10.  Conclusion

Sampling is now an accepted and integral part of census-taking. Its use can lead to substantial reductions in costs and respondent burden associated with a census, or alternatively, can allow the scope of a census to be broadened at the same cost. The price paid for these advantages is the introduction of sampling error to census figures that are based on the sample. The effect of sampling is most important for small census figures, whether they are counts for rare categories at the national or provincial level or counts for categories in small geographic areas. It should be noted that response errors and processing errors also contribute to the overall error of census figures and it is the same small census figures that are particularly susceptible to the effects of these non-sampling errors. Therefore, even with a 100% census, many small figures would be of limited reliability. As a general rule of thumb for the 2006 Census, figures of size 100 or less that are based on sample data are of very low reliability, while figures up to size 500 tend to have standard errors in excess of 10% of their size.

For many of the characteristics, a certain amount of bias was detected in the sample. A small portion of the bias may have been introduced during data processing and edit and imputation. The rest of the bias would have been due to one or more factors such as non-response bias, response bias or the selection of a biased sample by the enumerators. Calibrating sample estimates to known population counts as part of the census weighting procedures helped to reduce the impact of these biases.

   Previous page | Table of contents | Next page