INPUT SIGNIFICANCE EDITING TRIAL IN SIS YIELDS VERY GOOD RESULTS
A parallel run of Input Significance Editing (ISE) and Service Industry Survey (SIS) Intermediate Editing was done for the 01/02 Employment Services Survey. ISE is an editing approach applied at the input stage which is intended to direct resources to units that are expected to yield the most benefit from editing. Current applications of ISE are normally in surveys that have more recent historical data since the method needs expected values for all units. It was decided to do a trial of ISE in SIS to determine if the method will be applicable for surveys that don't necessarily have historical data.
For the SIS trial, 9 key variables or items which contribute to key outputs were identified. Expected values or imputes were calculated for each key item using current survey data after sufficient responses were received. Two types of imputes were tested: regression-based imputes; and a combination of means and medians of imputation classes.
Each unit was assigned an item score for each of the key items and a provider score which combined the item scores. Item scores were derived from the weighted difference of the unit's reported value and imputed value for the item.
ISE lists were generated at the provider level and for each item by ranking units according to the expected benefit in editing each unit. The scores were used to measure expected benefit. For each list, a cutoff was set for the level of cumulative benefit in editing all units above the cutoff. All units that were above the cutoff in the provider or any of the item lists were selected for editing. SIS kept snapshots of the survey data file before and after intermediate editing. These snapshots were used to analyse results.
Overall, the trial gave very good results. It showed that ISE was effective in prioritising units for editing and the cost-benefit trade off, as units were edited in the ranked lists, was quite strong. Provider scores performed well in summarising the nine key items. Mean and median imputes also performed reasonably well, which is a good outcome, since they are much easier to calculate than the regression-based imputes. Moreover, most of the units selected for editing in ISE matched with units flagged for editing by SIS using their current editing system. This was a desirable outcome since SIS have a very good editing strategy.
For information on the SIS trial, please contact Elsa Lapiz on (03) 9615 7364.
For general information on significance editing, please contact Keith Farwell on (03) 6222 5889.