Impact of using fiscal data on the imputation strategy of the Unified Enterprise Survey of Statistics Canada Ryan Chepita, Yi Li, Jean-S

1
Impact of using fiscal data on the imputation
strategy of the Unified Enterprise Survey of
Statistics CanadaRyan Chepita, Yi Li,
Jean-Sébastien Provençal, Chi Wai
YeungStatistics CanadaICES III, Montréal, June
2007
2
Goals

• To illustrate the challenges of applying a
  centralized E and I strategy to a broad range of
  industrial sectors
• To discuss the changes put in place due to the
  increasing use of fiscal data
• To discuss one approach used to quantify the
  overall E and I effect

3
Outline

• Overview of the Unified Enterprise Survey (UES)
• Survey content
• Imputation strategy
• Use of fiscal data
• Challenges
• Diagnostic tool
• Conclusion

4
Overview of the UES

• Annual business survey
• Initiated with 7 industries in 1997
• Presently integrates over 40 industries covering
  the major sectors of the economy
• 950K establishments in the population
• 127K establishments in the sample

5
Overview of the UES

• Stratified sampling design
• NAICS, province, and size in terms of revenue
• Data collection
• Mail out survey, fax and phone follow-up
• Edit and Imputation
• Estimation
• H.-T. for totals and provincial and industrial
  breakdowns

6
Survey content

• 2 or 3 Key variables
• Total revenue and total expenses
• Similar concepts from one industry to another
• A lot of details (over 50 variables)
• Totals breakdowns
• By province, type of expenses or source of
  revenue
• Industry specific
• Can be revised from year to year

7
Survey content

• Example manufacturing sector

VARIABLES
Sales oth. Goods and serv. produced
Total sales of goods purch for resale
Amount received for custom work
Amount received for repair work
Stumpage sales
Total sales of goods and services produced
Sales of logs and wood residue
Total sales
Details
Key
8
Imputation Strategy

• Categories of non-response
• Category 1 Partial response with at least 1 key
  variable reported
• Category 2 Total non-response with historical
  data
• Category 3 Total non-response without
  historical data

9
Imputation Strategy

• Historical data for some records
• Records sampled the year before
• Same questionnaire
• Administrative data for all records
• Stratification information
• NAICS, province, size in terms of revenue

10
Imputation Strategy

• Type 1 and type 2 non-response
• Missing key variables
• Historical Trend
• Ratio using current survey information
• Missing details
• Historical distribution
• Distribution from all respondent within a
  homogeneous group
• Distribution from a single donor

11
Imputation Strategy

• Type 3 non-response
• Donor imputation
• Closest neighbour based on administrative data

12
Use of fiscal data

• Use fiscal data as a proxy value for total
  non-response
• Use fiscal data as a proxy value for simple units
  randomly selected at the sampling stage
• Use to update the initial size in terms of
  revenue
• Number of survey variables for which we use
  fiscal data as proxy range from 7 to 25

13
Challenges

• Conceptual differences
• Questionnaire content review
• Variables for which there is no proxy value on
  the fiscal data base
• Modeling
• Industry specific needs
• Tailored strategy

14
Challenges

• Monitoring the effect
• Creation of a distinct path for records where we
  used fiscal data (category 4 of non-response)
• Creation of a diagnostic tool

15
Diagnostic tool

• Identification section
• Industry, province, variable description
• Weighted sums, share and percentages by category
  of non-response

Variable X Resp. Cat.2 Cat.3 Cat.4 Total
Sums 30M 5M 5M 10M 50M
Share 60 10 10 20 100
Percentages 20 20 25 18 20
Variable Y (Total)
150M 25M 20M 55M 250M
16
Conclusion

• Centralized E and I strategy vs industry specific
  needs
• Diagnostic tool
• Modeling

17
Thank you!Questions?