Hiding small numbers
Stop a table from pointing at two identifiable people.
What we have
tabulate: small cells, their partners, and small totalstabstat: a small group loses its whole rowsummarize: a small variable is hidden, and Min and Max always arecollapse: a small group's statistics are set to missing in the data- a self-audit that solves for the hidden cells and hides more until it cannot
Coming later
list,svy:,exportand the rest do not apply it- nothing looks across two tables at once, so two safe tables can still give the answer away between them
- this is the first step of disclosure control, not the last
set mincell 5 set mincell off
With this on, table cells resting on fewer than N observations are hidden, so a table can leave the building.
. set mincell 5
. tabulate condition sex
sex
condition | Male Female | Total
----------+---------------------+----------
Asthma | <5 <5 | *
Hiding the cell is only half the job
Print the totals beside a hidden cell and it is simply the total minus the others. So Virata also hides partner cells (marked *), and hides totals that are themselves small. After choosing what to hide, it tries to work every hidden number out the way an outsider would, and hides more until it cannot. That self-audit is the part that makes it worth having.
Three consequences worth knowing:
- "Small" counts observations, never weights. Three people each carrying a survey weight of 1,000 are still three people.
- A hidden cell loses its percentages too, because a percentage times a total gives the count back.
- The chi-squared statistic is not printed when cells are hidden, because it is computed from them and can be worked backwards.
What it covers
| Command | What happens |
|---|---|
tabulate | small cells, their partners, and small totals |
tabstat | a group under N loses its whole row |
summarize | a small variable is hidden, and Min and Max are always hidden while this is on, because they are single people's values however many people there are |
collapse | the statistics of a small group are set to missing in the data, since a collapsed dataset gets listed and saved later |
list, svy:, export and the rest do not apply it yet, and nothing looks across several tables at once.
This is the first step of disclosure control, not the last. It protects the totals printed in the same table. It cannot protect you from publishing two tables that together give the answer away.