To summarise the discussion in this paper, there are three main messages that we feel users of the data should keep in mind when making conclusions.

Firstly, researchers and policy-makers need to be aware of the different survey methodology used in CAs and NCAs. Given that no identifiers for CAs are publicly released on the CURF or the RADL, this effectively means that analysts have to be aware that a large proportion of remote data in the 2002 NATSISS was collected from CAs. It may be possible to purchase customised cross-tabulations from the ABS if people are particularly concerned about potential non-sampling error arising from the differing methodology in the two types of areas. ^[9] Whatever one’s position, some caution is warranted when making comparisons between the CAs and NCAs, and possible remote and non-remote areas.

The second issue that should be kept in mind, especially when making comparisons with either the 1994 NATSIS or the 2001 Census, is that 2002 NATSISS only collects information from Indigenous people aged 15 and over living in private dwellings. That is, the statistics collected do not provide information on the entire population of Indigenous Australians. Where other data sources are used, analysts need to ensure they are comparing the same age groups; non-private dwellings need to be excluded; and, in the case of the 1994 NATSIS, the re-weighted data set must be used once it becomes available on the CURF.

The final point is that given that the ABS has released information that enables users to reasonably accurately estimate standard errors, it is no longer acceptable to make statements about differences in means or proportions without taking into account sampling error. Analysts need to make themselves aware of the formula for calculating standard errors and confidence intervals. More sophisticated users will have to work out how to use replicate weights. Given that there is considerable room for programming errors, especially when using arrays in SAS, the ABS should consider releasing draft programming code (particularly, if such code is already available) to minimise the possibility of non-sampling error arising from analysts. Ultimately, though, analyst error is the responsibility of the analyst.