# F1-conf-02: Balanced-panel mean life expectancy, 2000–2022

# Question

How much did life expectancy at birth (WDI SP.DYN.LE00.IN) change between 2000 and 2022? The estimand is the simple unweighted mean across a fixed set of countries: those with a recorded value in every year from 2000 to 2022, each country counted once. As a descriptive baseline, the report also gives the naive 2022 mean over every country with a 2022 value.

# Data and exclusions

Input: data/life_expectancy_panel.csv, 439 data rows, 20 distinct iso3 codes, years 2000–2022. The file was assembled from two extractions. I found these problems:

1. Duplicate country (TUR). Türkiye appears twice, once as "Turkey" and once as "Turkiye", so there are 46 rows for 23 years. The values are identical in all 23 duplicated pairs. Countries are identified by iso3, so I kept one row per (iso3, year) and labelled the country with the later spelling, "Turkiye". This removed 23 rows.
2. Missing cell. Switzerland (CHE) has an empty value for 2007. It is treated as missing, with no imputation, so CHE is not in the balanced set. Its 2022 value is still used in the baseline.
3. Truncated series. Bangladesh (BGD), Ethiopia (ETH), India (IND) and Nigeria (NGA) only have rows for 2011–2022; there are no rows for 2000–2010. They are excluded from the balanced set but included in the 2022 baseline.

I screened the values for implausible entries (all fall between 51.5 and 84.6 years) and for large year-to-year jumps. The largest jumps are the 2020–2022 COVID-era drops and rebounds (for example IND 2021 −2.9 and 2022 +4.4, USA 2020 −1.8). These match known patterns, so I left them unchanged.

Exclusion ledger (out/exclusions.csv): 439 rows read → 439 after the year-window filter → 416 after removing duplicates (−23) → 415 after dropping the missing cell (−1) → 345 in the balanced panel (−70). The 345 rows are 15 countries × 23 years.

# Method

• Balanced set: the 15 countries with a non-missing value in all 23 years: AUS, AUT, BEL, CAN, DEU, ESP, FRA, GBR, ITA, JPN, KOR, NLD, SWE, TUR, USA.
• For each year, the unweighted mean over these 15 countries. Change = 2022 mean − 2000 mean.
• Baseline: the unweighted mean of 2022 values over all distinct iso3 codes with a 2022 value. That is 20 countries, with TUR counted once.
• No randomness was used. out/analysis.py regenerates every output deterministically.

# Results

In the balanced panel, mean life expectancy rose fairly steadily to a peak of 82.07 in 2019. It fell to 81.32 in 2020 and had only partly recovered by 2022 (81.55). By country, the 2000–2022 change ranges from +0.80 years (USA) to +6.77 years (KOR).

The naive 2022 mean (78.69) is about 2.9 years below the balanced 2022 mean. This is because it adds four lower-life-expectancy countries (BGD, ETH, IND, NGA) and CHE. It is a different population, so it should not be compared with the 2000 balanced mean.

Sensitivity checks (printed by analysis.py, not part of the estimand):

• If the TUR duplicates were not removed, the naive 2022 mean would be 78.6396.
• If CHE were kept despite its missing 2007 value, the means would be 78.1615 for 2000 and 81.6813 for 2022, a change of +3.5198.

# Uncertainty

These figures are exact descriptive statistics for a fixed set of countries, not estimates from a sample, so I report no sampling standard errors or confidence intervals. The only remaining uncertainty is measurement uncertainty in the WDI figures themselves, much of which are modelled estimates (UN WPP). That uncertainty is not quantified here, and the WDI may revise these values in later releases.

# Limitations and open issues

• The balanced set has 15 countries, mostly high-income OECD members plus Türkiye. The result does not describe the full 20-country set, and certainly not the world.
• The means are unweighted (each country counts equally), not population-weighted.
• I cannot tell from the file alone whether the missing CHE 2007 value and the missing 2000–2010 rows for BGD/ETH/IND/NGA are real gaps in the source or artefacts of the two-extraction assembly. The WDI normally does have values for these. Following the task rules, I treated them as missing and did not fill them from any other source.
• Choosing which Türkiye spelling to keep as the label does not affect any number, because the duplicated values are identical.