Quality of Income Data: Rounding Behaviour (1) CHINTEX The

advertisement
page 1
CHINTEX
The Change from Input Harmonisation to Ex-post
Harmonisation in National Samples of the European
Community Household Panel
– Implications on Data Quality –
Selected Slides and Tables from
Work-package 5 "Quality of Income Data"
Jens U. Hanisch, JWG University
email: jhanisch@wiwi.uni-frankfurt.de
Quality of Income Data: Rounding Behaviour (1)
Analysis
Measures for rounding
Stat. Literature on
rounding of incomes
Descriptive Analysis:
Comparison D-ECHP
with SOEP
Comparison Fin-ECHP
(1996) with register
Status
done
done
Remarks
Multiples of even numbers
Mixture models with unknown probability of
rounding to the nearest even number
at work
a) Earned income, household income
b) Dependency on level of income
– Analysis with nonparametric logit model
c) Apparent more rounding in the D-ECHP
at work
a) Survey monthly data, register annual data
b) Find monthly reference value
page 2
Rounding in D-ECHP and SOEP
Data: Net Household Income
Observations: Household Questionnaires
Rounding of Household Net Disposable Income
D-ECHP, 1994-1996, non-anonymized data
rounding HH net
income 1994
rounding HH net
income 1995
rounding HH net
income 1996
Count
507
%
12,6%
Count
566
%
14,5%
Count
509
%
13,3%
yes, 100
1742
43,1%
1691
43,3%
1757
46,1%
yes, 500
688
17,0%
632
16,2%
577
15,1%
yes, 1000
904
22,4%
798
20,5%
766
20,1%
yes, 5000
177
4,4%
191
4,9%
170
4,5%
yes, 10000
20
,5%
24
,6%
34
,9%
4038
100,0%
3902
100,0%
3813
100,0%
no
Total
Rounding of Household Net Disposable Income
SOEP, 1994-1996, Samples A+C
rounding HH net
income 1994
rounding HH net
income 1995
rounding HH net
income 1996
Count
902
%
17,0%
Count
939
%
18,1%
Count
852
%
16,5%
yes, 100
2265
42,6%
2203
42,5%
2156
41,7%
yes, 500
798
15,0%
774
14,9%
819
15,8%
yes, 1000
1115
21,0%
1028
19,8%
1043
20,2%
yes, 5000
209
3,9%
217
4,2%
269
5,2%
no
yes, 10000
Total
27
,5%
28
,5%
32
,6%
5316
100,0%
5189
100,0%
5171
100,0%
Comments: It is easy to see that the number of values which are not rounded is
approximately 3-4% higher in the SOEP than in the D-ECHP for the same reporting
periods. Since the SOEP is running longer, this indicates that people tend to tell
exact numbers more likely when panel participation duration increases.
However, for a further analysis and hypothesis testing it is necessary to take item
non-response and bracket usage into account.
page 3
Rounding in D-ECHP and SOEP
Data: Personal Wage and Earnings
Observations: Personal Questionnaires
Rounding of Personal Gross Earnings
D-ECHP, 1994-1996, non-anonymized data
rounding
personal wage
and earnings
1994
Count
no
%
rounding
personal wage
and earnings
1995
Count
%
rounding
personal wage
and earnings
1996
Count
%
609
16,5%
758
19,7%
647
17,3%
yes, 100
1977
53,5%
1986
51,7%
1954
52,3%
yes, 500
479
13,0%
510
13,3%
521
13,9%
yes, 1000
515
13,9%
504
13,1%
505
13,5%
yes, 5000
95
2,6%
79
2,1%
100
2,7%
yes, 10000
20
,5%
6
,2%
11
,3%
3695
100,0%
3843
100,0%
3738
100,0%
Total
Rounding of Personal Gross Earnings
SOEP, 1994-1996
rounding
personal wage
and earnings
1994
rounding
personal wage
and earnings
1995
rounding
personal wage
and earnings
1996
Count
1927
%
34,2%
Count
1940
%
34,4%
Count
1894
%
34,3%
yes, 100
2410
42,7%
2381
42,2%
2318
41,9%
yes, 500
584
10,4%
601
10,6%
588
10,6%
yes, 1000
577
10,2%
586
10,4%
588
10,6%
yes, 5000
113
2,0%
114
2,0%
105
1,9%
yes, 10000
30
,5%
24
,4%
34
,6%
5641
100,0%
5646
100,0%
5527
100,0%
no
Total
Comments: On personal level, the difference between the SOEP and the D-ECHP
is very easy to see. In addition to the absolute difference in rounding, it seems that
of those who do round, persons in the SOEP seem to round less likely to multiples
of 1000 and higher, more likely to 100. Thus, it can be assumed that rounding
causes a lower measurement error in the SOEP than in the D-ECHP.
page 4
Rounding of Personal Earnings vs. Earnings level
data: SOEP 1996 personal questionnaire
100%
90%
80%
type of
rounding
70%
60%
50%
40%
30%
20%
10%
0%
1-1750
DEM
17512250
DEM
22512750
DEM
27513250
DEM
32513750
DEM
37514250
DEM
42514750
DEM
47515250
DEM
5251> 7750
7750
DEM
DEM
34
yes, 10000
97
yes, 5000
33
yes, 1000
74
146
110
125
139
117
8
109
101
137
24
yes, 500
61
yes, 100
274
202
265
312
302
276
195
150
262
80
no
522
145
152
215
189
175
126
106
201
63
earnings level (grouped)
no
yes, 100
yes, 500
yes, 1000
yes, 5000
yes, 10000
Rounding of Personal Earnings vs. Earnings level
data: SOEP 1996 household questionnaire
100%
90%
80%
type of
rounding
70%
60%
50%
40%
30%
20%
10%
0%
up to
1750
- 2250 - 2750 - 3250 - 3750 - 4250 - 4750 - 5250 - 7750
32
yes, 10000
268
yes, 5000
yes, 1000
7751
and
higher
32
174
275
183
266
199
167
1
195
101
yes, 500
94
155
21
yes, 100
236
244
272
279
256
202
205
155
259
48
no
211
138
106
101
80
52
45
26
67
26
income level (grouped)
no
yes, 100
yes, 500
yes, 1000
yes, 5000
yes, 10000
page 5
Rounding of Personal Earnings vs. Earnings level
data: D-ECHP 1996 personal questionnaire
100%
80%
type of rounding
60%
40%
20%
0%
1-1750
DEM
17512250
DEM
22512750
DEM
27513250
DEM
32513750
DEM
37514250
DEM
42514750
DEM
47515250
DEM
5251> 7750
7750
DEM
DEM
11
yes, 10000
96
yes, 5000
23
yes, 1000
51
132
79
135
130
108
4
95
69
108
29
yes, 500
67
yes, 100
290
167
198
243
236
249
139
140
235
57
no
305
53
48
48
46
29
28
24
53
13
earnings level (grouped)
no
yes, 100
yes, 500
yes, 1000
yes, 5000
yes, 10000
Rounding of Personal Earnings vs. Earnings level
data: D-ECHP 1996 household questionnaire
100%
90%
80%
type of
rounding
70%
60%
50%
40%
30%
20%
10%
0%
up to
1750
- 2250 - 2750 - 3250 - 3750 - 4250 - 4750 - 5250 - 7750
34
yes, 10000
168
yes, 5000
yes, 1000
7751
and
higher
15
104
211
111
194
132
128
2
161
81
yes, 500
30
148
28
yes, 100
168
188
236
232
206
182
146
115
239
45
no
143
80
46
46
42
38
31
17
52
14
income level (grouped)
no
yes, 100
yes, 500
yes, 1000
yes, 5000
yes, 10000
page 6
Rounding in Fin-ECHP personal earnings data
Rounding of Personal Gross Earnings
SF-ECHP, 1996, responder data
responder rounding of
personal income
Count
424
%
9,6%
1044
23,5%
yes, 500
835
18,8%
yes, 1000
1510
34,0%
yes, 5000
251
5,7%
yes, 10000
373
8,4%
4437
100,0%
no
yes, 100
Total
Rounding of Personal Gross Earnings
Fin-ECHP 1996 (preliminary version)
yes,
100
yes,
500
yes,
1000
yes,
10000
no
up to 3250
16,9%
30,0%
16,4%
36,7%
100,0%
3251-4750
15,7%
26,3%
28,6%
29,4%
100,0%
4751-6250
15,1%
23,0%
10,2%
26,5%
6251-7750
9,0%
32,5%
30,2%
28,3%
7751-9250
8,9%
30,7%
13,3%
47,1%
9251-10750
6,2%
19,6%
23,4%
10751-12250
6,1%
17,4%
11,1%
65,4%
100,0%
12251-13750
9,4%
18,4%
37,1%
35,1%
100,0%
13751-15250
5,9%
13,5%
11,2%
34,9%
15251-16750
8,4%
14,8%
16,1%
60,6%
16751 and higher
5,3%
8,2%
10,2%
32,7%
15,5%
28,2%
100,0%
9,6%
23,5%
18,8%
34,0%
5,7%
8,4%
100,0%
Total
yes,
5000
SROUND3: rounding to multiple of "even" number in
1996 (cumulative = 100%)
Total
% within INCGRP3
25,1%
100,0%
100,0%
100,0%
50,8%
34,5%
100,0%
100,0%
100,0%
Comments: Compared to the german panels SOEP and D-ECHP, there is
absolutely (% of rounded values) and relatively (amount of rounding in NC) more
rounding in the Fin-ECHP. This may be due to the value of the national currency (1
EUR = 1.95583 DEM = 5.94573 FIM). An analysis of rounding in the ECHP with
respect to exchange rate is still at work.
page 7
Quality of Income Data: Rounding Behaviour (2)
Analysis
Impact of covariates
- gender
- panel participation
duration
Status
Remarks
just started
just started
Apparent differences
Separation of survey and panel effect
Effect of rounding
- FIN-ECHP (1996-2000)
vs. Register
deferred
deferred
Emphasis on stability (definition required)
blocked by data delivery (Autumn 2001)
second problem: monthly / annual data
- D-ECHP vs. SOEP
- UK-ECHP vs. BHPS
deferred
deferred
page 8
Quality of Income Data: Bracket Use
Analysis
Descriptive Analysis
- percentage
Status
Remarks
at work
Drop of rates from wave 1 to 3 in D-ECHP, LuxECHP (as predicted)
- pattern
Determinants for bracket use
Distribution of income within
brackets
at work
not started
temporal dependencies
Ordered probit model
Implication for Imputation package
- D-ECHP vs. SOEP
- Fin-ECHP vs. SOEP
not started
not started
Problems with comparability
Bracket Use in Panel Surveys
Number of Responses using Values or Brackets for Household Income
Dataset
SOEP
D-ECHP 1994
D-ECHP 1995
D-ECHP 1996
Fin-ECHP 1996
Lux-ECHP 1994
Lux-ECHP 1995
Lux-ECHP 1996
as Value
6554
4038
3902
3891
3186
523
526
597
as Bracket (%)
not an option
852 (17%)
682 (15%)
605 (13%)
870 (21%)
473 (47%)
408 (44%)
312 (34%)
page 9
1994
item nonresponse
1995
no
interview
9,3%
item nonresponse
0,4%
bracket
33,9%
Visualization of the
Change of Responder
Behaviour regarding
Bracket Usage for
Household Income
Data: D-ECHP 1994-1996
percent: Percentage of responders
starting in wave 1 with "bracket
usage"
bracket
52,5%
3,9%
value,
rounded
value
value, not
rounded
1994
1996
15,8%
0,4%
bracket
no
interview
item nonresponse
31,0%
bracket
49,0%
value,
rounded
3,9%
value, not
rounded
page 10
Visualization of the Change
of Responder Behaviour
regarding Bracket Usage
for Household Income
1995
1996
no
interview
Data: D-ECHP 1994-1996
upper chart:
conditional on bracket
usage in wave 2
5,3%
item nonresponse
0,7%
lower chart:
conditional on bracket
usage in wave 1 and wave 2
bracket
35,3%
bracket
54,7%
percent: Percentage of responders
starting in wave 1 with "bracket usage"
value,
rounded
3,9%
value, not
rounded
1994
1996
1995
no
interview
8,7%
0%
bracket
33,9%
bracket
item nonresponse
50%
bracket
37,8%
value,
rounded
3,5%
value, not
rounded
Download