1. ESSnet on common tools and harmonized methodology for statistical data confidentiality
Daniela Ichim, Luisa Franconi

2. Essnet on SDC harmonisation
ESSnet on common tools and harmonised methodology for SDC in the ESS
December 2010 – April 2012
CBS, Istat, Destatis, Statistics Austria, SCB
Task 1: Harmonisation of microdata release in multiple countries
Task 2: Case studies on tabular data
Task 3: Future directions of SDC software tools
Tasks on dissemination and management

3. Essnet on SDC harmonisation
Task 1
Task 1-1: Choice of measure for the output
definition of the set of objective measures to be maintained by all possible candidate methods
Task 1-2: User needs
analysis of the projects undertaken by researchers on the data: definition of the benchmarking statistics and to prioritise features
Task 1-3: Definition of methodologies
study anonymisation taking into account the new framework and the benchmarking statistics
Task 1-4: Implementation and reporting
a report on the implementation process: pros and cons and critical points

4. Dissemination strategy
Microdata  risk assessment
Apply SDL to reduce risk maintaining some utility
Evaluate utility
SDL methods
Disclosure risk
Original microdata
Utility
Anonimized microdata R U

5. Comparability HOW to achieve it? Bounded utility comparability
1 Assessment of effects of different practices on predefined statistics
2 Definition of a threshold to determine when action is needed
3 setting a process for choosing acceptable practices
Bounded utility comparability

6. SES-Benchmarking Setting of objectives: 1. Member States
a) Dissemination policy (Nace, Size, etc.)
b) Coherence
2. Users
a) High-priority variables:
(eg: NACE, SIZE, region, salary, etc.)
b) Minimum level of detail (NACE 2digits)
c) Types of analyses
Ratios, Weighted totals, salary change, etc.

7. SES-which predefined statistics???
Essnet SDC Harmonisation Deliverable 1:
Focus on consequences on SDL
Part A. Survey structure
a) quality
b) relationships between variables
c) classifications, etc
Part B. Scientific research on SES data
a) models
b) methods
c) breakdowns
d) minimum level of detail, etc
Input
Output

8. Optional variables  What to do???
SES: Requirements stated in the legislative framework and its implementing measures
Reg. (EC) No 1738/2005
5 themes:
information on local units,
employees,
working period,
earnings,
grossing up factors
Optional variables  What to do???

9. SES2006: required characteristics
Theme
Required characteristics
Data source
Tailored questionnaires, existing surveys, administrative sources or a combination of such sources; the information obtained must be of acceptable quality and be comparable between European countries
Reference period
Year 2006, month: October
In some countries the accounting year does not coincide with the calendar year; for these countries the financial year is the best match with the calendar year 2006.
Choice of another month is acceptable if appropriately justified.
Sampling design
Based on a sample of employees drawn from a stratified sample of local units.
reporting unit: the local unit or the enterprise
observation unit: local unit
Economic activities
Sections C-O excluding L of NACE Rev.1.1
Population: enterprises
Enterprises with at least 10 employees in the covered economic activities.
Population: employees
Employees in the observation unit which have an employment contract in the reference month

10. SES2006: minimum requirements1
1.1. Geographical location of the local unit
NUTS 1 level
1.2. Size of the enterprise to which the local unit belongs
1-9*, 10-49, , , , and more employees. *This first band is optional for the 2006 SES.
1.3. Principal economic activity of the local unit
2-digit level of NACE Rev.1.1 for sections C to O. NACE section L is optional for the 2006 SES 2
2.3. Occupation in the reference month
To be coded according to the International Standard Classification of Occupations, 1988 version (ISCO-88 (COM)) at the two-digit level and, if possible, at the three-digit level.
2.5. Highest successfully completed level of education and training
Six levels coded according to the International Standard Classification of Education, 1997 version (ISCED 97).
Share of a full-timer’s normal hours
For a part-time employee, the hours contractually worked should be expressed as a percentage of the number of normal hours worked by a full-time employee in the local unit 3
3.1. Number of weeks in the reference year to which the gross annual earnings relate
should correspond to the actual gross annual earnings (variable 4.1). 4
4.2. Gross earnings in the reference month
should be re-calculated so that it reflects the exclusion of such employees from the sample.
4.3. Average gross hourly earnings in the reference month
average gross earnings per hour paid to the employee in the reference month. 5
5.1. Grossing-up factor for the local unit
Within each sampling stratum, (Variable 5.1) = (Number of local units in the population) / (Number of local units in the sample)
5.2. Grossing-up factor for the employees
(Variable 5.2) = (Variable 5.1) * (Number of employees in the local unit / Number of employees in the sample)
Hierarchical classification
“Independent”
SDC sampling
relationship
formula

11. SES2006: constraints among variables
Variables that need to be consistent
Types of checks required
Variable 3.2, Number of hours actually paid during the reference month,
should be consistent with
variable 4.2, Monthly earnings
If the employee’s paid hours are affected by unpaid absence, then they should be adjusted to obtain paid hours for a full month. Where necessary, provide an approximate estimate of paid hours using: Adjusted 3.2 = Unadjusted 3.2 * (Adjusted 4.2/ Unadjusted 4.2). Where it is not feasible to adjust variable 4.2, then this employee should be excluded from the sample and the grossing-up factor (variable 5.2) re-calculated.
Variable 3.2.1, Number of overtime hours paid in the reference month
Variable 4.2.1, overtime earnings
If the employee’s overtime hours are affected by unpaid absence, then they should be adjusted to obtain the paid overtime hours for a full month. Where necessary, provide a rough estimate of paid overtime hours using: Adjusted = Unadjusted * (Adjusted 4.2.1/ Unadjusted 4.2.1). Where it is not feasible to adjust variable 4.2 or 4.2.1, then this employee should be excluded from the sample and the grossing-up factor (variable 5.2) re-calculated.
Variable 4.3 Average gross hourly earnings in the reference month
Variable 4.2, Average gross hourly earnings derived from gross earnings for the reference month, divided by Variable 3.2, the number of hours paid during the same period

12. SES2006: deviations from EU Regulation (quality reports)
Inclusion of the employees in the sample,
Due to calculation of the average month some consistency between variables are not met,
The definition of the variable gross annual earnings in the reference year
Classification of occupation and education
Enterprises, not local units

13. SES: main outputs Data producers
a) what is already published by MS and Eurostat
b) to be coherent
c) to avoid identification
d) don’t look for ways to increase the info to be published
Users/researchers
a) how SES data is used in scientific research (data, models, methods)
b) obstacles (not administrative)
c) without evaluating the scientific merit!

14. SES2006: indicators (Chronos)
1. Numbers of Employees
a) characteristics of the observation unit and employee.
b) specified bands of hours paid, annual holidays and of hourly/monthly/annual earnings
2. Gross Earnings, Paid Hours and Annual Days of Leave
a) hourly, monthly and annual earnings, monthly paid hours and annual days of leave.
b) several measures of location and of dispersion

15. SES2006: breakdowns (Chronos)
_ Region: restricted to the national level;
_ Economic activity: restricted to NACE,
_ Size of the enterprise: 1-9, 10-49, , , , 1000;
_ Age: restricted to 5 size classes;
_ Occupation: one digit level of the ISCO 88 (COM) classification.

16. Researchers/analysts: general
1. We do not evaluate the scientific merit!
2. More than 80 papers were consulted.
3. National and international comparisons
4. With or without using of the hierarchical employer-employee structure
5. Reporting some absence of information from the enterprise side (e.g. financial)
6. Few longitudinal studies (for the moment)
7. “Home-made” harmonisation (when a MS does not survey/disseminate info on a Nace division, the entire (EU-level) info on the division is excluded from analyses)
8. With or without sample weights.

17. Studies Wage differentials/wage dispersion Labour market policy
Determinants/decomposition
“classical” average gross earnings per enterprise or employee
Low(high)-pay dynamics
Bargaining regimes

18. breakdowns Wage differentials Gender pay gap
Gini coefficient, Quintile Share ratio
Region
Education
Age
Gender
Occupation
Employer (enterprise) impact –economic activity, size, productivity, policy, etc …
breakdowns

19. Models and methods Linear models Log(earnings) as response variable
Mixed-effects, multi-level, ANOVA, quantile
Log(earnings) as response variable
Assumption of normal distributions on error
Method: Ordinary least squares
Sometimes in two stages (enterprise and employee)
Role of local units?
No sampling weights
– earnings on enterprise level (exception???)

20. Selected benchmarking statistics
European dissemination
Breakdowns
NUTS, Gender, Education, Age, Occupation
Weighted means
Linear models
Relationships of earnings

21. Questionnaire to the LAMAS WG national representatives
Prioritization
Questionnaire to the LAMAS WG national representatives
11 questions aiming at the collection of information on preferences regarding dissemination of EU anonymised microdata file
19 MS answered (70%)

22. 100% existence of a legal dissemination framework
Prioritization
100% existence of a legal dissemination framework
58% - national requirements
100% - standard classifications
NACE
NUTS
ISCO
ISCED

23. 58% only the removal of the optional identifying variables
Prioritization
32% removal of all the optional variables, independently on their identification power
58% only the removal of the optional identifying variables

24. Prioritization 1. Principal economic activity 2. Number of employees
3. Geographical location
1. Gender
2. Occupation
3. Education,
4. Age
5. Length of stay in service

25. – hierarchical, relationships, etc Disclosure risk assessment
SDC Methodologies
Data structure
– hierarchical, relationships, etc
Disclosure risk assessment
Disclosure risk limitation
Individual ranking
Constrained regression
Flexibility

26. SDC Methodologies Part A: Risk assessment Part B: Protection
Part A: Risk assessment
Part B: Protection
Part C: Audit
ENTERPRISE
Population frequencies
Recoding
Quality indicators
Sample frequencies
Preliminary recoding
EMPLOYEE
Only outliers
Constrained regression
All records
Individual ranking

27. SDC Methodologies - enterprises
Name
Brief description
Default value St
Threshold for the sample frequencies 2 Pt
Threshold for the population frequencies.
ThresholdRiskStrata
Threshold for the percentage of the admissible strata at risk.
0.02
ChooseDirectlyTheMostDetailedCombination
The most detailed combination satisfying the criteria is considered.
“y”
Increase - severity
Large values – only sample information
0 means no risk
“n” – explorative analysis

28. SDC Methodologies – enterprises Testing
Country
Number of enterprises
Number of strata IT
19899
1029 NL
36762
867 AT
14368
105

29. SDC Methodologies – enterprises Testing
Country
Number of strata at risk
Number Enterprises Risk
Size
Region
Nace IT
139 (13.51 %)
201 (1.01%)
3 classes
NUTS0
2 digit NL
69 (7.96%)
80 (0.22 %) AT
49 (46.67%)
53 (0.37% )

30. Employees at Risk
Frequency criteria for each combination of key categorical variables:
Info on enterrpise (Nace, Nuts, Size)
Demographic variables (Age, Gender)
High AnnualEarnings: greater than a threshold
T = quantile

31. SDC Methodologies – employees Testing
Name
Brief description
Default value
SynthesizeAll
SynthesizeAll="y", how to evaluate the risk.
“n”
MinNbEmployeesPerStrata
More than MinNbEmployees per strata. 5 qq
Quantile value for the definition of the employees at risk.
0.99
qqOverall
How to compute the quantile.
“y”
threshold
Number of employees that could be considered at risk. 1

32. SDC Methodologies – employees Testing
Name
Value
SynthesizeAll
“y”
No risk evaluation, all units at risk.

33. SDC Methodologies – employees Testing
Name
Value
SynthesizeAll
“n”
threshold 1
Unique cases
Name
Value
SynthesizeAll
“n”
threshold 2
Unique and double cases

34. SDC Methodologies – employees Testing
Name
Value
SynthesizeAll
“n”
qqOverall
Risk threshold, by strata
Name
Value
SynthesizeAll
“n”
qqOverall
“y”
Risk threshold,
no stratification

35. SDC Methodologies – employees Testing

36. SDC Methodologies – employees Testing
Name
Value
SynthesizeAll
“y” qq
All units at risk.
Name
Value
SynthesizeAll
“y” qq
0.95, 0.99
If AE>qqV and if unique.
Name
Value
SynthesizeAll
“y” qq
100
No units at risk.

37. SDC Methodologies – employees Testing

38. SDC Methodologies – employees Testing

39. Protection Minimal Requirements
Protect with respect to the assumed scenarios.
Protect if needed.
 Depency on the disclosure scenario
Probabilistic method.

40. Dissemination strategy
Microdata  risk assessment
Apply SDL to reduce risk maintaining some utility
Evaluate utility
SDL methods
Disclosure risk
Original microdata
Utility
Anonimized microdata R U

41. Controlled Selective Masking
Perturb, but generate (control) quality:
coherence (already released statistics)
utility (users’ needs)
Add more (linear) constraints:
Weighted totals variation

42. SDL - implementation Individual ranking Model – based
Parameters: IR.param, stratification
Re-use
Ease of implementation
Flexibility

43. SDC Methodologies – employees Testing
Name
Value
SynthesizeAll
“y”
IR, 3 or 5, by strata, on all
Name
Value
SynthesizeAll
“n”
Method IR
IR, 3 or 5, by strata, on risk
Name
sdcMicro
IR, no strata, on all
IR:no control on

44. Model-based Disclosure Limitation
Assume a model.
Estimate the parameters.
Release the fitted values.

45. Model-based Disclosure Limitation
ConReg: control on

46. Utility Weighted Totals Analytical validity Correlations Variance
Linear models

47. Weighted Totals Constrained Regression

48. Weighted Totals IR

49. Correlation AE_AB

50. Correlation ME_OVER

51. Correlation ME_SHIFT

52. Variance

53. Linear Models By combs of Nace, Nuts and Size
Log(AE) = f(B21, B22, B23, B25, B26, B27)
Log(AE.pert) = f(B21, B22, B23, B25, B26, B27)
Compare the coefficients.
The same for:
B21 + B22
B23 + B25 + B26 + B27

54. Linear Models

55. Confidence intervals overlappingIT NL
Model
Method
Strat qq
OWithinP
PWithinO
B21 B22 IR
strat
0.992
0.984
0.985
B21 B22 B23 B25 B26 B27
0.998
0.952
B23 B25 B26 B27
0.968
no strat
0.945
0.929
0.936
0.947
0.960
0.958
0.949
0.970
0.946
0.956
ConReg
0.99
0.997
0.953
0.988
0.996
0.986
0.950
0.989
0.983
0.965
0.963

56. Data format – missing values, categories
Problems
Data structure
Data format – missing values, categories
(Open source) software knowledge
Data knowledge.
Documentation is a must!

57. – hierarchical, relationships, etc Disclosure risk assessment
Final issues
Data structure
– hierarchical, relationships, etc
Disclosure risk assessment
- national and subjective
Disclosure risk limitation
- protect w.r.t the scenario
Flexibility

58. Collaboration is necessary. Consultation is necessary.
Final issues
Collaboration is necessary.
Consultation is necessary.
Testing is necessary.
Comparability may be achieved.
- development of bounded-utility methods
Governance structure should be defined.

59. THANK YOU!