Socio-economische panelstudie van Belgische huishoudens (CSB-panel)

(Belgian Household Panel Study - CSP panel)

Administrative Unit responsible for the survey:

The CSP is a research center at the department of Sociology and Social Policy of the University of Antwerp (UFSIA). Funding for the survey is mainly provided for by the Belgian Government. The data are stored at the CSP at the address above. Copies of the original documentation and other documentation can also be obtained from the CSP. The following people, all at the above address, can be contacted for more information about various aspects of the survey: Prof. dr. Bea Cantillon (Director of the CSP), Mr. Rudi Van Dam, Mr. Karel Van den Bosch, Mrs. Diane Proost.

At the moment there are three waves of the Belgian panelstudy available (1985, 1988, 1992). The first wave of the panelstudy was conceived as a cross-sectional survey. In 1988 it was extended into a panel survey. The main purpose of the survey is the analysis of the income distribution, poverty and the effectivity of the Belgian social security system.

There are no restrictions on the use of this data by the public. A bibliography of the most important publications based on the CSP Panel is provided in section M. Also included in the bibliography in section M is a list of user documentation for the CSP data .

B. POPULATION AND SAMPLE SIZE, SAMPLING METHODS back

1) 1985 wave

The sample of the first wave of the panel was a stratified and clustered systematic EPSEM (equal probability of selection method) sample of private households.

The population of the survey consists of all private households, resident in Belgium. It therefore includes resident foreigners, and excludes people in institutions, as well as persons without permanent address. It is estimated that the survey-population covers more than 98% of the total Belgian population.

Sampling took place in two stages: first a number of municipalities were selected, secondly,within each municipality, a number of households were selected. All Belgian municipalities were grouped in 8 strata. First, municipalities were divided by region (Flanders, Wallonia, Brussels), secondly within the regions of Flanders and Wallonia, municipalities were stratified by average taxable income per head. In each of these regions, three strata were formed, which were equal in size as regards number of households. The city of Antwerp was treated as a separate, eight, stratum. From each stratum, except Antwerp, 10 municipalities were randomly selected, with a selection probability proportional to size (number of households), and with replacement. Because some municipalities were selected twice, and one even thrice, in all 61 municipalities were selected from a total of 589.

The sample consisted of 7.000 households. Each stratum was assigned a part of this sample, proportional to its size. This number was then equally divided over the selected municipalities in each of the strata (twice selected municipalities got a double share). In each municipality, the allocated number of household was selected from the municipal register of addresses, occupied by private households. This register contains the full address, and the name, age and sex of the reference-person of the household (a more neutral term for head of household). This was done by systematic probability sampling.

The procedure followed guarantees that ex ante each Belgian household had an equal probability to be selected. No groups were oversampled. However, for each selected address, one or two substitute addresses were taken in the same way from the same municipal register, to serve in case of nonresponse at the fist address.

2) 1988 wave

In principle, all members of wave one households were followed for the second wave, regardless of their family status in the first wave. It was attempted to collect information about all households in which wave one individuals lived.

Students going into universities are considered to be still part of their original household. This applies also to people who went to institutions like prisons and hospitals, if this is for a relatively short period. In fact, in these sort of cases, it is left up to the respondent to decide whom he or she regards as members of the household. In case of people moving to another town, the interview was assigned to another interviewer, who lived nearer (and Belgium is not a large country).

People who entered the population between waves 1 and 2, and who do not live in the same household as a wave 1 sample member, have no chance of being included in the wave 2 sample.

Interviewing started in September 1988 and ended in May 1989. It was administered by a mixture of personal interviews and mail questionnaires. About 500 households were not approached at all, because there was no name or address available. This state of affairs is partly explained by the fact that the 1985 survey was originally planned as a cross-sectional survey. Only afterwards came the idea to extend it into a panel.

Almost all of the names and addresses were checked by municipal services before interviewing started, which enabled us to correct many names and addresses, and to trace some movers to their present address.

3) 1992 wave

Same method as in 1988. However, to achieve a larger sample new households were added to the original panel sample in the '92 survey. These additional households were obtained via a new sample which had a design identical with the original panel sample.

Interviewing for the third wave started in December 1991 and ended in March 1992. It was administered by a personal visit by the interviewer after the interview had been announced by an introduction letter. Repsondents had the possibility to fill in the questionnaire themselves, in which case the interviewer only collected the interview after carefully checking if it had been filled in correctly.


C. MEASURES OF DATA QUALITY back

1) 1985 wave

Income information is always collected over the time interval that income from a certain source is actually paid out to households. This is monthly, except for study grants, which are paid out one a year. Data about labor incomes (wages, earnings of self-employed) and social security replacement incomes are collected on a personal basis. In the database they are available both on the household and on the individual level. Family allowances, study grants and social assistance have been asked for on the household level, and are not disaggregated.

About 13% (822) of households had missing data on at least one income component. In about 300 of these cases we had an estimate by the respondent of his total household income, and the missing income component was imputed from this, if the result appeared plausible. In the other cases, missing data were imputed using estimates of average income within classes. Control variables were: age, sex, (former) profession, region and position in household for labor income and pensions; sex, position in household and having ever worked or not for the other replacement incomes. For two categories (earned income of self employed and white collar workers) a hot-deck method was used, in order not to reduce the variance. If family allowances were missing these were estimated by a program incorporating the administrative rules for granting family allowances.

No government survey exists in which household and/or individual earned incomes are measured. Comparison with administrative data is very difficult, because they mostly measure gross income (before taxes), while we ask for net-income.

A research group at the University of Louvain (the HIVA) has estimated the distribution of

net-taxable incomes, on the basis of our survey-data, and has compared this to official statistics (table 1). It appears that both lower and higher incomes are overrepresented. These differences may be due to

1) underreporting of incomes to the tax authorities, and other forms of tax-evasion;

2) non-inclusion of certain people with low incomes in the official statistics;

3) different definitions of the tax-unit.

The age structure of the sample (individuals) has been compared to a population prognosis for 1985 of the National Institute of Statistics (NIS), made in 1983, on the basis of the 1981 general census (table 2). The distributions agree quite well for Flanders and Wallonia, except for the oldest age group. Brussels is very bad, which affects the CSP-distribution for the whole of Belgium.

Table 3 shows that, in comparison to the 1981 census, single people are underrepresented. More detailed analysis has shown that in particular elderly single women (mainly widows) are underrepresented (Brussels again appears to be a total failure).


TABLE 1:

Comparison of distribution of net taxable income: tax statistics vs. estimates on the basis of CSP-survey data.

Source: I. Nicaise a.o.: Methoden van Studiefinanciering, Deel III, HIVA, Leuven, 1987, p. 7.

In table 4 the labor-force participation of men and women according to the CSP-survey is compared to data from the labor-force survey by the NIS, held in 1985 on a large sample (> 70.000 individuals). Except for Brussels, there is close agreement.

TABLE 2:

Differences in age distribution between NIS-forecast and CSP-sample individuals (by region).

* deviation of distribution of CSP-sample from NIS distribution

TABLE 3:

Households by number of household members: CSP-survey 1985 and Census 1981.

TABLE 4:

Labor-force participation of men and women: comparison of CSP-survey with Labor Force Survey of NIS (1985).

B: WOMEN

b) 1988 wave


c) 1992 wave

Following categories appeared to have specifically low response rates between the second and the third wave:

- households from the Walloon part of the country
- households with two or more people employed
- households with a head aged 75+
- lower (standardised) income deciles
- households which had moved between the secon and the third wave
- households consisting of young and single people

To evaluate the quality of income data two comparisons were made:

1. the (simulated) gross income and taxes were compared with tax-statistics
2. social security allowances were grossed up and compared with (aggregate) administrative statistics

1. the (simulated) gross income and taxes compared with tax-statistics

N= The number of fiscal units with a gross income lower than the upper border of the fiscal decile

2. Comparison of (grossed up) social security allowances with administrative statistics.

Both comparissons should be interpreted with some caution as in the survey only monthly amounts are asked.
Specifically concerning table 2 caution is necessary for following reasons:

- Administrative data contain gross amount, whereas the survey results are net-amounts;
- the persons in collective households and institutions are excluded from the survey, but expenditure for this group is not deducted from the administrative amounts. Particularly in the pension-sector this may have a significant effect
- the administrative categories could not be completely reconstructed with the survey data. Some allowances for early retirement wre asked in the survey as pensions, in accordance with the repsondents perception, but in fact these allowances come under unemployment scheme.

Item non-response on some questions:

D. DATA COLLECTION AND ACQUISITION back

1) 1985 wave

The actual interviewing was done by a commercial firm (Dimarso - Gallup - Belgium). Interviewing began in May 1985 and ended in May 1986, and was wholly administered by personal visits. Interviewers had to try at least three times to contact a household. If they failed to contact a sampled household, or if it refused to cooperate, the interviewer had to try to use one of the substitute addresses. If these also did not respond, the interviewer was instructed to find a similar household (same neighborhood, same age-groups, same sex head of household), preferably the neighbors.

Eventually, 6.471 households were successfully interviewed. This is equal to 92,4% of 7.000 households, but this figure is not equal to the response rate, because of the various kind of substitute addresses. Table 1 shows the situation. In all 53% of respondents were on the original list of sample addresses and substitute addresses. 47% are selected by the interviewers themselves. The situation is worse in Wallonia and in Brussels. The households successfully interviewed are distributed as follows over regions (between brackets distribution of population): Flanders: 58,5% (55,4%), Wallonia: 32,2% (32,3%), Brussels 12,3% (9,3%). Flanders is overrepresented, Brussels is underrepresented.

No attempt has been made to relate the response rate to other information available in the sampling frame. However, addresses selected in various kinds of ways have been compared with each other. Households on substitute addresses selected by the interviewer are somewhat younger, bigger and richer than other households.

2) 1988 wave

Mail questionnaires were sent to all households, except the very old (head + 75 years) and households of which the head had only primary education. Households, who did not qualify for a mail questionnaire (about one third of the sample), as well as households who did not respond to it, were approached for a personal interview. The mail questionnaire was administered in the following way:

1st week: letter announcing the questionnaire
2nd week: questionnaire with accompanying letter
3rd week: reminder (printed out)
5th week: 2nd questionnaire.

Table 6 shows what response rates we had. "Main" households responded better than "new" households. In Brussels the response rate was very low. Flanders households were more likely to respond to the mail questionnaire; households in Wallonia responded better to personal visits. It is practically impossible to distinguish between geographically mobile and immobile households.

TABLE 6:

Responses rates in second wave.

*Including households which have not been approached, due to lack of name and/or addresses (about 500), and households who left the population (death, move to institution).

3) 1992 wave

Interviewing was administered by a personal visit by the interviewer after the interview had been announced by an introduction letter. Repsondents had the possibility to fill in the questionnaire themselves, in which case the interviewer only collected the interview after carefully checking if it had been filled in correctly.

E. WEIGHTING PROCEDURES back

1) 1985 wave

No weights have been developed for the first wave. Because, in principle, all households had an equal probability of selection, and information was collected about all individuals in each household successfully interviewed, the sample was in principle self-weighing for both

households and individuals.

2) 1988 wave

Because of the, possibly selective, non-response and the panel design, a system of weights was developed.

The weighing system applied here must perform two functions:

a. to correct for non-response.
b. to correct for unequal selection probabilities of households, created by the sample design.

In particular, follow-up rules state that all households in which wave-1 sample members live, must be included in the 2nd wave. This means that a household composed of a wave-1 sample member and a non-wave-1 sample member, has a double selection probability.

The population represented by the sample is the non-institutionalized resident population of Belgium, i.e. all private households. Leaving the population occurs because of death, entering an institution, and emigration. Leavers are represented without problem. People enter the population through birth, leaving an institution and (re)immigration. Births are well represented, because they mostly happen in private households. Immigrants and people returning from an institution are represented only if they join previously existing households.

A distinction was made between 'original' households, i.e. first-wave households, and split-offs. In the case of a broken-up household, the household that lived at the original address was regarded as the original household. Though this does not seem a very relevant criterion for analysis, it suits our present purposes, as the selection of a split-off depends on the successful interviewing of the corresponding original household.

To be able to calculate a selection probability for each second wave original household, we first had to distinguish between responses, non-response and households that were no longer part of the population. To identify households of which all members had left the populations and to distinguish them from non-response, we relied on municipal services, which have checked almost all available names and addresses. In this way we learned that 78 households had ceased to exist because of death or entry in an institution, and another 16 had emigrated. The first of these figures agreed well with estimates made previously,and therefore this information was considered to be correct and complete.

Thus, of all first-wave households 94 (1,5%) had left the population, while 2.806 (43,4%) did not respond, either because the address was lost or not correct, or because no member of the household was found at home or was willing to cooperate. Non-response includes 79 "problem-cases" of which the link between 1st and 2nd wave was uncertain. This leaves a response of 3.565 households (55,1% or 56,0% of all households still in the population), excluding split-offs.

The probability of response of original households was estimated using an additive model with Multiple Classification Analysis, and with several demographic, social and economic characteristics as independent variables. The results show that age, composition of family, employment of husband and wife and region have a serious influence on the probability of response.

The estimated probability of response of household i can then be calculated with the following equation:

Pi = P +b1. X1i + b2. X2i. ... bn.Xni

In which: P: average rate of response (0.56):

b1 ... bn: estimated coefficients

X1i ... Xni: dummy variables, indicating whether a certain characteristic is present or not.

If an "original" household included persons which had not been household members at the time of the first wave, the selection probability is greater than Pi, because the household might have been included through these additional persons, if these had been a member of another sample household. For various reasons, the probability of this actually happening was judged to be rather low, and set at 0.17, which is the total response rate of split-offs.

The number of such households is 143. The selection probability of split-off household (i.e. households, at least some of whose numbers were, at the time of the first wave, part of another household at another address) is a more complicated affair. It depends in the first place on the response probability of the part of the household that is regarded as the original household (further termed the "parent household"), because without its cooperation we could not be aware of the existence of the split-off.

Secondly, it depends on the response rate of the split-off household itself (including the probability that the parent household provides a correct address.

The probability of response of the "parent household" has been calculated as explained in the previous section. The average response rate of known split-offs was 0.30, which, assuming that non-responding original households did break up at the same rate as responding ones, corresponds to an overall selection rate of split-offs of 0.56 x 0.30 = 0.17. Because of the fairly low number of interview split-off households (128), this response rate was not further differentiated by subgroup. The selection probability of a split-off household i was estimated in the following way:

___

Pi = P(s) = P(p.h.)i

___

with: P(s): average response rate of known

split-off households (= 0,30)

P(p.h.)i: estimated response probability of parent household

(average: 0.56)

If the split-off household included non-wave 1-sample members, it was assumed that the selection probability through these persons was equal to the selection probability through the wave 1-sample members. Hence for such a household:

___

Pi = 2 x P(s) x P (p.h.)i

There are 101 cases which could not be matched with certainty to a wave-1 household. For these households, the probability of selection has been set equal to 0.56 (which means their weight will be equal to 1.00, see below).

To ensure that the weighted sample size is about equal to be unweighted sample size, which implies an average weight of about one, the weights W have been calculated with the formula:

Wi = 0.56 / Pi

The weights have been limited to the range 0.56 - 2.80. Three weights were recoded from below to 0.56; 42 weights were recoded down to 2.80.

c) 1992 wave

As the 1992 survey is the third wave of a panel survey, weights had to be developped to correct for panel attrition and for differential cross-sectional selection probabilities caused by the imposed follow-up rules. Weights were assigned to each survey household and to each person in the survey.

In a first step response probabilities were estimated on the household level, using a log additive model.

separate response probabilities were estimated for splitt-off households.

To obtain cross-sectional weights as weight-halving procedure is used for households with new members.

For analysts on the individual level the household weights were assigned to each household member.

On household level as well as on individual level, the weights were scaled such that the average value of weights equals 1.

The household and persons of the new sample were assigned a weight of 1.

Correction for Item Non-repsonse:

Different methods were used to impute values for item non-response on income questions.

In cases with a missing personal income and a known total household income the latter was used to deduct the missing personal income. In these cases the (missing) personal income is equal to the total household income minus income from other sources and income from other household members.

In the case of missing pensions, and with aknown pension in the previous wave and with an unchanged family composition, the pesnion of the previous wave was used to impute a pension in the 1992-file, taking account of price- and welfare adaptions.

In the case of missing social security allowances (allowances for the disabled, allowances for occupational disability, and unemployment allowances) a mean value within classes imputation was used because of the relatively small group size.

In all other cases a sequantial hot-deck imputation within classes was used.

Overall, 7.8% of the cases with apersonal income have an imputed value. 26.6% of the cases with a missing income had an income imputed with deduction from the total household income, 6% had a pension imputed on the basis of the pension in the previous wave, 6% had amean value imputation and in 61.3% of the cases the hot-deck procedure was used.

On the household level, in the case of item-non-respons on child allowances a simulation program was used to impute a value. 5.4% of the households with child allowances have an imputed value. The cases with an imputed income are flagged.

F. DETERMINATION OF SURVEY UNIT MEMBERSHIP back

The basic unit of aggregation used in this survey was the household. Households were defined as a housing unit comprised of people eating together, living of the same budget or sharing a large part of their income (including children living at home, working or not, students living outside the family, but still living of the same budget, hospitalized members of the family, aged persons living with the family, etc. ...). The survey unit head was the male, in the case of married or unmarried couple, in the case of a single person living with children, the male or the female, in all other cases, the person that by the respondent is considered as the head of household. Individuals other than the sampling unit can be identified. The relationship of each member to the head of household is encoded.

G. CHILDREN AND SPOUSES back

In this survey children are defined as all persons, not active and not elderly (in practice this amounts to all persons below 16 years, plus those between 16 and 25 years who are in full- time education). They are not necessarily descendants of the head of household and/or partner (grandchildren, nephews may be included). In this survey the definition of spouses includes persons who are legally married , and cohabiting partners of the head.

H. AVAILABILITY OF BASIC SOCIAL AND DEMOGRAPHIC INFORMATION back

In Table 7 are summarized the basic social and demographic information which is available in the CSP Panel..

Table 7
Availability of Basic Social And Demographic Information

I. AVAILABILITY OF LABOUR MARKET INFORMATION back

Labor force participation was measured as of the time of interview, and since last month. This and other labor market information is summarized in Table 8.2.

Table 8
Availability of Labor Market Information

J. AVAILABILITY OF GEOGRAPHIC INFORMATION back

It is possible to identify the geographic location of the sampling units in the CSP Panel. In the original survey the geographic location can be identified by:

1) region (Flanders, Wallonia and Brussels);
2) province and;
3) a codenumber of the municipality.

K. SOURCES AND AMOUNTS OF CASH INCOME back

Sources and amounts of income are recorded for a monthly period. Income sources and amounts are recorded for each person.

L. TAXES back

M. BIBLIOGRAPHY OF MAIN PUBLICATIONS BASED ON THE CSP PANEL back

Deleeck, H., Indicateurs de la securite sociale 1976-1985, in: Courrier Hebdomadaire du CRISP (Centre de recherche et d`information socio-politiques), nr. 139, 15 dec. 1986, 37 p.

Deleeck, H., Indicateurs sociaux et evaluation de la securite sociale en Belgique, in: Europe Sociale, nr. 3, 1987, p. 69-74.

Deleeck, H., Cantillon, B., Indicators of subsistance insecurity and the evaluation of social security in Belgium, in: The role of Research in Social Security, Studies and Research, International Social Security Association, Geneva, nr. 25, 1988, p. 89-108.

Deleeck, H., Research note: The adequacy of the social security system in belgium, 1976- 1985, in: Journal of Social Policy, 1, 18, 1989, p. 91-117.

Deleeck, H., De Lathouwer, L., Van den Bosch, K, Regional differences in the distribution of Social Security Benefits in Belgium: Facts and Causes, in: Cahiers Economiques de Bruxelles, Nr. 123, 3e trimester 1989, p. 265-310.

Cantillon, B., Mutations socio-demographiques et securite sociale, in: Bulletin de documentation, Belgian Ministry of Finance, november-december 1990, p. 228-257.

Cantillon, B., The socio-economic situation of single women in Belgium, Report for the EC- single women study, Centre for Social Policy, Antwerp, August 1989.

Cantillon, B., Soziodemographische Wandlungen und Soziale Sicherheit, in: Internationale revue fuer Soziale Sicherhet, 1990/4, p. 407-442.

Cantillon, B., Socio-demographic changes, income distribution, and poverty, bevolking en Gezin, 1, 1991, p. 95-122.

Deleeck, H., Cantillon, B., Meulemans, B., Van den Bosch, K., "Some longitudinal results of the Belgian Socio-Economic Panel", Journal of Income Distribution, Vol. 2, no. 2, 1992, p. 57-73.

Delhausse, B., "La Pauvrete en Belgique: Utilisation de donnees de panel", Universite de Liege, 1992.