IMPACT OF LARGE WEIGHTS ON ONE PERSON PER HOUSEHOLD METHODOLOGY
A number of collections within the ABS household survey program enumerate one individual within selected households. The initial weight is calculated as the number of in-scope persons in the household, multiplied by the household selection weight. This weight is then post stratified. Thus random persons in large households can have relatively large weights and, if they have a relatively low or high value for particular data items, could "unduly" influence the estimate for those data items.
The usual method for addressing this situation would be some form of outliering technique but there may be better ways to adjust for this which does not involve "tampering" with what is quite possibly valid data. As the problem of unstable estimates arises from the selection method rather than data outliers a better approach might be to address the issue of weight outliers.
One approach to reducing the impact of extreme weights is to select more persons from those households which would likely generate the large weights. This can be achieved by selecting two random persons from large households thereby effectively halving the weight and subsequently halving the impact of selected persons with extreme, but valid data values and large weights.
An annual file from the 1998/99 Population Survey Monitor was analysed to examine the impact of outliers on estimates. This showed there were 127 households that had greater than five persons per household, and hence were classified as 'large households'. Weight analysis showed that the weights increase as household size increases. However the size of the increase is not as large as would be expected for the maximum weights. This suggests that the post-stratification also has a large influence on the size of the weight.
A sample was synthesised to reflect the selection of two persons from large households by randomly selecting additional persons from the 127 households and adjusting the weights accordingly. Income data was imputed for the new selections and estimates of deciles were derived. The effect on the deciles was small, and the differences were limited to only the 4th, 6th and last deciles. Even then, the differences are minor, with the greatest change in decile cutoffs being $2000 (annual income).
Looking more closely at the weights showed that what made the weight large had more to do with the demographic variables and area that the respondent was from, than the size of the household. This is reflected in the maximum and minimum weight analysis, where the weights for respondents in households of 7 people range from 357 to 6,678.
The results of this investigation do not suggest that we need to be overly concerned about the impact of large weights arising due to the selection of one person from large households. However, it remains a possibility that a combination of large weights and extreme values could "unduly" influence estimates.
For further information please contact Julie Watkinson on (08) 8237 7539