Preserving Confidentiality and Quality of Tabular Data: Are Safe Data Necessarily Inferior Data?

Slide 1

Lawrence H. Cox, Associate Director
National Center for Health Statistics
LCOX@CDC.GOV     

Bureau of Transportation Statistics Confidentiality Seminar
Washington, DC
September 17, 2003

PRESENTATION HANDOUT–DO NOT QUOTE OR CITE

Slide 2
Statistical Disclosure Limitation (SDL) for Tabular Data

Tabular data

• frequency (count) data organized in contingency tables
• magnitude data (income, sales, tonnage, # employees, ..) organized in sets of tables

Tables

• there can be many, many, many tables (national censuses)
• tables can be 1-, 2-, 3-, .........up to many dimensions
• tables can be linked
• table entries:  cells (industry = retail shoe stores & location = Washington DC)
• data to be published:  cell values (first quarter sales for shoe stores in Washington DC = $17M)

What is disclosure?

Count data: disclosure = small counts (1, 2, ...)

Magnitude data: disclosure = dominated cell value

Example:

# 2 can subtract its contribution from cell value and infer contribution of #1 to within 10% of its true value = DISCLOSURE

Cells containing disclosure are called sensitive cells

How is disclosure in tabular data limited by statistical agencies?

• identify cell values representing disclosure
• determine safe values for these cells

Example: If estimation of any contribution to within 20% is safe (policy decision), then a safe value above would be $18M

• traditional methods for statistical disclosure limitation
• Count data:
• rounding
• data perturbation
• swapping/switching
• cell suppression
• Magnitude data:
• cell suppression

What is cell suppression?

• replace each disclosure-cell value by a symbol (variable)
• replace selected other cell values by a symbol (variable) to prevent narrow estimates of disclosure-cell values
• process is complete when resulting system of equations divulges no unsafe estimates of disclosure-cell values

Some properties of cell suppression:

• based on mathematical programming
• very complex theoretically, computationally, practically
• destroys useful information
• thwarts many analyses; favors sophisticated users

How does cell suppression addresses data quality?

Cell suppression employs a linear objective function to control oversuppression

Namely, the mathematical program is instructed to minimize:

• total value suppressed
• total percent value suppressed
• number of cells suppressed
• logarithmic function related to cell values
• etc.

These are overall (global) measures of data distortion

Further, individual cell costs or capacities can be set to control individual (local) distortion

These are all sensible criteria and worth doing

However, they do not preserve statistical properties (moments)

Moreover, suppression destroys data and thwarts analysis

Slide 3
Controlled Tabular Adjustment (CTA)

• new method for SDL in tabular data
• perturbative method–changes, does not eliminate, data
• alternative to complementary cell suppression
• attractive for magnitude data & applicable to count data

Original CTA Method (Dandekar and Cox 2002)

• identify sensitive tabulation cells
• replace each disclosure cell by a safe value–namely, move the cell value down or up until safety is reached
• use linear programming to adjust nonsensitive values in order to restore additivity (rebalancing)
• if second and third steps are performed simultaneously, a mixed integer linear program (MILP) results. MILP is extremely computationally demanding
• otherwise (most often), the down/up decision is made heuristically, followed by rebalancing via linear programming (LP). LP computes efficiently even for large problems

Slide 4
(Nearly) Actual Example of Magnitude Table with Disclosures

Example 1: 4x9 Table of Magnitude Data & Protection Limits for the 7 Disclosure Cells (red)

Example 1a: After Optimal Suppression: 11 Cells (30%) & 2759 Units (7.5%) Suppressed

Example 1b: After Controlled Tabular Adjustment

Example 1: 4x9 Table of Magnitude Data & Protection Limits for the 7 Disclosure Cells (red)

Example 1b: Table After Controlled Tabular Adjustment

Example 1c: Table After Optimal Controlled Tabular Adjustment (Regression)

Slide 5
MILP for Controlled Tabular Adjustment (Cox 2000)

Original data:  nx1 vector a

Adjusted data:  nx1 vector a + y ⁺ - y ^-

T denotes the coefficient matrix for the tabulation equations

Denote  y = y ⁺ - y ^-

Cells i = 1, ..., s are the sensitive cells

Upper (lower) protection for sensitive cell i denoted P_i(-P_i)

MILP for case of minimizing sum of absolute adjustments

Subject to:

T (y) = 0

y_i^- = p_i(l-I_i)
y_i⁺ = p_iI_i          i = 1, ... , s (sensitive cells)

0 ≤ y_i^- , y_i⁺ ≤e_i , i = s+1, ..., n

(nonsensitive cells)

I_i  binary,  i = 1, ..., s

Capacities e_i  on adjustments to nonsensitive cells typically

small, e.g., based on measurement error

Slide 6
Data Quality Issues

Based on mathematical programming, just like cell suppression CTA can minimize:

• total value suppressed
• total percent value suppressed
• number of cells suppressed
• logarithmic function related to cell values
• etc.

In addition, adjustments to nonsensitive cells can be restricted to lie within measurement error

Still, this may not ensure good statistical outcomes, namely,

analyses on original vs adjusted data yield comparable results

Slide 7
Towards Ensuring Comparable Statistical Analyses

Verification of “comparable results” is mostly empirical

Many, many analyses are possible: Which analysis to choose?

Instead, we focus on preserving key statistics and linear models

• mean values
• variance
• correlation
• regression slope

between original and adjusted data

Can do this using direct (Tabu) search

I will describe how to do so well in most cases using LP

For simplicity, assume that the down/up decisions for sensitive cells have already been made (by heuristic)

Slide 8
Preserving Mean Values

When the LP holds a total fixed, it preserves the mean of the cell values contributing to the total e.g., fixing the grand total preserves the overall mean

In general, to preserve a mean, introduce (new) constraint: Σ (adjustments to cells contributing to the mean) = 0

A criticism of CTA is that it introduces too much distortion into the values of the sensitive cells

In general the intruder does not necessarily know which cells are sensitive nor cares to analyze only sensitive data, so focusing on distortions to sensitive values may be a bit of a red herring

Still, it is useful to demonstrate how to preserve the mean of the sensitive cell values, as the method applies to preserving the mean of any subset of cells

Preserving the mean of the sensitive cell values is equivalent to constraining net adjustment to zero:

If, as in the original Dandekar-Cox implementation, we allow only two choices for y_i , this is unlikely to be feasible

However, satisfying this constraint is not a problem if we simply expand the set of possible y-values viz., if we permit slightly larger down/up adjustments

The MILP is:

min c(y)

Subject to:

T(y) = 0

p_i(l - I_i) ≤ y_i^- ≤ q_i(l - I_i)
       p_iI_i ≤ y_i⁺ ≤ q_il_i          i = 1, ... , s

0 ≤ y_i^- , y_i⁺  ≤e_i   i = s+1, ..., n

 q_i are appropriate upper bounds on changes to sensitive cells

c(y) is a linear cost function, typically involving sum of absolute adjustments

If the down/up directions are pre-selected, this is an LP

Slide 9
Preserving Variances

Seek: Var(a + y) _ Var(a), assuming

Var(a + y) = Var(a) + 2Cov(a,y) + Var(y)

Define L(y) = Cov(a,y)/Var(a)

L(y) is a linear function of the adjustments y

Var(a + y)/Var(a) = 2L(y) + (1 + Var(y)/Var (a))

|Var(a + y)/Var(a) - 1 |=| 2L(y) + (Var(y)/Var(a))|

Var(y) is nonlinear, but can be linearly approximated

Alternatively:  typically Var(y)/Var(a) is small

Thus, variance is approximately preserved by minimizing | L(y) |

The absolute value is minimized as follows:

  * incorporate two new linear constraints in the system:

w ≥ L(y)

w ≥ - L(y)

* minimize w

Slide 10
Assuring High Positive Correlation

Seek: Corr(a,a + y) _ 1

Corr (a, a + y) = Cov(a, a + y) ÷ √ Var(a) Var(a + y)

After some algebra,

Corr (a, a + y) = (l + L(y)) ÷ √ Var(a + y) / Var(a)

Again: min | L(y) |  yields a good approximation because it drives both numerator and denominator to one

Slide 11
Assuring Slope of Regression Line(s)

Seek:  under ordinary least squares regression

Y = β₁ X + β₀

of adjusted data Y = a + y on original data X = a,

we want: β₁ _ l and β₀ _0

As , then β₀ _ 0 if β₁ _ l

This corresponds to L(y) _ 0 (if feasible)

Note again:  this is achieved via min | L(y) |

Slide 12
The Compromise Solution

Variance is preserved by minimizing L(y)

Correlation is preserved by minimizing L(y)

Regression slope preserved by L(y) _ 0  (if feasible)

All subject to  

If Var(y)/Var(a) is small (typical case), imposing objective

function min | L(y) | assures good results simultaneously

• for variance
• for correlation
• for regression slope

Shortcut is to incorporate the constraint L(y) = 0 (if feasible)

Choosing L(y) _ 0  is motivated statistically because it implies (near) zero correlation between values a and adjustments y viz., as solutions  y and  -y are interchangeable, this correlation should be zero

Slide 13
Examples

4x9 Table

Table 1: 4x9 Table of Magnitude Data and Protection Limits for Its Seven Sensitive Cells (in red)

Table 2: Original Table After Various Controlled Tabular Adjustments Using Linear Programming to Preserve Statistical Properties of Sensitive Cells Only

Slide 14
Results for 4x9 Table

Summary: 4x9 Table Linear Programming

Table 3: Summary of Results of Numeric Simulations on 4x9 Table Using Linear Programming

* = compromise solution

Slide 15
Results for 13x13x13 (Dandekar) Table

Summary:   13x13x13 Table Linear Programming

Table 4: Summary of Results of Numeric Simulations on 13x13x13 Table Using Linear Programming

* = compromise solution

Slide 16
Concluding Comments

• statistical agencies have responsibilities
• to respondents (to maintain confidentiality)
• to data users (to deliver high-quality data products)

• these responsibilities
• are often in opposition
• nevertheless, are not mutually exclusive
• have, in the past, been approached separately

• research indicates these responsibilities can be addressed
• simultaneously
• using systematic, computationally efficient methods