Use of Administrative Data in Statistics Canada

1
Use of Administrative Data in Statistics Canadas
Annual Survey of Manufactures

• Steve Matthews and Wesley Yung
• May 16, 2004
• The United Nations Statistical Commission and
  Economic Commission for Europe
• Conference of European Statisticians

2
Outline

• Introduction
• Tax data programs at Statistics Canada
• The Annual Survey of Manufactures (ASM)
• Overview
• Strategy for use of tax data
• Analytical studies
• Conclusions and Future Work

3
Introduction

• Desire to increase use of tax data
• Reduce respondent burden
• Reduce survey costs
• Can be used at many stages of survey process
• Stratification
• Survey data validation
• Edit and imputation
• Estimation

4
Tax Data programs at Statistics Canada

• Tax data available to Statistics Canada
• Collected by Canada Revenue Agency (CRA)
• Access via a data-sharing agreement
• To be used only for statistical purposes
• Two extensive tax data programs
• Unincorporated businesses (T1)
• Incorporated businesses (T2)

5
Tax Data programs at Statistics Canada (contd)

• T1 - Population
• Unincorporated businesses
• Account for small share of revenues
• Administrative Data
• Sample-based
• Limited set of variables
• Edit and imputation is applied
• Weighted benchmarked estimates

6
Tax Data programs at Statistics Canada (contd)

• T2 - Population
• Incorporated businesses
• Account for large share of revenues
• Administrative Data
• Census-based
• Extensive set of variables
• Edit and imputation is applied
• Micro-data is produced

7
The Annual Survey of Manufactures

• Manufacturing is an important sector of Canadian
  economy
• 17 of GDP
• Annual Survey of Manufactures
• Take-none Portion and Survey Portion
• Extensive questionnaire (financial and commodity)
• Data requirements (pseudo-census)

8
The Annual Survey of Manufactures (contd)

• Target population
• Drawn from Statistics Canadas Business Register
  (BR)
• All businesses classified to manufacturing
• Sample design
• Non-survey portion
• Administrative data
• Survey portion
• Stratified SRS (Stratum NAICS Province
  Size)
• Small take-some / Large take-some / Take-all
• Collected via mail-out / mail-back, follow-up via
  telephone

9
The Annual Survey of Manufactures (contd)

• Edit and Imputation
• Edits applied to ensure accuracy and coherence
• Extensive imputation to produce pseudo-census
  dataset
• Historical imputation
• Ratio imputation
• Nearest-neighbour donor imputation

10
The Annual Survey of Manufactures (contd)

• Estimation
• Non-survey portion (tax data)
• Total Expenses only
• T1 weighted domain estimates
• T2 aggregates from administrative census dataset
• Survey portion (survey data and imputed data)
• Aggregates from pseudo-census dataset
• Domains of interest NAICS and Province

11
Analytical Studies

• Motivation for two studies
• Which variables should be replaced?
• What are the effects of the strategy on final
  estimates for all variables?

Study 1 Data comparison
Study 2 Impact Analysis
12
Analytical Study 1

• Study to select appropriate variables
• Comparison of reported data collected via survey
  and tax
• Simple businesses only
• Assess suitability for substitution of survey
  data

Based on 6,000 businesses
13
Analytical Study 1 (contd)

• Correlation Analysis
• Wide range of correlations
• Total Expenses 0.9
• Total Energy Expenses -0.10
• Reporting Patterns
• Same pattern (zero or positive) for individual
  businesses
• Total Expenses 99
• Total Energy Expenses 50

14
Analytical Study 1 (contd)

• Distribution of Ratios
• Examined histograms, fraction between 0.9 and 1.1
• Total Expenses 60
• Total Energy Expenses 16
• Population Estimates
• Relative difference between tax and survey-based
  estimates
• Total Expenses 3
• Total Energy Expenses 28

15
Analytical Study 1 (contd)

• Selected several variables for direct
  substitution
• Section totals and sub-totals
• expenses, revenues, inventories, etc.
• Remaining variables are imputed
• Imputation gt assign distribution of details
  within each total

16
Analytical Study 1 - Conclusions

• Distinctively different results for different
  variables
• Direct substitution seems feasible for totals
• Direct substitution not recommended for details
• Use standard methods to impute other variables

17
Analytical Study 2

• Analysis to evaluate impact of tax data strategy
• Bias
• Comparison of estimates from different scenarios
• Variance
• Shao-Steel approach for variance estimation
• Reflects variance from sampling and imputation
• Assume equal probability of response within
  imputation class

18
Analytical Study 2 (contd)

• Scenarios

Tax Data Used in Imputation Estimator Variance
HT No Tax None (ratio imputation based on frame revenues) Horvitz-Thompson Sampling Imputation
PC No Tax None (ratio imputation based on frame revenues) Pseudo-census Imputation
PC - Tax Non-response (in or out of sample) Direct substitution Ratio imputation Pseudo-census Imputation
19
Analytical Study 2 (contd)

• Comparison of resulting estimates for Total
  Expenses
• Relative Difference from HT No Tax Total
  Expenses
• Median value for all such domains

All Manufacturing NAICS3 x Province
PC No Tax 1.8 0.0
PC Tax 0.5 1.3
20
Analytical Study 2 (contd)

• Comparison of estimated CVs for Total Expenses
• Co-efficient of Variation Total Expenses
• Median value for all such domains

All Manufacturing NAICS3 x Province
HT No Tax 0.3 1.5
PC No Tax 0.3 1.5
PC Tax 0.1 0.7
21
Analytical Study 2 (contd)

• Comparison of resulting estimates for Total
  Energy Expenses
• Relative Difference from HT No Tax Total
  Energy Expenses
• Median value for all such domains

All Manufacturing NAICS3 x Province
PC No Tax 1.2 0.0
PC Tax 0.8 1.2
22
Analytical Study 2 (contd)

• Comparison of estimated CVs for Total Energy
  Expenses
• Co-efficient of Variation Total Energy Expenses
• Median value for all such domains

All Manufacturing NAICS3 x Province
HT No Tax 0.3 1.8
PC No Tax 0.4 1.8
PC Tax 0.4 1.8
23
Analytical Study 2 - Conclusions

• Bias
• Small relative difference between estimated
  totals from scenarios
• Variance
• Relatively low CV for all options
• Tax substitution variables Scenario 3 most
  efficient
• Non-tax substitution variables Scenario 1 most
  efficient
• Analytical capabilities
• Scenarios 2 and 3 provide most detail

24
Conclusions

• Results used to select 2004 strategy PC Tax
  
• Meets needs of data users
• Reduced cost and response burden
• Maintain (improve) quality
• Striving to further increase use of tax data
• Increased portion of population
• Increased number of variables

25
Future Work

• Editing of tax data
• Similar approach to survey data approach
• Potential to expand list of direct substitution
  variables
• Indirect use of tax data
• More adaptive models
• Quality indicators
• Account for increased variance and potential for
  bias due to imputation