The data are synthetic. The EU-SILC public microdata are simulated by Eurostat from models of the real survey: they contain no real person, and Eurostat states that they cannot be used to draw conclusions about the population. This project treats the file as if it were confidential, to show the full disclosure-control workflow. All risk figures and indicator values below illustrate the method; they say nothing about real people or regions.

Summary

This project plays the role of a statistical office that must release a public-use microdata file and a published table from a household survey without disclosing information about any respondent. It uses the Italian file of the EU-SILC public microdata (2013, 37,209 persons aged 16 and over in 18,487 households) as if it were the confidential source.

  • In the original file, 24,337 records (65.4%) have a combination of eight publicly knowable characteristics that nobody else in the sample shares. The model-based estimate is 178.0 expected re-identifications.
  • Three protection scenarios were compared. All of them remove every sample unique. They differ in how much detail they sacrifice to get there.
  • The recommended scenario, S3: NUTS 1, 10-year age bands, grouped codes, k = 5, leaves no combination shared by fewer than 5 records, also when the released weight is added to the keys, and reduces expected re-identifications to 1.78. It changes the income and poverty indicators by at most 2.98%, almost all of it because the released weight can no longer differ between the NUTS 2 regions of a NUTS 1 area. Only this scenario is released; the other two are comparison runs, shown here only through their aggregate risk and loss measures.
  • In the table of employee income by region, occupation and sex, 27 cells are sensitive and 64 more are hidden so that no sensitive cell can be estimated too closely from the published cells. With the primary suppressions alone, 0 of the 27 sensitive cells are protected; with the secondary suppressions, 27 of 27 are.

Why disclosure control

European statistics are produced under a legal duty of statistical confidentiality (Regulation (EC) No 223/2009, Article 20, and Principle 5 of the European Statistics Code of Practice). Removing names and identifiers is not enough: a person can be recognised from a combination of ordinary characteristics, such as region, age, household size and occupation. Statistical disclosure control (SDC) measures that risk and reduces it while keeping the data useful.

Data and scenario

Source. Eurostat, EU-SILC public microdata for Italy, file IT_PUF_EUSILC.zip, SHA-256 97135d88df36ebd1…, downloaded from the Eurostat website. The public microdata are fully synthetic: they are simulated from models of the real survey and contain no real respondent. Eurostat states that they cannot be used for inference about the population. Here they are used for what they are meant for: training and method development. The risk figures below describe this synthetic file, not real people.

Intruder scenario. An intruder knows some facts about a person and looks for records that match them. The facts considered are the key variables (quasi-identifiers):

Key variable EU-SILC variable Detail in the source
Region DB040 NUTS 2 (21 regions)
Sex PB150 2 categories
Age PX020 single years, 16 to 80+
Marital status PB190 5 categories
Citizenship PB220A Italy / other EU / non-EU
Education PE040 ISCED 1997, 7 levels
Occupation PL051 ISCO-08 major group, 9 groups
Household size HX040 1 to 10

Survey identifiers (PB030, PX030) are direct identifiers and are replaced by random numbers. Employee income (PY010G) and equivalised disposable income (HY020 / HX050) are confidential numerical variables; household disposable income and the equivalised household size are not released, because together with equivalised income they would reveal household size. The personal weight (PB040) is replaced by a released weight, described below, because the original weight takes one value per NUTS 2 region and would reveal the region in the scenarios that publish only NUTS 1.

Risk in the original file

Two families of measures are used.

  • k-anonymity counts records whose combination of keys appears fewer than k times in the sample. It needs no model and is easy to explain.
  • Individual risk (Benedetti and Franconi, the method used by Istat and implemented in sdcMicro) estimates the probability that a match is correct, using the sampling weights to estimate how many people in the population share the same combination. The sum over records is the expected number of re-identifications. Household risk is the probability that at least one member of a household is re-identified, since identifying one member exposes the others.
Measure Value
Records 37,209
Sample uniques 24,337
Records violating 3-anonymity 30,860
Records violating 5-anonymity 34,387
Expected re-identifications 178.0
Global risk 0.48%
Household risk 1.17%

The two families disagree, and the reason matters. About two thirds of the records are unique in the sample, but each sampled person represents about 1,400 people (the mean weight), so the model expects few of those sample uniques to be unique in the population as well. Agencies usually require both: a k-anonymity threshold for the release file, and a check that the model-based risk is low.

Protection methods

Each scenario applies three standard methods in this order (Hundepool et al., Handbook on Statistical Disclosure Control, 2012):

  1. Global recoding of the key variables into broader categories.
  2. Local suppression: individual key values are blanked, choosing the values that break the most risky combinations, until every combination is shared by at least k records. Region, sex and age are suppressed last because they matter most to users. Household size is never blanked: the number of members a household has in the release would give it back. Region describes the household, and the release keeps a household identifier, so a region blanked for one member is blanked for every member of the household.
  3. Microaggregation (MDAV, groups of 3) of the two income variables: each value is replaced by the mean of a group of three similar values, so no exact income is released. Equivalised income is treated once per household, so all members of a household keep the same value.

The released weight is the original weight post-stratified to the released region: one value per released region (NUTS 2 in S1, NUTS 1 in S2 and S3) and one common value for records whose region is blanked, so that regional and national totals are unchanged. The common value is the national mean weight, or the mean of a region with blanked records if that is lower, so that no region is left with a weight below its mean. It carries no information below the released region. The price is that, in S2 and S3, it no longer corrects for the different sampling rates of the NUTS 2 regions inside a NUTS 1 area: any weight that does so must differ between NUTS 2 regions, and so reveals them.

Scenario Region Age Other keys k
S1 NUTS 2 5-year bands source detail 3
S2 NUTS 1 10-year bands grouped 3
S3 NUTS 1 10-year bands grouped 5

“Grouped” means: occupation in four skill groups, education as low / medium / high, marital status as never married / married / previously married, household size top-coded at 5 or more.

Risk after protection

Scenario Step Sample uniques < 3-anonymity < 5-anonymity Expected re-identifications Household risk
S1 after global recoding 16,697 23,728 29,839 133.19 0.869%
S1 after local suppression 0 0 7,217 7.52 0.048%
S2 after global recoding 2,669 4,882 8,373 22.15 0.151%
S2 after local suppression 0 0 2,592 2.89 0.019%
S3 after global recoding 2,669 4,882 8,373 22.15 0.151%
S3 after local suppression 0 0 0 1.78 0.012%

Recoding alone does most of the work for S2 and S3. In S1, which keeps detailed codes, recoding leaves many combinations rare, so local suppression has to blank many more values to reach the same guarantee.

Release checks

Before a release file is written, five checks run on it; the pipeline stops if one fails.

  • Smallest group with the weight as a key (min_fk_with_weight): the smallest number of records sharing a combination of the eight keys and the released weight. It must be at least k.
  • Weights that point to one NUTS 2 region (weights_spanning_one_nuts2): released weight values whose records all come from a single NUTS 2 region. It must be 0 in S2 and S3; it does not apply to S1, which publishes NUTS 2.
  • Households with inconsistent keys (households_inconsistent_keys): households where region or household size is blank for some members and not for others. It must be 0.
  • Blank household size (hh_size_blanked): records whose household size is blank in the release. It must be 0.
  • Forbidden columns (forbidden_columns): equivalised household size, household disposable income, survey identifiers or exact age in the release. It must be 0.
Check S1 value S1 pass S2 value S2 pass S3 value S3 pass
min_fk_with_weight 3 yes 3 yes 5 yes
weights_spanning_one_nuts2 n/a n/a 0 yes 0 yes
households_inconsistent_keys 0 yes 0 yes 0 yes
hh_size_blanked 0 yes 0 yes 0 yes
forbidden_columns 0 yes 0 yes 0 yes

With the released weight added to the keys, the smallest group has 3 records in S1, 3 in S2 and 5 in S3. Because the released weight is one value per released region plus one common value for blanked regions, this check gives the same result as k-anonymity on the eight keys alone: it shows that the weight adds no detail to the keys, not a second, independent guarantee. Records with a blanked region all share one weight, which says only what the blank already says, so in the group check their weight matches any value, like their region, and in the NUTS 2 check that value is not counted (were their weights different, they would be checked as they are).

Information loss

Suppressed values

Scenario Values suppressed Share of key values
S1 26,885 9.03%
S2 5,163 1.73%
S3 9,588 3.22%

Indicators

Each indicator is computed on the original and on the protected file with the same code. The at-risk-of-poverty (AROP) rate is computed as in the Eurostat indicator ilc_li02 (threshold at 60% of the national median equivalised disposable income), but on persons aged 16 and over with a personal weight, because only the personal file is released. Because the file is synthetic, the levels are not real Italian statistics; only the differences matter.

The tables separate two sources of change. Weight only computes the indicator on the original file with each record’s weight replaced by its released weight; incomes only replaces employee and equivalised income by their microaggregated values; released file is the protected file as published. The two partial effects do not add up exactly to the total, because they interact and because records with a blanked region drop out of the regional figures.

Microaggregation is the same in every scenario and moves no indicator by more than 0.27%. In S1 the weight changes only through the records with a blanked region, which share one common weight while the rest of their region absorbs the difference; its effect is at most 0.46%. In S2 and S3 the weight is the larger source: it is constant within each NUTS 1 area, so it no longer corrects for the different sampling rates of the NUTS 2 regions inside an area, and it moves indicators by up to 2.98% (AROP rate, ITI, Centro). That is the price of not revealing NUTS 2 through the weight.

S1: NUTS 2, 5-year age bands, detailed codes, k = 3. Relative difference from the original value
Indicator Domain Original value Weight only Incomes only Released file
AROP rate ITC 0.270 0.46% 0.00% 0.52%
AROP rate ITF 0.396 0.11% 0.00% 0.11%
AROP rate ITG 0.441 0.12% 0.00% 0.14%
AROP rate ITH 0.225 0.42% 0.00% 0.44%
AROP rate ITI 0.282 0.08% -0.17% 0.10%
AROP rate Italy 0.312 0.23% -0.03% 0.23%
Gini (equivalised income) Italy 0.470 0.09% 0.04% 0.13%
Mean employee income Italy 22,466 0.12% -0.12% -0.00%
Mean employee income Men 25,645 0.14% -0.27% -0.10%
Mean employee income Women 18,566 0.07% 0.13% 0.19%
Median employee income Italy 19,017 0.13% -0.04% 0.12%
Median equivalised income Italy 15,947 0.10% -0.01% 0.10%
S2: NUTS 1, 10-year age bands, grouped codes, k = 3. Relative difference from the original value
Indicator Domain Original value Weight only Incomes only Released file
AROP rate ITC 0.270 -0.86% 0.00% -0.89%
AROP rate ITF 0.396 -0.05% 0.00% -0.07%
AROP rate ITG 0.441 -1.44% 0.00% -1.38%
AROP rate ITH 0.225 0.08% 0.00% 0.08%
AROP rate ITI 0.282 -2.98% -0.17% -2.98%
AROP rate Italy 0.312 -0.97% -0.03% -0.97%
Gini (equivalised income) Italy 0.470 -1.39% 0.04% -1.36%
Mean employee income Italy 22,466 -2.42% -0.12% -2.43%
Mean employee income Men 25,645 -2.13% -0.27% -2.21%
Mean employee income Women 18,566 -2.42% 0.13% -2.32%
Median employee income Italy 19,017 -1.47% -0.04% -1.47%
Median equivalised income Italy 15,947 0.10% -0.01% 0.10%
S3: NUTS 1, 10-year age bands, grouped codes, k = 5. Relative difference from the original value
Indicator Domain Original value Weight only Incomes only Released file
AROP rate ITC 0.270 -0.86% 0.00% -0.89%
AROP rate ITF 0.396 -0.05% 0.00% -0.07%
AROP rate ITG 0.441 -1.44% 0.00% -1.38%
AROP rate ITH 0.225 0.08% 0.00% 0.08%
AROP rate ITI 0.282 -2.98% -0.17% -2.98%
AROP rate Italy 0.312 -0.97% -0.03% -0.97%
Gini (equivalised income) Italy 0.470 -1.39% 0.04% -1.36%
Mean employee income Italy 22,466 -2.42% -0.12% -2.43%
Mean employee income Men 25,645 -2.13% -0.27% -2.21%
Mean employee income Women 18,566 -2.42% 0.13% -2.32%
Median employee income Italy 19,017 -1.47% -0.04% -1.47%
Median equivalised income Italy 15,947 0.10% -0.01% 0.10%

Regression

A user fitting a wage equation (log employee income on sex, age band and NUTS 1 area, weighted) should reach the same conclusions on the protected file. The confidence-interval overlap (Karr et al., 2006) is 1 when the 95% intervals coincide and 0 when they do not overlap at all. The model is fitted with lm(weights = weight), which treats sampling weights as precision weights, so its intervals are not design-based; they serve here only to compare the two files. In S2 and S3 the overlap is lower for the same reason as the indicators: the released weight changes the relative weight of the NUTS 2 regions inside each NUTS 1 area, which shifts the regional and age coefficients.

Scenario Mean overlap Lowest overlap Coefficients
S1 0.94 0.86 18
S2 0.79 0.67 12
S3 0.79 0.67 12

Risk-utility trade-off

Points lower down are safer; points further left lose less detail. Global recoding is not captured on the horizontal axis: S2 and S3 publish coarser categories than S1 for every record.

Recommendation. S3 is the release candidate. It reaches 5-anonymity, with the weight included, with far fewer suppressions than S1 because recoding removed the rare categories first. Its cost is in the weight: S2 and S3 publish NUTS 1, so the weight cannot reflect NUTS 2, and indicators move by up to 2.98% instead of 0.52% in S1. Moving from S2 to S3 (k = 3 to k = 5) leaves the indicators as in S2, roughly doubles the suppressions and removes the remaining records that share their combination with only 3 or 4 others. S1 keeps regional detail but must blank many more values, mostly occupation, and offers only 3-anonymity. Users who need NUTS 2 detail should use the scientific-use file under a research contract rather than a more detailed public file.

Table protection

Published tables leak information too. The table is the estimated total employee income by region (Italy, NUTS 1, NUTS 2), occupation (ISCO major group) and sex: 810 cells including all margins.

  • Primary suppression. A cell is sensitive if it has fewer than 3 contributors (threshold rule) or if the second-largest contributor could estimate the largest within 10% by subtracting its own value from the total (p% rule, p = 10).
  • Secondary suppression. A suppressed cell could be recomputed from its row or column totals, so additional cells are hidden across the region hierarchy and both other dimensions. sdcTable’s quick heuristic (SIMPLEHEURISTIC) is tried first, then its optimisation method (OPT); the first pattern that passes the audit below is published.
  • Audit. For every sensitive cell, a linear program, independent of sdcTable, computes the lowest and highest income the cell can have given the published cells and the sums of the table. That interval must keep the true value at a distance of at least its p% protection level on both sides: the smallest error the second-largest contributor must be left with, so that it cannot estimate the largest contribution to within 10%.
  • Additions driven by the audit. sdcTable’s protection levels act on the number of contributors, not on income, so its pattern can leave an interval too narrow. When that happens, further cells are hidden around the failing cell, as rectangles or cubes of neighbouring cells (a sibling region, another occupation, the other sex), choosing the group that hides the least income and still protects the cell, and the audit is run again until every sensitive cell is protected.
Item Value
Cells 810
Published 719
Primary suppressions 27
Secondary suppressions 64
of which added by the audit 13
Income in suppressed detailed cells 4.9%
Sensitive cells protected, primary suppressions only 0 of 27
Sensitive cells protected, with secondary suppressions 27 of 27

Secondary suppression method used: OPT + audit-driven additions, with 13 cells added by the audit. With the primary suppressions alone, 0 of 27 sensitive cells are protected; with the secondary suppressions, 27 of 27 are. The published table and the audit files give only which cells are suppressed and whether each sensitive cell is protected; the values, contributions and intervals of the suppressed cells are not published.

21 of the 25 sensitive detailed cells are for women, in occupations with few female employees: 8 for skilled agricultural workers (group 6), 6 for plant and machine operators (group 8), 5 for managers (group 0-1), 2 for craft workers (group 7). For men there are 4 for skilled agricultural workers (group 6). The secondary suppressions, including the cells added by the audit, keep the interval of every sensitive cell at least as wide as its p% protection level on both sides.

Limitations

  • The source is synthetic, so the risk levels are illustrative. On a real confidential file the same code would give different numbers.
  • The intruder scenario fixes eight key variables. A different scenario (for example an intruder who knows income) needs its own key variables.
  • Microaggregation by itself does not guarantee that a very high income cannot be linked to an outlier; a production release would add top-coding of income checked against the actual distribution.
  • The released weight is constant within each released region. Real EU-SILC weights vary within regions by design; a production release would calibrate the weights to several published margins (for example sex and age band within NUTS 1) and run the same release checks on the result.
  • The released weights keep each region’s total, and records with a blanked region share one common weight. Anyone who knows the official NUTS 1 population totals can therefore work out how many records with a blanked region belong to each NUTS 1 area, though not which ones.
  • The cells added after the audit are chosen one failing cell at a time; the final pattern is protected but not minimal. τ-ARGUS’s modular method with income-based protection levels would usually hide less.

Reproducibility

Source file pinned by SHA-256 in data/manifest.json; sdcMicro 5.8.2, sdcTable 0.34.0; run at 2026-10-28 22:01:53 UTC. Code: https://github.com/simones99/eu-silc-sdc.