Referenced by 1 other explainer
Perturbation is how the Australian Bureau of Statistics protects people's privacy in published tables. Before a table is released, the ABS shifts each cell up or down by a small random amount. The shape of the data is kept. But no single cell can be trusted as an exact count of real people.
For a council planner, an ABS table is best read as close to the truth, not the truth itself. Two cells that look like they should add up to a third often will not. The same group of people, counted in two different ABS tables, can come back with slightly different figures. This is on purpose.
Under the Census and Statistics Act 1905, the ABS is not allowed to release information that could identify any one person or business. Small cells are the main risk. A cell that shows one person of a rare age, sex, country of origin, and address could give that person away. Perturbation removes that risk. The ABS sees it as the best way to keep small-area detail open without breaking confidentiality.
Non-zero cells, including totals, are shifted by a small random amount. Each cell is shifted on its own. It is not shifted to match its row or column total. A true count of 4 might come out as 3, 4, or 5. The same cell, used in two different ABS tables, will always come out as the same shifted value across products. So the figures stay consistent across products.
The shifts are fair on average. Across a whole table, the error is small. It tends to cancel out. For a single small cell, the error can still be a big share of the value.
Cells do not always sum to their totals. Adding up the rows of a shifted table will give a number close to, but not equal to, the total shown. The ABS suggests using the total they publish, not adding the rows up yourself.
The same population looks a bit different at different geographies. The count of one-parent families in an SA1, added across an SA2, will not match the SA2 figure the ABS publishes on its own. Each is shifted on its own. The ABS suggests pulling tables at the geography you actually need. Do not build them up from smaller areas.
Small cells should not be relied on. A cell that shows 3 may have been 0, 1, 2, 3, 4, 5, or 6 in the real data. Patterns across many cells stay useful. Single small cells do not.
In older releases, the ABS used cell suppression. Risky cells were replaced with "np" (not published). Other cells were sometimes hidden too, so the missing value could not be worked out. That left holes in the table.
Perturbation took the place of suppression for most Census products from 2006 on. Every cell now has a value. But every value carries a small random error. A "0" in a shifted table is not always a true zero. A low count near zero may still be shifted down to zero. An "np" still means the cell was held back, not shifted.
All ABS data that flows into Place Forecast is already shifted before it arrives. This covers Census counts, dwelling counts, household types, and age and sex breakdowns. Place Forecast does not undo perturbation. Place Forecast does not add its own.
So a small drift between Place Forecast figures and ABS figures for the same area is expected. It is built in, not a bug. When Place Forecast adds up mesh-block or SA1 cells to a small area, the result will not exactly match the SA2 or LGA figure that ABS publishes for the same population. The two paths through perturbation are different.
Place Forecast pages anchor on per-category cells when an ABS-published LGA cell exists for that category. They accept a small gap on the totals. When an exact match with an ABS-published value is needed, Place Forecast scales the parts against that value. It does not rebuild the value from shifted parts.