1# F1-conf-02 — Balanced-panel mean life expectancy, 2000–202223## Question4How much did life expectancy at birth (WDI `SP.DYN.LE00.IN`) change between 2000 and 2022? The estimand is the unweighted mean over a fixed, balanced set of countries, meaning countries with a recorded value in every year from 2000 to 2022. As a descriptive baseline, the report also gives the naive 2022 mean over all countries with a recorded 2022 value.56## Data and exclusions7Input: `data/life_expectancy_panel.csv`, with 439 data rows covering 20 distinct iso3 codes. Data problems found:891. **Duplicate Türkiye series.** All 23 TUR country-years appear twice: once labelled "Turkey" and once "Turkiye". This is an artefact of merging two extractions. The values in each pair are identical. Countries are identified by iso3, so one row was kept per (iso3, year), and the label was set to "Turkiye". 23 rows were removed.102. **Missing value for CHE 2007.** The `life_expectancy` cell is empty. It was treated as missing and not imputed, as the task rules require. Switzerland is therefore not in the balanced set. 1 row was removed at the missing-value step, and CHE's other 22 rows were removed at the balanced-panel step.113. **Truncated series for BGD, ETH, IND and NGA.** These countries have no rows at all for 2000–2010; they are present only for 2011–2022, probably because they came from the narrower extraction. They were excluded from the balanced panel (48 rows) but kept in the 2022 baseline.1213Exclusion ledger (`out/exclusions.csv`): 439 rows → dedupe 416 → drop missing 415 → year window 2000–2022 415 → balanced panel **345** (15 countries × 23 years).1415There are no other problems: no out-of-window years, no non-numeric values, and no implausible values. The largest year-on-year changes fall in 2020–2022 and are consistent with COVID-19.1617## Method18- Balanced set: the iso3 codes with a non-missing value in every year 2000–2022, after deduplication.19- For each year, the simple mean of `life_expectancy` over those 15 countries.20- Change = balanced mean for 2022 minus balanced mean for 2000.21- Baseline: the simple mean over all distinct iso3 codes with a recorded 2022 value (20 countries, each counted once after deduplication).22- No weighting, imputation or randomness was used. Running `out/analysis.py` reproduces every output.2324## Results25- Balanced set (15 countries): AUS, AUT, BEL, CAN, DEU, ESP, FRA, GBR, ITA, JPN, KOR, NLD, SWE, TUR, USA. Five countries are incomplete: BGD, CHE, ETH, IND, NGA.26- Balanced mean, 2000: **78.0603 years**.27- Balanced mean, 2022: **81.5531 years**.28- Change 2000–2022: **+3.4928 years**.29- Naive 2022 mean over all 20 countries: **78.692 years**. This is 2.86 years below the balanced 2022 mean, because it includes the low-life-expectancy countries (NGA, ETH, IND, BGD) that drop out of the balanced set. It should not be compared with the balanced 2000 mean.30- Change by country, 2000–2022: from +0.797 years (USA) to +6.770 years (KOR).3132## Uncertainty33These values are exact descriptive statistics for a fixed, purposively chosen set of countries, so there is no sampling uncertainty in the usual sense. The changes vary across countries with SD = 1.41 years. For illustration only: if the 15 countries were treated as a random sample, the standard error of the mean change would be 0.36 years. That figure has no formal inferential meaning here. The WDI estimates themselves are modelled values, and they get revised between releases; no uncertainty for them is available in the input.3435## Limitations and open issues36- The balanced set is almost entirely high-income countries; Türkiye is the only exception. The result therefore says nothing about the excluded lower-income countries (BGD, ETH, IND, NGA), nor about Switzerland, which was dropped for a single missing year.37- The CHE 2007 gap and the missing 2000–2010 rows for four countries are probably assembly errors and not true gaps in the WDI data. The task rules forbid looking up or imputing values, so these countries are excluded. If the original extraction were recovered, the balanced set (and the estimates) could change.38- The unweighted mean is not a population-weighted mean.39- 2022 values come from a single WDI release (updated 2026-10-08) and may be revised later.40