Income enrichment of a synthetic population of France
Application to social leasing for electric vehicles

Sebastian Hörl

19 November 2027

AFET Conference 2026

Context

  • A synthetic population digitally represents the population of a territory
     
  • For France, eqasim provides a well-tested framework to generate such data sets



     
  • They allow for a detailed sociodemographic and geographic analysis of patterns, policies and are used in agent-based simulation
     
  • Goal: evaluate spatially and socially eligibility for the 2026 social leasing campaign in France.
     
  • Problem: correctly impute household income information to the synthetic population

0:00 - 9:00

10:00 - 17:30

17:45 - 21:00

22:00 - 0:00

x 60,000,000

A comment on Filosofi

  • The Filosofi dataset provided income distributions by household type and / or geographic zone - an integration into eqasim exists
     
  • Decommissioned in 2023, so we need to find a new method!
     
  • Disclaimer: Filosofi 2 exists now, didn't look at it in detail (but only median values seem to be given

Two alternative data sources

ERFS (2023)

IRCOM (2024)

  • Provides the Standard of Living S of households in Metropolitan France
     
  • Link to socio-demographic attributes of a household
    • By household type (single M, single F, couple, ...)
    • By age group of the reference person* (<30, 30 - 39, ...)
    • By socio-professional group (CSP)
       
  • For each uni-dimensional modality (30 - 39, couple, ...) the percentage of households is given (on ten unaligned income brackets)
     
  • Households refer to people living in the same housing unit
  • Provides statistics on collected income tax, including the fiscal reference revenue R of households (révenue fiscal de réference; RFR)
     
  • Link to geogaphical zones
    • By department
    • By municipality (for some with sufficient inhabitants)
       
  • Number of households within a given bracket of R
  • Brackets globally aligned
     
  • Households refer to tax households (foyer fiscal)

Two alternative data sources

ERFS (2023)

IRCOM (2024)

Two alternative data sources

ERFS (2023)

IRCOM (2024)

S     =      T    /     U

Standard of living

Disposable household income

Consumption units

After taxes, social benefits, ...

1 for first adult + 0.5 for additional adult + 0.3 per child (<14)

R     =      f(H) = ?

Reference fiscal income

No direct correspondance to S or H!

Basis for income tax calulation

Census data as a basis

Census (2021)

  • Provides weighted household observations across France with a sampling rate of about 30%
     
  • Already integrates a flag for the reference person of a household
     
  • Allows us to count household observations
    • with reference person in age group A
    • with reference person in CSP C
    • with household type H
    • inside zone Z
      • either Z = M for municipalities
      • or Z = D for deaprtments
N(A, C, H, Z)

Census tensor

Census data as a basis

Census (2021)

  • Provides weighted household observations across France with a sampling rate of about 30%
     
  • Already integrates a flag for the reference person of a household
     
  • Allows us to count household observations
    • with reference person in age group A
    • with reference person in CSP C
    • with household type H
    • inside zone Z
      • either Z = M for municipalities
      • or Z = D for deaprtments
    • with consumption units U
      • Need to create finite number of brackets
N(A, C, H, U, Z)

Census tensor

Standard of living model

Census tensor

N(A, C, H) = \sum_{U, Z} N(A, C, H, U, Z)

Standard of living model

Census tensor

N(A, S), N(C, S), N(H, S)

Standard of living tensors

  • Agree on aligned S brackets over all distributions
     
  • Resample the provided ERFS distributions
     
  • Then count number of househods in each bracket for given modality
N(A, C, H) = \sum_{U, Z} N(A, C, H, U, Z)

Resampled ERFS

Standard of living model

N(A, C, H) = \sum_{U, Z} N(A, C, H, U, Z)

Census tensor

N(A, S), N(C, S), N(H, S)

Standard of living tensors

M_S(A, C, H, S)

Standard of living model

\sum_{S} M_S(A, C, H, S) = N(A, C, H)
\sum_{C, H} M_S(A, C, H, S) = N(A, S)
\sum_{A, H} M_S(A, C, H, S) = N(C, S)
\sum_{A, C} M_S(A, C, H, S) = N(H, S)

s.t.

  • Construct a new tensor M that fits
    • Census marginal tensor
    • ERFS marginals
       
  • Can be solved using Iterative Proportional Fitting!

Standard of living model

M_S(A, C, H, S)

Standard of living model

Standard of living model

M_S(A, C, H, S)

Standard of living model

  • Obtain (A, C, H) for each household in the synthetic population
  • Sample from P(S | A, C, H)

Spatialized model

N(A, C, H, Z)

Census tensor

N(A, S), N(C, S), N(H, S)

Standard of living tensors

N(Z, R)

Geographical tensor

M_Z(A, C, H, U, Z, S, R)

Spatialized model

N(A, C, H, Z)

Census tensor

N(A, S), N(C, S), N(H, S)

Standard of living tensors

N(Z, R)

Geographical tensor

M_Z(A, C, H, U, Z, S, R)

Problem 1

S is normalized by consumption units, R is absolute for a household

Problem 2

There is no link between R and S in our model

Spatialized model

N(A, C, H, U, Z)

Census tensor

N(A, S), N(C, S), N(H, S)

Standard of living tensors

N(Z, R)

Geographical tensor

M_Z(A, C, H, U, Z, S, R)

Problem 1

S is normalized by consumption units, R is absolute for a household

Problem 2

There is no link between R and S in our model

N(S, U, R)

Linking tensor

Spatialized model

Establishing a distribution of T

  • Sample N = Large observations of households


     
  • Sample standard of living for each obs.


     
  • Calculate disposable income T for each obs.


     
  • This gives us an empirical approximation of the distribution P(T) in France
(a, c, h, u)_i \sim N(A, C, H, U)
t_i = s_i \cdot u_i
s_i \sim M_S(S\ |\ A = a_i, C = c_i, H = h_i)
F_T(t) = P(T \leq t)

Constructed disposable income

Spatialized model

Establishing a distribution of T

  • Sample N = Large observations of households


     
  • Sample standard of living for each obs.


     
  • Calculate disposable income T for each obs.


     
  • This gives us an empirical approximation of the distribution P(T) in France
(a, c, h, u)_i \sim N(A, C, H, U)
t_i = s_i \cdot u_i
s_i \sim M_S(S\ |\ A = a_i, C = c_i, H = h_i)

Piecewise-linear approximation of F(R)

  • Given the brackets of R and France-wide household counts, we can construct a piecewise-linear distribution F(R)


     
  • Major assumption: A household that is at the kth quantile of F(R) is roughly at the kth quantile of F(T).


     
F_R(r) = P(R \leq r)
F_T(t) = F_R(r)
F_T(t) = P(T \leq t)

Spatialized model

Establishing a distribution of T

  • Sample N = Large observations of households


     
  • Sample standard of living for each obs.


     
  • Calculate disposable income T for each obs.


     
  • This gives us an empirical approximation of the distribution P(T) in France
(a, c, h, u)_i \sim N(A, C, H, U)
t_i = s_i \cdot u_i
s_i \sim M_S(S\ |\ A = a_i, C = c_i, H = h_i)

Piecewise-linear approximation of F(R)

  • Given the brackets of R and France-wide household counts, we can construct a piecewise-linear distribution F(R)


     
  • Major assumption: A household that is at the kth quantile of F(R) is roughly at the kth quantile of F(T).


     
  • We locate each t on the distribution of R


     
  • Going back to categorical S and R: count how many observations i fall in each combination of (U, S, R):
F_R(r) = P(R \leq r)
F_T(t) = F_R(r)
r_i = F_R(F_T^{-1}(t_i))
N(S, U, R)
F_T(t) = P(T \leq t)

Spatialized model

Distribution matching

Linking tensor

Distribution matching

Linking tensor

(consumption units marginalized out)

Spatialized model

M_Z(A, C, H, U, Z, S, R)

Geographical model

\sum_{S, R} M_Z = N(A, C, H, U, Z)
\sum_{C, H, U, Z, R} M_Z = N(A, S)

s.t.

\sum_{A, H, U, Z, R} M_Z = N(C, S)
\sum_{A, C, U, Z, R} M_Z = N(H, S)
\sum_{A, C, H, Z} M_Z = N(S, U, R)
\sum_{A, C, H, U, S} M_Z = N(Z, R)
  • Department version: all departments have information
     
  • Municipality version: substract known municipality (M, R) counts from (D, R) to obtain "unknown" municipality counts, treat them as one group

Spatialized model

M_M(A, C, H, U, M, S, R)
  • Department version: all departments have information
     
  • Municipality version: substract known municipality (M, R) counts from (D, R) to obtain "unknown" municipality counts, treat them as one group
M_D(A, C, H, U, D, S, R)

6 x 7 x 6 x 81 x 96 x 11 x 8 = about 172M cells

6 x 7 x 6 x 81 x 1,944 x 11 x 8 = about 3.5B cells

Running on ~32 GB memory applying sparse matrix calculation

~4h

~24h

Spatialized model

Convergence

Fitting the geographical model

Spatialized model

Some first validation results

  • Percentage of mismatched observations for each marginal, relative to the stratum size


 

Spatialized model

M_S'(A, C, H, U, Z, S) = \sum_R M_Z(A, C, H, U, Z, S, R)

Application to a synthetic population

  • Create a spatially biased model of S:



     
  • For each household (A, C, H, U, Z) sample



     
  • Calculate disposable income, then r:




     
s_i \sim M_S'(S\ |\ A, C, H, U, Z)
t_i = s_i \cdot u_i

(here, S is already spatialized)

r_i = F_R(F_T^{-1}(t_i))

Spatialized model

M_S'(A, C, H, U, Z, S) = \sum_R M_Z(A, C, H, U, Z, S, R)

Application to a synthetic population

  • Create a spatially biased model of S:



     
  • For each household (A, C, H, U, Z) sample



     
  • Calculate disposable income, then r:




     
s_i \sim M_S'(S\ |\ A, C, H, U, Z)
t_i = s_i \cdot u_i

(here, S is already spatialized)

r_i = F_R(F_T^{-1}(t_i))

Useful: Fiscal reference income per part

  • One part for each adult
  • 0.5 parts for the first and second child
  • One part for third and more child

Application case

Social leasing eligibility

  • Commuting distance > 10km
    • Obtained from synthetic population's activity chains
    • Simplification (here): Euclidean distance!
    • Add a distance eligibility to each person
       
  • Fiscal reference income per part <= 16,880 EUR
    • Impute income on the synthetic population
    • Calculate the RFR par part for each household
    • Add an income eligibility flag for each household 
       
  • A person is eligible if they are both distance and income eligible!
     
  • We focus on working persons with at least one car in the household!

Application case

Income eligibility

Distance eligibility

Combined

Eligibility [%]

Spatial eligibility analysis

Application case

Social eligibility

Example: Social eligibility analysis

Application case

Focus on Île-de-France

Income eligibility

Distance eligibility

Combined

Eligibility [%]

Application case

Focus on Île-de-France

Eligibility [%]

Eligibility [%]

Next steps

Validate our main assumptions

  • Quantile-matching of disposable income and fiscal reference income
  • Quantile matching between ménages in the census sense and foyers fiscal in tax decalaration sense

 

Validation with reference data

  • Filosofi 2? Household travel surveys?

 

Technical

  • Ensemble runs to test sensitivity over multiple random seeds
  • More benchmarking for RAM and computation

 

Social leasing use case

  • Routed distance all across France (osmr)
     
  • Cross with open data on public charging stations: Where is the leasing potential blocked because the infrastructure is not available?

Thank you!

sebastian.horl@irt-systemx.fr

Icons throughout the presentation: https://fontawesome.com

Website

Contact