Total

Male

Female

Age0-17

Age18-64

Age 65+

Cross sectional

·Over-coverage

·Under-coverage

·Misclassification

-information system

-interviewer

-For building up the questionnaire we took the blue print questionnaire of Eurostat as the basis (documents SILC055, SILC065 and EU-SILC65/02 Addendum II). The order of the questions and the groups (themes of) questions is taken from this blue print.  The majority of the questions are almost literally copied (and translated), other questions are changed, however, because experiences in Belgium gave better results posing the questions in another way (The questionnaires were developed in collaboration with the universities that have the experience of the ECHP/PSBH project in Belgium). 

Overall we had the impression that the working-experience of the interviewers with EU-SILC starts to pay of. In our opinion the basis data has improved since 2010. All new interviewers have to follow a two day formation. All trained interviewers followed a formation for an hour and half.

They both had to complete a test-interview before they could download their data. So we can be sure they can completely manage the use of the PC and that they know the questionnaire before they go on the field.

A training group for new interviewers consisted of minimum 5 to maximum 20 interviewers, and according to the size of the training group there were 1 or 2 trainers.

Even though the accent was given to the practical side of the training (getting to know the questions and mastering the CAPI-program by imitating interview situations), three manuals were distributed and explained during the training:

-A general manual (‘Manuel general aux enquêteurs’) containing information about the objectives of the survey, the organisation of the survey, legal and administrative aspects around the survey, fieldwork aspect (how to contact the household, how to introduce oneself, who answers which questions, time delays, …) and the content of the questionnaires.

-A second manual (‘Manuel contenu’) with all kinds of additional explanations and examples for certain questions/answers.

-A third manual (‘Manuel CAPI’) about the use of the portable PC for the SILC Computer Assisted Personal Interviews and about the data entry program itself.

The first day of the training there was half a day for learning about and discussing the first two manuals.  In the afternoon the trainees received their laptop and got to know the survey and the tool to carry out the interview in practice. One test-interview was simulated collectively.  The second day of the training a small part of the time was dedicated to testing to send the data electronically after carrying out the interview.  All the rest of the day interviewers practiced several interviews and interview situations with each other on the basis of household profiles that were given.  There was also a lot of time for questions and discussions in between these test-interviews.

At the end of the training sessions the instructors had a good image on the degree in which each interviewer ameliorated during the training and on the degree in which they mastered the work.  For certain interviewers two days of training was more than enough to master the work, for others it was necessary that they practiced some more at home on specific aspects of carrying out this survey (for example using of the CAPI-program itself, working on the content of the survey, …).  They were recommended to do so before carrying out their first real interview.  They were often also recommended to start interviewing one-person households.

• Skills testing before starting the fieldwork

Interviewers were selected from the interviewer database that Statistics Belgium has centralised for all the survey’s that are carried out by the institute.  For each interviewer a basic curriculum vitae is present in the database (mentioning for example for which surveys they have experience, their language knowledge, their knowledge of pc, …). A specific unit at Statistics Belgium (‘Unité Corps Enquêteurs’) is occupied with the selection of the interviewers for each survey; they have good contact with and knowledge of the interviewers.  They try to find the best interviewer for each of the geographical areas to cover for SILC.  This is not always an easy task because for certain geographical areas several interviewers are candidate, but for other geographical unit there are few or no candidates.  Note that interviewers in Belgium most often carry out this work as a second or casual occupation.

• Skills control during the fieldwork

During the fieldwork we controlled the work of the interviewers by looking at some of their completed questionnaires. We gave extra attention to all new interviewers and to some trained interviewers that we suspected to be less accurate. Remarks (positive as negative) resulting from these controls were immediately communicated to the interviewer so they could improve their way of working and interviewing.

• Number of households by interviewer

% of household having received an amount

% of household with missing values (before imputation)

% of household with partial information (before imputation)

% of household having received an amount

% of household with missing values (before imputation)

% of household with partial information (before imputation)

The definition of household membership is the same as mentioned in the Eurostat document EU-SILC065/03 about the description of target variables (Chapter ‘Units’).

F

F

F

F 

F

F

F

F

F

F

NC

F

F

F

F

F

F

F

F

F

F

F/L/P/NC

F/L/P/NC

F/L/P/NC

F/L/P/NC

Stratification and sub stratification criteria

Sample selection schemes

 Sampling units and 2-stage sampling in 2004

 In 2004, when organizing EU-SILC for the first time (ignoring the pilot survey in 2003), 2-stage sampling has been applied in each sampling stratum.

Stage 1 – Primary Sampling Units

The primary sampling units (PSUs) in stage 1 are the municipalities, or parts thereof in the larger ones. In each stratum, the PSUs in the frame are first descendingly sorted by average income; next, a fixed number of times a PSU is drawn according to a systematic PPS (probability proportional to size) selection scheme, where size is measured as the number of private households. This systematic sampling method generally causes some PSUs being selected repeatedly (e.g. Schaerbeek, a rather large municipality in stratum  BE10, turns out to be drawn 6 times).  In total, i.e. in all 11 sampling strata together, 275 PSU draws were made in 2004, once and for all (i.e. for the whole duration of EU-SILC).

Stage 2 – Secondary Sampling Units

The secondary sampling units (SSUs) in stage 2 are private households.  According to each single PSU draw, a group (generally of fixed size) of households is selected in this stage; notice that a group of households corresponds to each PSU draw.

In 2004, 40 households have been selected for each PSU draw (i.e. in each group); e.g. in Schaerbeek, 6 times 40 households were drawn. Systematic selection of households has been applied, after sorting the households in selected PSUs by age of reference person. Within each group, the selected households were numbered 1 to 40; households 1-10 constitute the first rotational group or replication, households 11-20 constitute the second rotational group or replication, and so on. The first replication was meant to participate in 2004 only, the second until 2005, and so on.

The initial household sample in 2004 was self-weighting, by the combination of (systematic) PPS sampling of sub-municipalities (PSUs) – size of PSUs being the number of private households – and (systematic) sampling of private households (SSUs), as explained.

Renewal of the sample by rotation, since 2005

Since 2005, a rotation scheme has been applied. Details for each year, from 2005 to 2010, can be found in the corresponding Quality Reports .

The rotation pattern is such that the overlap between samples in any two successive years is roughly 75%, and that the sample is completely renewed after 4 years. Hence four replications or rotational groups in each year, one of which is replaced the year after. Since 2005, each new replication remains in the survey during the next 4 years, and since 2007, each of the four replications is in the survey during four consecutive years.

At the start of 2011, the replication that is in the survey since 2007, is entirely (i.e. irrespective of whether the households are responding or not) dropped. The three replications which entered into the survey in 2008, 2009 and 2010, respectively, are retained (including their split-offs); the households belonging to these three replications will be designated ‘old’ hereafter.

The supplementary sample, i.e. the new replication that replaces the just dropped replication, is obtained by selecting, for each PSU draw, a fixed number of new households from the corresponding PSU. This selection is done again by systematic sampling, after sorting the households in each PSU on age of reference person. The number of new households for each PSU draw, is determined by considering some (expected) attrition of old households, some (expected) nonresponse for new households, and the required/desired minimum and maximum numbers of responding households, given some precision and budget constraints.

Hence, the (cross-sectional) sample of SILC 2011 consists of

•           “old” households: drawn between 2008 and 2010; and

•              “new” households: drawn in 2011, staying until 2014

Following Eurostat’s suggestion (see Document 065, WEIGHTING II. WEIGHTING FOR THE FIRST YEAR OF EACH SUB-SAMPLE), we replaced the homogeneous response groups (based on household size crossed with urbanity) ratio by a multiple regression model (based on the same dummy variables).  By “responding”, we mean only those households whose results were accepted (DB135=1).  Since 2009 we used logistic regression.

The file was split by NUTS1 and the following variables were used

-         Everywhere: Household size, recoded into the four values “one”, “two”, “three” and “four or more” (so three dummies)

-         Out of Brussels: DB100 = urbanity

-         In Brussels = BE10: median fiscal income of municipality

The regression produced a new variable “expresp”, allowing us to define

NRwei = INIwei/expresp

2.1.8.3 Attrition for the old households

Before “sharing” the 2010 weights, a correction for attrition should be introduced.  This year, we elected to perform this correction at the level of individuals, since a 2010 sample person either stays in the panel or leaves it (rotated out, left population, noncontact, refusal or inability to respond, while the structure of a household can change.   Note that all household characteristics (e.g. HH021) can be distributed to the members.

We separated the “Children” (for which only basic personal information from the R-file and the distributed H-file is available) from the “Adults” (present in the 2010 P-file as well), i.e. those persons born in 1992 or before.

In the children’s model, the following predictors (all, except the last, from the 2010 file – although this does not matter much for group A) were used, grouped by type :

A.       individual demographic information: age  from RB080, sex = RB090,

B.       housing information: dwelling type = HH010 and tenure = HH020

C.       household type: a limited number of dummies, as there is at least one dependent child;

D.       monetary indicators: we refrained from taking the equivalised income (outliers), but took a transform of it, as well as the dummy “poor or not” and the subjective ability to make ends meet = HS120

E.       sampling and rotation: number of years in panel (from DB075) and urbanisation (=DB100)

F.       one variable (paradata) related to fieldwork in 2010 (computed from HB040 and HB050)

For the adults, the same predictors were used, and moreover :

G.       variables from the P-file (related to education level and health);

H.       country of birth (dummy Belgium Yes/No)

were integrated.

We used logistic regression.

Recall that 11 sampling strata were used (provinces= NUTS2); we use 3 extrapolation strata (the 3 NUTS1 regions BRUssels=BE1, VLAanderen=BE2 and WALlonia=BE3)

Calibration model:

VLA, WAL:

SIZE4+(AGE8XSEX2)+PROV5       20 individual  + 4 household constraints

BRU:  

SIZE4+(AGE8XSEX2)                  16 individual + 4 household constraints

Prov = province where interviewed (differs from DB040 in two cases)

Individual constraints                              27=16+11          (age*sex + prov; note that each province belongs to one single region (extrapolation stratum), for the other two regions, the total is set to 0 and the condition is vacuous)

Household  constraints                4                      (size: "1", "2", "3 or "4 & more",)

N : 5910

Minimum: 89.24

Maximum : 5958.57

 Mean : 800.42 

Std Dev :  402.05

We (i.) corrected (not imputed – in fact) a greater number of cases and if correction was not desirable or possible, but information on a directly comparable variable was present anyway (see section on internal information above), we (ii.) resorted to direct imputation, via a regression model. 

i.          Corrections.

Corrections were also mainly done on basis of information in other comparable variables. Gross-net ratio of 12 - yearly income entered as monthly or vice versa - lead to simple corrections of the gross or the net, for example. 

ii.          Regressions.

If correction was not desirable or possible but information on a directly comparable variable was present anyway, we resorted to direct imputation, via a regression model, of the variable for which input was missing. Below we describe how this was done for net –gross imputation, which were the most prevalent instances of that sort. The method was extended, however, to other imputations (imputations of the 2011 income based on the current income, for example).     

Missing values on gross income variables (PY010G, PY020G, … and components) were, if collected, imputed on the basis of the corresponding net variables (PY010N, PY020N, … and components). The implementation of this imputation procedure was quasi-similar for almost all (income) variables on which it was applied.  The procedure implied 6-steps:

1.       Identification of the ‘reference cases’ (both gross and net collected) and identification of the cases to be imputed (net collected – gross missing).

2.       Calculation of the gross/net ratio for the reference cases. Cases with an extreme value on this ratio were excluded from further use in the procedure.

3.       Curve estimation of the relation (regression model) between gross and net income. The best fitting model (linear, logarithmic, quadratic, exponential) was being implemented. 

4.       Implementation of the regression model for the reference cases to identify outliers.

5.       Re-implementation of the regression model for the reference cases after removal of the outliers.

6.       Actual imputation step:  missing (gross) values are imputed on the basis of

a)       net values and

b)       the estimates for the relation between gross and net income assessed in the steps above.      

In step 1 the cases of which both gross and net income were collected are identified. We refer to these cases as ‘reference cases’ (step 1). The relationship between their net and gross income serves as reference for the imputation of the gross incomes for the cases where only the net was collected (cases to be imputed).

To avoid bias in this imputation model atypical reference cases (both outliers and errors) were identified and removed at several steps in the procedure (step 2 and 4).  

In step 2 (reference)cases for whom the ratio between gross and net income exceeded what can be considered typical for the taxation regime applicable to the income concerned, were excluded.

In the case of almost all variables the boundary value of this ratio was set at 2,5. This boundary was arbitrary chosen.

Scrutiny of the excluded cases, however, validates this value’s potential to discriminate between incomes which were subjected to real(istic) taxation and outliers or errors.

The latter category seldom counted more than a few percent of the total population in the survey and their gross/net ratio often exceeded the 2,5 considerably.

Further exploration also revealed that the exclusion of these cases from the procedure results in a dramatic increase of the fit of the regression model on which the imputation is based.

In step 4 outliers in the regression model were identified and removed using default regression diagnostics.    

The underlying probability model of the net-gross relation was assessed with ‘curve-estimation’ procedure (step 3). It can be hypothesised that in most taxation schemes this relation will not be linear as higher revenues will be subjected to disproportionate higher taxes. The concern therefore is that application of a linear regression model may lead to biased result. Step 3 is an answer to that concern, which turned out to be unfounded, however. In fact, for most variables the linear model fitted the data well. For a few variables the fit of the quadratic model was slightly better, however. Overall, and we underline this, the fit was very good and R-squares very high (always > 0.85)

The estimates of this regression model (step 5) served as direct input for the implementation of the actual imputation (step 6).

iii.         Other techniques.

Although we preferred the techniques above we were in some instances forced to resort to other techniques (due to lack of information – for example).

From 2007 onwards a measure for ‘imputed rent’ needs to add to the data.

Since 2005, we decided to work with the national rules of the tax authorities. The benefit for individuals of using a company car for private goals was not directly assessed at the interview but afterwards calculated by applying the applicable taxation rules.

The fiscal benefit of all nature that a person has - due to disposition of a company car for private goals - is calculated by multiplying a fixed amount of kilometres driven for private use by a coefficient. To calculate the latest we need the fiscal cylinder capacity of the car. This fixed amount of kilometres driven for private use is for the tax authorities 5000 km if the distance home-work is less than 25 km, and 7500 if it’s more than 25 km.

Since 2005, we asked directly the fiscal cylinder capacity and the distance between work and home. In case of non response of the cylinder capacity, we asked the mark, type and registration year of the car.  Than we had to use an imputation method.

Imputation: To calculate the cylinder capacity, we did the following. We assumed that a company car is mostly diesel driven. We looked up for each mark, type and diesel engine what the corresponding cylinder capacity is. If we had several cylinder capacities for the type of the mark, we calculated the weighted mean of the cylinder capacity. If there is not diesel version for a type of car, we did the same logic but than for petrol.

Once we had that we could easily find the corresponding fiscal coefficient. Than we only had to multiply it by the fixed amount of kilometres driven for private use to obtain the fiscal benefit of all nature

Example:

Type of car       Fiscal cylinder capacity      Forfait   Distance home work         Fixed amount      Fiscal benefit of all nature

Smart fortwo    5          0,1898   < 25 km           5000      949 €

Smart fortwo    5          0,1898   > 25 km 7500      1423.5 €