Methods

The EuroMOMO model is based on certain principles and assumptions, which are the following:

General assumptions:

The primary purpose of the network's weekly monitoring is to detect signals of elevated excess mortality in a timely manner.
A key factor is that all countries in the network must be able to participate in the monitoring. Therefore, any improvement/change must not exclude any countries currently participating in the weekly monitoring of mortality.
A methodological shift should be epidemiologically or statistically justified.
The statistical estimation of expected mortality should be simple and robust.
The national EuroMOMO algorithm should be flexible and transparent, include version control and execute relatively quickly.

Model assumptions for estimating expected mortality:

Deaths follow a Poisson-distributed time series, influenced by trends and (sometimes) seasonal cycles
Winter and summer may introduce external mortality drivers (e.g., flu, heatwaves)
Spring and autumn are used to establish baselines, as they are less likely to have excess mortality

Modelling the Expected Mortality (Baseline)

The baseline for expected mortality is calculated using a generalized linear Poisson model, corrected for overdispersion. The model is fitted on up to five years of historical data, excluding:

Periods requiring delay correction
Data from the years 2020, 2021 and 2022 (to avoid bias due to disruption of usual mortality patterns)

Sample Selection for Baseline Fitting:

To reduce bias from excess mortality events (e.g., influenza or heatwaves), the baseline is fitted using data from:

Spring: Week 15–26
Autumn: Week 36–45

These periods are considered less likely to include excess mortality drivers and provide a stable foundation for modelling.

Model components:

The expected number of deaths is modelled with a trend, capturing demographic and structural changes, and an optional seasonal component. Over relatively short time spans, demographic and structural changes can be approximated by a linear trend, while seasonality is represented by a sine curve with yearly frequency. Model components are specified separately for each group under analysis, with a distinct model chosen for each group rather than a single model applied across all analyses.

Current monitored age groups:

Ages 0–14 years
Ages 15-44 years
Ages 45-64 years
Ages 65-74 years
Ages 75-84 years
Ages 65+ years
Ages 85+ years

Monitoring Requirements

The EuroMOMO algorithm must:

Estimate weekly observed, expected, and excess mortality (total and by age group)
Work across diverse national mortality patterns
Function reliably even with limited historical data or low death counts
Adjust for reporting delays
Produce standardized estimates to compare across subpopulations

Optional functionalities include:

Sex-specific and regional breakdowns
Cumulative excess mortality over selected periods
Retrospective performance testing for delay correction

Correction for delays in registration

The delay correction assumes that the number of deaths reported is proportional to the number of working days within a given registration period. Only valid historical data with known registration dates is used.

A binomial GLM is used to model the proportion of deaths already reported, considering weekends and holidays ("days off") and trends. The final number of deaths is then estimated using a Poisson GLM based on the registered deaths and calculated proportions. Corrections are adjusted based on the day of the week when aggregation occurs (e.g., Wednesday aggregations analyze historic data accordingly).

Validity: The model performs well with smooth and regular data transmission, even with long delays. However, irregular or batch reporting can reduce accuracy. Control graphs help determine the consistency of data flow and identify the periods needing delay correction.