.3% | −5.6%−6.2% to −5.1% | −3.0%−3.3% to −2.6% | reference | +2.6%+2.1% to +3.2% | +3.8%+2.8% to +4.7% |
| 5 | −9.1%−10.0% to −8.1% | −5.4%−5.9% to −4.9% | −2.7%−3.0% to −2.3% | reference | +2.4%+2.0% to +2.8% | +3.7%+3.0% to +4.4% |
| 6 ★ | −9.1%−10.0% to −8.1% | −5.3%−5.8% to −4.9% | −2.7%−3.0% to −2.4% | reference | +2.5%+2.1% to +2.8% | +3.7%+3.1% to +4.4% |
| 7 | −4.9%−6.2% to −3.7% | −3.3%−3.9% to −2.8% | −1.9%−2.2% to −1.5% | reference | +1.9%+1.6% to +2.3% | +3.0%+2.3% to +3.7% |
estimate, with the 95% interval beneath
| Group | ≤10 °C | 10–15 °C | 15–20 °C | 20–25 °C | 25–30 °C | >30 °C |
|---|---|---|---|---|---|---|
| 1 | −2.6%−9.7% to +4.6% | −3.2%−6.2% to −0.2% | −3.3%−5.2% to −1.3% | reference | +1.6%−0.6% to +3.9% | 0.0%−3.3% to +3.2% |
| 2 | −2.6%−11.0% to +5.8% | −5.3%−9.7% to −0.8% | −3.9%−7.1% to −0.6% | reference | +1.3%−1.6% to +4.2% | −6.0%−10.5% to −1.5% |
| 3 | −22.3%−30.9% to −13.7% | −8.0%−12.8% to −3.1% | −4.3%−7.3% to −1.4% | reference | +0.8%−3.5% to +5.1% | −4.8%−11.3% to +1.7% |
| 4 | −6.1%−11.1% to −1.0% | −3.0%−5.7% to −0.3% | −3.8%−5.3% to −2.2% | reference | +4.2%+2.2% to +6.2% | +0.1%−2.8% to +3.0% |
| 5 | −16.0%−19.6% to −12.3% | −12.0%−14.1% to −9.8% | −5.2%−6.6% to −3.9% | reference | +4.4%+2.7% to +6.0% | +2.6%+0.3% to +4.9% |
| 6 ★ | −16.0%−19.7% to −12.4% | −11.8%−13.9% to −9.7% | −5.3%−6.6% to −3.9% | reference | +4.5%+2.8% to +6.1% | +2.7%+0.4% to +5.0% |
| 7 | +2.5%−2.6% to +7.6% | +0.1%−2.7% to +3.0% | −0.3%−2.0% to +1.4% | reference | +0.5%−1.3% to +2.2% | −3.2%−5.7% to −0.8% |
The seven specifications
| Fixed effects | National day dummies | Same day | 30 days | |
|---|---|---|---|---|
| 1 | agem + date | yes | −5.5%*** | −2.6% |
| 2 | agem^yday + agem^year + date | yes | −5.6%*** | −2.6% |
| 3 | agem^week + state^year + weekday | no | −10.0%*** | −22.3%*** |
| 4 | agem + state^year + yday + weekday | no | −9.4%*** | −6.1%** |
| 5 | agem^year^month + weekday | no | −9.1%*** | −16.0%*** |
| 6 ★ | agem^weekday + agem^year^month | no | −9.1%*** | −16.0%*** |
| 7 | agem^year^month + date | yes | −4.9%*** | +2.5% |
★ the paper's preferred specification
- The same-day effect survives everything. Cold reduces visits in all 7 specifications, at better than one percent significance in every one. The magnitude splits cleanly: 9.1% to 10.0% in the 4 specifications without national day fixed effects, 4.9% to 5.6% in the 3 that include them.
- The 30-day cumulative effect does not survive everything, and this is the honest limitation of the design. Without national day fixed effects it runs from 6.1% to 22.3% and is significant throughout. With them it is small, insignificant, and in one specification changes sign.
- There is a reason rather than an accident. A full set of national day dummies absorbs everything common to a given date across the country, and cold spells are largely synchronous national events. What is left identifies the contemporaneous response perfectly well, but strips out much of the across-day variation that a thirty-day accumulation needs. The paper's preferred specification, marked with a star, is one of those without day dummies, and gives 9.1% on the day and 16.0% cumulated.
Change the thermometer
estimate, with the 95% interval beneath
| Group | ≤10 °C | 10–15 °C | 15–20 °C | 20–25 °C | 25–30 °C | >30 °C |
|---|---|---|---|---|---|---|
| ERA5 (used) | −9.1%−10.0% to −8.1% | −5.3%−5.8% to −4.9% | −2.7%−3.0% to −2.4% | reference | +2.5%+2.1% to +2.8% | +3.7%+3.1% to +4.4% |
| CONAGUA, interpolated | −8.6%−9.7% to −7.6% | −4.7%−5.2% to −4.2% | −2.1%−2.4% to −1.9% | reference | +2.3%+2.0% to +2.7% | +3.3%+2.7% to +3.8% |
| CONAGUA, stations only | −8.7%−9.9% to −7.6% | −4.9%−5.4% to −4.3% | −2.2%−2.6% to −1.9% | reference | +2.3%+1.9% to +2.6% | +3.4%+2.8% to +4.0% |
estimate, with the 95% interval beneath
| Group | ≤10 °C | 10–15 °C | 15–20 °C | 20–25 °C | 25–30 °C | >30 °C |
|---|---|---|---|---|---|---|
| ERA5 (used) | −16.0%−19.7% to −12.4% | −11.8%−13.9% to −9.7% | −5.3%−6.6% to −3.9% | reference | +4.5%+2.8% to +6.1% | +2.7%+0.4% to +5.0% |
| CONAGUA, interpolated | −17.1%−21.7% to −12.5% | −11.7%−14.2% to −9.3% | −4.7%−6.5% to −3.0% | reference | +5.8%+4.1% to +7.5% | +5.8%+2.7% to +8.8% |
| CONAGUA, stations only | −18.2%−22.9% to −13.6% | −13.1%−15.6% to −10.5% | −5.7%−7.7% to −3.8% | reference | +5.8%+3.8% to +7.7% | +5.2%+2.1% to +8.4% |
The paper measures temperature from ERA5 reanalysis. Rebuilding the exposure from CONAGUA's physical weather stations instead, both interpolated to population-weighted municipal centroids and used raw, moves the cumulative cold effect from 16.0% to 17.1% and 18.2%. All three are significant at better than one percent. The finding does not depend on the choice of weather data.
Which visits disappear
estimate, with the 95% interval beneath
| Group | ≤10 °C | 10–15 °C | 15–20 °C | 20–25 °C | 25–30 °C | >30 °C |
|---|---|---|---|---|---|---|
| Discharged, not hospitalized | −10.2%−11.1% to −9.4% | −5.7%−6.2% to −5.2% | −2.8%−3.1% to −2.4% | reference | +2.6%+2.1% to +3.0% | +3.7%+3.1% to +4.4% |
| Hospitalization or death | −3.5%−6.2% to −0.8% | −3.6%−4.3% to −2.9% | −2.1%−2.7% to −1.6% | reference | +2.1%+1.7% to +2.6% | +3.4%+2.2% to +4.5% |
estimate, with the 95% interval beneath
| Group | ≤10 °C | 10–15 °C | 15–20 °C | 20–25 °C | 25–30 °C | >30 °C |
|---|---|---|---|---|---|---|
| Discharged, not hospitalized | −17.3%−21.4% to −13.1% | −12.2%−14.5% to −9.9% | −5.5%−7.0% to −4.0% | reference | +4.2%+2.6% to +5.9% | +1.7%−0.8% to +4.3% |
| Hospitalization or death | −8.2%−13.6% to −2.8% | −9.1%−12.3% to −5.9% | −4.0%−6.5% to −1.5% | reference | +4.5%+1.9% to +7.2% | +4.8%+0.3% to +9.4% |
Splitting by what happened to the patient is informative about mechanism. Over thirty days a cold day reduces visits that ended in discharge by 17.3%, and visits that ended in hospitalisation or death by 8.2%. Both fall, so this is not only people with minor complaints staying home, but the reduction is concentrated in the less severe cases, which is what a demand-side channel would predict.
What this design cannot tell you
- It cannot tell you where. Every estimate is national. The paper produces no municipality-level effect, so no local number can be read off it.
- It cannot tell you why. The mechanisms in the paper are suggestive: the design identifies that visits move, not which of behaviour, physiology or ecology moves them.
- It cannot follow people. The data are counts of visits, not patient records, so nothing here tracks whether the same person came back later or never came at all.
- It cannot see private care, or anyone who never presented. Only public hospitals report, and only 657 municipalities host one.
Everything behind this
The full replication package contains the code that produced every estimate on this site, and the archived data it runs on. Each figure here also links the exact numbers plotted.
github.com/FPavanello/tmp_er_admissionsdoi:10.5281/zenodo.21873500