IMPUTATION FOR NON-SAMPLING ERROR IN THE FREIGHT MOVEMENT SURVEY
The Freight Movement Survey (FMS) was conducted out of the Brisbane office and covered four separate modes of transport involved in freight movement: rail, sea, air and road. The road component involved a fortnightly sample survey of vehicles used for freight movement purposes. The reference period was the year beginning April 2000 through to the end of March 2001, with 26 non-overlapping fortnightly samples selected to cover the whole reference period. This article details the non-sampling error in the road component and the subsequent investigations and the final adjustment made to the data.
Selected vehicles in the road component of the FMS were asked to keep a log of trips (both laden and unladen) for the reference fortnight. During the production of estimates for the FMS, concerns were raised about the comparison between FMS results and results from the Survey of Motor Vehicle Usage (SMVU). These concerns were related to FMS results being considerably lower for the three major variables of total distance, weight carried and tonnes kilometres.
Investigations carried out into the possible reasons for this discrepancy identified the possible under-reporting of trips made during week 2 of the fortnightly cycle. Subsequent time series analysis conducted on five variables, (namely laden distance travelled, total distance travelled, weight carried, number of laden trips, and total number of trips) produced results showing strong statistical significance and consistent direction. This analysis concluded the under-reporting in week 2 for these 5 variables ranged from 5-10% with a standard error of between 1% and 2%.
Given the significance of the under-reporting in week 2, a decision was made to compensate for the discrepancy in the estimates. Conceptually, an adjustment for each of the five variables of interest was required.
Imputation of additional trip records for week 2 using donor records from week 1 was the preferred method of adjustment. An analysis of possible imputation classes was undertaken with the best performing imputation classes being based on the variable of laden distance travelled, split into three groups (less than 900 km, 900 km to 2000 km, and 2001 kms and greater).
Allocation to imputation classes was achieved using a method of minimising the difference between the intended adjustment and the expected adjustment at the imputation class level, constrained to the total level of intended adjustments as obtained from the previous analysis. A Lagrangian multiplier technique was employed to obtain a solution which enabled the allocations in each imputation class to be resolved. This technique was applied using the variables laden distance and weight carried, with imputation classes of laden distance outlined above.
The final adjustments made based on this method of imputation were within one standard error of the under-reporting identified in the time series analysis for all of the five variables with the exception of laden trips.
For more information, please contact Brett Frazer on (07) 3222 6028
Email: brett.frazer@abs.gov.au
Print Page
Print All