Isolated Countries, Not Isolated Years: Social Contact, Loneliness and Mortality in Fifteen European Countries, 2002–2019

Social isolation predicts mortality across fifteen European countries and not across the years within them. Using nine rounds of the European Social Survey aggregated to country-years and matched to Human Mortality Database age-standardised death rates, the between-country coefficient on log mortality is 0.0104 and the within-country coefficient is 0.0016, with a wild cluster bootstrap p-value of 0.125 against a minimum detectable effect one quarter the cross-sectional size. The cross-sectional association is confined to one measure of structural isolation: neither the share lacking a confidant nor the share living alone tracks mortality between countries, and living alone carries two and a half times more within-country variation while showing nothing at either level. CONTENTS. The derived country-year panel, every intermediate series, the Stata do-files for all estimation and robustness work, the Python script that rebuilds the panel from source, the script that produces the five figures, and the figures themselves. NOT INCLUDED. The raw European Social Survey and Human Mortality Database files. Both licences prohibit redistribution and HMD asks that users be referred to the site rather than passed copies, so that nobody works from a stale version. Both are free after registration, and build_panel.py regenerates everything from them together with a World Bank World Development Indicators extract. RUN ORDER. build_panel.py, then the do-files, then make_figures.py. Set the working directory manually; no do-file contains a cd command. Requires Python 3.10 or later with pandas, numpy and pyreadstat, and Stata 17 or later with boottest, xtcd2, xtscc, estout and coefplot. FOUR THINGS A REPLICATOR SHOULD KNOW. The ESS confidant question changes definition at round 6, from a binary to a count, with no overlap round, so the two are never pooled. France in round 6 is excluded from that measure because it records no respondents without a confidant out of 1,968 valid answers while every other cell carries a mass at zero. Rounds 10 and 11 are excluded because Human Mortality Database life tables end before the relevant year for three and six of the fifteen countries respectively. Mundlak group means must be rebuilt on whatever subsample is being estimated, since carrying full-sample means into a subsample breaks the decomposition. DATA SOURCES. European Social Survey ERIC, integrated files, rounds 1 to 9, distributed by Sikt. Human Mortality Database, Max Planck Institute for Demographic Research, University of California Berkeley, and French Institute for Demographic Studies. World Bank, World Development Indicators.