Home

Wealth and Assets Survey User Guide

1. 220 HRP economic unit not complete one of 2 eligible adults missed either interview incomplete 230 No individual interviews with HRP economic unit but household interview completed Although the dataset exclusively consists of responding households not every individual in every household responds The variable Out1W3 indicates the interview outcome of individuals Full interview in person or by proxy Partial interview in person or by proxy Ineligible for interview child aged O to 15 Ineligible for interview adult aged 16 to 18 in full time education Eligible adult refused to be interviewed Eligible adult non contact DnuOBWN PE Please note Although individuals with an outcome code of 5 or 6 did not give an interview they can still be included in the analysis because their values for wealth component variables have been imputed Also analysts should be aware that although children have not been interviewed for this survey the data on children assets has been recorded against their person number in the household not against the adult who responded to the relevant questions in this section Longitudinal data linkage All final data files are linked files have a single variable to use for linking cases between waves For household linking there are separate variables for each wave each case may have up to three variables with a valid code For person level there is one variable used for matching a case
2. Face to face interviewers Interviewers working on the Wealth and Assets Survey undergo training in two stages prior to starting any WAS interviews Firstly they are provided with a home study pack to work through which provides detailed information on the purpose and design of the survey as well as the questionnaire content Following completion of the home study interviewers complete an electronic learning questionnaire or ELO This Blaise supported questionnaire is designed to test interviewer s knowledge of the survey and identify areas where interviewers require further support The results of the ELQ are submitted to the HQ field team for review ahead of a face to face briefing of up to 12 interviewers This briefing reviews the content of the home study pack in more detail and offers the opportunity for interviewers to ask questions The briefing day is tailored to address areas highlighted by results from the ELQ The briefing is led by one or two field managers sometimes with support from research and field team HQ staff Interviewers do not start WAS work until their field manager is assured that they are fully briefed and ready to undertake the survey Respondent contact Once the sample has been selected either from the small users Postcode Address File new cohort or by maintaining panel address details old cohort advance letters are issued to sampled households respondents Advance letters are issued approximately ten d
3. Plewis 2007 Non Response in a Birth Cohort Study The Case of the Millenium Cohort Study International Journal of Social Research Methodology 10 5 p 325 A Laurie H Smith R 8 Scott L 1999 Strategies for Reducing Nonresponse in a Longitudinal Panel Survey Journal of Official Statistics 15 p 269 16 Questionnaire Content Overview The Wealth and Assets Survey WAS collects data on a wide range of assets and liabilities that private individuals and households in Great Britain have The primary aim of the survey is to derive overall estimates of wealth and monitor how these change over time WAS broadly splits wealth into four categories 1 Financial wealth 2 Pensions wealth 3 Physical wealth 4 Property wealth The questionnaire is designed to collect relevant information across these four domains of wealth to provide aggregated measures of wealth but also to afford significant potential for analysis within these four domains The questionnaire is therefore both broad and detailed in coverage with a wide range of stakeholders interested in the data WAS provides The wave one questionnaire content was determined by the requirements of the WAS consortium of government departments at that time namely the Department for Business Innovation and Skills BIS Department for Work and Pensions DWP HM Revenues and Customs HMRC HM Treasury HMT and the Office for National Statistics ONS the Department for Com
4. or could have made contributions to at the time of the survey e defined benefit DB e additional voluntary contributions AVCs to DB schemes employer provided defined contribution DC e personal pensions The distinction between employer provided DC pensions and personal pensions is as reported by the respondent So for example if an individual had a Stakeholder Pension facilitated by their employer and chose to report that as an employer provided occupational scheme this is counted as an employer provided DC pension Conversely if an individual reported this simply as a Stakeholder Pension it would be included in personal pensions In addition to these four categories of current pension scheme wealth from five other types of pension was calculated pensions already in receipt e retained rights in DB type schemes e retained rights in DC type schemes e pension funds from which the individual is taking income drawdown e pensions expected in future from a former spouse How the wealth for each of these components was calculated is described in detail in the following sections Current defined benefit occupational pension scheme wealth Individuals could report up to two current defined benefit pensions The wealth in each of these schemes was calculated separately as described below and then summed to derive total wealth in current defined benefit DB occupational schemes Wealth in these schemes wa
5. 50 000 gt 99 999 7 100 000 or more Age Group 0 1 1 16 24 8 No Pension in Wave1 Wave2 2 25 44 3 45 59 Female Banded 1 Less than 2 500 45 64 Male Estimate Wave2 2 2 500 gt 4 999 4 59 74 Female 3 5 000 gt 9 999 65 74 Male 4 10 000 gt 19 999 5 75 5 20 000 gt 49 999 6 50 000 gt 99 999 Sex 0 1 1 Male 7 100 000 or more Wave2 2 Female tSample 3 month sampling time NS SEC 1 Professional frame Wavel 8 Wave2 2 Intermediate Wave2 3 Routine To impute missing data in wave two based on 4 Never worked rates of growth decay between waves donors 5 Unclassified were selected with reference points in wave one similar to the recipient record based on an Employment Sector 0 1 1 Private imputation class derived from the cross Wave2 2 Public classification of observed Banded Values in 3 Oth r wave one observed Banded Estimates in wave Education 0 1 1 Degree level two and tSample in both waves The category Wave2 2 Other level No Pension in wave one helped differentiate 3 Level unknown between new and established pensions 4 No qualifications Topic expert review also indicated that changes in Gross Salary and Employment Status were likely to contribute to the variance in rates of change between waves Consequently wave one and wave two data for these variables were included in the donor selection process Other notable variations in the processing strategy appli
6. AVCs Individuals who reported being members of an occupational DB scheme were asked whether they had made any AVCs and if so what the value at the time of interview of their AVC fund was Current AVC wealth is therefore simply defined as the fund value reported by the respondent at the time of the interview Definition of current defined contribution occupational pension scheme wealth Individuals could report up to two current defined contribution pensions The wealth in each of these schemes was calculated separately as described below and then summed to derive total wealth in current defined contribution DC occupational schemes This procedure was also followed for those who reported that their employer provided scheme was a hybrid scheme or that they did not know the type of scheme Individuals were asked to report the value of their fund at the time of the interview and were encouraged to consult recent statements where available Current occupational DC pension wealth is therefore simply defined as the fund value reported by the respondent at the time of the interview Definition of current personal pension wealth Individuals could report up to two current personal pensions current being defined as schemes to which the individual was or could have been contributing at the time of interview The wealth in each of these schemes was calculated separately as described below and then summed to derive total wealth in personal pe
7. an in built hierarchical block structure that effectively makes all questionnaires modular The ability to handle the associated routing of a modular questionnaire is core to Blaise s architecture In addition to its hierarchical block structure Blaise also allows the creation of blocks which can be accessed in parallel allowing interviewers to switch out of one set of hierarchical blocks to another set This provides valuable flexibility as it for instance allows an interviewer to pause an interview with one household member initiate an interview with another household member e g a household reference person and then resume the interview with the original household member at a convenient time in the future e Blaise meets the requirement of being able to split the sample geographically or by sample identifiers Separate questions can be allocated to these different sections of the sample or to randomly selected sub samples of different sizes e Handling complex routing including loops and repeated events applying automatic logic and consistency checks in real time during the interview and using text fills where required are all core to Blaise s architecture They are functions that we make extensive use of on the Wealth and Assets Survey Blaise allows interviewers to exit and restart interviews at any point which allows interviews to be suspended and resumed The Wealth and Assets Survey questionnaire records the length of
8. d tof m m deviati variation on W1 W2 43 338 1 341 642 47 9 203 3 900 longitudinal W1 W2 W3 28 696 2 026 1 208 59 6 267 7 000 longitudinal W2 W3 31 472 1 870 1 236 66 1 155 7 000 longitudinal Cross sectional Weights A W3 pseudo cross sectional weight has been created pseudo because the data used contains any W3 respondent regardless of the sample that they were selected in Therefore some of the sample that were selected in W1 is not representative of the W3 population The original panel and the new panel have had weights calculated separately The two panels were then combined Cross sectional weights for the original panel There are several subgroups within the cross sectional population and so there were many different methods implemented when producing the cross sectional weights Those who had previously responded had their most recent weight as their base weight For cases responding for the first time the weights were created as follows 35 W3 SSMs including births based on a weight share derived from the base weights for individuals within their household W3 EOSMs original design weights constructed as the inverse of the selection probabilities The final stage for each subgroup is a rescaling of the weights to a specified total Each of these steps are described in more detail throughout this document The first stage for constructing the W3 cross sectional weight was to assign all previously re
9. have a higher probability of selection than others This reflects the fact that wealth has a heavily skewed distribution with a relatively small number of addresses holding considerable wealth This skewed distribution of wealth and the fact that it is often harder to secure response from wealthier households is the reason for the over sampling of wealthy addresses For year one of wave one addresses identified as having high wealth were 2 5 times more likely to be sampled than other addresses This factor was increased to 3 0 for the second half of wave one in order to further increase the number of achieved interviews with high wealth addresses High wealth addresses are identified after the postcode sectors have been established A limited amount of information is available about the type of household resident at a particular address on the PAF and what is generally available relates to the area around the address rather than being specific to an address However HMRC collects data on income and certain components of wealth in order to administer the tax system and the Self Assessment regime Data from HMRC on tax returns at an address level in conjunction with 9 average FTSE350 dividend yields from the previous calendar year are used to estimate the value of share holdings at a household level Those addresses estimated to be in the 90 percentile of shareholding value were then oversampled at a rate of 2 5 wave one or 3 0 waves th
10. independently test the questionnaire along with staff in ONS Survey Operations team Questionnaire testing is done every month prior to the questionnaire scatter for the next fieldwork period 20 Editing An extensive range of computer edits were applied to both the household and individual questionnaires during data entry in the field and to the aggregate data file in the office These edits checked that o logical sequences in the questionnaire had been followed o all applicable questions had been answered e specific values lay within valid ranges e there were no contradictory responses e that relationships between items were within acceptable limits Edits were also designed to identify cases for which values although not necessarily erroneous were sufficiently unusual or close to specified limits as to warrant further examination Once an interview had taken place the WAS data were transmitted back to ONS and were aggregated into monthly files Further editing occurred at this stage and included e recoding text entries if an appropriate response category was available e investigating interviewer notes and utilising the information where applicable e confirming that overridden edit warnings had been done correctly e broad data consistency checks The next stage involves checking that the routing of the questionnaire output is correct using a process referred to as base checks SPSS programmes are run to emulate the r
11. number of addresses included on the sampling frame in Great Britain For an address in the predicted high or low wealth stratum respectively the selection probabilities are nN 3X Npsy 12 P high stratum address e NM 2 5xM N 13 P lowstratum address b psu 7 N MP 25xM where ns is the number of addresses selected from each PSU 26 in year 1 and 13 in year 2 M and M are the number of addresses in the low and high stratum in the ith PSU The design weights for the sampled addresses are then the reciprocal of the appropriate address selection probability Once the new panel have design weights a non response adjustment was applied Non response can bias the estimates if not accounted for therefore responders are weighted up to represent the non respondents from the new panel sample This was carried out in a similar way to the attrition modelling in the longitudinal weights A logistic regression model was used to produce the response propensity for each case denoted by or see formula 1 The adjusted weights for the new panel were calculated as The term address is taken here to refer to the delivery point as listed on the Postcode Address File sampling frame 38 where UN represents the predicted probability of response from the new panel using a logit model with region output area classification and quarter interviewed as predictor variables and w is the design weight Calibrating the cross
12. of children plus the value of endowments purchased to repay mortgages less the value of non mortgage debt Some points to note e informal financial assets exclude very small values less than 250 e money held in Trusts other than Child Trust Funds is not included e financial liabilities are the sum of current account overdrafts plus amounts owed on credit cards store cards mail order hire purchase and loans plus amounts owed in arrears e private pension wealth is the sum of the value of current occupational pension wealth retained rights in occupational pensions current personal pension wealth retained rights in personal pensions Additional Voluntary Contributions AVCs value of pensions expected from former spouse or partner and value of pensions in payment Note that while net property wealth physical wealth and net financial wealth are calculated simply by adding up the value of assets minus liabilities if applicable for every household in the dataset private pension wealth is more complicated because modelling is needed to calculate the value of current occupational pension wealth retained rights in occupational pensions etc for each household As with all models the results depend on the assumptions made 46 Private pension wealth measures Nine separate components of private pension wealth were calculated based on the WAS survey responses There were four categories of pension to which respondents were making
13. sectional weights Once all cases in both panels have a weight it is necessary to calibrate the cross sectional weights to population totals at W3 As the sampling design was different for the two years in the new panel these were also considered as two separate samples to reduce the variance of the weights Therefore calibration was carried out three times once for the original panel once for year one of the new panel and once for year two of the new panel The aim of the cross sectional weights is to create a single weight to cover both households and individuals In order to achieve this aim an integrative calibration Lemaitre amp Dufour 1987 approach was used simultaneously to create both household and person level W3 pseudo cross sectional weights This results in all people in the household having the same weight which is also the household weight The population totals to which the weights were calibrated were based on interpolations of ONS mid year estimates taken from the midpoint of the W3 fieldwork period June 2011 The result at this stage is three sets of cross sectional weights each calibrated to the W3 population totals one for the original panel and one for each year of the new panel As users will want to carry out analysis on the cross section as a whole it is necessary to combine these three sets of weights Combining the cross sectional weights In order to do this the constructed weights for individuals in
14. separately The following table provides a detailed breakdown of the outcome of cases included in the set sample for both waves one and two Outcome Wave one Wavetwo Wave 3 old Wave 3 new W3 all Issued cases 61917 32195 25234 12683 37917 Eligible cases 55835 29584 21397 11297 32694 Co operating households 30511 20009 15517 5734 21251 Non contacts 3889 2717 1503 988 2491 Refusal to HQ 3805 1268 809 876 1685 Refusal to interviewer 15397 4527 2868 3296 6164 Total Refusal 19202 5795 3677 4172 7849 Other non response 1770 1063 700 403 1103 Response rate 55 68 73 51 65 Non contact 7 9 7 9 8 Refusal to HQ 7 4 4 8 5 Refusal to interviewer 28 15 13 29 19 Other non response 3 4 3 4 3 13 Keeping in Touch WAS is a longitudinal survey that follows all adults interviewed in wave one original sample members or OSMs The survey is biennial i e two years in between each interview WAS like other longitudinal surveys experiences attrition which may occur for inevitable reasons such as death or for reasons that can be minimised such as failure of tracing failure of contact or refusal The longitudinal design of WAS requires following OSMs over time in order to be able to measure changes in wealth It is evident that tracing and following sample members becomes difficult when circumstances of sample members in particular their location change over
15. similar to those outlined in Table 1 In contrast for respondents with longitudinal data the processing strategy was tuned more towards the observable interdependencies and rates of change in the data between wave one and wave two To this end when imputing each variable respondents with longitudinal data were divided into four imputation groups as outlined in Table 2 Table 2 Wave one and wave two longitudinal Imputation groups Data Status Wave one Wave two Imputation Group Observed Observed Potential donor 0 0 Missing Missing Missing both Waves M M Missing Observed Missing Wave1 M 0 Observed Missing Missing Wave2 O M For each variable potential donors were selected only from records with valid observations in both waves 0 0 When imputing values for respondents with data missing in both waves M M discrete values for both waves were drawn from a single donor This strategy served to preserve any implicit interdependencies between waves for categorical data and any implicit rates of growth and or decay for data with continuous characteristics To maintain the principle of the longitudinal processing strategy when imputing missing data in records where data was observed in one wave but missing in the other M O or O M categorical data was treated slightly differently than continuous data For categorical data a discrete value observed in one wave was employed to serve as a constraint on donor selection in the sa
16. tested on a variety of collection vehicles WAS is a relatively long and complex survey and reporting errors may also have been introduced due to interviewer and or respondent fatigue While efforts were made to minimise errors arising from deliberate misreporting by respondents some instances will have inevitably occurred Lack of uniformity in interviewing standards can also result in non sampling error as can the impression made upon respondents by personal characteristics of individual interviewers such as age sex appearance and manner In ONS thorough training programmes the provision of detailed supporting documentation and regular supervision and checks of interviewers work are used to encourage consistent interviewing practices and maintain a high level of accuracy One of the main sources of non sampling error is non response which occurs when people who were selected in the survey cannot or will not provide information or cannot be contacted by interviewers Non response can be total or partial and can affect the reliability of results and introduce a bias The magnitude of any bias depends upon the level of non response and the extent of the difference between the characteristics of those people who responded to the survey and those who did not It is not possible to accurately quantify the nature and extent of the differences between respondents and non respondents However the level of non response bias was mitigated through ca
17. time To minimise attrition caused by the loss of sample members due to the failure of tracking WAS has a number of measures implemented in the survey design to maximise the likelihood of contact being made with the sample member at the next wave Firstly the WAS questionnaire asks respondents at the interview to confirm their address details as well as further contact details such as phone numbers email address and contact details of two nominated persons not resident at the same address that are authorised to provide ONS with the respondent s new address in case the respondent has moved and cannot be traced Secondly a few weeks after the interview all respondents receive a Change of Address card together with the posted incentive alternatively this will be sent by email which aims to encourage respondents to inform the ONS if their contact details change Thirdly a brief telephone interview is conducted prior to the next wave s interview This telephone interview is referred to as the Keep in Touch Exercise or KITE During this interview information about household members as well as their address and contact details are confirmed or updated It provides the opportunity to identify movers from the household and their new contact details as well to identify joiners to the household A Portanti M Attrition on Longitudinal Survey Literature Review ONS Working Paper Social Survey Division November 2009 pg 2
18. with the linked data The following person level flags are only included on the person level datasets Type Indicator for linkage status 1 W1 W3 Linked cases regardless interview eligibility and response status 2 W2 W3 Linked cases regardless interview eligibility and response status 3 W3 HAK Joiner Individual joined the household when keep in touch exercise was conducted 4 W3 HAD Joiner Individual joined the household when debtor survey was conducted 5 W3 New respondents Individual joined the household when W3 interview was conducted 6 W3 New Household Individual is part of a household that responded at W3 for the first time 7 Individual no present at W3 This person was part of a responding household in W2 but left the household at W3 and did not respond 8 Household not present at W3 Individual was part of a responding household in W2 but the whole household did not respond at W3 P_Flag1W3 Flag for wave member status 1 LOSM Longitudinal original sample member individual was a member of a responding household in W2 and W3 2 EOSM Entrant original sample member individual was a member of a responding household in W3 but household did not respond in W2 3 SSM Secondary sample member individual was not a member of any household in W2 but joined a longitudinal household in W3 4 NSM Non responding sample member individual was a member of a responding household in W2 but left t
19. 1 207 877 72 7 69 9 999 40 The mean weight for W3 is quite similar to that of W2 the slight decrease is due mostly to the increase in the sample size from W2 to W3 The variation in the weights has increased reflecting the additional non response and unknown eligibility adjustments However it is clear that when the sample is combined proportional to effective sample size as appose to achieved sample size the variation is decreased as desired Table 5 4 2 Descriptive statistics for the W3 cross sectional weight combined proportional to achieved sample size split by sub samples Sub sample n Mean Standard Coefficient Minimu Maximu deviatio of variation m m n All cases 49 447 1 207 963 79 8 74 9 999 Longitudinal cases 32 485 1 436 1 067 74 3 74 9 999 W3 SSMs 1 697 1 391 1 177 84 6 74 9 999 W3 EOSMs 1 420 669 178 26 7 205 1 302 New panel 13 845 700 233 33 3 149 1 908 The mean weights of the different subgroups vary considerably because of the different adjustments applied to certain groups and the length of time the cases have been in the sample i e the longitudinal cases have the largest mean weights as most of these respondents have been in the sample since W1 and so have had unknown eligibility status and non response adjustment applied twice In line with this the variation is larger for those cases who have been in the sample the longest Table 5 3 Descr
20. a Office for ANS National Statistics Wealth and Assets Survey User Guide CONTENTS OVERVIEW A aa di 3 USING WEALTH AND ASSETS SURVEY DATA cscscsscscessccccscsccscscscess 4 CONTENT OF DATA FILES A oia pic 4 VARIABLE NAMING CONVENTIONS ssseeeceeececscececseeeeesesesceeeeseseseseeceeeeseseseseseseseeeecececececececececececeeeeececeeeeeeers 4 WEIGHTS Stes stoves aeaolodes A a A a an 4 INTERVIEW OUTCOME CODES a dd dd ti 5 LONGITUDINAL DATA LINKAGE cscceecedccoacceccessccccccoascedecoaecedcaeadcdacaseecdcdecsscedaceseceddecs ccccessccdcdecstccdeceseceddacs ccds 6 LONGITUDINAL ELLAS a a as 7 VARIABLE SPECIFIC NOTES esses ccscsdacssesscccccistalsciscsvsccedadievssccssccswedecevedess 8 SURVEY DESIGN iii iii init ea 9 SAMPLING STRATEGY coe cceeusisicetececcdosesdia cc ceetstuss iaa cairo Tis 9 SAMPLESIZES OFEACH WAVE a oa iii abi 10 AVES UC UREA poo 10 MODE OF DATA COLLECTION 0 00 a i 11 Mainstage TS ARA itiiti sute ba cadaudeeicavate sae decddcdvabe CasutandcdcaVatecacaddaddddavutetacalantenda dace 11 Keep in Touch Exercise int rviCW csssccccccccsseessssscesecsseesesseaesecsceesesseaesecsesesessaaesessesseessaasceseeseeenes 11 FIELDWORK PROCEDURES cui id 12 INTERVIEWER TRAINING a Ae 12 Generic interviewer trQiNiNQ ococonininonononnnnnnnnnnnnnnnananananarararnnr 12 SUTVEV SPECITIC EOI Asia A A A Eat ana iveieas lalecnete 12 RESPONDENT CONTACT E a A Ei aa cito 13 FIELD SAMPLING PROCEDURES 3 cccccceranacdec
21. anel was calculated The sample size was then divided by the design effect to create the effective sample size The three panels were then combined proportional to their effective sample size by multiplying by the factors Ogg for the original panel and 9 and 0 for the new panel where newy n n es orig 0 newy1 orig newy1 newy2 orig newy1 newy2 ni 0 newy2 O y o newy nf n n orig newy1 newy2 As this method gives more weight to those panels with a smaller variance the new panel have more power and the old panel slightly less This process results in the sum of the weights changing quite dramatically Therefore the weights are calibrated a final time Descriptive statistics for the resulting weights are given in Table 5 4 1 as for achieved sample size and es for effective sample size The equivalent statistics for the W2 cross sectional weights are also provided for comparison purposes Table 5 4 2 and table 5 4 3 contain descriptive statistics of the weights broken down by key subsamples of the population when the panels were combined in proportion to achieved sample size and combined proportional to achieved sample size respectively Table 5 4 1 Descriptive statistics for the W3 cross sectional weight Weight n Mean Standard Coefficient Minimum Maximum deviatio of variation n W2 46 347 1 277 730 57 2 106 3 700 W3 as 49 447 1 207 963 79 8 74 9 999 W3 es 49 447
22. ariables to each person in the household Note Person level variables cannot be added to the household file unless they are aggregated first Linking End User Licence EUL data As the EUL datasets are anonymised the variables HHSerial and PIDNo have also been anonymised To link household files the variable Case replaces HHSerial therefore Case will need to be used when linking cases To link person files the variable Person replaces PIDNo therefore variables Case and Person will need to be used when linking cases Linking using EUL datasets e Use CaseW1 CaseW2 and or CaseW3 for household linking e Use PersonW1 PersonW2 and or PersonW3 and CaseW1 CaseW2 and or CaseW3 for person level linking To add W2 variables to the W3 person file keeping only W3 cases e Sort W2 by CaseW2 and PersonW2 e Sort W3 by CaseW2 and PersonW2 e Match W2 and W3 using PersonW2 and CaseW2 to MATCH e This will add linkable W2 cases to the W3 file and add W2 variables to W2 cases e Select the required cases e g for W3 cases including linked W2 cases use CaseW3 gt 0 To add W3 variables to the W2 household file keeping all cases e Sort W2 household file by CaseW2 e Sort W3 household file by CaseW2 e Match files using CaseW2 e This will produce a linked W2W3 file with W2 and W3 variables Longitudinal Flags A number of longitudinal flags have been produced that may help to understand changes in the data when conducting longitudinal analysis
23. asset liability x was followed by a question to specify the value what is the amount income expenditure of asset liability x In cases where an exact amount was not known participants were asked to provide a banded estimate from a range of bound values such as 0 to 100 101 to 500 and so on For imputation the structure of the survey questions gives rise to several important distributional properties in the data Data from routing questions are categorical Data from amount income expenditure questions can be highly skewed Furthermore distributions are often characterised by discrete steps or clustering This can emerge through constraints imposed by implicit laws or regulations governing the absolute value of an asset or liability or through respondents able only to provide a banded estimate The key analytical aim of the survey was to provide longitudinal estimates of change over time as well as cross sectional single year estimates To meet this aim the imputation must account not only for the distributional properties of the data associated discretely with each variable but also the distributional properties of the rate of growth and or decay over time At this point data users should be aware that the previously released wave one data only included imputation for item non response Over the course of processing wave two data the decision was made to also impute missing data from unit non response to minimise the underestimation of
24. ays prior to the start of the monthly fieldwork period The advance letters are intended to inform eligible respondents that they have been selected for an interview provide information on the purpose of the interview explain the importance of respondent s participation and to provide contact details in case eligible respondents want to find out more New cohort households are issued one advance letter addressed Dear resident which assumes no prior knowledge or involvement in the survey For the old cohort each eligible respondent is sent an advance letter addressed specifically to them thanking for their help in the previous interview and inviting them to take part again The exception to this is the old cohort where the respondent was a proxy interview in the previous wave these respondents are sent a named advance letter but the letter assumes no prior knowledge or participation in the survey ONS recognises that some sectors of the community can be difficult to contact These include but are not limited to metropolitan areas flats London ethnic minorities and gated estates ONS recently reviewed and updated the interviewer guidance on calling patterns designed to maximise contact This strategy is known as Achieving Contact Efficiently and is underpinned by a Calling Checklist The calling strategy which achieves the highest contact rate at the lowest cost is to vary calling times Many households will be easily contacted withi
25. ct follow up propensity first for known eligibility status and second for aresponse This gives us an estimated propensity for each case denoted by Generically i e ignoring subscripting this is calculated as I _ exp P x 1 exp B x where is a vector of coefficients estimated by the regression model and X is a vector of response predictors in the regression model The first model predicted the log odds of known to unknown eligibility status using a set of characteristics taken from the W2 survey data and using the W2 weight in the analysis In many cases both respondents and non respondents to W3 have data from W2 so a rich set of response predictors is available The unknown eligibility status weights were calculated as follows oO oO 2 w gt lEs where A is the predicted probability that eligibility status at W3 is known from the logit model regressing eligibility status on various W2 individual and household level characteristics for person i S3 is the sample of people who have a known eligibility status at W3 This is those who are 1 W1 W3 longitudinal cases for the survivor weights or 2 W2 W3 longitudinal cases for the T 1 to T weights These are the cases which are weighted up to represent the cases with an unknown eligibility status at W3 This often but not exclusively occurs when interviewers are unable to trace people who have moved address either whole households or household spl
26. d be calculated from the many different combinations of the waves It would not be appropriate to calculate all possible sets of weights and we have considered how to make the number of weights we calculate manageable Additionally limiting the number of weights produced minimises the chance of the weights being used incorrectly We sought to choose those weights which will be most useful to users We propose from W3 onwards to produce three types of weights calculated at each wave these are as follows longitudinal weight for the survivors from wave 1 to wave T longitudinal weight from wave T 1 to wave T cross sectional weight for wave T Figure 3 1 indicates the types of respondents which are included in each of the weighting strategies at W3 Wave 1 Wave 2 Wave 3 SSM W3 A re rer gree Tere EOSM W3 KEY OSM Original sample member SSM Secondary sample member EOSM Entry original sample member W Wave Figure 3 1 Constituent respondent groups in each of the three weighting procedures At W3 the longitudinal weight for the survivors consists of responders to all three waves i e W1 W2 and W3 indicated by the red box in figure 3 1 The longitudinal weight for the latest two consecutive waves is applied to all responders in W2 and W3 demonstrated by the purple box in figure 3 1 This includes OSMs from W1 as well as SSMs and EOSMs from W2 Finally the cross sectional weight incorporates all responders to W3 de
27. d on NS SEC 0 1 1 Professional observed data tSample was also significant 2 Intermediate as research had indicated that private 3 Routine pensions were particularly sensitive to 4 Never worked economic trends over a short time frame 5 Unclassified The matching variable set consisted of variables related to the observed data Employment 0 1 1 Private identified through modelling and domain Sector expert review Annual Gross Salary and 2 Public Employment status were given higher weights 3 Other when calculating the distance between the recipient record and the potential donor as Education 0 1 1 Degree level the strength of association was stronger for 2 Other level these variables 3 Level unknown 4 No qualifications 1 Applied only to cross sectional data where the respondent was new to the survey and did not have observed data for other waves Typically the final potential donor pool was set to contain between 10 and 20 records It is important to note that through the first stage of constructing a potential donor pool the two nested processes used to establish this pool provide an implicit distributional model of the frequency range and variance of the set of discrete values observed in the data for records with characteristics similar the record being imputed In the last stage of the process the final donor was selected at random Consequently the probability of a particular category or value being selected was proportional
28. dded at wave 3 Interviews in all waves were conducted using Computer Assisted Personal Interviewing CAPI Wave one achieved approximately 30 000 household interviews wave two achieved approximately 20 000 household interviews and wave three achieved approximately 21 000 household interviews The economic well being of a households is often measured by its income and yet a household s resources are composed of its stock of wealth as well as its flow of income To more fully understand the economic well being of households it is necessary to look beyond measures of household income The WAS addresses this gap in data about the economic well being of households by gathering information on the ownership of assets financial assets physical assets and property pensions savings and debt The WAS is funded by a consortium of government departments Department for Work and Pensions HM Revenues and Customs HM Treasury Office for National Statistics and the Scottish Government Fieldwork is undertaken by the Office for National Statistics Using Wealth and Assets Survey data Content of data files The data are split into two linked files 1 a household level file containing all property and physical wealth component variables as well as all derived variables DV used for the calculation of aggregated household wealth and income 2 A person level file consisting of all person level financial wealth pension wealth and income compone
29. ed to the Wealth and Assets data described to this point were associated typically with samples too small to implement imputation classes based on complex multivariate cross classification In such cases variables that would have been included in donor selection as an imputation class were included instead as a matching variable Accordingly the weights applied to the matching variables were adjusted to best suit a preferred priority order In extreme cases where for 29 instance a variable contained less than twenty observations and a small number of missing values imputation was based on deterministic editing The range and variance of values imputed this way was guided by topic expert review and was often based on the mean median or mode of the observable data Quality Assurance and Evaluation Without exception the imputed data for all Wealth and Asset variables was examined and tested before being formally accepted The overarching aim of each evaluation was to ensure that the distributional properties of the observed data had not been distorted inappropriately by the imputation process Fundamentally evaluation was based on a comparing the observed data prior to imputation with the fully imputed data In all cases any notable departures from the observed data based on statistical measures such as shifts in central tendency or variance and or the introduction of unexpected changes in the shape of the distribution had to be justified Ju
30. esign based weights Users can re scale and normalise if they so wish in order to get the weights to sum to the sample size However the majority of popular statistical packages available on the market now account more accurately for the weights so we recommend using survey based procedures where possible when using weights in the analysis of survey data It is important to note that using the weights will help to reduce bias However reducing bias comes at the cost of increasing variance of the estimates As the variance of the weights increases so too does the estimated sampling variance 42 Data Quality All reasonable attempts have been made to ensure that the data are as accurate as possible However there are two potential sources of error which may affect the reliability of estimates and for which no adequate adjustments can be made These are known as sampling and non sampling errors and should be kept in mind when interpreting the WAS results Sampling error Sampling error refers to the difference between the results obtained from the sample population and the results that would be obtained if the entire population were fully enumerated The estimates may therefore differ from the figures that would have been produced if information had been collected for all households or individuals in Great Britain One measure of sampling variability is the standard error which shows the extent to which the estimates should be expected to
31. f a suitable donor In general donors were selected based on information specified by other auxiliary variables in the data Typically auxiliary variables are employed to constrain donor selection in two ways Primarily they serve to identify donors with similar characteristics as the respondent with missing data Importantly the auxiliary variables should be related with the data observed in the variable currently being imputed to help estimate accurately the missing value Auxiliary variables can also be applied to ensure donor selection is tuned towards the key analytical aims of the survey and planned outputs For all imputed variables in the Wealth and Assets Survey appropriate auxiliary variables were identified through traditional regression based modelling supplemented by guidance from experts familiar not only with a particular subject domain but also with the analytical program designed to provide outputs that meet customer needs Imputation was implemented in CANCEIS a Nearest Neighbour imputation tool designed and developed by Statistics Canada Cancies 2009 The CANCEIS platform was configured to select a suitable donor for each record needing treatment in two stages In the first stage a pool of potential donors was established through two nested processes The first process divided all records in the survey into imputation classes based on cross classification of auxiliary variables chosen for this stage Potential dono
32. hat respondents understand the questions to mean and the terminology used 17 e understand how respondents formulate their answers and by so doing ensure that the questions are interpreted as key users and stakeholders intended e ensure that response options are comprehensive e ensure that respondents are willing to provide answers e ensure that respondents are able to provide answers e ensure that the order in which the questions are asked does not affect the answers given e address issues relating to the collection of proxy data if proxy information can be collected The quantitative piloting aims to provide a test run of the new questionnaire and to identify any issues with the questionnaire before the next wave s data collection starts An interviewer de brief is held following the pilot to seek feedback on the questionnaire and any areas for improvement The pilot also provides the opportunity to produce survey metrics such as interview length broken down by topic area and indicative response and data linkage consent rates Length of questionnaire The table below shows the mean interview lengths for the first three waves of WAS WAS wave Mean interview length mins 75 percentile 90 percentile WAS wave 1 88 103 135 WAS wave 2 85 104 137 WAS wave 3 75 91 119 The mean wave one interview length was 88 minutes and has remained relatively consistent for wave two of the survey T
33. he W3 population total minus the sum of the EOSMs weights see formula 10 This ensures that the EOSM s weights have not been double counted In other words the OSM and SSM base weights were multiplied by the factor 0 where rs4 8 This is a Consequence of the assumption that population entrants were in the original sample which we made because the data do not allow us to distinguish population entrants 37 10 0 Nys Wezosm rs4 E a Y Weru gt Wiese N y is the W3 population total and w is the weight of individual i Cross sectional weights for the new panel The new panel have design weights This is the reciprocal of the selection probability of an address As there is oversampling of wealthier households this needs to be incorporated into the design weight construction The primary sampling units PSUs were sampled using a standard probability proportional to size method where the size is measured as the number of addresses per PSU Within each PSU those addresses flagged as being from the predicted high wealth stratum were sampled at 3 times the rate of other addresses in the predicted low wealth stratum The address selection probabilities for addresses from the ith PSU are thus P address sampled P PSU sampled P address sampled PSU sampled 11 nN ar P address sampled PSU sampled where n is the number of sampled PSUs 324 N is the number of addresses in the this PSU and N the total
34. he introduction of a new random sample of around 12 000 addresses Wave structure The following diagram illustrates the longitudinal design of the Wealth and Assets Survey Wave one started in July 2006 with fieldwork being spread over a two year period Wave two a follow up to wave one was conducted between July 2008 and June 2010 The introduction of a new cohort of addresses in wave three is shown in blue All interviews have a two yearly interval between waves therefore providing estimates of change in relation to the same period of time For example wave one interviews conducted during July 2006 would be repeated for wave two in July 2008 It is important that this gap remains constant so that estimates of change are comparable wave on wave 10 Jul 06 Jul 07 Jul 08 Jul 09 Jul 10 Jul 11 Jul 12 Jul 13 Jul 14 Jul 15 Wave 3 Wave 3 new cohort Wave 4 Wave 4 new cohort Wave 5 Wave 5 new cohort Mode of data collection The Wealth and Assets Survey has two interview stages in the longitudinal panel design The primary interview is where the WAS questionnaire is utilised this is referred to as the mainstage interview The second is the Keeping in Touch Exercise KITE which is used to maintain respondent s contact details between waves Mainstage interview The mainstage interview is conducted using Computer Assisted Personal Interviewing CAPI Face to face interviewing is the preferred choice for the Wealth a
35. he mean wave three interview length has reduced to 75 minutes However the mean interview length is a slightly misleading metric when considering respondent burden The WAS questionnaire uses extensive routing in order to ensure that respondents are only asked questions that are relevant to them For example a one adult household with no or little assets and liabilities would be routed to a relatively small number of questions and therefore have a short interview Conversely a two adult household with a lot of different assets and or liabilities would be routed to a lot of questions and therefore have a much longer interview This range is reflected in the variance of interview lengths In wave one ten per cent of all interviews lasted at least two and a quarter hours This decreased in wave three to just below two hours Programming and testing The Wealth and Assets Survey data is collected using Computer Aided Personal Interviewing CAPI The software loaded into interviewer s laptops is called Blaise All face to face ONS social surveys use Blaise for interviewing as ONS feel that it has the flexibility and technical capability to cope best with the complexity of social research surveys Blaise s powerful 18 programming language offers numerous features and its data entry program supports a variety of survey processing needs A number of features of Blaise are particularly advantageous for this survey e Blaise CAPI scripts have
36. he sample at W3 P_Flag2W3 Flag for wave entrant status 1 OSM birth entrant Child entrant 15years or younger born to OSM household member 2 SSM birth entrant Child entrant 15years or younger born to SSM household member 3 Other SSM entrant Adult entrant 16years or older P_Flag3W3 Flag for wave eligibility status 1 Eligible adult Aged 16 years or older and not in full time education 2 Ineligible adult Aged 16 to 18 years in full time education 3 Ineligible child Aged 15 years or younger P_Flag4W3 Flag for HRP status 1 HRP in W2 W3 Individual was HRP in both waves 2 HRP in W2 not W3 Individual was HRP in second but not in third wave 3 HRP in W3 not W2 Individual was HRP in third but not in second wave 4 Never HRP Individual was never the HRP inc children Variable specific notes Refer to the wave 3 paper questionnaire for notes on specific variables Survey design Sampling strategy The Wealth and Assets Survey collects information about private household wealth in Great Britain The survey uses the small users Postcode Address File PAF as the sample frame for residential addresses in Great Britain that is England Wales and Scotland excluding North of the Caledonian Canal and the Isles of Scilly The ONS copy of the PAF is updated twice a year to ensure that recently built addresses are included and demolished or derelict properties are removed quickly The survey es
37. here significant differences were found ensure that reference periods populations geography samples modes of collection questions concepts and derivations were comparable Results from these analyses indicated that estimates from the Wealth and Assets Survey were broadly in line with results from other administrative and survey sources Further work to produce more detailed analyses and comparisons is ongoing and any data quality issues which are identified with WAS variables will be fully documented and made available on the ONS website 45 Wealth estimates The wealth estimates in this report are derived by adding up the value of different types of asset owned by households and subtracting any liabilities Total wealth with pension wealth is the sum of four components e net property wealth e physical wealth e net financial wealth and e private pension wealth Total wealth without pension wealth is the sum of the first three of these components The components are in turn made up of smaller building blocks e net property wealth is the sum of all property values minus the value of all mortgages and amounts owed as a result of equity release e physical wealth is the sum of the values of household contents collectibles and valuables and vehicles including personalised number plates e net financial wealth is the sum of the values of formal and informal financial assets plus the value of certain assets held in the names
38. household wealth In order to make data records comparable on longitudinal level all longitudinal records that were a unit non response in wave one were also imputed in the recent imputation exercise for the wave two report However this means that when conducting cross sectional analysis based on wave one data only part of the data was imputed for unit non response 23 In general because of the distributional properties of the data elicited by the Wealth and Assets question structure missing data was best treated using a non parametric imputation method To this end all item non response and unit non response was imputed using a Nearest Neighbour approach Bankier Lachance amp Poirier 1999 Durrent 2005 Waal Pannekoek amp Schltus 2011 In this approach missing data was replaced with plausible values drawn from other records in the data set referred to as donors For categorical data and skewed or clustered continuous data donor based methods are advantageous in that they use only values actually observed in the data Significantly this helps to avoid the distributional assumptions associated with parametric methods such as regression modelling Importantly if applied correctly imputation will estimate the distributional properties of the complete data set accurately Rubin 1987 Chen amp Shoa 2000 Durrent 2005 Donor Selection The key to a successful application of Nearest Neighbour imputation is the selection o
39. in any wave e Always used the linked file as a base when matching variables across waves e Use HHSerialW1 HHSerialW2 and or HHSerialW3 for household linking e Use PIDNo for person level linking this remains the same over the survey life time of a sample unit e When GETting the files only KEEP the variables you need to add to the file including the one needed to match cases This makes it easier when matching To add W2 variables to the W3 person file keeping only W3 cases e SortW1W2W3 person level linked files by PIDNo W1W2W3_UKDA sav e Sort W2 Person file by PIDNo e Sort W3 Person file by PIDNo e Match W2 and W3 files with the linked file being used as a look up TABLE use PIDNo to MATCH e This will add W2 variables to W2 cases and W3 variables to W3 cases e Select the required cases e g for W3 cases including linked W2 cases use HHSerialW3 gt 0 To add W3 variables to the W2 household file keeping all cases e Sort the linked file by HHSerialW2 W1W2W3_UKDA sav e Sort W2 household file by HHSerialW2 e Match files using HHSerialW2 e Sort the new file by HHSerialW3 e Sort W3 household file by HHSerialW3 e Match files using HHSerialW3 e This will produce a linked W2W3 file with W2 and W3 variables Linking within a wave To add household variables to the W3 person file e Sort both files by HHSerialW3 and use this variable to MATCH e Use the household file as a look up TABLE This will add the household v
40. ing values Users of WAS data need to distinguish between item non response which typically occurs when a respondent does not know or refuses to answer a particular survey question and unit non response missing units where an individual in a responding household refuses to be interviewed or a contact cannot be made Item and unit non response can be problematic in that many standard analytical techniques are not designed to account for missing data More significantly missing data can lead to substantial bias and inconsistencies in estimates and publication figures Imputation is a statistical process that serves to counter these problems by replacing missing values with valid plausible data To avoid distorting the data through this process inappropriately the method applied must account for the survey question structure and the distributional properties of the observable data that structure yields lt must also take into account the possibility that unrecorded data is not missing completely at random It is important to note that as the overarching aim of imputation is to improve the utility of the data the key analytical aims of the survey should also be factored into the design of the imputation process Information about discrete assets or liabilities recorded by the Wealth and Assets Survey was collected through a relatively consistent question structure Typically an affirmative response to routing questions designed to determine do you have
41. iptive statistics for the W3 cross sectional weight combined proportional to effective sample size split by sub samples Sub sample n Mean Standard Coefficient Minimu Maximu deviatio of variation m m n All cases 49 447 1 207 877 72 7 69 9 999 Longitudinal cases 32 485 1 339 1 001 74 7 69 9 999 W3 SSMs 1 697 1 300 1 108 85 2 69 9 999 W3 EOSMs 1 420 619 167 26 9 187 1 212 New panel 13 845 945 317 33 5 204 2 598 The second version of the weights is very similar especially when comparing the variability within each subgroup However the new panel have larger weights and the longitudinal cases have smaller weights which is an improvement Additionally the overall variance of these weights is smaller as desired We recommend that the second version of the W3 cross sectional weights those where the panels are combined proportional to effective sample size are the best weights 41 Additional Considerations The weighting process for future waves should be very similar to that used for W3 It will therefore be important to analyse the additional considerations discussed below such that suitable recommendations can be made and the most efficient weighting strategies can be developed and taken forward for future waves A methodological change was made between W2 and W3 In W2 births to OSMs were given their mother s weight rather than simply having their weight allocated thro
42. is a true increase in the population and so in this case the sum of the weights does increase This method also holds when households split between interviews The sum of the W2 weights within a W2 household are shared across the two associated W3 households Finally the weights are scaled to the W3 population total Nw3 using the scaling factor 0 where rs2 Nys 8 6 9 WisW2 W3cases W1landW 3cases The weight share allows longitudinal OSMs and SSMs to be treated together as a single sub sample but the construction of the original panel cross sectional weight requires an amalgamation of this sub sample with EOSMs The EOSMs weights are their original design weights constructed as the inverse of the selection probabilities These are then rescaled to the proportion of the responding W1 sample that they represent multiplied by the W1 population total Therefore the weights were multiplied by the factor 0 where rs3 n EOSM x Nr 9 respondents rs3 y WieEOSM is the number of EOSMs at W3 n is the number of W1 respondents N is the n EOSM W respondents W1 population total and w is the weight of individual This subsample then needs to be combined with the weights for the OSMs and SSMs The EOSM s weight would have previously been captured in the weights of the OSMs at W1 during the non response adjustments In order to account for this the weights for the OSMs and SSMs were scaled to t
43. its as it is not known whether they remain in the target population or not 7 The regression model accounts for the clustered survey design with the nesting of observations people within households households within PSUs using the PSU as the ultimate cluster for the purposes of calculating standard errors of coefficients 33 In equation 3 below VA represents the predicted probability of response from the known eligibility status sample base again using a logit model with W2 individual and household level characteristics as predictor variables S is the sample of individuals within a responding household at W3 this is those who are 1 W1 W3 longitudinal cases for the survivor weights or 2 W2 W3 longitudinal cases for the T 1 to T weights These are the people who are weighted up to represent the non respondents at W3 For individuals in a respondent household at 1 both W2 and W3 for the T 1 to T weights or 2 W1 W2 and W3 for the survivor weights the longitudinal base weight w is the product of the W2 weight we and the two loss to follow up adjustment weights For individuals known to be part of movements out of the target population at W3 Out wy is the product of w and the unknown eligibility status weight 0 i bni W Ws Waj 5 LES longitudinal responders W 4 L cal o A W W gt l Out loss to population at W3 The base weight w is trimmed at the 99 percen
44. key international longitudinal surveys As WAS is concerned with enabling estimation for all population members and weighting is based on calibrated population totals it seemed desirable and appropriate to ensure sharing was across all cases enumerated within households The WAS weight share was constructed following Kalton amp Brick 1995 This standard approach is based on the W2 household member s weights and sharing these weights between all associated W3 household members A key challenge for the weight share method is being able to distinguish between those SSMs who are new population entrants and those who were in the original population but not originally in the sample Unfortunately it is not possible to make this distinction with WAS data and consequently except for births we treated all SSM entrants as if they were in the population at the time the sample was drawn First we sum the base weights of the individuals i in each W2 household j Then we divide this value by the number of individuals in the associated W3 household minus the number of 36 births as shown in formula 7 where bw is the base weight and N is the number of individuals in household j Ni J bw ij ie j l er ae N births J This ensures that all respondents within a household have a weight but that the sum of the weights does not increase when SSMs excluding births enter the sample On the other hand a birth either to an OSM or an SSM
45. l development This is done with the assistance of their field manager They are also assigned a mentor who is an experienced interviewer New interviewers shadow mentors as well as having a mentor accompany them when they begin working on a survey Interviewers also participate in specific training events such as Achieving Cooperation Training known as ACT and Achieving Contact Efficiently ACE Both of these training packages have been reviewed and rolled out to the entire field force face to face and telephone interviewing This is managed through training events and interviewer support group meetings Quarterly meetings of field managers and their teams are held throughout the year where training issues and refresher training are regularly addressed Telephone interviewers and ONS help desk operatives receive equivalent training and can very often convert refusals following the receipt of an advance letter Survey specific training Telephone interviewers ONS telephone interviewers working on the Wealth and Assets Survey receive an annual briefing on how to administer the Keep in Touch Exercise KITE questionnaire This briefing delivered by research staff covers the importance of the KITE interview and the importance of collecting contact details and ensuring these are reported correctly KITE interviewers are trained to try and turn around refusals should panel respondents express concerns over future involvement in the survey 12
46. ling error are e response errors such as misleading questions interviewer bias or respondent misreporting e bias due to non response as the characteristics of non responding persons may differ from responding persons e data input errors or systematic mistakes in processing the data Non sampling errors are difficult to quantify in any collection However every effort was made to minimise their impact through careful design and testing of the questionnaire training of interviewers and extensive editing and quality control procedures at all stages of 43 data processing The ways in which these potential sources of error were minimised in WAS are discussed below Response errors generally arise from deficiencies in questionnaire design and methodology or in interviewing technique as well as through inaccurate reporting by the respondent Errors may be introduced by misleading or ambiguous questions inadequate or inconsistent definitions or terminology and by poor overall survey design In order to minimise the impact of these errors the questionnaire accompanying documentation and processes were thoroughly tested before being finalised for use in the first wave of WAS To improve the comparability of WAS statistics harmonised concepts and definitions were also used where available Harmonised questions were designed to provide common wordings and classifications to facilitate the analysis of data from different sources and have been well
47. me way as an imputation class when imputing the missing value in the other wave For continuous data an appropriately banded range was used in a similar way However instead of taking a discrete value from the donor the ratio that described the rate of growth or decay in the donor between waves was transferred to the record to be imputed The ratio was then used in conjunction with the observed value in one wave to calculate missing value in the other This strategy is typically referred to as ratio based roll back M 0 or roll forward O M imputation Table 3 shows a typical example of a longitudinal auxiliary variable set used to impute a missing value for a respondent s private 27 pension in Wave2 in the presence of observable data in wave one Comparing Table 3 and Table 1 will help identify the subtle differences between cross sectional and longitudinal imputation processing strategies 28 Table 3 Imputation Classes and Matching Variables used for the longitudinal imputation of Private Pensions in wave two in the presence of observed data in wave two Imputation Class Matching Variable Variable Classification Variable w Classification Banded Value 1 Less than 2 500 Annual Gross 0 3 Various amounts Salary observed in 2 2 500 gt 4 999 Wavel Wave2 Wavel 3 5 000 gt 9 999 4 10 000 gt 19 999 Employment 0 2 1 Employee Status 5 20 000 gt 49 999 Wavel amp Wave2 2 Self Employed 6
48. monstrated by the green box in figure 3 1 This consists of OSMs from W1 SSMs and EOSMs from W2 SSMs and EOSMs from W3 as well as the new panel introduced in W3 Longitudinal Weights The weighting strategy is based on a principle of maintaining the link between the initial selection probability and the ongoing loss to follow up adjustments that remain for the evolving respondent subset over time This is achieved through developing the longitudinal base weight see e g Verma et al 2007 This principle enables the weights to refer back to 32 the desired population as closely as is possible with the current sample design and respondent follow up procedures Different longitudinal base weights are used to construct the two longitudinal weights The product of the relevant W2 weight and the W3 attrition weights creates the W3 longitudinal base weight The relevant W2 weight is different for the two longitudinal weights it is 1 the longitudinal weight for the survivors and 2 the W2 cross sectional weight for the T 1 to T cases as cases that do not appear in W1 do not have a W2 longitudinal weight Construction of the W2 weights is detailed in Ashworth et al 2012 The first step in the weighting process is to develop the attrition models for W3 Two separate steps were used to adjust for attrition a model for unknown eligibility status a model for non response non contact In both cases logistic regression was used to predi
49. munities and Local Government DCLG and the Cabinet Office CO The primary focus of the questionnaire is to provide for estimates of wealth however some additional information is collected on non wealth topics such as socio demographic characteristics income and financial acuity This allows for aggregate and component analysis of wealth with other factors Questionnaire changes WAS is a longitudinal survey and therefore in order to measure change over time the questionnaire needs to be as stable as possible so as to reduce discontinuities in the outputs However there is scope to make changes to the questionnaire between waves in order to adopt harmonised question standards and or emerging information requirements Changes between waves are made with consortium agreement Sponsoring departments provide their information requirements and specify any requested changes These changes are discussed by the WAS Technical Group TG with recommendations for questionnaire changes being submitted to the WAS Steering Group SG The WAS SG is formed from senior representatives of the consortium departments Recommended questionnaire changes have previously been subject to cognitive question testing and quantitative piloting The cognitive question testing has the following objectives e ascertain whether the proposed questioning will address the information needs identified by key users and stakeholders from the respondents perspective e establish w
50. n the first couple of calls but for those which are not it is important to make sure that successive visits are at different times of the day including evenings and on different days of the week 13 ONS Methodology conducted a review of interviewer calling patterns and the success of these as the time of day and day of week varied This report recommended a set of calling patterns for interviewers to follow in order to maximise the likelihood of establishing contact with respondents Interviewers were required to attempt to complete each monthly quota of 13 addresses within five visits to the area and up to 28 working hours excluding travel time Best practice procedures whereby interviewers varied their calling times and days in the area were also employed in an attempt to maximise response to the WAS Field sampling procedures Where an interviewer discovered a multi household address in England and Wales or a Scottish address with an multi occupancy MO count less than two up to a maximum of three randomly sampled households from the address were included in the sample For Scottish addresses sampled with an MO count of three or more a single household was sampled if the MO count equalled the actual number of households present If the number found differed from the MO count the number of households sampled was adjusted but again to a maximum of three The number of additional households that could be sampled was subject to a maximum
51. nd Assets Survey due to the complex subject matter of the survey and the need for the interviewer to support the respondent in answering the questions The interviewer respondent interaction is much greater on a face to face survey compared with other modes such as paper and telephone Another reason for face to face interviewing is the need to interview everyone aged 16 and over in the household This is more challenging with some alternative modes of data collection The interview length of the WAS questionnaire also means that CAPI is a good approach Face to face contact with respondents allows interviewers to identify when respondents are becoming fatigued during the interviews This allows interviewers to suggest a break from the interview or perhaps for them to continue the interview at another time in some cases Identifying respondent fatigue picking up on body language is best done when the interview is face to face CAPI was also considered the best approach to maximise cooperation with the survey Response rates to face to face surveys tend to be higher than telephone paper and web alternatives Keep in Touch Exercise interview Conversely the KITE interview aims to collect much less information and only from one person in each household The questionnaire is set up to establish whether the household circumstances have changed In the vast majority of cases there is no change to the household s address or composition so the intervie
52. ndividuals were asked to report the value at the time of interview of their retained DC fund Rights retained in schemes from which individuals are drawing down Individuals could also report that they were already drawing down assets from a retained pension scheme In these cases individuals were asked to report what the remaining fund value for their scheme was at the time of interview The wealth in each of these schemes was then summed to derive total wealth held in schemes of this type Pensions expected in future from former spouse partner Individuals were asked to report in total how much they expected to receive in the future from private pensions from a former spouse or partner Respondents were given the choice to report this either as a lump sum wealth figure or as an expected annual income Two slightly different approaches were followed depending on how the respondent answered For those who reported a total lump sum value this figure was taken as the relevant wealth measure and discounted back to the time of the interview For those who reported an 49 expected future annual income wealth was calculated in much the same way as for DB schemes described above AY 1 py Where Ar is the age and sex specific annuity factor at retirement age R see above Y is expected annual pension r is the real investment return assumed to be 2 5 per cent a year R is assumed to be 65 or the individual s current age if he she
53. nsions Individuals were asked to report the value of their fund at the time of the interview and were encouraged to consult recent statements where available Current personal pension 48 wealth is therefore simply defined as the fund value reported by the respondent at the time of the interview Retained rights in defined benefit occupational pension scheme Individuals could report up to three pensions in which rights have been retained These could be either DB of DC schemes The wealth in each DB retained scheme was calculated separately in much the same way as for current DB schemes described above and then summed to derive total wealth held as retained rights in defined benefit DB occupational schemes Wealth in these schemes was defined as y AS tbe l Where Ar is the age and sex specific annuity factor at retirement age R see above Y is expected annual pension Li is the lump sum that the individual expects to receive at retirement r is the real investment return assumed to be 2 5 per cent a year R is assumed to be 65 or the individual s current age if he she was already aged over 65 a is the individual s age at interview Retained rights in defined contribution occupational pension scheme The wealth in each DC retained scheme was calculated separately in much the same way as for current DC schemes described above and then summed to derive total wealth held as retained rights in DC schemes Specifically i
54. nt variables and DVs Variable naming conventions e Wave suffix Variables in both datasets are given the suffix W3 to indicate that they contain values collected in wave 3 There are a few variables with the suffix W2 or W1 and these contain values from wave 2 or wave 1 Most of the W2 and W1 variables are present to allow the datasets to be matched to those from previous waves of the survey e g HHSERIALW2 e Imputation suffix All variables used as components for wealth DVs were subject to imputation Variables that have had missing data imputed appear in the datasets in two versions The version that contains only the values observed at interview will end with the suffix W3 as described above e g FSInValW3 The version that contains both observed and imputed values will end with the suffix _i e g FSInValW3_i e Aggregation suffix To calculate total household wealth all component DVs were aggregated to household level To enable data users to use aggregated household level DVs on person level relevant DVs are also provided on the person level file Weights To carry out cross sectional analysis based on the individual wave data the following table has the appropriate variable weight to apply for cross sectional analysis Wave Cross sectional Calibration Weight 1 XS_wgtW1 2 XS_calwgtW2 3 w3xswet As opposed to cross sectional analysis longitudinal analysis can only be carried out
55. of four per PSU Some occupied dwellings are not listed on the PAF This may be because a house has been split into separate flats only some of which are listed If the missing dwelling could be uniquely associated with a listed address a divided address procedure was applied to compensate for the under coverage In these cases the interviewer included the unlisted part in the sample only if the associated listed address had been sampled Any sampled addresses identified by the interviewer as non private or non residential were excluded as ineligible l Hopper N An analysis of optimal calling pattern by Output Area Classification ONS Working Paper Methodology Division 2008 14 Response rates The following graph provides household response for waves one two and three by the monthly field periods Wave 3 old wave 3 new WAS achieved an average response rate of 55 per cent for wave one with fieldwork being conducted between July 2006 and June 2008 The achieved sample for wave one was issued for re interview between July 2008 and June 2010 yielding an improved response of average response rate of 68 per cent In wave 3 interviews were attempted with the responding households and non contacts from waves one and two In addition to this a new random sample of around 12 000 addresses was added Response rates for these old and new cohorts in wave three are shown
56. on person level The following table has the longitudinal variable weight to apply for longitudinal analysis Wave Longitudinal Calibration Weight W1W2 Longitudinal Weight Longit_calwgtW2 W1W3 longitudinal weight wlw3wgt W2W3 longitudinal weight w2w3wet Interview Outcome codes The datasets include responding households only The variable HOutW3 gives an indication of the type of interview outcome of the household Fully co operating 110 Complete interview by required respondent s in person 120 Fully co operating household one or more interviews completed by proxy 121 HRP economic unit interviewed in person one or more other interviews by proxy 122 HRP and or spouse partner interview by proxy 130 Complete interview by proxy Partially co operating 212 Non contact with one or more respondents 213 Refusal by one or more respondents all contacted 214 All adults interviewed but one or more interviews was incomplete 222 Non contact with one or more respondents and some proxy information 223 Refusal by one or more respondents all contacted and some proxy information 224 All adults interviewed but one or more interviews incomplete and some proxy information 211 Full response in person from HRP economic unit HRP and spouse partner One or more other interviews missing or incomplete 212 Full response from HRP economic unit one or both by proxy One or more other interviews missing or incomplete
57. outing performed in Blaise This process is used to identify where Blaise has incorrectly routed respondents This can either be corrected for by recoding data or where cases haven t been routed as they should have been imputation requirements are specified Where errors in routing are discovered the Blaise questionnaire is corrected to enhance the quality of future data collection The sooner base checks are performed the sooner the Blaise questionnaire can be corrected thus leading to lower levels of data imputation Editing and validation processes for the second wave of WAS were similar to those used for wave one more details are provided in section 10 4 of the wave one report However due to the longitudinal component of the survey design part of the achieved sample size in wave two is linkable to wave one data Therefore it was important to introduce longitudinal edit checks to the existing editing and validation processes The edit and validation checks were run in two stages whereby first cross sectional checks were carried out on the second wave to validate or edit outliers As opposed to checks for the property and physical wealth data checks for financial and pension wealth data were exclusively done on individual level because of the way the data had been collected The investigation of outliers largely focused on the top and bottom ten per cent of the distribution of each wealth component although for some variables this propo
58. ree and four new cohort sample relative to other addresses within a given postcode sector Sample sizes of each wave The following table provides a summary of the sample sizes rounded both issued and achieved for each of the first three waves of the Wealth and Assets Survey Wave Issued addresses Achieved Achieved adults households One 62 800 30 500 53 300 Two 32 200 20 000 34 500 Three 37 900 21 450 40 400 Respondents aged 16 and over In developing the survey precision targets for change on key estimates were agreed in consultation with funding departments From these it was estimated that an overall achieved sample of approximately 32 000 households spread evenly over the two years of wave one was required In addition to the above precision targets there was a further target to achieve a two year sample of 4 500 households above the top wealth decile for wave one This was well above the 3 200 households that would be above the top wealth decile for an equal probability sample Oversampling the wealthiest households allows for more detailed analysis of this group and gives more precise estimates of the levels of wealth across the whole population For wave two the achieved wave one sample was issued plus all of the non contacts A total of 32 200 addresses were issued for wave two In wave three follow up of the respondents and non contacts at wave one and wave two was supplemented by t
59. reful survey design and compensation during the weighting process the latter having been discussed earlier To further reduce the level and impact of item non response resulting from missing values for key items in the questionnaire ONS undertook imputation see Imputation Non sampling errors may also occur between the initial data collection and final compilation of statistics These may be due to a failure to detect errors during editing or may be introduced in the course of deriving variables manipulating data or producing the weights To minimise the likelihood of these errors occurring a number of quality assurance processes were employed which are outlined elsewhere in this guide External source validation In the final stages of validating the WAS data comparative checks were undertaken to ensure that the survey estimates conformed to known or expected patterns and were 44 broadly consistent with data from other external sources This work was undertaken by ONS and analysts from the funding departments as well as a number of academics who had expertise in the various topics included in WAS The following guidelines were recommended by ONS when undertaking the external source validation process e identify alternate sources of comparable data e produce frequencies and cross tabulations to compare proportions in the WAS dataset to those from external sources e if differences were found assess whether these were significant e w
60. rs could only be selected from the sub population of records in the same class as the record currently being imputed The second process served to refine the potential donor pool by ranking all of the records within class Ranking was determined by calculating the distance between the potential donor and the recipient record based on a second set of auxiliary variables referred to as matching variables Where appropriate the calculation included differential weighting to account for cases were some auxiliary variables were more important than others In general one of two distance functions were used to calculate the distance between the potential donor and the recipient record depending on the characteristics of each particular auxiliary variable Xf the recipient record with n auxiliary variables Xq the potential donor record with n auxiliary variables 24 n Dsa gt wiDi i 1 w the weight for the it variable D the individual distance for the it variable For categorical data with no ordinal relationship between categories 0 where xf xa D d 1 where xp Xq For categorical or continuous data with an ordinal and or ratio relationship between categories or values 1 if xp xal gt y 2 i 1 mE el otherwise y desired minimum x xal at which point and beyond D 1 A subset of records with the smallest distance values were selected for the final potential donor pool as these were mos
61. rtion was reduced if the number of cases highlighted for investigation was particularly high When http www ons gov uk ons rel was wealth in great britain main results from the wealth and assets survey 2006 2008 report wealth in great britain pdf 21 outliers were investigated in the pensions or the financial section various variables within the same wealth component section or even different sections of the questionnaire were included to establish whether particularly large outliers could be explained by the circumstances of respondents The majority of investigated cases proved to be genuine and only a small number of cases had to be edited whereby data was only edited if sufficient information was recorded by interviewers to establish the correct response The second stage of checks was conducted after the linkage exercise was completed At this stage the change of wealth components between the two waves was calculated and subsequently outliers of change were highlighted To investigate these outliers the circumstances of relevant respondents in both waves had to be considered to decide whether the value in either wave one or wave two was correct As with the cross sectional checks only a small number of corrections were made for each wealth component variable where sufficient information was available 22 Imputation General Methodology In a way similar to all social surveys data from the Wealth and Assets Survey contained miss
62. s defined as _ Bae Ph r Where Ar is the age and sex specific annuity factor at normal pension age R based on single life annuity rates quoted by the Financial Services Authority assuming average age and sex specific life expectancies as estimated by the Government Actuary s Department and a discount rate of 2 5 per cent etfs Y ans Y is annual pension income defined as ioii 47 a is the accrual fraction in the individual s scheme n is the individual s tenure in the scheme si is the individual s gross pay at the time of interview Li is the lump sum that the individual expects to receive at retirement r is the real investment return assumed to be 2 5 per cent per annum R is the normal pension age in the pension scheme a is the individual s age at interview Since these are individual not household pension wealth measures and due to the complexity of the calculations and the information that would have been required from respondents survivor benefits are not modelled In practice this would lead to a underreporting of pension wealth for women since the expected future survivor s benefits that they will receive when they on average outlive their husbands will not be measured To the extent these survivors benefits will be sometime in the future for most women their omission will have only a small effect on the calculations Definition of wealth from Additional Voluntary Contributions
63. sponding cases a base weight The most recent weight for each of these respondents was used For respondents who were in W2 and W3 W2 W3 longitudinal weights were used W1 and W3 W1 cross sectional weights were used This ensures that where possible weights which have been previously adjusted for non response and unknown eligibility status were used as base weights for W3 The W1 and W3 cases weights would have previously been captured in the longitudinal cases weights by the non response modelling undertaken during W2 In order to account for this the weights for the W2 W3 cases were scaled down to the sum of the W2 W3 cases minus the sum of the W1 and W3 cases see formula 6 This ensures that the W1 and W3 cases weights have not been double counted The scaling is carried out by multiplying the longitudinal cases weights by the factor 0 where rs1 6 0 e gt ed a gt SW E N rsl w e W2 W3cases w is the weight of individual i The next challenge was to assign a weight to people entering the sample as SSMs It is common to use a weight share method to approximate these weights e g Huang 1984 Ernst 1989 Kalton amp Brick 1995 rather than attempting to work out selection probabilities directly Different surveys use different approaches to weight sharing Some surveys restrict the sharing to adults or use other criteria for example see Schonlau amp Kroh 2010 who detail the methods used by
64. stification was based on the identification of sub populations in the data with proportionally higher non response rates that would correspond with an appropriate observable change in the properties of the data This preliminary evaluation was supplemented by a more detailed review of the utility of the data by topic experts familiar not only with the analytical aims of the survey but also with expected data trends and characteristics inferred from other reliable external data sources 30 Weighting Overview The weighting strategy embeds two important principles The first principle is to maintain the link between the initial selection probability and the ongoing loss to follow up LTFU adjustments that remain for the evolving respondent subset over time This is achieved through developing the longitudinal base weight from the wave one cross sectional weight The second principle is that SSMs in the survey receive a temporary share of the base weight appropriate to their status at any given time point These principles enable the weighting to refer back to the desired populations as closely as is possible with the current design Terminology As the survey develops there are numerous combinations of responding patterns e g responded in all waves or responded in the first and second wave but not the third etc The number of categories is extended further when the wave in which the participant was sampled is considered as well as the wave in
65. t similar to the record being imputed For non categorical data extreme outliers were excluded from the donor pool to prevent propagation of values likely to have a significant impact on estimates derived from the data These were identified through expert review and routinely represented values greater the 95 percentile of the observed data s distribution Table 1 shows a typical example of an auxiliary variable set This particular set was used to impute an unknown value for a respondent s private pension All Wealth and Asset variables were treated in a similar way 25 Table 1 Imputation Classes and Matching Variables used for imputing values for Private Pensions Imputation Class Matching Variable Variable Classification Variable w Classification Banded 1 Less than 2 500 Annual Gross 0 3 Various amounts Estimate Salary 2 2 500 gt 4 999 3 5 000 gt 9 999 Employment 0 2 1 Employee Status 4 10 000 gt 19 999 2 Self Employed 5 20 000 gt 49 999 6 50 000 gt 99 999 Age Group 0 1 1 16 24 7 100 000 or more 2 25 44 3 45 59 Female tSample 3 month sampling time 45 64 Male frame 4 60 74 Female 65 74 Male To impute missing values for private pensions 5 75 donors were selected from an imputation class derived from the cross classification of Sex 0 1 1 Male observed Banded Estimates and tSample The 2 Female Banded Estimate provided an important constraint on donor selection base
66. tavceudaren Ee e EAEE E eE AEREE cada bno 14 RESPONSE RATES ui 15 KEEPING INT o1 EAEE E E E AA AE E E E N 16 QUESTIONNAIRE CONTENT scsccscscsccccccscsccccscsccccccccsssccccscscesceceees 17 QUESTIONNAIRE CHANGES ita tddi i 17 LENGTH OF QUESTIONNAIRE roserne a e a a e a eia paa a a takaan i raean ga 18 PROGRAMMING AND TESTING a aiiai 18 EDITING osuna ia 21 IMPUTATION iii ida 23 GENERAL METHODOLOGY sec eli is 23 DONOR SELECTION ur ia ai tones iio 24 PROCESSING TRATE dicos 27 QUALITY ASSURANCE AND EVALUATION cccccccesesescecececesnssescecececeesssesceceseaeeueneececeseceegeseeceeeseaeeusneeceeeesauenes 30 WEIGHTING add 31 DATA QUALITY aii acid 43 SAMPLING ERROR aida ala Da dei di il 43 NON SAMPLINGiERROR ccoo E a dlaitochashevedeasashochastccden sade 43 EXTERNAL SOURCE VALIDATION da as 44 WEALTH ESTIMATES suda Raid 46 CONTACT DETAILS icon 50 Overview The Wealth and Assets Survey WAS is a longitudinal survey that interviewed across Great Britain England Wales and Scotland excluding North of the Caledonian Canal and the Isles of Scilly Respondents to wave one July 2006 June 2008 of the survey were invited to take part in a wave two follow up interview two years later July 2008 June 2010 Respondents to wave 2 were then invited to take part in a wave three follow up interview another two years later July 2010 June 2012 In addition to these follow up interviews a new random sample of addresses was also a
67. the original and new panels were re scaled in 2 ways in proportion to the achieved sample size for the three panels in proportion to the effective sample size for the three panels as proposed by Chu et al 1999 Korn and Graubard 1999 In the first option the constructed weights were multiplied by the factors 6 for the orig Oi as as PS original panel and 0 and Ofw fOr the new panel where as Noy ig ga E Prowl 15 orig SE newyl orig Prewy M rewy2 Norig Aaewyl N newy2 ga Prewy2 newy2 Norig Few Fh cpp 39 Nig S the sample size of the original panel n is the sample size of year 1 of the new ori newy1 panel and n is the sample size of year 2 of the new panel newy2 This scaling approach will result in each panels weights being in the same proportion to each panels sample size The weights were then re scaled to the W3 population total by multiplying by 0 where rsd ie a W 1 The second option accounts for the variance within each panel and combines the weights such that the variance is minimised As the new panel is more representative of the W3 population as it was selected more recently and the weights are less variable as they have had fewer adjustments made to them there is a potential improvement that could be made by combining the weights in proportion to effective achieved sample sizes for the three panels For this method firstly the design effect for each p
68. tile and scaled to the W2 population total for the T 1 to T weights and the W1 population total for the survivor weights The two longitudinal sub samples eligible respondents and ineligible outflows are after adjustment for attrition representative of the populations to which they relate back to W2 for the T 1 to T cases and W1 for the survivors so it is possible to calibrate the longitudinal base weight to the relevant population totals This procedure should have the advantage of making a further correction for any attrition not already accounted for by adjusting the weights to calibration control groups The calibration weights are calculated to sum to a set of known calibration totals t minimising the distance between the pre calibration weight ww and the calibrated weight which we write as an adjustment of the pre calibration weight g w47 If the membership of the calibration groups is represented by a vector of auxiliary values x then the problem can be represented as On on Jon 5 min ase E ZW such that Y Xg WE 34 The final longitudinal calibration weight is the product of the g weight and the initial longitudinal base weight where the g weights are defined as the solution to 5 The g weight helps to rebalance the sample towards the population values of the variables included in the calibration model Basic descriptive statistics relating to the W3 longitudinal weights for both the T 1
69. timates are designed to be representative of the GB population therefore WAS like most social surveys uses a probability proportional to size or PPS method of sampling cases This means that the probability of an address being selected is proportional to the number of addresses within a given geographic area with a higher number of addresses being selected from densely populated areas WAS uses a two stage or clustered approach to sampling Firstly postcode sectors are randomly selected from the PAF The postcode sectors are the primary sampling units PSUs for the survey Within each of these postcode sectors 26 addresses are randomly selected The selection uses a stratified ordered PAF where addresses are listed by postcode and street number The list of 26 addresses is split into two quotas of 13 addresses to ease the allocation to interviewers and management of fieldwork The sampled PSUs were allocated to months at random This was done using a repeating random permutation which ensured that PSUs allocated to the same quarter and month were evenly spread across the original sample while still ensuring that each sampled PSU had an equal chance of being allocated to each month This even spread meant that monthly and particularly quarterly samples were balanced with respect to the regional and census based variables used in the stratification Although the address selection within postcode sectors is random some addresses
70. time spent on different questions during interviews by placing time stamps at the start and end of different questions We can use the session log file called the audit trail in Blaise to time individual questions This method affords us the ability to monitor how different questions contribute to the overall length of the questionnaire which is essential when conducting questionnaire content reviews Other features of Blaise which make it excellent for undertaking the Wealth and Assets Survey include e the ability for interviewers to back track in instances where later sections of an interview highlight an error made earlier e flexibility over styles fonts font sizes and colours Blaise allows these to be specified for all text or for individual words questions etc This helps ensure the screen seen by the interviewer is as well designed as possible with effective interviewer prompts This in turn helps promote interviewer respondent rapport thereby contributing to better data quality http www blaise com capabilities 19 e the ability to interact with a question by question QbyQ help facility This provides interviewers with real time access to guidance on specific questions during the interview This is an electronic programme that operates in conjunction with Blaise The Wealth and Assets Survey questionnaire is tested extensively prior to being scattered to field interviewers Currently staff in the research team
71. to T and survivor sub samples are provided in Table 4 1 Descriptive statistics for the W1 W2 weights are also provided for comparison purposes It is clear that due to attrition the sample size has decreased from W2 to W3 The sample size for the survivor group who were present in W1 W2 and W3 is smaller again as this is a subsample of the W2 W3 longitudinal cases As the sample sizes get smaller the mean weight increases as each individual represents a larger proportion of the population The variation of the weights has increased since last wave as the weights have been further adjusted to account for more characteristics of the evolving sample non response and unknown eligibility status It is interesting to note that the variation in the weights is higher for the W2 W3 longitudinal cases than it is for the survivors even though the survivors are a smaller subset This has occurred because the minimum weight is smaller for the W2 W3 population whilst the maximum weights are the same for both subsamples The maximum weight reflects the trimming and the constraints imposed in the calibration process i e the gg percentile of the unadjusted distribution The minimum weight is greatest for the survivors which is expected as each individual counts for a larger proportion of the population Table 4 1 Descriptive statistics for the longitudinal weights Weight n Mea Standar Coefficien Minimu Maximu n
72. to the number of times that category or value was observed with respect to the total number of observation This strategy served to support the aim of ensuring that the imputation did not have an unwarranted impact on the distributional properties of data 26 Processing Strategy The Wealth and Assets Survey data were processed in three Sections Property Physical Pensions and Financial For all variables imputation followed a basic processing strategy First missing routing was imputed against an appropriate set of auxiliary variables Following that where the routing indicated a missing value for the amount associated with a particular asset liability the value was imputed against its own set of auxiliary variables To meet the key analytical aim of the survey to provide longitudinal estimates of change over time as well as cross sectional single year estimates the detail of the basic processing strategy varied for cross sectional data belonging to respondents new to the survey compared to the longitudinal data belonging to respondents who had been in the survey for both Wave1 and wave two In general for respondents with cross sectional data only processing focused on imputing a discrete category or value drawn from the range and distribution of categories values observed directly in the data of records reaching the final potential donor pool For these respondents donors were selected against a set of auxiliary variables in a way
73. ugh the weight share In W3 this has not yet been carried out due to time constraints Further planned work involves recalculating the cross sectional weights using the W2 method and analysing the impact on the weights We believe that the impact will be relatively small as it is unlikely that the mother s weight will vary much from the household average Depending on the results we will conclude that either 1 the impact on the weights is negligible and does not warrant making changes to the existing W3 weights or 2 the impact on the weights is substantial so the existing W3 weights will be updated accordingly The conclusion will also form the basis of a recommendation on the method to be used for all future waves Typically pre calibrated weights were carried forward from the previous waves during the weighting strategies described above It would be desirable to carry out further analysis in order to quantify the differences between the final W3 weights produced using pre calibrated and calibrated W2 weights to see if there is a significant difference The principle aim of WAS is the estimation of gross change but it is also important to produce cross sectional estimates of wealth over time which in turn enables the computation of net change ONS will produce weights for each new release of data arising from each completed survey wave to assist both longitudinal and cross sectional analysis WAS weights are model assisted d
74. vary over repeated random sampling In order to estimate standard errors correctly the complexity of the survey design needs to be accounted for as does the calibration of the weight to population totals see Weighting WAS has a complex design in that it employs a two stage stratified sample of addresses with oversampling of the wealthier addresses at the second stage and implicit stratification in the selection of PSUs Although data users should produce standard errors with the outputs of their analysis with the WAS datasets available at UKDA this is not possible without design information details of weights stratification clustering and calibration Such information could not be provided with the datasets for statistical disclosure reasons However methodologists in ONS are planning to develop and test the generation of appropriate standard errors Note that some initial estimates of standard errors for key variables are available in the supporting tables to the report referred to above but imputation effects need to be taken account of so these should be treated as preliminary more accurate estimates would be likely to be larger Non sampling error Additional inaccuracies which are not related to sampling variability may occur for reasons such as errors in response and reporting Inaccuracies of this kind are collectively referred to as non sampling errors and may occur in a sample survey or a census The main sources of non samp
75. w is very short about five minutes The requirements of KITE are much simpler than the mainstage interview therefore in order to reduce costs and maximise value for money the interviews are conducted using Computer Assisted Telephone Interviewing CATI 11 Fieldwork procedures The following provides a summary of interviewer training prior to starting a HAS quota of interviewing how progress is monitored and performance benchmarked during data collection and how contact is maintained with HAS respondents between waves Interviewer training Interviewers working on the Wealth and Assets Survey have received both generic field interviewer and survey specific training Generic interviewer training New interviewers to ONS are placed on a six week training programme the Interviewer Learning Programme ILP where they are equipped with the skills required for social survey interviewing The programme coordinates the activities of managers trainers and interviewers into a structured programme that ensures all interviewers can meet the high standards expected of an ONS interviewer The training adopts a blended learning approach Methods used include classroom training instructional and activity based workbooks instructional and activity based e learning applications activity based applications that test the interviewer s skills and knowledge base At the end of the six weeks interviewers continue to be supported in their persona
76. was already aged over 65 a is the individual s age at interview Definition of wealth from pensions in payment In order to calculate the value of the future stream of income provided by pensions from which the individual was already receiving an income the lump sum which the individual would have needed at the time of interview to buy that future income stream from a pension provider was calculated Wealth from pensions in payment was therefore defined as W AY Where Aais the age and sex specific annuity factor based on respondent s current age a Yp is reported current annual private pension income For those age groups for whom no market annuity factor was available ages 75 and over we predicted a hypothetical annuity factor based on the information from those ages where annuity prices were available Contact details For further information or to provide feedback on the Wealth and Assets Survey documentation please contact Craig Orchard Phone 01633 455755 Email Craig Orchard ons gov uk 50
77. which the participant first responded and whether respondents are original sample members OSMs or SSMs These different categories are treated differently throughout the weighting process and so it is necessary to categorise individuals before weighting A variable called sumstat has been created to identify the different categories see table 2 1 The classification of each sumstat is included which describes the sampling and responding behaviour of individuals in each sumstat This includes terms such as OSM an Original Sample Member which refers to an individual who was sampled and who responded in the first wave EOSM an Entrant Original Sample Member which refers to an individual who lives at an address which was sampled in the first wave but the household did not respond until a later wave SSM a Secondary Sample Member which refers to an individual who joined a previously responding household Sumstat Wave 1 Wave 2 Wave 3 Classification o Non productive Y OSMnotinw2orw3 2 Y w2EOSMnotinwz 3 7 WsSMnotinw 8 OsM s J 7 w sa Y Wsnewcohot 7 vV 7 OSMnotinw_ s 7 7 weosm o v 7 jwsm_ ao 7v Y OSMnotinw Table 2 1 Categories of key subsets of respondents 31 Different Types of Weights Along with the numerous combinations of responding patterns as the survey develops there are also numerous longitudinal weights that coul

Wealth and Assets Survey User Guide

Contents

Download Pdf Manuals

Related Search

Related Contents