How baselines, confidence intervals and excess mortality are computed.
Weekly all-cause mortality counts and population estimates are drawn from Eurostat, Human Mortality Database (HMD) and from National statistical offices. Counts are converted to age-standardised mortality rates (ASMR) per 100,000 population using the 2013 European Standard Population (ESP13) with 5-year age bins (90+) and weekly interpolated populations.
A number of country-specific adjustments apply: Israel uses the 0–19 age group only; Germany uses 10-year bins below age 40 (20–44 observations); All HMD countries (Excluding Israel- None European) are given as 0-14, 15-64, 65-74, 75-84 & 85+. This limits their main explorer processing to 65+ and all ages observations and may cause some distortion to their ASMR. Sweden includes deaths recorded with an unknown week of death (W99). All 2025 data are preliminary and subject to revision. Some data for 2024 or even 2023 should be considered preliminary.
ASMR is not a death count. The age-standardised mortality rate is a derived measure that adjusts for differences in age structure across populations and over time. It is the appropriate metric for comparing trends across countries and age groups, but it does not directly represent the number of deaths. Users seeking absolute death counts should consult the source databases listed in the Sources section.
Baseline exceedance is not confirmed excess mortality. A week or year falling above the upper confidence interval indicates that observed mortality was statistically unusual relative to the pre-pandemic trend — it does not identify a cause, confirm that deaths were preventable, or constitute a formal excess mortality estimate. Formal excess mortality analysis requires additional modelling steps, validation against multiple baseline specifications, and expert epidemiological interpretation.
Baselines are extrapolations. All baselines project a pre-pandemic trend forward into years not used for fitting. Structural changes in mortality trends — demographic shifts, healthcare improvements, long-term behavioural changes — that occurred after the training period will not be captured. The further the projection extends beyond the training window, the greater the uncertainty.
Data quality varies across countries. Timeliness, completeness, and coding practices differ substantially. Some countries have reporting lags of several weeks; 2025 data and in some cases 2024 data should be considered preliminary. The tool notes where data are incomplete, but users should apply additional caution when making cross-country comparisons, particularly for smaller countries with higher week-to-week statistical noise.
Suppressed baselines are not errors. For certain country-age combinations — primarily small populations with very low death counts in younger age groups — the baseline and CI are suppressed because the signal-to-noise ratio is too low for a meaningful result. This is a deliberate quality gate, not a data gap.
The explorer offers three levels of temporal aggregation, each suited to a different question.
Weekly ASMR values are summed across the calendar year (W01–W52/W53). Years with 53 ISO weeks are scaled by 52/53 so that all annual sums are comparable on a 52-week basis.
Seasons run from W26 of one year to W25 of the next (mid-summer to mid-summer), keeping winter mortality — which spans the calendar year boundary — within a single observation. 53-week seasons are similarly scaled by 52/53.
Individual weekly ASMR values plotted directly, with the baseline and confidence interval described below overlaid.
The baseline represents the mortality level that would have been expected in the absence of any unusual event, based on the pre-pandemic historical trend. All baselines are intentionally anchored to the pre-pandemic period so that post-2019 years can be evaluated against a stable counterfactual.
For each country and age group, annual (or seasonal) ASMR sums are computed for every year in the chosen baseline window. A simple ordinary least squares linear regression is then fit:
The fitted line is projected forward to produce the expected level in any given year. The user can select from several pre-defined baseline windows (e.g. 2011–19, 2012–19) to assess sensitivity to the choice of training period. A Baseline Type toggle switches between three modes:
Trend — the OLS regression line projected forward. Accounts for the long-run decline in age-standardised mortality, so post-2019 years are judged against where the trend would have been, not where it was during the baseline period. This is the primary mode and the appropriate framing for evaluating recent years.
Mean — a flat horizontal line at the average of the baseline-period annual (or seasonal, or weekly) values. Simpler and more transparent; useful as a cross-check and for audiences unfamiliar with trend adjustment. Because it ignores the largely declining trend, the mean baseline will likely appear too high for years well beyond the training window and too low for years preceding it.
Log-linear — fits the regression on the natural logarithm of ASMR rather than the raw value, equivalent to modelling a constant percentage decline per year rather than a constant absolute decline. The projected baseline curves gently (it is convex rather than straight) and cannot cross zero, which is physically correct. Confidence intervals are computed in log space and back-transformed via exponentiation, giving a slightly asymmetric band that is wider above the baseline than below — again, physically appropriate.
A naive per-week OLS regression on weekly ASMR is noisy because it treats each week in isolation and ignores the strong seasonal structure of mortality. Instead, the weekly baseline follows a simplified EuroMOMO-style approach: decompose the signal into a seasonal shape and an annual trend level, fit the trend on stable shoulder weeks only, and apply a Newey-West autocorrelation-corrected standard error for the confidence interval.
For each week-of-year (W01–W52), the mean ASMR across 2015–2019 is computed. These five years form a stable, pre-pandemic reference period. Together the 52 weekly means constitute the seasonal shape — a profile of how mortality typically distributes across the year.
For each calendar year, the total annual mortality level is estimated from shoulder weeks 15–26 and 35–46 only. These mid-spring and early-autumn periods are epidemiologically stable: the winter flu peak has cleared, the summer trough has not yet begun. Using only shoulder weeks insulates the annual level estimate from years with unusually severe flu seasons or summer isolated heat events, such as they are.
Rather than always fitting a linear trend, the tool selects the best-fitting model for each country-age-baseline combination using the Bayesian Information Criterion (BIC). Three candidate models are evaluated against the shoulder-week residuals:
BIC penalises model complexity, so the linear or log-linear trend only wins when the data provide clear evidence of a trend over the baseline window — otherwise a flat mean is preferred. This ensures model choice is driven by the data rather than manual intervention. Note that slope clamping is still used as a suppression trigger — if the BIC-winning model would require its slope to be clamped to remain physically plausible when extrapolated, the entire slice is suppressed rather than displayed in distorted form (see Suppression below).
BIC selection is performed once per slice — where a slice is a unique combination of country, age group, and baseline window. The winning model is then applied uniformly across every week and every year in that slice. There is no per-week or per-year model switching. For example, if BIC selects log-linear for Germany all ages with the 2015–19 window, the Trend tab shows a log-linear projection for every week from the chart start through to 2025.
The Mean and Log-linear tabs always show those specific models regardless of BIC, allowing direct comparison with the data-selected Trend baseline.
When the BIC-winning model's slope is severe enough to project the baseline to implausible levels outside the training window — detected when extrapolation would require clamping the slope — the baseline and CI for that slice are suppressed entirely rather than shown in a misleading form.
The predicted weekly baseline for any year and week is the BIC-selected model's projected annual level, scaled by the week's share of the seasonal shape:
The shape of the expected seasonal curve is always drawn from the stable 2015–19 reference period, while its level follows whichever trend model BIC selected for that slice.
The 95% confidence interval is a standard regression prediction interval:
The leverage term (year − year̄)²/Sxx grows as a year moves further from the centre of the training window. This correctly reflects that extrapolating a regression line further into the future carries genuinely greater uncertainty — the confidence band widens with each post-baseline year. With 8–9 training years, t(n−2) ≈ 2.365, somewhat wider than the large-sample 1.96, appropriately penalising the small sample.
The weekly CI is a flat-width band — constant across all 52 weeks and all years — derived from pooled shoulder-week residuals using a Newey-West heteroskedasticity and autocorrelation consistent (HAC) standard error.
The SE is estimated exclusively from residuals in the stable shoulder weeks (W15–26 and W35–46) — the same weeks used to estimate the annual level in Step 2. This mirrors the EuroMOMO approach: shoulder weeks are the least volatile part of the year, so their residuals give the cleanest measure of model fit. The resulting CI is intentionally calibrated to stable-season variation. Weeks outside this range — including winter flu peaks — that exceed the upper bound are flagged as genuine excess, not absorbed by a wider seasonal envelope.
Weekly residuals within a year are serially autocorrelated — an unusually high week tends to be followed by another high week. Ignoring this would overstate the effective sample size and produce a CI that is too narrow. The Newey-West estimator corrects for this by computing a weighted sum of lagged autocovariances using the Bartlett kernel:
The resulting SE is larger than the naive pooled SE by a factor of approximately 1.07–1.55× depending on the autocorrelation structure of the slice. The large-sample critical value t(∞) = 1.96 is used; with N ≈ 120 the difference from t(118) is negligible.
The CI is fixed-width — neither a leverage penalty nor a per-week scaling factor is applied. The leverage-inflated prediction interval used in the annual and seasonal views is appropriate there because each data point is a single annual aggregate and extrapolation uncertainty genuinely grows over time. For the weekly view, where the SE is estimated from pooled cross-sectional residuals at large N, adding leverage would double-count the already-embedded extrapolation uncertainty and produce bands that balloon without justification as the chart is scrolled further from the training window.
For the log-linear baseline tab, residuals are computed in log space (log(observed) − log(predicted)) before applying the Newey-West estimator. This gives a dimensionless SE that is applied symmetrically in log space:
A minimum half-width of 5% of the mean shoulder ASMR is applied. This prevents the CI from becoming visually imperceptible for very large, stable populations (notably the United States all-ages series) where the NW SE is genuinely very small.
The weekly baseline and CI are suppressed for a given country-age slice when the signal-to-noise ratio is too low to produce a meaningful result. Two conditions trigger suppression:
High CV. If the Newey-West half-width exceeds 65% of the mean shoulder ASMR (coefficient of variation > 0.65), the CI is wider than the signal itself and cannot usefully distinguish excess from noise. This affects all age groups for Liechtenstein, the 15–44 and 45–64 groups for Iceland and Malta, and the 15–44 group for Cyprus and Luxembourg.
Slope clamping triggered. BIC may correctly select a linear or log-linear trend within the training window that, when extrapolated, projects the baseline to implausible levels outside it — for example to near-zero values for a steeply declining small-population series. When the fitted slope would require clamping to remain within physically plausible bounds, the entire slice is suppressed rather than rendered with a distorted baseline. This currently affects Estonia 15–44.
In both cases the baseline line is also suppressed, not just the CI bands — a misleading trend line is more harmful than no trend line.
Six baseline windows are provided. Each trains the regression on a different set of years, allowing the user to assess how sensitive conclusions are to the choice of training period.
| Label | Training years (annual) | Training seasons | n |
|---|---|---|---|
| 2009–19 | 2009–2019 | 2009/10–2018/19 | 11 / 10 |
| 2010–19 | 2010–2019 | 2010/11–2018/19 | 10 / 9 |
| 2011–19 | 2011–2019 | 2011/12–2018/19 | 9 / 8 |
| 2012–19 | 2012–2019 | 2012/13–2018/19 | 8 / 7 |
| 2013–19 | 2013–2019 | 2013/14–2018/19 | 7 / 6 |
| 2015–19 | 2015–2019 | 2015/16–2018/19 | 5 / 4† |
† The 2015–19 baseline produces only 4 seasonal training points (df = 2, t ≈ 4.3), so seasonal confidence intervals for this window are very wide. Baselines with fewer than 5 training years are suppressed in the weekly (& annual) view.
The Causes of Death Explorer is a companion page that displays annual age-standardised mortality rates broken down by ICD-10 cause chapter for 28 countries from 2013 to 2023. Unlike the weekly all-cause data, these are calculated annually with static, mid-year or average populations.
Users may select any combination of countries, any age group, and one or more causes of death simultaneously. When multiple causes are selected, their rates are summed for each country and year to produce a single composite series.
In Separate mode, users can instead view each selected cause as its own independent series. A single country is shown at a time, with each cause plotted as a distinct line coloured to match its selection chip — making it straightforward to compare the trajectory and magnitude of individual causes, and to see which contribute most to any period of excess mortality. The baseline and confidence interval are computed independently for each cause series using the same OLS trend method, and are suppressed per-cause where the training data are zero or incomplete.
The causes page applies the same annual OLS trend baseline described above — a linear regression of the annual ASMR sum against year, fit over the chosen training window (2013–19, 2014–19, or 2015–19), with the same full prediction interval formula including the leverage penalty:
Here the regression input is already an annual rate, so no weekly summation or 52-week normalisation is required. The projected baseline and widening CI band behave identically to the Annual view on the main explorer.
The baseline and CI are suppressed for a given country when the summed ASMR across selected causes is zero or missing for any year within the training window. For example, selecting only (U071) Covid-19 identified produces a zero sum for every pre-2020 training year, so no trend can be meaningfully fitted. Adding a second cause whose values are non-zero throughout the training period — such as (V01–Y89) External — raises the summed total above zero for all training years, restoring the baseline. This ensures the trend is never fit on a series that is structurally zero in the training window, which would produce a spurious projection.
The Sub-National Life Expectancy Explorer displays annual life expectancy losses or gains in months at NUTS2 regional level for up to 252 regions across 33 European countries, sourced from Eurostat. Three measures are available — life expectancy at birth, at age 65, and at age 80 — broken down by sex.
The displayed value for each region is not the raw life expectancy but the deviation from the expected trend in months: a positive number means life expectancy exceeded what the pre-pandemic trend would have predicted (a gain); a negative number means it fell short (a loss). Hover over a region to see actual LE value and its deviation from the selected trend.
The method is identical to the annual trend baseline used on the Mortality Explorer and the Causes of Death page. For each region, sex, and age group, an OLS linear regression is fit on the observed life expectancy values across the selected training window (2013–19, 2014–19, or 2015–19), and the fitted line is projected forward using the same full prediction interval formula:
The excess in years is converted to months by multiplying by 12. The confidence interval on the tooltip likewise reflects the range of plausible excess values given the uncertainty in the projected baseline.
A region's baseline is suppressed if any year within the selected training window is missing. For the 2013–19 window, a small number of regions (predominantly in Hungary and Poland) lack 2013 data and are therefore shown without a baseline for that window only. Switching to the 2014–19 or 2015–19 window restores their baselines, as all regions have complete data from 2014 onward.
Eurostat — Deaths by week, sex and 5-year age group:
demo_r_mwk_05
Eurostat — Deaths by week, sex and 10-year age group:
demo_r_mwk_10
Eurostat — Population on 1 January by age and sex:
demo_pjan
Eurostat — Methodology and quality report for weekly death statistics:
demomwk_esms
Israel Central Bureau of Statistics:
cbs.gov.il
Eurostat — Causes of death by age and sex (annual):
hlth_cd_aro
US CDC WONDER — Multiple Cause of Death:
wonder.cdc.gov/mcd
US Census 2010-19:
Census 2010-19
US Census 2000-24:
Census 2020-24
ONS — Deaths by sex, 5-year age group and underlying cause, England and Wales, 2010–2021:
ONS ad hoc — 2010 to 2021
ONS — Deaths by sex, 5-year age group and underlying cause, England and Wales, 2022:
ONS ad hoc — 2022
ONS — Analysis of population estimates tool for UK:
ONS ad hoc — 2022
Te Whatu Ora (Health New Zealand) — Mortality Data Web Tool:
tewhatuora.govt.nz
Stat NZ Infoshare:
Infoshare
Eurostat — Life expectancy by age, sex and NUTS2 region:
demo_r_mlifexp