TL;DR. Using completed cohort fertility (CCF) for ~30 countries, I link each birth cohort’s early-life life expectancy and income per person (log GDP)-averaged over childhood/adolescence-to that cohort’s completed fertility. Within the same country, cohorts that grew up safer (higher life expectancy) and richer (higher log GDP) tend to end up with lower completed fertility. Results are robust across exposure windows (0–14, 0–18, 0–25 years) and to a “placebo” test.
Subscribe free to get weekly insights that don’t make it into mainstream coverage. Paid members unlock deep dives + exclusive apps for exploring the data yourself.
Safer childhoods, slower life strategies. Life-history theory says that when survival prospects improve and returns to skill are high, species shift toward “slower” life strategies: later reproduction, fewer births, and more investment in each child (Oli, 2004). Using data on 30–34 countries, I built cohort-level exposure measures that average life expectancy and income per person over the first 14–25 years of life, and then followed each cohort’s fertility year by year from age 15 to 45. Within countries, cohorts who experienced higher early-life survival and higher incomes finished with lower completed fertility. Cross-country differences are small by comparison. In other words, when and where childhood conditions improved within a country, the cohorts that grew up in those years ultimately had fewer children (Whether this reflects later start, fewer births overall, or both requires age-specific data; I explore that in the premium post).
What’s different about my approach?
Most charts you see relate period TFR (a year-by-year rate) to GDP, education, or life expectancy in the same year across countries. That’s a useful snapshot, but it mixes countries, periods, and ages, and it’s hard to say whether today’s income is driving today’s fertility or just riding the same trend. Ourworldindata.org has a neat article titled “Why the total fertility rate doesn’t necessarily tell us the number of births women eventually have” which I recommend reading if you want to dig deeper.
I do something different. I follow birth cohorts and align them to the conditions they actually grew up in. For each cohort, I average log(GDP) per capita and life expectancy at birth over their early-life window (ages 0–18), and I use completed cohort fertility (CCF) as the outcome-how many children women from that cohort ultimately had. Then I estimate within–between mixed-effects models that separate:
Within-country shocks: did cohorts from the same country who grew up in unusually good times (relative to their own country’s long-run level) end up with different completed fertility?
Between-country levels: do countries that are, on average, richer/longer-lived differ in completed fertility from other countries?
To check I’m not just picking up trends, I also run placebo windows (averaging exposures before the cohort was even born, years -19 to -1) and compare effect sizes; and I repeat the analysis across multiple exposure windows (0–14, 0–18, 0–25) for robustness.
By moving from cross-sectional correlations to cohort-aligned, early-life exposures, I’m much closer to the developmental mechanisms we care about. The design asks a cleaner question: when childhood conditions improve within a country, do the cohorts who experienced those improvements ultimately have more or fewer children?
Why not ML feature hunts? Unlike semi-blind ML forecasts that optimize error over many features (see Lessons from Forecasting Fertility), my design tests an a priori life-history prediction with cohort-aligned exposures. For comparability, I also include a male–female education ratio covariate—motivated by prior findings by uncorrelated.xyz that this gap improves out-of-sample fertility forecasts and is positively associated with TFR in recent cross-country data. My main results are unchanged with or without this control.
What I measured
Outcome: Completed cohort fertility, CCF (children per woman) by birth cohort.
Exposures (cohort-level):
log(GDP) per capita (
pcGDPstored as ln GDP) and life expectancy at birth, averaged over childhood/adolescence.Exposure windows: 0–14, 0–18, 0–25 (years after birth).
Design: A within–between (Mundlak) linear mixed model with country random intercepts. I separate:
Within-country variation (cohorts from the same country growing up in different conditions).
Between-country differences (country means across cohorts).
Main result (0–18 window)
The 0–18 window uses 1,034 cohort-country observations across 30 countries. Standardizing predictors (1 SD = one unit).
Within-country log(GDP): -0.072 (***). When a cohort grows up in years that are richer than their own country’s long-run average (ages 0–18), its completed fertility is lower by ~0.07 children per woman per 1 SD. That’s the timing-sensitive signal: better early-life economic conditions within a country go with smaller families.
Between-country log(GDP): +0.071 (n.s.). Comparing country averages to each other, richer countries do not have reliably higher or lower cohort fertility once we account for the within signal; the cross-section is small and statistically indistinguishable from zero here.
Within-country life expectancy: -0.160 (***) Cohorts raised during periods when their country’s life expectancy is higher than its own average end up with ~0.16 fewer children per woman per 1 SD - a larger within effect than GDP. This matches the life-history intuition: higher expected longevity correlates with a shift toward lower fertility.
Between-country life expectancy: -0.112 (*). Countries with higher average life expectancy tend to have lower cohort fertility even in the cross-section, though this effect is smaller than the within signal.
Model fit: R²(marginal)=39.4% (fixed effects alone) and R²(conditional)=71.7% (adding country random intercepts). So the early-life GDP and life-expectancy terms explain about 40% of the variance in CCF; including stable country differences brings total explained variance to ~72%.
Takeaway: the strongest and most consistent associations are within countries over time-cohorts exposed to higher log(GDP) and especially higher life expectancy from birth to age 18 complete fertility at lower levels. Cross-country averages play a smaller role once those within-country dynamics are accounted for.
Figure: fixed-effect estimates (0–18 window)
Robustness to exposure window
I repeat the analysis with exposure windows 0–14 and 0–25 (rows and country counts below are from the script):
0–14: 1,013 rows, 30 countries
0–18: 1,034 rows, 30 countries
0–25: 1,073 rows, 31 countries
The within-country effects remain negative in sign for both life expectancy and log(GDP). Magnitudes shift modestly; the 0–25 window often sharpens the income effect (consistent with longer adolescence/education raising the opportunity cost of childbearing).
Figure: cross-window comparison
Additional results: Life expectancy specification with added covariates (0–18 window)
To complement the main life-expectancy model, I estimate a "life only" specification , adding national IQ (NIQ) and the male–female education ratio as covariates, while retaining the within–between structure (see variable definitions above). Coefficients are standardized (1 SD units).
Results mirror the headline pattern: cohorts exposed to higher early-life income and higher life expectancy within country complete with fewer births; between-country life expectancy is modestly negative. NIQ and absolute latitude are not distinguishable from zero. A higher male–female education ratio within country is associated with slightly lower CCF, but the effect is very small (-0.014) and barely significant (p= 0.035). The marginal R² for this expanded covariate set is 40.6%.
These within-country cohort results do not replicate the cross-sectional findings in Lessons from Forecasting Fertility—namely that (controlling for NIQ) a larger male–female education ratio is positively associated with TFR and helps forecasting—nor do they support a robust NIQ effect. uncorrelated.xyz
Placebo (“pre-birth exposure”) test
To check that I’m not just picking up slow-moving trends, I ran a placebo version of the design: instead of averaging log(GDP) and life expectancy over birth…birth+18, I averaged them over birth-19…birth-1 (i.e., before the cohort is even born). I then fit the same mixed model and compared the within-country effects from the main window to the placebo window using Wald tests inside a joint model.
Figure: side-by-side forest of the within-country coefficients for the main (0–18) window vs the placebo (-19…-1) window.
Results. The main 0–18 within-country effects are significantly more negative than the placebo effects (which are actually slightly (but not significantly) positive.
Because predictors are standardized within country, a +1 SD increase in a cohort’s log(GDP) during ages 0–18 is associated with a reduction in completed cohort fertility that is about 0.24 children per woman more negative than the corresponding pre-birth association. The same logic holds for life expectancy (about 0.17 more negative). This pattern supports the intended timing story-early-life conditions matter in the right window-rather than generic drift or country-specific trends driving the results.
Concluding thoughts
Spain and Italy are a handy real-world example. They top Europe for life expectancy and sit near the bottom for births per woman. That doesn’t prove cause and effect, but it fits the story here: when childhoods get safer and richer, people tend to have fewer kids. The same goes for South Korea and Japan. I could also try and replicate these results within China, as there is large fertility variation across Chinese provinces.
This design is a useful lens on cohort fertility, but it’s not the whole picture. I’m not claiming causal identification, only carefully aligned associations. Many other forces shape fertility that I don’t model here-policy changes (family benefits, childcare, parental leave), contraception and abortion access, education expansion, female labor markets, housing costs, marriage/union timing, norms and religion, migration, shocks (war, recession, epidemics), and country-specific histories. Data limitations also matter: early-life measures are noisy; country coverage is uneven; my exposure windows (0–14/18/25) and functional forms are choices; period autocorrelation and unobserved cohort trends can bias estimates; and standardized coefficients can obscure scale. The placebo test strengthens the timing story but isn’t definitive proof. I see this as a starting point: a transparent, cohort-aligned baseline that future work should extend with richer covariates, policy instruments, alternative window definitions, dynamic specifications, and more rigorous causal strategies.
References
Oli, M. K. (2004). The fast–slow continuum and mammalian life-history patterns: an empirical evaluation. Basic and Applied Ecology, 5(5), 449–463. https://doi.org/10.1016/j.baae.2004.06.002
The data used in this study were downloaded from ourworldindata.org