|Page tools: Print Page Print All|
Winsorisation of Rates for Average Weekly Earnings
In business surveys, it is possible for a sample to include a small number of units with highly unusual values. These units are referred to as outliers. Their selection in the sample, combined with the application of the specified weights for these units, can mean that survey estimates significantly over- or understate the true population value. In the ABS, the two main methods of addressing outliers are surprise outliering and winsorisation.
Surprise outliering reduces the design weight of an outlier unit to one, while adjusting the weights of the other units in the stratum upwards to compensate. While surprise outliering is relatively simple to implement, there are some issues with this method. These include:
· The identification of surprise outliers can be a subjective choice, and require a large amount of effort.
· Setting a unit as a surprise outlier can have unexpected effects for rate estimates.
· If surprise outliering efforts focus on the impacts to a rate estimate, there may be little control over the effects on the components of the rate.
Winsorisation outliering identifies extreme reported values and replaces these with more reasonable values, design weights are left unchanged. The identification of winsorised outliers depends on objective cut-offs calculated using historical survey data. Reported values that exceed a given cut-off will be modified to a more central value. The theory for winsorising level estimates is reasonably well established. However, the theory for directly winsorising rate estimates has not yet been developed.
Recently, Business Survey Methodology (BSM) undertook an investigation to assess whether winsorisation for rate estimates can be effectively accomplished by winsorising the components of the rate separately. Data from the Average Weekly Earnings (AWE) survey was used for this investigation. The AWE survey is a biannual collection that obtains data on employment and earnings from businesses in order to produce estimates of average weekly earnings (total earnings divided by total employment). These estimates are produced for males, females and all persons, for particular breakdowns. AWE undertakes surprise outliering of units in the sample.
The approach taken for this investigation was to winsorise the main earnings and employment data items for AWE separately. Winsorisation was applied to historical AWE survey data (excluding the effect of surprise outliers), and the resulting estimates were compared with publication estimates.
Results from the investigation have been positive. In general, the impact on estimates when winsorisation was used instead of surprise outliering was non-significant, with most winsorised rate estimates lying well within the 95% confidence interval of the original published estimates.
One benefit we have seen with winsorisation of rate components has been that while we only target units with unusually large earnings or employment values, the winsorisation process will pick up units that have either an unusually high or unusually low earnings-to-employment ratio.
Winsorisation outliering is scheduled to be placed into production in the near future, and will be used in conjunction with a reduced effort on surprise outliering.
These documents will be presented in a new window.