An Application of Selective Editing to the US Census Bureau Trade Data

Maria GarciaUS Census Bureau

UNECE/SDE, Oslo, Norway, 24-26 September 2012

An Application of Selective Editing to the US Census

Bureau Trade Data

Foreign Trade Statistics Programs

• Official source of US international merchandise trade statistics

• Electronic data collection• Complete enumeration • Pre – editing: check for fatal errors• Micro editing

– Range and ratio edits– Automatic imputation – “Rejects” – imputation not successful

Foreign Trade Data Processing

• Rejects– Distribute among analysts for manual correction. – Analysts review large number of records under tight

time constraints• Goal: Use selective editing to identify highly

suspicious errors having a high potential effect on the estimates– Value (V)– Quantity (Q)– Shipping weight (SW)

Hidiroglou-Berthelot Method (HB)

• Latouche and Berthelot (1992) used HB when developing their score functions.

• HB uses historical ratios () to detect outlying observations in periodic data (1986).

• Our data: cannot test historical ratios.• For record , instead of using HB to identify

errors in or , identify errors in unit prices

• Apply a series of transformations:– Identify outliers at both ends of the distribution

of unit prices

– Size transformation, where

HB Method for Our Trade Data

• Measure distance of quartiles of from median

• Measure displacement from median

weighted by appropriate distance

HB for Our Trade Data (Cont’d)

Effect on Publication Totals• Examine effect of changes on final publication

totals (Adapted from Latouche and Berthelot, 1992)

• If no anticipated value is available, use median of ratios and reported data, e.g., for Value

• For every record (Similar to Jäder and Norberg, 2005)

Simulation and Evaluation

• Simulation– Extract from a one-year exports data file– Archived raw and edited (final) data

• Evaluation– Absolute Pseudo-bias = (Latouche and

Berthelot, 1992)

Evaluation Results

• Examining results at lowest level of aggregation: – Data users may closely scrutinize the

statistics for particular types of products– Ex: import/export of rough diamonds – Kimberley Process - joint governments,

industry and civil society initiative to stem the flow of conflict diamonds

Absolute Pseudobias for Exported Non-industrial Diamonds (>0.5 carats)

Using Ratio to Measure Suspicion

0

0.2 0.4

0.6

0.8

1 1.2 1.4

1.6

0.0% 20.0% 40.0% 60.0% 80.0% 100.0%

Percent of Erroneous Records Corrected

Abs

olut

e P

seud

obia

s

Evaluation Results

Evaluation Results

Customer’s Feedback

• Subject matter experts questioned:– High ranking given to records that by

experience they consider insignificant to final cell estimates

– Low ranking given to records that would have been flagged for manual correction

 Total Value (V)

 Total Quantity (Q)

 Unit Price

(V/Q)  Ratio 

Bounds 

Reported cell total $102,190 7,217 $14.15 90 3000

Reported suspicious Record $3,024  7,144 $0.42 90 3000

Final suspicious record $3,024 10 $302.40 90 3000

Final cell total $102,190 83 $1,231.20 90 3000   

Commodity XXXXXXXXXX

Suspicious record is correctly identified by selective editing as having a large effect on the total quantity:

Customer’s Feedback

 Total Value (V)

Total Quantity (Q)

 Unit Price

(V/Q)  Ratio  

Bounds 

Lower Bound

Upper Bound

128 records, 87 records imputed, three rejects $3,142,622 129,973,502 $0.02 0.25 50

Final cell totalAll three rejects corrected $3,142,622 1,230,629  $2.55 0.25 50Selective editing cell totalTwo highest ranked records corrected $3,142,622 1,804,699 $1.74 0.25 50     

Commodity YYYYYYYYYY

Customer’s Feedback

Concluding Remarks• Combination of Hidiroglou-Berthelot and

Latouche-Berthelot methods. • Tried alternative ways to calculate : Quartile

method and Resistant fences method.• Looking at alternative evaluation methods,

determining optimal levels of aggregation, and including seasonality in calculation of simple statistics.

Thank you!