European Conference on Quality in Survey Statistics

1
European Conference on Quality in Survey
Statistics
2
Exploring the statistical utilisation of
financial statements to estimate sub-regional
economic variables

• P. Calia - C. Filippucci
• University of Bologna - Italy
• Filippucci_at_stat.unibo.it

3
Aim of the work

• Obtain statistical data at local level SLL
  (sub-regional level) and by economic sector not
  provided by official statistics (value added,
  output, labour cost) referring to a digital data
  base of companies financial statements AIDA
• SLL cluster of municipalities based on
  commuting flows
• Emilia-Romagna Region and manufacturing sector
  are considered - year 2002

4
Aim of the work

• THE PROBLEM
• to face unit selection by an appropriate
  estimation strategy for variables

5
Why to use accounting data

• Moreover
• Information is available about a large number of
  business - much more than sample surveys,
• Reduce the respondent burden
• Supply micro data for modelling economic
  behaviour

6
Outlines of the paper

• Does business accounts provide necessary
  information to estimate the economic variables of
  interest?
• AIDA reliability
• AIDAs coverage of the population
• Strategy for estimation

7
AIDA

• Edited by Bureau Van Dijk
• Collects financial statements of Italian
  companies since 1996
• Accessible via web after subscription
• ----------------
• 18,796 firms located in Emilia Romagna out of
  365,000
• year 2002

8
AIDA cnt

• It is possible
• Select firms by different criteria (economic
  activity, location, identification codes, legal
  form, etc)
• select variables of interest (accounting items
  and ratios, user defined indexes, structural
  economic variables, stakeholders, etc.)
• perform peer group comparisons and simple
  statistical analysis
• export data to be processed with other softwares

9
AIDA cnt

• Companies compelled to present the financial
  statements are 26 but 72 in terms of
  employment.
• In AIDA they are more than 99
• Total coverage in the manufacturing sector is
  14 (firms) and 61 (employees) it improves
  considering the above group 24 (firms) and 84
  (employees)
• In AIDA only firms having an estimated 2002
  production value gt500,000 Euros

10
From Business accounts to SNA concepts

• The problem has been addressed by UN (UN, 2000)
  success depends on the formats used for
  compilation of financial statements
• In Italy, business accounting structure, format
  and contents are established by law into the EU
  framework
• Output components and costs are pointed out

11
From business accounts to SBS concepts cnt

• Exact correspondence cannot be established for
  all variables (items detail and aggregation)
• In Italy, the adequacy of accounting data to EC
  regulations for business statistics has been
  studied by National statistical Institute -
  Istat
• accounting data covers about 65 of the variables
  considered by structural business surveys
• high correspondence between accounting and
  surveys values
• Actually, accounting data are used by Istat to
  impute for missing data in structural business
  surveys

12
From business accounts to SBS concepts cnt

• Macrovariables in the frame of SNA are not
  exactly the same
• BUT
• Financial statements are the only source for
  obtaining fairly good indicators of firms
  perfomances at sub regional and activity sectors
  levels

13
Production value (output)
14
Value added at factor costs
15
Reliability

• two strategies
• Internal checks
• External checks comparing with ASIA 2002
• 2002 is considered because Asia is referred to
  2002
• ASIA permanent archive of all active firms in
  industry and services, yearly updated. It is the
  frame for all the business surveys

16
Issue 3 avalaible information

• From ASIA E-R (2002)
• Fiscal code (unique id)
• Name of the enterprise
• Activity code (ateco 2002 Nace rev. 1.1)
• Geographical location
• Legal form
• Number of employees

• From AIDA
• Fiscal code and business register code (CCIAA)
• Name of the enterprise
• Activity code (ateco 2002 Nace rev. 1.1)
• Geographical location
• Legal form
• Date of birth
• End date of financial statement
• Accounting items from income statement

17
Preliminary check

• Internal Checks Codes duplication and Missing
  data Essentially no problems
• 2. External checks
• Linking AIDA and ASIA (by fiscal code, Nace rev.
  1.1 codes, location and legal form)
• Some problems very few or easy to recover and
  to correct using ASIA

18
Issue 3 Preliminary check cnt.

• ONLY 35 records are duplicated by fiscal code
  (same company)
• Different cciaa code only 2 are clearly
  recording error
• Different geographical location (33 companies)
• Different legal form (2 companies)
• Different economic activity code (17 companies)
• Differences in accounting items (5 companies)
• Retain records whose geographical location is
  equal to that of asia 18761 unique records

19
Issue 3 Preliminary chek cnt.

• Missing data in
• Fiscal code (12 records recovered in aida)
• Geographical location (15 record 14 recovered
  in aida)
• Legal form (859) Easy to obtain from ASIA
• Economic activity (43 record)

20
Issue 2 Preliminary chek cnt.

• LINKING BY FISCAL CODE WITH ASIA
• 18023 companies out of 18761 match
• 738 are not in ASIA NOT CONSIDERED IN OUR
  ANALYSIS
• reasons
• not eligible according to ASIA (excludes
  businesses in sectors A, B, L, P, Q, division 91
  and legal unit classified as institutions or
  belong to the private non profit sector)
• not resident in Emilia Romagna at 2002 and moved
  later (location and other characteristics in AIDA
  refers to the last balance sheets available)
• no more active in 2002 but compelled to present
  the balance sheet (i.e. liquidation, bankruptcy )
• acquired by other enterprise during that year

21
Issue 3 Preliminary chek cnt.

• Problems in comparing (Because point 2 -
  previous slide)
• Ateco code 7982 records in AIDA with different
  values from ASIA (reduce to 3245 considering
  2-digits code)
• Geographical location after normalisation, 814
  business are located differently (371 change SLL
  location)
• Legal form after normalisation, 1416 enterprises
  with different legal form (whose 859 with missing
  data)
• In the following analysis data from ASIA have
  been chosen

22
Framework for estimation

• Sample is not random predictive approach is
  needed for estimation (Royall, 1988 Royall and
  Pfeffermann,1982 Särndal, 1996)
• The predictive approach is based on a
  statistical model in the population relating the
  variable of interest to auxiliary variables.
• The model should be able to make the selection
  process ignorable
• It is not a behavioural model but its only task
  is to incorporate the relationships between
  observed and non-observed units

23
Framework for estimation cnt

• The model is the only source for inference
• From this model predictions for non observed
  units are obtained
• Auxiliary variables are chosen in order to
  identify groups where the selection process is
  not informative (Estevao et al., 1995)

24
Framework for estimation cnt

• Finite
  population
• variable to be estimated
• Auxiliary variables known for each
  unit
• subsets of of units observed and
  non observed
• sample units in a sample s
• selection mechanism for a sample s

25
Framework for estimation cnt

• In the model approach it i necessary to specify
• function connected to the
  selection
• Y distribution in the
  finite population on the basis of the a priori
• Joint distribution
  of Y and
• 
  distribution based on
  observations

26
Framework for estimation cnt

• Making inference on selection process
  can be ignored if , given ,
  (Rubin -1976)
• If in a sample the variable of interest would be
  independent from the unit selection but
  dependent from the Z konwn
• Then
  would be
  equal to the distribution of observations
  obtained ignoring the selection (inference is
  possible referring only to observations and the
  model)

27
Framework for estimation cnt

• In a non random sample selection is represented
  by where Z is a variable indicator
• Given s, inference on ,for the group
  selected by mean of Z, is obtained from
  , ignoring the selection
• Making inference on others groups also
  should be known
• If X are variables known for all the units in the
  population and it is true
  then
  the selection on Z can be ignored
• Inference on quantities of the finite population
  needs also to know

28
Framework for estimation cnt

• The first task for model specification is
• to find out the variables driving the selection

29
Variables driving the selection

• Legal form is the crucial variable for
  partecipating in AIDA and it is not possible to
  use it as auxiliary variable
• Possible auxiliary variables are Size (number of
  employees) Sector of activity Sub-regional area
  SLL
• To define their relevance, coverage and
  distributions in AIDA are compared with
  population

30
Tables by legal form
31
Coverage by size

• SMALL FIRMS ARE NOT REPRESENTED
• - gt 20 employees more than 74 (firms) 76.5
  (employees)
• - 1 employee 1.2
• - 2-9 employees 6 (firms) and 9 (employees)
• - Distributions in AIDA 22 has less than 10
  employees vs 80 in the population.

32
Coverage by size

• Excluding proprietorships and partnerships,
  coverage rise up
• - gt 20 employees more than 90 (firms and
  employees)
• - 1 employee 9
• - 2-9 employee 30-40
• - distributions get closer, especially for
  employees

33
Tables by size
34
Coverage by sector

• Coverage by sector ranges between 5 to 60
  (firms) and between 30 to 80 (employment)
• 50 of sectors present a coverage greather than
  17 (firms) and than 58 (employment)
• AIDA and ASIA distributions are not too much
  different taking into account the simple index
  of dissimilarity is 0.22 and 0.11
• Considering the subgroup of companies and
  cooperatives, coverage improves and distributions
  gets closer

35
Table by sector (2-digit code)
36
Coverage by SLL

• Coverage by SLL is very variable it ranges
  between 3 and 23 (firms) and between 19 and
  78 (employees)
• coverage in 50 of SLL is greater than 10.6
  (firms) and 53 (employees)
• AIDA and ASIA distributions are not too much
  different simple dissimilarity indexes are very
  low (respectively 0.12 for firms and 0.06 for
  employees)
• Considering the subgroup of companies and
  cooperatives, coverage improves and distributions
  gets closer

37
From the above analysis

• From the evidence presented it is clear that the
  variables considered to study coverage are
  important in driving the selection hence are
  relevant to implement a model-based estimation
  strategy
• Moreover they are important from an economic
  point of view.

38
From the above analysiscnt

• Because of severe undercoverage of small firms,
  only firms with more than 1 employee (excluding
  mostly proprietorships and partnerships) are
  considered
• As consequence, totals will be underestimated
  even if economic weight of firms with one
  employee is negligible they are 29 of
  population but represent only 2.8 of employees

39
BLU predictor

• The total of Y
• The predictor of T
• We need a predictor only for the non observed
  units of the population
• The form of the predictor its unbiasedness and
  optimality depend on the model
• General specification for a BLU predictor as well
  as unbiasedness conditions are obtained assuming
  a general regression model

40
Model Groups

• In our case, SLL are the estimation domains they
  form a partition of the population of N units
• The population is to be divided in H groups able
  to explain most of the variance in Y.
• Groups are based on size and economic activity
  according to strategies
• 1 - groups are obtained cross classifying by
  firm size (number of employees) and economic
  activity
• 2 - economic activity defines the groups and
  size enter the model as continuos variable with
  different parameter for each group

41
Model Groups cnt

• SLL (estimation domain) could also be a relevant
  auxiliary variable to establish the model groups
• Only if SLL effect is captured by activity sector
  (many SLL are specialised in one activity) they
  are not relevant
• As consequence also SLL could be used to form
  groups

42
Superpopulation models
Group h- (size by sector) effect equal across
domains
Group (size x sector) effect plus indipendent
domain (q) effect
Interaction effect of size (continuous) with
sector plus indipendent domain effect
Interaction effect of size (continuous) with
sector
Group effect (size by sector) different across
domains
43
Superpopulation models cnt

• Group effects are fixed
• The model specification ask for the definiton of
  variance and covariance structure of observations
  
• The simplest structure is Homoskedasticity
• We do not expect Homoskedasticity but variability
  should depend on size and/or economic activity

44
Superpopulation models cnt

• The analysis of residual of the model estimation
  for Value added confirms our expectation
• We concetrated on two main heteroskedasticity
  structures
• A) x size, k firm
• B) m activity
  sector or size
• A further alternative consider x2 in A and B

45
Superpopulation models cnt

• Model selection is based on usual criteria to
  evaluate model optimality
• - model fitting indicators (AIC, Likelihood
  ratio test)
• - standard errors

46
Main findings cnt

• the estimation under the model 5 is not suitable
  because it is possible only by high aggregation
  of groups considered no useful
• The heteroskedasticity structure has a
  foundamental role and in particular the two
  alternative tested are relevant.
• When x2 is considered further improvements are
  obtained

47
Main findings cnt

• Under the assumption (A)
• in the models 2 and 3 SLL effect is not
  significative (according to usual tests) hence
  the models 1 and 4 (SLL not considered) should be
  preferred

48
Main findings cnt

• 2) Under the assumptions (B)
• a) many specifications meet estimation problems
  (likehood does not behave properly)
• b) when estimation is possible there are cases
  where SLL effect is significative
• c) this happens in the model 3 when variance is
  modelled by m activity
  sector or size
• d) according to AIC the best result is using
  activity sector

49
Main findings cnt

• As consequence
• The models to be considered are (1) and (3) but
  looking at AIC the variance specification (A)
  gives the worse results
• model (3) is to be preferred also because it
  includes a significative SLL effect

50
Conclusions

• AIDA is the only source to estimate main economic
  variables at sub-regional level and by economic
  activity sector
• It is possible to reach a quite good
  approximation of national accounting variables
  and use it to obtain estimates of firms
  perfomances at sub regional and activity sectors
  levels

51
Conclusions cnt

• Problems of estimation due to undercoverage and
  selection bias are managed by predictive
  inference
• The approach has been implemented referring to
  different model specifications and variance
  assumptions
• The
• is to be preferred with the assumption that
  variance depends on firm size and activity sector

52
Further work

• Further check of model robustness will be carried
  out
• - Effect of influential observation
• - Effects of some different groups
  aggregations
• - Sensitivity analysis