EUSILC Helsinki November 2006

1

Representative wealth data for Germany The
impact of methodological decisions around
imputation and the choice of the aggregation unit
Joachim R. Frick, Markus M. Grabka Eva M.
Sierminska Luxembourg Wealth Study (LWS)
conference, 14-15 December 2006, Luxembourg.
SOEP at DIW Berlin, Technical University
Berlin (TU Berlin) and IZA Bonn. c/o DIW Berlin,
Department SOEP, Königin-Luise-Str. 5, 14195
Berlin, Germany. ltjfrick_at_diw.degt
2
Contents

• Motivation
• SOEP Wealth Data
• Empirical Results
• The impact of imputation
• The aggregation unit
• Concluding Remarks

3
Motivation

• Measurement error in economic outcome measures
  (income, wealth)
• Survey specific decisions on
• Pre-data collection How and whom to survey ?
• Individuals versus Households
• Post-data collection How to deal with
  measurement error
• Inconsistencies ? Editing
• Item-non-response ? Imputation
• Impact of methodological and surveying decisions
  on substantive results (inequality, mobility)
• What does this mean for cross-national datasets
  (LWS) ?

4
Surveying wealth information at individual level
5
Data SOEP ? LWS

• SOEP Wealth module in 2002
• Individual level (all HH members gt16) n23.900
  (12.500 HH)
• Own Property
• Other Property
• Financial Assets (gt2.500 ) Total Assets
• Private Pensions
• Business Assets
• Tangible Assets (gt2.500 ) Net Worth
• Main Property Debt
• Other Property Debt Total Debt
• Consumer Debt (gt2.500 )
• Not included cars, public pension entitlements,
  durables

6
Data SOEP ? LWS

• Non-Response
• Unit-NR ? weighting
• Partial-NR 5 ? imputation
• Item-NR 15-33 ? imputation
• Other Measurement error
• Inconsistencies lt10 ? editing

7

Item non-response, editing and imputation
(population share affected)

8
Correlates of INR on Total Assets TA (Heckman
selection correction)

• (1) Selection model ? Prob(TA 1)
• male, higher age, high educated, self-employed
• - unemployed, pensioners
• (2) Probability model ? Prob(INR 1) (TA
  1)
• Low education, self-employed,
  self-administered interview
• - male, civil servants, number of interviews

9
Principles of the Imputation strategy

• Imputation of missing information (INR/PUNR)
• Logit Filter, Share
• Regression Market Value, Debt
• Heckman sample selection model
• Controlling for regional clustering effects
  (market value of private property)
• Maintaining variance by adding random residuals
  (taken from the true distribution)
• Incorporating uncertainty of imputation process
  ? Multiple imputation (k5)

10
Market value for own property Observed and
prediction considering residuals
11
Market value for own property (PR) MI for INR
and prediction for observed cases vs. observed
cases
12
Empirical Results
a) The impact of imputation
13
Population share holding wealth componentsbefore
and after editing imputation
Source SOEP 2002 Population Adult population
(17 years and over) with interview 1 Only those
with observed value are included. 2 After
editing and imputation 3
(final-obs)/obs
14
Mean wealth before and after editing imputation
(individual level, weighted)
Source SOEP 2002 Star () indicates means are
significantly different. Standard errors are
bootstrapped (100 reps). 1 Only
those with observed personal share and value are
included. 2 After editing and
imputation 3 (final-obs)/obs
15
Wealth inequality before and after editing
imputation (individual level, weighted)
Source SOEP 2002 Star () indicates means are
significantly different. Standard errors are
bootstrapped (100 reps). 1 Only
those with observed personal share and value are
included. 2 After editing and
imputation 3 (final-obs)/obs
16
Empirical Results
b) The aggregation unit
17
Effect of the choice of the aggregation unit on
the distribution (net worth)
Source SOEP 2002. Asset poverty threshold 50
median net worth
18
Effect of the choice of the aggregation unit on
the distribution (net worth)
Source SOEP 2002. Asset poverty threshold 50
median net worth
19
Effect of the choice of the aggregation unit on
subgroup indices (net worth)
Source SOEP 2002. Basis all individuals with
completed interview (n23135)
20
Effect of the choice of the aggregation unit on
subgroup indices (net worth)
Source SOEP 2002. Basis all individuals with
completed interview (n23135)
21
Individual vs. HH-perspective Gender wealth gap
22
Individual vs. HH-perspective Gender wealth gap

Total

Married
Not married
23
Concluding Remarks

• Multiple imputation effective means to cope with
  selective NR
• significant impact on share of wealth holders,
  mean, aggregate, inequality
• Survey design wrt aggregation Individual vs
  Household
• significant redistribution effect within
  households ? gender wealth gap
• but missing wealth held by children !
• Outlook Cross-national harmonization of LWS data
• comprehensive wealth measure by simulating
  public pension entitlements
• imputation strategy matters !
• sacrifice the superior information at ind.
  level for the sake of comparability?

24
at Last we Will Suceed

25
Appendix

26
Comparison of total wealth of private households
with national balance sheet 2002
27
Share of net wealth and decile wealth imputed by
deciles
28
Net worth by GenderIndividual vs.
HH-perspective (SOEP 2002)