In This Issue · newsletter. This issue is the last under her watch. Our next editor will be Steve Heeringa ([email protected]). I thank him for agreeing to keep the Survey Statistician

EditorsLeyla MohadjerJairo Arrow

Section EditorsJohn Kovar — Country ReportsJames Lepkowski — Software Review

Production EditorTherese Kmieciak

CirculationClaude OlivierAnne-Marie Vespa-Leyder

The Survey Statistician is published twicea year in English and French by theInternational Association of SurveyStatisticians and distributed to all itsmembers. Information for membership inthe Association or change of address forcurrent members should be addressed to:

Secrétariat de l’AISE/IASSc/o INSEE-CEFILAtt. Mme Claude Olivier3, rue de la Cité33500 Libourne - FRANCEE-mail: [email protected]

Comments on the contents or suggestionsfor articles in The Survey Statisticianshould be sent via e-mail [email protected] or mailed to:

Leyla MohadjerWestat1650 Research Blvd., Room 466Rockville, MD 20850 - USA

In This IssueNo. 48, July 2003

1 Letter from the President

3 Charles Alexander, Jr.

4 Software Review

15 Country Reports

22 Articles

22 Training Needs for Survey Statisticians inDeveloped and Developing Countries

27 Special Articles: Censuses Conducted Aroundthe World

27 – The Possible Impact of Question Changes on Dataand Its Usage: A Case Study of Two South AfricanCensuses (1996 and 2001)

32 Discussion Corner: Substitution

38 Announcements

38 IASS General Assembly38 IASS Elections38 Joint IMS-SRMS Mini Meeting38 IASS Program Committee for ISI, Sidney 200539 Publication Notice40 Association for Survey Computing 200341 IASS Web Site

42 In Other Journals

42 Survey Methodology43 Journal of Official Statistics45 Statistics in Transition48 The Allgemeines Statistisches Archiv

Change of Address Form

List of IASS Officers and Council Members

List of Institutional Members

The Survey Statistician July 20031

Lett

er fr

om th

e Pr

esid

ent

Time goes by quickly�too quickly. Ithas already been four years since youelected me to succeed Kirk Wolter asPresident of the IASS and two yearssince I actually took up my duties. Andalready I am writing you this fourth andfinal �Letter from the President,� beforeI, in turn, yield my position to LuigiBiggeri at the ISI session in Berlin.This, then, is a letter of goodbye that Iam sending you today. It is also a sortof summing up of these two years; ittherefore includes topics that havealready been discussed in previousissues.

During these two years that I was thepresident of our association, I believethat a number of things have beenaccomplished, while there has beenlittle change for some others. Whatcould be done was achieved inconjunction with not only the Council�in particular the Executive�but alsothe committee chairs, some especiallyactive and devoted members, and theSecretariat. To all of them, I wish toexpress my warmest appreciation.

Survey StatisticianYou have been regularly receiving ourbeloved Survey Statistician, of whichthis is the 48th issue. Somewhatirregular in its frequency and content atthe beginning, it has gradually found itsplace. I wish to thank the editor, LeylaMohadjer, who has accomplished, withconsistency and discretion, aformidable task in terms of both formatand content over the past four years.She has had her share of problems,and she has effectively surmounted agreat many of them.

After these four years, Leyla is passingthe torch, as was planned when shetook on responsibility for thenewsletter. This issue is the last underher watch. Our next editor will be SteveHeeringa ([email protected]). Ithank him for agreeing to keep theSurvey Statistician going for thecoming years.

However, some problems remain. Thecountry reports, overall, are the objectof some criticism. Also, the �Questions-Answers� section has been inactive forseveral years now. There will be moreto say about this in connection with the

Web site.

I must once again thank those who areinstrumental in bringing out this publication:for translations, Statistics Canada; and forprinting and distribution, the AustralianBureau of Statistics for the English versionand INSEE for the French version.

Web SiteFollowing a change in assignment, FredVogel has asked to be relieved of hisresponsibilities as Webmaster. Hisreplacement is Eric Rancourt([email protected]).

Considering that the Internet is in generaluse throughout the entire world, I think thatthere should be a broad discussion as tothe place and role of the Survey Statisticianand the web site respectively, with aparticular focus on the two problemsmentioned above regarding the newsletter.This discussion could begin at the BerlinSession. Some have already given somethought to these topics, and the session willbe an opportunity to exchange viewpointson the matter. But what happens after thatwill be the responsibility of my successor,Luigi Biggeri.

PublicationsIn the last two years there have been morepublications than usual.

Aside from the traditional publications�theSurvey Statistician, which I have alreadymentioned, and the Proceedings of theSeoul Session, prepared by the then chairof the IASS Programme Committee for thatsession, David Binder, who is now Vice-President of the Association, and by EricRancourt�two exceptional works werepublished: the �IASS JubileeCommemorative Volume � LandmarkPapers in Survey Statistics,� prepared by acommittee chaired by Gad Nathan, and�Leslie Kish, Selected Papers,� prepared byGraham Kalton and Steve Heeringa.

The latter work was publishedcommercially by Wiley and Sons, and theAssociation purchased enough copies forall its members to receive one. TheProceedings of the Seoul Session and theJubilee Volume, which the Associationpublished on its own, were printed anddistributed to the membership by theAustralian Bureau of Statistics.


I want to express hearty thanks to all thoseinvolved in preparing these works, as well as to theABS and its Director, Dennis Trewin.

The membership directory, which could not beproduced in 2001, was distributed at the end of2002, thanks to the contribution of INSEE. Also, theISI has just brought out a directory that includes allmembers of the Institute itself and the members ofthedifferent sections, including, for the first time,those who only belong to the IASS.

Cochran-Hansen PrizeThe 2003 Cochran-Hansen Prize has beenawarded to Krishna Mohan for his contribution on�A Multilevel Approach to Explore the Influences ofInterviewers on Item Nonresponse in ComplexSurveys�. The jury assembled to analysecandidates� contributions was chaired by ChrisSkinner. I wish to thank him and his entire team fortheir efforts.

Administrative MattersThe Secretariat has made new progress inoverhauling its membership database. We hope totake this opportunity to make the database easierto manage, and also to bring ourselves more in linewith the principles and practices of the ISIPermanent Office in the management of its file ofmembers of the Institute and the other sections.

Berlin SessionThe statutory meetings of the Association will takeplace at the Berlin Session:

! Council, August 14, 7:30 a.m. to 9:00 a.m.;! General Assembly, August 14, 11:15 a.m. to

1:00 p.m.; and! Executive Committee, August 19, 7:30 a.m. to

9:00 a.m.I hope that all Association members present inBerlin will make a point of participating in theGeneral Assembly.

Also, the Programme Committee for the SydneySession, chaired by Pedro Luis do NascimentoSilva, will meet on August 15, from 11:15 a.m. to1:00 p.m. As I indicated earlier, ad hoc meetingswill be organized on the following two themes:

! the Survey Statistician and the web site; and! the IAOS/IASS Joint Conference on Poverty,

Social Exclusion and Development.The place and time of these meetings will beannounced on-site in Berlin.

Lastly, I would like mention here once again the�short courses� organized by our Scientific

Secretary, Seppo Laaksonen. They were listed inthe last issue of the Survey Statistician.

IAOS/IASS Joint ConferenceI have already made several references in thiscolumn to the conference, organized jointly with theIAOS, to be held in 2004 in Abidjan, Côte d�Ivoire.Because of that country�s internal situation, theIAOS President Paul Cheung and I have decided topostpone the Abidjan Conference. Also, the Chairof the Programme Committee, Alain Azouvi, whohas devoted considerable personal effort, has notreceived the support needed to obtain trulyinternational validation of his scientific programmeor to identify leaders for meetings. We have not yettaken a definite position on where to go from here,but on this point too, we will have in-depthdiscussions in Berlin with the new Council and allthose concerned with this conference.

I hope that I will see many of you in Berlin. Fornow, let me conclude by saying au revoir to all ofyou. I have been both honoured and happy topreside over the Association these past two years,and now I pass the torch to Luigi Biggeri and theCouncil that you have just elected, whose makeupis not yet known at the time of writing.

CHANGE OF ADDRESS

Members are encouraged to inform the IASSSecretariat of changes of address as soon aspossible. Mailings of the proceedings of theIASS papers presented at the ISI sessions, andThe Survey Statistician will be delayed and maybe lost if the Secretariat does not have yourcorrect address.

You may notify Ms. Claude Olivier of yourchange of address by completing and mailingthe Change of Address form given at the end ofthis newsletter. Alternatively, you can providethe same information to Ms. Olivier by email [email protected].

mailto:[email protected]


In MemoriamCharles Alexander, Jr.

“Chip”1947 - 2002

Chip Alexander, statistician at the United StatesCensus Bureau and Fellow of the AmericanStatistical Association, died on September 1, 2002in a drowning accident. He is survived by his wifeDiane and children David and Emily. Chip receivedhis B.A. degree in Mathematics in 1965 fromPrinceton University. In 1974, he earned a Ph.D. inStatistics from the University of North Carolina atChapel Hill. He taught Statistics at the StateUniversity of New York at Binghamton for fiveyears. When his father became ill, he moved toMaryland and began his 23-year career at theCensus Bureau.

Over the past decade, Chip Alexander developedthe methodology for what is now known as theAmerican Community Survey, using the rollingsample methods of Leslie Kish and others. Hecontinued to shepherd all statistical aspects of theAmerican Community Survey until it was adoptedas the �key to the Census Bureau�s future.� Whenfully implemented, it will provide demographic,economic, and housing profiles of America�scommunities every year based on a nationalsample of 3,000,000 housing units and willeliminate the need for a long form in the 2010census.

The development of continuous measurementmethodology in the 1990s was one of the mostsignificant contributions in census taking in severaldecades. This field of study involves integratingmultipurpose survey designs with a variety ofstatistical methods for combining information fromdifferent sources, to create data products thatsimultaneously serve multiple purposes. Under thedirection of Dr. Alexander, estimation methodswere developed using multi-year moving averagesintegrated with the Census Bureau�s annualpopulation estimates.

Dr. Alexander was the Census Bureau�s chiefspokesperson on the American Community Surveymethods. He was invited to speak at manystatistical meetings in the United States and aroundthe world. He was also a frequent speaker at datauser groups, and at meetings with legislators andother policymakers. He had an unusual knack forpresenting esoteric, technical, statistical informationin a manner conducive to consumption by a broadconstituency. According to two of his colleagues,Cynthia Taeuber and Paula Schneider, �He waspatient and never condescending when we missedimportant points or didn't get things quite right. Hepolitely but strongly suggested how we couldimprove our contribution to the goal. He wouldn't letus misrepresent, oversell, or undersell his conceptand plan. His legacy to the nation is hiscontributions to the design of the AmericanCommunity Survey.�

A session on the American Community Survey inhonor of Dr. Alexander will be held at the JointStatistical Meetings in San Francisco, U.S., August3-7, 2003. The participants include Jay Waite,Freddie Navarro, Debbie Griffin, and Sue Lovefrom the United States Census Bureau, Joe Salvofrom the NYC Planning Commission, Eric Slud fromthe University of Maryland, and Alan Zaslavskyfrom Harvard University.

Al Tupek will include a tribute to Dr. Alexander inhis presentation on �Combining Years of Data froma Rolling Sample Survey� by Alexander and Tupekat the ISI Meeting in Berlin, Germany, August 13-20, 2003 as part of an invited session on �Newapproaches to population censuses (a specialmemorial to Leslie Kish).�

Prepared by Alan R. Tupek, USA


IVEware: Imputation and Variance Estimation Software

John Van HoewykSurvey Research Center, University of Michigan

OverviewIVEware is an imputation and variance estimation software application developed by the Survey MethodologyProgram at the University of Michigan�s Survey Research Center (SRC). The SRC makes IVEware availablewithout charge.

IVEware performs single and multiple imputations of missing values using the Sequential RegressionImputation Method (Raghunathan et al., 2001), analyzes multiply imputed data sets, and takes into accountcomplex sample design features such as clustering, stratification, and weighting.

IVEware�s four modules are SAS® -callable applications; users work within the SAS system. IVEware setupfiles (command files) use keywords, commands, and instructions similar to those used in SAS. Setup files aresubmitted directly from the SAS program editor (or by way of a SAS batch file). IVEware imputation andanalysis modules require SAS data sets. IVEware users are expected to have a moderate amount of SASexperience, including familiarity with basic file concepts, naming conventions, and command file structures.

IVEware is currently available for personal computers using the Microsoft Windows® or Linux operatingsystems, and UNIX workstations using Sun Solaris®, IBM AIX®, or Compaq/DEC Alpha Tru64Unix® operatingsystems. IVEware requires SAS version 6.12 or higher.

IVEware ModulesIVEware’s four modules are IMPUTE, DESCRIBE, REGRESS, and SASMOD.

IMPUTE uses a multivariate sequential regression approach to imputing item-missing values. Imputations aredraws from the predictive conditional distribution of the variable with missing values, given all other variables.Imputations are repeated over several iterations, each time overwriting previously imputed values, buildinginterdependence among imputed values and exploiting the correlational structure among covariates.

IMPUTE can impute missing values for any of five variable types: continuous, binary, categorical, count, ormixed (continuous variables with a nonzero probability mass at zero), executing the appropriate regressionmodel�linear, logistic, generalized logit, Poisson, or a mixed logistic/linear�once the type of variable typehas been declared.

IMPUTE permits users to specify interaction terms in the imputation model and execute stepwise regressions.

IMPUTE can also accommodate two common features of survey data that add to the complexity of themodeling process: the restriction of imputations to subpopulations and the bounding of imputed values.IMPUTE can produce a single or multiply imputed data set. (A technical discussion of the sequentialregression imputation procedure can be found in the appendix.)

The following illustration of the IMPUTE setup file and resulting printed output is abbreviated from a largersetup used in the imputation of the public use National Health and Nutrition Examination Survey III(NHANES).

SOFTWAREREVIEW


Setup File

The input data file is identified by the keyword DATAIN. DATAOUT indicates the name of the output file thatwill contain the imputed data. ITERATIONS specifies the number of cycles of imputation to be undertaken;MULTIPLES indicates the number of imputations to be performed; and SEED specifies a seed for the randomdraws from the posterior predictive distribution.

DATAIN ADULTXR;DATAOUT ADULTOT;

ITERATIONS 5;MULTIPLES 5;SEED 11111;

DEFAULT CATEGORICAL;TRANSFER SEQN ADULT DMPFSEQ HFAGERR BMPTRI BMPSUB BMPSUP;CONTINUOUS HFA8R HSAGEIR DMPPIR HAM5S . . . ;MIXED HAN6HS HAN6IS HAN6JS HAR15 . . . ;

BOUNDS HFA8R (>=0,<=17), HAM5S (>=41 ), HDP (>=8), . . . ;RESTRICT HAR3 (HAR1 =1), FPPSUDRU (HSAGEIR >=40),. . . ;RUN;

The keywords DEFAULT, TRANSFER, CONTINUOUS and MIXED identify the variables in the input data set.In this example, DEFAULT classifies undefined variables as categorical. Variables declared TRANSFER arecarried over to the imputed data set but are not imputed or used as predictors in the imputation model.Continuous variables are identified by the CONTINUOUS keyword, while MIXED variables treat a value ofzero as a discrete category and values greater than zero are considered continuous. The mixed variableHAN6HS (amount of beer consumed per month) will be imputed using a two-stage model. First, a logisticregression model is used to impute zero (nondrinker) vs. nonzero (drinker). Once a nonzero (drinker) hasbeen imputed, a normal linear regression model imputes nonzero values (amount of beer consumed).

The BOUNDS statement restricts the range of the values to be imputed. HFA8R (Years of Education) will beimputed with values ranging between 0 and 17.

RESTRICT limits the imputation of a variable to a subpopulation within the data set. The imputation of HAR3(currently smoke cigarettes) is restricted to those respondents who report having ever smoked 100 cigarettesin their lifetime�HAR1.

Printed Output

For each imputation, IMPUTE provides means (continuous variables) and proportions (categorical variables)for the observed cases, imputed cases, and all cases. The printed results for the variables HFA8R, HAR1,and HAR3 are shown below.


VARIABLE HFA8R OBSERVED IMPUTED COMBINED NUMBER 19827 223 20050 MINIMUM 0 0.323875 0 MAXIMUM 17 16.9898 17 MEAN 10.7901 9.36205 10.7742 STD DEV 3.88592 3.48302 3.88448

VARIABLE HAR1 OBSERVED IMPUTED COMBINED CODE FREQUENCY PERCENT FREQUENCY PERCENT FREQUENCY PERCENT 1 9799 48.91 10 62.50 9809 48.92 2 10235 51.09 6 37.50 10241 51.08 TOTAL 20034 100.00 16 100.00 20050 100.00

VARIABLE HAR3 OBSERVED IMPUTED COMBINED CODE FREQUENCY PERCENT FREQUENCY PERCENT FREQUENCY PERCENT 1 4990 50.93 4 0.04 4994 24.91 2 4807 49.07 8 0.08 4815 24.01 3 0 0.00 10241 99.88 10241 51.08 TOTAL 9797 100.00 10253 100.00 20050 100.00

For the continuous variable HFA8R (years of education), 223 observations were imputed. The setup called forthe imputation of HFA8R to be bounded by the range 0 through 17. Accordingly, the resulting range ofimputations is 0.323 to 16.98.

The categorical variable HAR1 (ever smoke 100 cigarettes) had 16 observations with missing data. Ten wereimputed to the value 1 (yes), while the remaining 6 were imputed to the value of 2 (no).

The imputation of HAR3 (currently smoke cigarettes) is restricted to those observations with a value of 1 onHAR1 (yes to ever smoke 100 cigarettes). HAR3 had 12 observations with missing data. Ten of the 12 werefirst imputed to the value 1 on HAR1. The remaining 2 had an observed value of 1 on HAR1, but were missingon HAR3. Observations outside the restriction are given the value 3.

DESCRIBE estimates population means, proportions, subgroup differences, contrasts, and linearcombinations of means and proportions. A Taylor Series approach is used to obtain variance estimatesappropriate for a user-specified complex sample design (Kish and Frankel, 1974).

Under DESCRIBE, multiple imputation analysis can be performed. If multiple data sets are specified as inputs,DESCRIBE performs the analysis for each data set and combines the inferences (Rubin, 1987).

The following is an example of the DESCRIBE setup file and printed output.

Setup File

The DATAIN keyword identifies the five imputed data sets that are to be used in the analysis. The five wereextracted from the file ADULTOT produced in the above IMPUTE example. The next three keywords�WEIGHT, STRATUM and CLUSTER�specify the survey design variables.


DATAIN ADULTOT1 ADULTOT2 ADULTOT3 ADULTOT4 ADULTOT5;

WEIGHT WTPFHX6; STRATUM SDPSTRA6; CLUSTER SDPPSU6;

TABLE HAR1; MEAN HFA8R; RUN;

TABLE produces estimates of population proportions for HAR1 (ever smoke 100 cigarettes) and MEANestimates the population means for HFA8R (years of education).

Means and proportions can also be estimated for subpopulation with a BY statement. For example, BYGENDER would produce separate estimates for males and females in addition to estimates for males andfemales combined.

Printed Output

DESCRIBE provides weighted and unweighted estimates of the proportions and means, along with standarderrors that take into account survey design features. The printed output for HAR1 and HAF8R appears below.

��(��continued

��)


Problem 1

Degrees of freedom Infinite

Factor Covariance of denominator None 0.03317

Table Number of Sum of Weighted Standard HAR1 Cases Weights Proportion Error 1 8850 9.882268e+007 0.52702 0.00771 2 9312 8.869123e+007 0.47298 0.00771

Lower Upper T Test Prob > |T| Bound Bound 1 0.51189 0.54214 68.35834 0.00000 2 0.45786 0.48811 61.35014 0.00000

Unweighted Bias Design Proportion Effect 1 0.48728 -7.53945 4.33045 2 0.51272 8.40071 4.33045

All imputations

Problem 2

Degrees of freedom Infinite

Factor Covariance of denominator None 0.03317

Mean Number of Sum of Weighted Standard HFA8R Cases Weights Mean Error 18162 1.875139e+008 12.24597 0.08058174

Lower Upper T Test Prob > |T| Bound Bound 12.08784 12.4041 151.96949 0.00000

Unweighted Bias Design Mean Effect 10.77538 -12.00875 11.65533

REGRESS fits linear, logistic, polytomous, Poisson, Tobit, and proportional hazard regression models tosurvey data with complex sample designs. The jackknife repeated replication approach is used to estimate thesampling variances (Kish and Frankel, 1974).

REGRESS can output predicted values, model estimates, and variances/covariances into SAS-type outputfiles. Another feature allows users to produce a series of plots�residuals, standard deviation and normalprobability�that are useful for regression diagnostics.

If there are missing values, multiple imputation analysis can be performed. When multiple data sets arespecified, REGRESS will perform the analysis for each data set and combine the inferences.


Setup File

The DATAIN keyword identifies the five imputed data sets; WEIGHT, STRATUM and CLUSTER specify thesurvey design variables.


WEIGHT WTPFQX6; STRATUM SDPSTRA6; CLUSTER SDPPSU6;

PREDOUT PRED_DIA;

LINK LINEAR;

DEPENDENT HDP; PREDICTOR LDRINK SMOKER ACTIVE BMI NONWHITE HISP FEMALE AGE HFA8R; CATEGORICAL SMOKER; RUN;

PREDOUT instructs REGRESS to output the predicted values to the file PRED_DIA. LINK defines the type ofregression model to be fit.

Keyword DEPENDENT names HPD (serum cholesterol) as the outcome variable, and PREDICTOR identifiesthe independent variables. CATEGORICAL further defines the predictor SMOKER as a categorical variable.REGRESS will create dummy variables for each of the three categories of SMOKER.

��(

��continued

��)


Printed OutputThe results include the combined estimates of the imputed data sets and standard errors produced by thejackknife repeated replication approach.

NUMBER OF VALID CASES: 20050SUM OF WEIGHTS: 187647206.32

DEGREES OF FREEDOM: 8.3107193138

SUM OF SQUARES: MODEL: 9634751445.5 ERROR: 34344022347 TOTAL: 43978773792 R-SQUARE: 0.21908 F-VALUE: 0.21195 P-VALUE: 0.99009

VARIABLE ESTIMATE JACKKNIFE T TEST PROB > |T| STD ERRORINTERCPT 55.096624473 1.4817833693 37.18264 0.00000LDRINK 2.7560503214 0.15092331379 18.26126 0.00000SMOKER.1 -3.2556773734 0.35164732811 -9.25836 0.00001SMOKER.2 -0.54375056663 0.39316430078 -1.38301 0.20268ACTIVE 1.0023113894 0.44933488718 2.23066 0.05502BMI -0.6443368471 0.032059454534 -20.09818 0.00000NONWHITE 3.4416654566 0.50416970404 6.82640 0.00011HISP 0.86577743998 0.5431173425 1.59409 0.14817FEMALE 10.429910556 0.36754005066 28.37762 0.00000AGE 0.057795637071 0.010428137771 5.54228 0.00048HFA8R 0.09519964185 0.053121777683 1.79210 0.10948

VARIABLE ESTIMATE 95% CONFIDENCE INTERVAL LOWER UPPERINTERCPT 55.096624473 51.701743978 58.491504968LDRINK 2.7560503214 2.4102733173 3.1018273255SMOKER.1 -3.2556773734 -4.0613286448 -2.4500261021SMOKER.2 -0.54375056663 -1.4445204387 0.35701930545ACTIVE 1.0023113894 -0.027149647717 2.0317724266BMI -0.6443368471 -0.71778754112 -0.57088615307NONWHITE 3.4416654566 2.2865736059 4.5967573073HISP 0.86577743998 -0.37854646775 2.1101013477FEMALE 10.429910556 9.5878478257 11.271973286AGE 0.057795637071 0.033903965554 0.081687308588HFA8R 0.09519964185 -0.026506466241 0.21690574994

VARIABLE DESIGN SRS % DIFF EFFECT ESTIMATE SRS V ESTINTERCPT 3.0504572909 56.550513434 2.6387986103LDRINK 3.153969167 2.5549735698 -7.2958301989SMOKER.1 1.7748105226 -2.4735989803 -24.021986929SMOKER.2 2.0910889982 -0.37947585571 -30.211409606ACTIVE 3.9332530261 0.53650958126 -46.472764161BMI 2.7991974738 -0.66034405523 2.4842919052NONWHITE 3.4890495224 5.350109188 55.451169074HISP 3.2984201436 0.86321216685 -0.29629706448FEMALE 2.0868771613 9.5436271006 -8.4975173115AGE 2.2302588605 0.058163151143 0.63588549416HFA8R 2.9480248483 0.041161687673 -56.762770455


SASMOD allows SAS users to take complex sample design features into account. Currently the following SASPROCS can be called: CALIS, CATMOD, GENMOD, LIFEREG, MIXED, NLIN, PHREG, and PROBIT.

After specifying sample design features with IVEware commands, the user simply adds the SAS PROCstatements to the setup file. SASMOD executes the user-specified SAS PROC for every replicate andcombines the results to compute the proper variance estimates.

SASMOD, like DESCRIBE and REGRESS, is capable of accepting multiply imputed data sets. Each data setis analyzed separately and the results are combined.

The following example illustrates how to use SASMOD to account for survey design features when using theSAS CATMOD procedure.

Setup File

The five imputed data sets are identified with the DATAIN keyword and survey design variables are indicatedby the WEIGHT, STRATUM, and CLUSTER keywords. The SAS CATMOD statements follow.

CATMOD calls for HAB1, a five category self- assessment of health status variable, to be regressed onrespondent gender, smoking status (current, former, never), and race.


WEIGHT WTPFQX6; STRATUM SDPSTRA6; CLUSTER SDPPSU6;

PROC CATMOD; MODEL HAB1=FEMALE SMOKER NONWHITE; RUN;

��(��continued

��)


Printed Output

SASMOD produces the following output for the SAS CATMOD procedure.

Valid cases 14058Sum weights 155428259.8

Degr freedom Infinite

-2 LogLike 428649627.2

Variable Estimate Std Error Wald test Prob > Chi[1]Intercept 1.8975252 0.0963519 387.84189 0.00000[2]Intercept 2.2929562 0.0879903 679.08194 0.00000[3]Intercept 2.5117710 0.0881987 811.02740 0.00000[4]Intercept 1.5661782 0.0991464 249.53341 0.00000[1]FEMALE.1 0.2137564 0.0833359 6.57921 0.01032[2]FEMALE.1 0.1503387 0.0718700 4.37569 0.03646[3]FEMALE.1 0.0783372 0.0748061 1.09663 0.29501[4]FEMALE.1 -0.0239401 0.0918233 0.06797 0.79431[1]SMOKER.1 -0.5242103 0.1358919 14.88072 0.00011[2]SMOKER.1 -0.3006035 0.1347148 4.97918 0.02565[3]SMOKER.1 -0.0838631 0.1339074 0.39222 0.53113[4]SMOKER.1 -0.0739939 0.1433337 0.26650 0.60569[1]SMOKER.2 -0.0983358 0.1331557 0.54539 0.46021[2]SMOKER.2 -0.1725536 0.1190886 2.09946 0.14735[3]SMOKER.2 -0.1382790 0.1172381 1.39115 0.23821[4]SMOKER.2 -0.0335599 0.1199688 0.07825 0.77968[1]NONWHITE.1 0.4142672 0.1078764 14.74716 0.00012[2]NONWHITE.1 0.4785610 0.1059228 20.41249 0.00001[3]NONWHITE.1 0.1833801 0.1132303 2.62288 0.10533[4]NONWHITE.1 0.0383872 0.1146092 0.11218 0.73767

Variable Odds 95% Confidence Interval Ratio Lower Upper[1]Intercept 6.6693685 5.5203944 8.0574816[2]Intercept 9.9041733 8.3335432 11.7708214[3]Intercept 12.3267413 10.3676923 14.6559665[4]Intercept 4.7883130 3.9417254 5.8167273[1]FEMALE.1 1.2383210 1.0515049 1.4583277[2]FEMALE.1 1.1622279 1.0093483 1.3382632[3]FEMALE.1 1.0814873 0.9338323 1.2524891[4]FEMALE.1 0.9763442 0.8153570 1.1691173[1]SMOKER.1 0.5920227 0.4534468 0.7729483[2]SMOKER.1 0.7403713 0.5683825 0.9644027[3]SMOKER.1 0.9195571 0.7070626 1.1959128[4]SMOKER.1 0.9286774 0.7009881 1.2303229[1]SMOKER.2 0.9063445 0.6979320 1.1769919[2]SMOKER.2 0.8415132 0.6661458 1.0630473[3]SMOKER.2 0.8708557 0.6918812 1.0961269[4]SMOKER.2 0.9669970 0.7641583 1.2236773[1]NONWHITE.1 1.5132614 1.2245535 1.8700368[2]NONWHITE.1 1.6137506 1.3108867 1.9865875[3]NONWHITE.1 1.2012709 0.9619267 1.5001681


Availability

IVEware application files, Installation Guide, and User Guide, along with example setup and SAS data filescan be downloaded at no cost from www.isr.umich.edu/src/smp/ive/.

Appendix: Sequential Regression Multiple Imputation

Imputation methods to replace item missing values in surveys have a long history, including a theoreticallyappealing Bayesian framework (Rubin 1976). The Bayesian approach requires an explicit model conditionedon the fully observed variables and unknown parameters, a prior distribution for the unknown parameters, anda model to impute the missing data mechanism.

But survey data sets have large numbers of variables with diverse distributional forms (such as continuous,dichotomous, and censored variables). Survey data have restrictions that define subsets of cases to receiveimputations for a given variable. They also have logical or consistency bounds that must be incorporated intoan imputation process. Developing a full Bayesian model in the survey setting is a difficult and complex task.

A general purpose multivariate imputation procedure has been developed for survey data with complexstructure which develops full multivariate models that are fully conditional on the observed values(Raghunathan et al., 2001). The approach imputes through a sequence of multiple regressions, variable byvariable, using as covariates all other variables observed or imputed for a given case. The imputed values aredraws from the posterior predictive distribution given in the regression model, and multiple imputed values aredrawn for each missing value. Survey variables can have a continuous, binary, categorical (more than twocategories), count, or mixed (continuous with a mass at zero) distributions.

Let X denote a ×n p matrix of p continuous, binary, nominal, count, or mixed variables, and let …1 2 kY ,Y , ,Ydenote k variables with missing values ordered from least to most amount missing. The joint conditionaldensity of the …1 2 kY ,Y , ,Y given X can be factored as

( ) ( ) ( ) ( )θ θ θ θ θ θ−=… … " …1 2 1 2 1 1 1 2 2 1 2 1 2 1k k k k k kf Y ,Y , ,Y | X , , , , f Y | X , f Y | X ,Y , f Y | X ,Y ,Y , ,Y where the f areconditional density functions and θ j is a vector of parameters. Each conditional density is modeled through anappropriate regression model with unknown parameters θ j , with a diffuse prior distribution for the parameters.

Each imputation consists of c rounds, starting by regressing 1Y on X, imputing missing values under theregression model, using draws from a corresponding predictive posterior distribution. The matrix X is thenupdated by appending 1Y , and 2Y is regressed on the updated X. The imputation is repeated for 2Y , and theregression and imputation process continues through …3 4 kY ,Y , ,Y . The imputation process is then repeatedin rounds 2 through c, modifying the predictor set X to include all jY except the variable being imputed.Cycles are repeated until stable imputed values occur.

The procedure is modified to incorporate restrictions and bounds. In addition, since it is difficult to drawparameter values directly from their posterior distribution with truncated normal likelihoods, a Sampling-Importance-Resampling algorithm (Raghunathan and Rubin, 1988) is used to draw from the actual posteriordistribution.

The entire imputation process for k variables can be repeated m times, yielding m imputed values for eachitem missing value. Standard multiple imputation estimation procedures are then used to estimate the

precision of values estimated from the imputed data set. For example, if ( )α lˆ denotes a vector of regression

coefficients estimated in a regression model from the lth imputed data set, and ( )lV denotes the

corresponding covariance matrix, the multiply imputed estimate of α is ( )α α=

= ∑1

ml

MIl

mˆ , with covariance

matrix ( )

=

+= +∑

1

1ml

MI ml

mV V m Bm

, where ( )( ) ( )( )α α=

= − −∑1

m tl lm MI MI

lˆ ˆB a a mˆ ˆ .

The Survey Statistician

References

Kish, L. and Frankel, M. R. (1974). Inference from complex samples. Journal of Royal Statistical Society,Series B, 36, 1-37.

Raghunathan, T. E., Lepkowski, J. M., Van Hoewyk, J., and Solenberger, P. (2001). A multivariate techniquefor multiply imputing missing values using a sequence of regression models. Survey Methodology, 29, 85-95.

Raghunathan, T.E., and Rubin, D.B. (1988) An application of Bayesian statistics using sampling/importanceresampling to a deceptively simple problem in quality control. In G.E. Liepins and V.R.R. Uppuluri (Eds.), Dataquality control: theory and pragmatics. New York:Marcel Dekker.

Rubin, D.B. (1987). Multiple imputation for nonresponse in surveys. New York:Wiley.

Rubin, D.B. (1976) Inference and missing data (with discussion). Biometrika, 63, 581-592.

! The IASS needs your contributio! Please do not forget to renew yo! As of January 2002, French Fran

dues and subscriptions must be

To All Members

n.ur membership.cs are no longer accepted. As a consequence, the payment of

made in either Euros or U.S. dollars.

July 200314


ALGERIANacer-Eddine Hammouda

Thus far, estimation of infant mortality remains asubject of controversy between Algerianstatisticians and demographers, owing at least totwo phenomena: the rate of coverage of vitalevents and the assumption of false stillbirths. Grossinfant mortality rates based on vital statistics dataare approximately 37 per thousand. Demographerssubject this estimate to a double adjustment forcoverage rates (births, and deaths under 1 year ofage) and for the proportion of false stillbirths. Thisyields an infant mortality estimate greater than 50per thousand. The two adjustment factors werethemselves estimated on the basis of a samplesurvey dating back more than twenty years. Anindirect estimate of the infant mortality rate basedon data from the last census of population appearsto support the demographers� estimate, since it iswithin the same ballpark. But does it apply to theAlgerian context in which a drastic decline in fertilityhas been observed for more than a decade?Sample surveys conducted in the past fifteen yearsdo not settle the matter, since they provide amedian estimate which, moreover, is not veryprecise given the sample sizes used. For thisreason, the WHO, as part of the family healthproject run by the Arab League, conducted asurvey on a sample of 20,000 households in 2002.(SOUABER Hassen [email protected]).

For years, child labour has been a taboo subjectin Algeria. Because of its nature and the method ofcollecting information about it, the phenomenoncannot be gauged adequately by the nationalstatistics available. The Institut national du travail(INT) has arranged for a first survey of child labour,to be conducted in 2003. At the outset, the teamresponsible for technical aspects raised theproblem of the objective pursued by this survey,since the sampling strategy depends on it.Estimating the extent of the phenomenon was ruledout at the outset so as to focus more on the formsthat child labour takes and what it entails.Accordingly, researchers put forward certain

hypotheses stating that these forms of labour differaccording to at least two criteria: the sex of thechild and where the child lives. Based on these twocriteria, the sampling strategy will be to target areaswith low school enrolment, since some form of childlabour is more likely to exist among children whoare unschooled or have left school. The targetpopulation would therefore be children of bothsexes who are between 10 and 15 years of ageand not in school. Age 10 is when schoolattendance starts to drop, and 15 is the age whencompulsory schooling ends. (Kamel [email protected]).

AUSTRALIAGeoff Lee

The West Australian Aboriginal Child HealthSurvey (WAACHS) is a large-scale epidemiologicalstudy of the health, development and well-being ofaboriginal children and their families in WesternAustralia. It has been conducted by the WestAustralian (WA) Institute of Child Health Research(ICHR). The survey selected a 1 in 6 sample offamilies with aboriginal children under 18 years ofage. In total, 2,000 families and 5,300 childrenhave provided information for the survey.Interviews were conducted with both primary andsecondary carers where possible, as well as withyouths aged 12-17 years. For children attendingschool, schools were contacted to obtaininformation on academic competence andperformance. The data have been linked with theWA Maternal and Child Health Research Databaseand the WA Linked Database system giving accessto midwives records and records of all hospitaladmissions for survey children and hospitaladmissions for survey carers. The resulting datasetis complex, hierarchical and a valuable resource formany investigations into influences on the health ofAboriginal Children in WA.

Extensive consultation occurred with theindigenous communities to ensure their support forthe survey. Wherever possible indigenousinterviewers were recruited and trained to


enumerate the survey. Particular attention was paidto explaining the survey objectives and processesto selected respondents to ensure that theirparticipation and agreement to the subsequent datalinkages was genuinely an informed decision.

The survey design was an area-based clustereddesign with clustering within Census Districts andwithin families. The sample design was undertakenby the ABS, with the original sample selected fromthe 1996 census frame. ABS outposted officershave worked alongside ICHR for the duration of theproject. A number of interesting lessons have beenlearnt about surveying the indigenous population,and will be discussed in future evaluation reports.Both the sample design clustering and thehierarchical data structure will need to beaccounted for in the data analysis.

ICHR is collaborating with the ABS to produce aseries of three publications reporting the results ofthe survey. The first publication is scheduled to bereleased in September 2003. There are severalanalytical challenges that need to be addressedwithin the timeframe of the first publication.Principal among these is the development ofmodelling techniques that can account for both theclustered survey design and the hierarchical datastructure simultaneously. While techniques havebeen developed to address one or the other ofthese challenges individually, new methods need tobe developed to address both problemssimultaneously.

For more information about the survey contactProfessor Steve Zubrick ([email protected]).

CANADAJohn Kovar

The primary goal of the Youth in TransitionSurvey (YITS) is to provide policy-relevantinformation on the school-to-work transitions ofCanadian youth. For the 2000 panel, thislongitudinal survey comprises two targetpopulations, represented by a cohort of 15-year-olds and a young adult (18-20 years old) cohort,while the 2003 panel comprises 15-year-olds only.The first survey cycle of both panels of the 15-year-old cohort is integrated with the OECD Programmefor International Student Assessment (PISA).

Panel 2000Cycle 1 data for both cohorts of this panel werecollected in 2000. For the 15-year-old cohort theschool participation rate (with replacements) was96 percent; student participation among theseschools was 87 percent. The final response rate forthe 18-20 year-old cohort was 81 percent, reflectingdifferences in the target population, frame andmethod of collection.

Both cohorts have presented estimationchallenges. Given multiple data collection vehiclesof the 15-year-old cohort, several options wereconsidered for defining a respondent and weightingthe data. Ultimately a separate set of weights wascreated for analysis of parent characteristics, dueto 11 percent parent nonresponse amongresponding students. For variance estimation,although PISA data for every participating countrywere released with a set of 80 BRR replicateweights, with the large number of PSUs in theCanada sample we opted instead for the bootstrapmethod. For the 18-20 year-old cohort, in additionto the normal types of weighting adjustments forsurveys based on the Labour Force Survey frame,the initial YITS weights had to be modified to reflecta sample comprising 36 rotation groups withoverlapping clusters.

Development of scales for both cohorts requiredextensive analysis using Item Response Theoryand factor analysis. High-school engagementscales were developed for both cohorts. For the 15-year-old cohort, several scales of psychologicaland social functioning of the student as well as aparenting-style scale were created.

A preliminary data release of the 15-year-old cohortand a national report occurred in December 2001,which included all of the PISA data and most of theYITS student questionnaire variables. In January2002 a national report on the 18-20 year-old cohortwas released, accompanied by a preliminary file ofvariables. Final data for this cohort becameavailable in March 2003; the final cycle 1 productfor the 15-year-old cohort, which includes studentand parent variables, was released later in thespring.

For cycle 2 of panel 2000, a single CATIquestionnaire in BLAISE was developed andadministered concurrently to the two cohorts inJanuary to May 2002. Among the total sample of


52,000, all of whom were cycle 1 respondents, theoverall preliminary response rate was 88.1 percent.However, split by cohort, the respective rates of 90percent and 85 percent for 17-year-olds and 20-22year-olds confirmed the challenge of tracing andcontacting the more mobile older cohort.

Panel 2003Although similar in design to the 2000 panel, thePISA 2003 cognitive tests will focus on the domainof mathematics rather than reading. A pilot surveywas conducted in April-May 2002. A sample of1,425 schools was selected for the main survey,which is designed to meet estimation criteria forPISA as well as the YITS longitudinal component.This is expected to yield a sample of approximately37,000 students in participating schools. Datacollection will occur in two phases, with the studentcomponent in May to June 2003, followed by CATI-administered parent interviews.

For more information on the survey methodology ordata products, contact Johanne Denis (613-951-0402, e-mail: [email protected]), SocialSurvey Methods Division or Lynn Barr-Telford (613-951-1518, e-mail: [email protected]),Centre for Education Statistics, Statistics Canada,Ottawa, Ontario K1A 0T6.

The annual Survey of Household Spending(SHS) collects detailed information on householdexpenditures, dwelling characteristics andhousehold facilities from a sample of about 23,000households. In the past, the survey weights werecalibrated to population counts for three age groupsand by household size. Two major considerationsled to the revision of the calibration strategy.Studies on the representativeness of the samplesuggested that the use of more detaileddemographic controls and the addition of income-related controls would be beneficial. Furthermore aTask Force on Income Statistics identified the needto harmonize the calibration methods used bysurveys on personal and household finance inorder to reduce differences in estimates fromdifferent surveys as well as those from the Systemof National Accounts and administrative incomefiles. As a result, demographic calibration wasstandardised. The use in calibration of tax datafrom the Canadian Customs and Revenue Agencywas also assessed. The objective was to solve theproblem of representativeness by income classobserved in these surveys.

The statement of remuneration paid by anemployer to an employee is called a T4 form. In thefirst step in the introduction of income data incalibration, wages and salaries from the T4 filehave been added to the SHS calibration. Theapproach, which controls on the number ofindividuals within income intervals, has improvedthe income distribution and the precision of theestimates. A reduction of CVs was obtained formost expenditure estimates and this reduction islarger than 10 percent for one quarter of them. Afew problems were experienced with the T4 filesand a clean-up strategy had to be developed. Areconciliation exercise between the T4 file and theincome tax and benefit returns file is in progress.This will improve the clean-up strategy and provideinformation on the coverage of these files.Investigations are still required on the use of othersources of income for SHS. Furthermore, for othersurveys, additional work is needed before incomecomponent are introduced in their calibration.

For further information on the Survey of HouseholdSpending, please contact Johanne Tremblay([email protected] or at 613-951-0682).

ITALYClaudio Quintano and Linda Laura Sabbadini

From Claudio QintanoThe National Survey on violence against womenoriginates from the need to disclose the problem onviolence against women in Italy in terms of itsprevalence, incidence rate and nature. Thisdedicated survey is the first one in Italy and will beconducted in its final phase in 2004 with a nationalrepresentative sample of 30 000 women; allinterviews will be done by adopting CATI(Computer Assisted Telephone Interview) method.The questionnaire that will be used aims to addressseveral aspect of violence against women:

! Prevalence and incidence rate of differenttypes of violence (physical, psychological andsexual). Specific attention will be addressedtowards domestic violence by current or formerpartner;

! Characteristics of those involved andconsequences of violence;

! Risk and protective factors related toindividuals as well as to the socio-demographical domain. Special attention will


be paid to psychological/emotional andeconomical abuse which are usually difficult toreveal, disclose and assess.

Only a limited number of countries haveimplemented a dedicated survey like this one.Statistics Canada in 1993 has for the first timeaddressed this problem, at the end of the nineties,the UNICRI (United Nation Interregional CrimeResearch Institute) under the auspices of theUnited Nations started addressing the problem.Also the National Institute of Statistics in Italy(ISTAT) in the past years (97/98 and 2002)addressed the problem of victimization, anddeveloped a survey on citizen�s safety whereattention is also paid towards measurement ofsexual harassment and rape.

This project is the result of a joint agreementbetween the Department of Equal opportunities andISTAT, under the auspices of United Nations,UNICRI, HEUNI, and Statistics Canada, whichcame up with a project that can be internationallycomparable according to the International Violenceagainst Women Survey.

This type of survey addresses several problemsfrom the procedural and methodological point ofview not faced before by ISTAT. The questionnairehas to be designed in such a way that helps theinterviewee to disclose the violence. This meansthat it has to be easy to administer and tounderstand, not judgmental but clear in the aim itwants to reach and it has to tackle several aspectsof the violence. For this reason, in order toundertake this complex type of survey, severalphases are followed: 1- Qualitative research and 2-Quantitative research.

1) The Italian National Institute of Statistics, inpreparation of the National Survey on violenceagainst women, has conducted a series of FocusGroups (FG) with different groups of peopleworking in the field: women working in shelters forbattered women, social workers, women who havebeen abused, to learn from them the differentpatterns of such violence. The content analysis ofthe FG revealed different aspects of violencebetween intimates, especially in terms of the moresubtle types of psychological violence. In addition,a Group Interview (Delphi) with experts working inthe field (judges, lawyers, police officers) will helpdisclose further aspects of violence against women.

2) The International Team, consisting of expertsfrom different countries (such as Canada, Poland,Denmark, Australia, Costa Rica, Italy) during thefirst semester of the year 2002, has tested, on alimited number of women, a first version of thequestionnaire. This has been done in order to testthe accuracy of the instrument and relevance of theissued addressed. A pilot survey with a sample of1000 women will be interviewed in 2003 to testboth the questionnaire and the procedure. In year2004, the final survey will be conducted with asample of 30000 women. Special attention will bededicated to the selection and the training of theinterviewers and to their supervision and follow-up.The relationship between the interviewee and theinterviewer is essential in helping women disclosingthe violence.

The 2004 final survey will become part of theintegrated social system of multipurpose surveyson households. For more information on the aboveinitiatives, please contact Muratore MariaGiuseppina ([email protected]).

From Linda Laura SabbadiniOfficial poverty estimates are calculated yearlyusing the data of the Household Budget survey.Studies and applications are in progress in order to:obtain reliable poverty values at regional level andevaluate the sampling errors of poverty estimatesin the proper way.

According to the European Council Regulation,ISTAT is going to put in place a new ContinuousLabour Force Survey. During 2003, ISTAT willcarry out simultaneously the traditional quarterlysurvey and the continuous survey, in order to havethe possibility to re-build time series for previousyears. The new survey uses new techniques togather data: a face to face interview with CAPIsystem for the first interview and a telephoneinterview (CATI) for the other three interviews withthe same family. For the first time, ISTAT will useits own team of 311 interviewers for the CAPIinterviews. The data are transmitted daily throughinternet connections.

Within the Italian Time Use Survey an innovatingmonitoring system of PAPI surveys has beentested. We have planned a small questionnaire, tobe filled by interviewers, that contains someimportant information about each survey unit. Inparticular, on the designated date for the


conclusion of interview and diary collection, theinterviewers communicate to ISTAT what is thesituation about each household: has the householdrefused to collaborate, has it filled the diaries, has itpostponed diary day and other specific information.The indicators derived from this information areupdated weekly and analysed in order to have acomplete picture of the fieldwork trend and, whereor when it may be necessary to correct someproblematic aspects. In this way we havemonitored more than 90 percent of the sampleunits and we have offered to ISTAT regional officessome tools to better intervene in the field.

For more information please contact Linda LauraSabbadini ([email protected]).

PHILIPPINESGervacio G. Selda, Jr.

The Bureau of Labor and Employment Statistics(BLES) recently completed the conduct of a projectentitled “Development of a Design for theConduct of a National Migration Survey” underthe Re-engineering the Government StatisticalServices Project�Phase II. The project aimed to fillthe data gaps on migration characteristics of thePhilippine population including their inter-relationships with national and regionaldevelopment patterns. It came up with a sampledesign that will generate reliable estimates onmigration with concern for minimizing cost andmaximizing comparability of estimates across time.In general, the activity�s outputs were:

! sampling design;! questionnaire design; and! estimation procedure.The proposed survey would shed light on theinternal migration and international migrants in thecountry. Some of the information that would becollected in the proposed study are:

! migration history;! education;! training;! income;! occupation;! industry;! household characteristics;! housing;

! asset ownership; and! marriage and family.However, the conduct of a migration survey in themagnitude proposed by this study is a difficult ideato sell at this time of dwindling governmentresources to finance statistical activities. Also, afull-blown migration survey is not in the priority listof government statistical offices. This is due mainlyto the fact that there is no single governmentdepartment or agency which deals directly with theissue of migration as this cuts across the concernsof local governments and national governmentagencies that deals with urban planning,population, labor and employment, and the deliveryof basic needs such as housing, education, health,infrastructure, peace and order. This is why thestudy also included a list of proposed options toaddress the issue of funding support for the survey.For more details on the study, please contact: Ms.Editha B. Rivera, OIC-Director of BLES [email protected].

The Philippine National Statistics Office (PNSO)has completed the preparation for the conduct of aNational Time Use Survey (TUS) to obtaininformation on how people aged 10 years old andover allocates their time to various activities.Specifically, the survey is geared towards:

! determining whether significant differences intime use exist between population groups;

! gathering data that can serve as basis forvaluation of paid and unpaid work for men andwomen of the Philippines;

! providing inputs for the estimation of thecontribution of women and men to country�seconomy;

! providing information of the informal sector,child work activities, etc.; and

! providing information for market research use.The design of the survey was based on the resultsof the 2000 Pilot TUS and subsequent workshopconducted on October 1-2, 2002. The national TUSshall be a rider to the Labor Force Survey (LFS);hence, shall adopt the sampling design of themaster sample using the multi-stage sampling forone panel or one replicate. One panel sample isabout one-fourth of the expanded samples witharound 12,500 households. Reliability of estimatesis at the national level only. All sample householdswill be interviewed during survey period. The LFSquestionnaire will be administered first followed bythe TUS questionnaires. Demographic and socio-


economic characteristics, time-use diary portion,estimates of self-rated valuation of unpaid work andhousehold time saving/devices/equipment will beamong the major items that will be collected usingpersonal interviews; self-administeredquestionnaires (SAQ) or left-behind diary method;and diary days. For more information, contactAdministrator Carmelita N. Ericta, PhilippineNational Statistics Office, [email protected].

POLANDJanusz Wywial

The survey research of the hog population wasreorganized in Poland. The survey is evaluatedevery year in April, August and December by theCentral Statistical Office. Mainly, the number ofhogs is estimated in the population of Polish farms.The population consists of farms whose area is atleast 15 hectares. The number of such farms isequal to 833,000. The size of the stratified sampleis 15,000 farms. The sampling frame wasconstructed on the basis of the 2002 census. Thenumber of hogs (on the farms in the year 2002)was used as an auxiliary variable to optimize thesample sizes drawn from the strata. The populationof the farms was partitioned into 16 strataaccording to an administrative rule. Each stratumwas a suitable region of Poland, partitioned into twoparts. The first part consisted of large farms. Theremaining farms were in the second part. All farmsfrom the first part were selected in the sample. Asimple sample without replacement was drawnfrom the second part of the strata. Hence, thesample drawn from the strata consisted of two sub-samples. The first one is the purposive sub-sampleand the second one is the random sub-sample. Thefollowing allocation problem was evaluated. It wasassumed that the sum of the sizes of samplesselected from the strata was fixed. On the basis ofthe auxiliary variable the sizes of both sub-samplesdrawn from the appropriate strata were determinedin such a way that the precision of the estimation ofthe total value in each stratum was the same. Theex-post analysis of the precision estimation will beevaluated. For more details, please, get in touchwith Bronislaw Lednicki, [email protected],Central Statistical Office, 208 NiepodleglosciStreet, 00-925 Warsaw, Poland,http://www.stat.gov.pl.

SPAINMontserrat Herrador

From 2002 the Spanish Statistical Office (INE) hasstarted conducting a new annual Survey ofInformation and Communication Technology(ICT) which aims to collect information on theavailability, distribution and utilisation of ICT amongSpain households. This survey is adapted to therecommendations made by the EuropeanStatistical Office (EUROSTAT) which, in its turn,has followed the suggestions from the Organisationfor Economic Co-operation and Development(OECD).

In June 2002 this survey was carried out for thefirst time using a subsample of 20,000 householdscorresponding to two rotation groups of the LabourForce Survey (LFS) sample. These householdswere interviewed in the LFS during six consecutivequarters and, in the first and second quarters of2002, they were going out of the LFS sample. Fromthe second quarter of 2003 the ICT survey is goingto be implemented by using a new sampling designwhich is described below.

A rotation pattern will be used. Any householdentering the sample is expected to provideinformation for four consecutive years, after whichthe household will leave the sample. For theselection of the sample, a stratified three-stagesampling will be applied and for each region aseparate sample has been designed to enable thepublication of regional data. The first-stage unitsare the census sections (geographical areas); thesecond-stage units are the households. In the third-stage, a person aged 15 years and over is selectedfrom each household.

The sampling frame of geographical areas is builtfrom the list of census sections dated at the first ofJanuary of 2001. For the second-stage samplingunits, the Spanish Register of Population is used inorder to obtain a list of households for each censussection selected in the first-stage. Information onthe household will cover all resident members with10 years or more and be collected by telephone,whereas data on Internet use by householdmembers will be obtained by personal interview. Amatching process between the household frameand telephone directories is performed in order toinclude phone number in the household sample.


For those selected households without telephone,the data will be collected face-to-face.

The results of the survey will be made available bythe end of every year. For more information, pleasecontact Montserrat Herrador ([email protected]).

UNITED STATES OF AMERICARonald Fecso

The census of the United States� population andhousing was conducted as of April 1, 2000. Wellbefore then the United States Census Bureauinitiated the Census 2000 Testing,Experimentation, and Evaluation Program, themost comprehensive research program in itshistory. This program is informing the CensusBureau and our customers about the effectivenessof Census 2000 and providing valuable informationto plan the next decennial census in 2010, theAmerican Community Survey, the Master AddressFile, and other Census Bureau censuses andsurveys. The program consists of several tests andexperiments and over 90 evaluations on theimplementation of Census 2000. Relevant findingsfrom this body of research will be synthesized inbroader topical reports on key aspects of Census2000. To assure the reporting of accurate results,the Census Bureau also initiated the Census 2000Quality Assurance Process for all program reports.

Topic Reports summarize and integrate relevantfindings from tests, experiments, evaluations, andother research on related census subjects, andmake recommendations for the next decennialcensus in 2010. Reports will be prepared on thefollowing subjects: Address List Development,Automation of Census Processes, Content andData Quality, Coverage Improvement, CoverageMeasurement, Data Capture, Data Collection, DataProcessing, Ethnographic Studies, Partnership andMarketing, Privacy, Puerto Rico, Race andEthnicity, Response Rates and Behavior Analysis,and Special Places and Group Quarters.

Tests and experiments were conducted duringCensus 2000 because the decennial censusenvironment provided the best conditions to learnabout new methodologies. As examples, the effectsof the following were measured: a) questionnairecoverage questions and the �select more than one�race option; b) administrative records as a datacollection tool and as a means to assess imputationmethods; and c) alternative electronic modeoptions (with and without incentives), including theInternet and interactive voice recognition.

Evaluation reports assess the effectiveness ofCensus 2000 operations, systems, and processes.Evaluations on some of the newer programs orsystems to the decennial census include: a) thenational paid advertising campaign; b) OpticalCharacter Recognition of the Data Capture System;c) address list development operations; d) fieldfollowup operations to improve coverage; and e)the new race reporting�this evaluation includescomparability issues to survey and pre-2000census data.

The Census 2000 Quality Assurance Process wasdesigned to meet overall guidelines established bythe Census Bureau�s quality assurance policy forreport (evaluations and experiments) development,review, and writing. It involved building proceduresinto all stages of report writing, from projectdevelopment through issuance of the final report,and providing a framework so reports haveconsistent designs and formats. To help makequality integral to every stage of an evaluation,tools such as checklists and templates weredeveloped for project managers as they preparedstudy plans, specified programs and procedures,analyzed data, and prepared final reports for reviewby key operational customers and high-levelmanagers.

For information on the Census 2000 Testing,Experimentation, and Evaluation Program TopicReports, or to access reports currently available,visit: www.census.gov/pred/www.


Training Needs for Survey Statisticians inDeveloped and Developing Countries

Graham KaltonWestat, USA

Presented at an Invited Session, dedicated to the memory of Leslie Kish, Survey Research Methods Section,American Statistical Association, Joint Statistical Meeting, 2002

1. IntroductionThe session in which this paper is presented isentitled �Training of Survey Statisticians: A SessionDedicated to the Memory of Leslie Kish.� The topicof the session is highly appropriate given Kish�slifelong dedication to training survey statisticians,especially those from developing countries. It isalso appropriate that the session is cosponsored bythe International Association of Survey Statisticians(IASS) and by four ASA sections�Health PolicyStatistics, Survey Research Methods, GovernmentStatistics, and Social Statistics. Kish was a keymember of the committee that founded the IASS,and he served as President of both IASS and ASA.

Kish established a major program for trainingsurvey statisticians from developing countries atthe Survey Research Center of the University ofMichigan in 1961. That highly successful two-month summer Sampling Program for SurveyStatisticians (SPSS) continues to this day and hasnow trained more than 500 survey statisticians frommore than 100 countries. Kish�s choice of�Developing Samplers for Developing Countries� ashis topic for the fourth Morris H. Hansen MemorialLecture is an indicator of the importance that heattached to this activity (Kish, 1996). His papercontains many valuable observations on thetraining needs of survey statisticians in developingcountries.

This paper considers the training needs for surveystatisticians in both developed and developingcountries. The term �survey statistician� refers inthis paper to those engaged in the statistical designand analysis of surveys, primarily survey samplers,rather than to the broader interpretation of the termthat encompasses those engaged in any aspect ofsurvey methodology. For those in developedcountries, the focus here is on the educationneeded to prepare them for careers as surveystatisticians, mainly in terms of a master�s degree.For those from developing countries, the focus ison short-term programs that can provide appliedtraining in survey sampling, programs like the

SPSS and shorter programs. Throughout theemphasis is on training in applied sampling ratherthan in sampling theory.

Educational needs for survey statisticians indeveloped countries are discussed in Section 3,and the training needs for survey statisticians indeveloping countries are discussed in Section 4. Aspreparation for the discussion in those sections,Section 2 contains observations about generaltraining needs in survey research.

2. Training in Survey Research MethodsSurvey research is by its nature an interdisciplinaryactivity. It is generally conducted as a teamoperation, with the team members providingexpertise in the survey�s subject matter,questionnaire design, survey sampling, surveyoperations, data collection, computing, dataprocessing, and statistical analysis. In large-scalesurvey organizations, the survey team may becomposed of specialists in each of the separatecomponents of the survey procedures. In small-scale survey organizations, however, teammembers may need to cover more than onecomponent. For example, the team member whodesigns the questionnaire may also manage thedata collection, and the team member who selectsthe sample may also direct the data processing andsurvey analysis. Regardless of whether teammembers have responsibility for one or more thanone of the survey components, they need to have abasic understanding of all components. Designingand implementing an effective survey requires acareful blending of its various components. Thusthe members of the survey team need to be able towork together to formulate a survey design that canbe implemented successfully. They should fashiontheir individual components to be the best in thecontext of an overall design, rather than optimal insome narrower sense. To interact effectively withthe other members of the survey team, eachmember needs to appreciate the issues involved inthe overall design of the survey.


An alternative way of demonstrating this need forsurvey team members to understand the wholesurvey process is through the widely used conceptof total survey error. Kish (1965, Chapter 13)describes the diverse sources of error in surveyestimates, including sampling error and thenonsampling errors that arise throughnonobservation (noncoverage and nonresponse)and through observation (data collection andprocessing errors). Survey design often requirestrade-offs to be made between different sources oferror. For instance, a larger sample size will reducesampling error but perhaps at the cost of moreerror from nonresponse and data collection, or amore burdensome data collection protocol maylower measurement error at the cost of greaternonresponse error. All members of a survey teamneed to appreciate such interrelationships. Thus,an important part of training for all surveyresearchers should be to show them how theirspecialist area fits into the big picture.

There is a notable lack of educational and trainingprograms in survey research around the world. Fewuniversities provide, or are equipped to provide,these programs. They seldom have the full range offaculty needed, and most lack instructors with thewealth of experience required to teach the appliedaspects of survey research effectively. Inrecognition of the shortage of such programs in theUnited States, in 1992 the Federal statisticalagency heads, the head of the Office ofManagement and Budget Statistical Policy Office,and the chair of the Council of Economic Advisorsinitiated the establishment of a center for surveymethods that would provide graduate educationand research to serve the needs of the Federalagencies. As part of the development of thatinitiative, the Committee on National Statistics(CNSTAT) at the National Research Councilconvened a workshop on the proposed center. Thesummary of the workshop (Wolf, 1992) providesmany important ideas about the training needs forsurvey researchers from the perspectives both ofthose working in survey research in Federalagencies and of faculty involved in teaching surveyresearch. Here are a few of those ideas:

! There should be a focus on the interdisciplinarynature of survey research and training for it,along the lines discussed earlier.

! The center should engage in both degree andnondegree programs.

! Training could include core courses, specialtycourses, and electives.

! Apprenticeships or practicums are important inproviding real-world experience.

! A mechanism to respond to new training needswill be necessary.

! A wide variety of courses might be part of thecenter�s curricula.

The center was established at the University ofMaryland in 1992 after an open competitiveprocess. It involves a collaboration of the Universityof Maryland, the University of Michigan, and Westatand is known as the Joint Program in SurveyMethodology (JPSM). The ways that JPSM hasinstituted the ideas given earlier are on its web site:http://www.jpsm.umd.edu. The JPSM has set upPhD and MS degree programs, Citation andGraduate Certificate nondegree programs, andnoncredit short courses. The JPSM MS degreeprogram requires 46 credit hours. It has two areasof concentration, statistical science and socialscience, with all students taking certain corecourses.

It is instructive to make note of the collaborationinvolved in JPSM. The three-way collaborationbrings together the range of faculty and experienceneeded to run the program and, in addition, theprogram is supported by teaching contributionsfrom staff in the Federal agencies. Collaborations intraining in survey research between a university, asurvey research organization, and government alsooccur in the United Kingdom. The Department ofSocial Statistics at the University of Southamptonoffers Diploma/MSc programs in Social Statisticsand in Official Statistics, the latter offered incollaboration with the Office for National Statistics.The department also collaborates with a surveyresearch organization�the National Centre forSocial Research�and the University of Surrey inthe Centre for Applied Social Surveys (CASS),which offers short courses in survey methods. Seehttp://www.socstats.soton.ac.uk for details.

3. Education for Survey StatisticiansThis section focuses on graduate-level educationfor survey statisticians, mainly survey samplersand, in particular, on the issues involved in settinga curriculum for a master�s level program. Like Kish(1996), the focus here is on survey sampling andsample design, not the foundations of samplingtheory.

In general, also like Kish, I favor a flexible programto accommodate the needs of a variety of students.Some may want to concentrate specifically onsurvey sampling, others may want to take a moregeneral program in survey methodology with aspecial emphasis on sampling, while yet othersmay want to combine survey sampling with aparticular subject-matter area (e.g., agriculture,


education, health, economics). Some may plan tomove into managerial positions as their careersprogress, whereas others may want an educationthat prepares them for a lifetime career in surveysampling. The MS education for this latter groupneeds to provide a firm foundation in the subject onwhich they can build as new developments occur.

My criteria for a master�s program for surveystatistics are the following:

! Strong foundation in statistical theory andmethods,

! Specialized training in the theory and practiceof survey sampling,

! Broad knowledge of the whole survey process,! Development of good computing skills, and! Training in oral and written communication.Statistical theory and methods are listed firstbecause they provide the basic foundation forsurvey sampling. Students need a strongfoundation in this area to be able to read thecurrent sampling literature. They will no doubt haveeven more need for this foundation to upgrade theirskills as advances in survey sampling occur in thefuture. As an illustration, consider the recentdevelopments in small area estimation and invariance estimation with data sets containingimputed values. Both of these areas illustrate hownew methods in survey sampling are using morehigh-powered statistical methods. A strongfoundation in statistical theory and methods todayrequires at least four courses (two in probabilityand statistical theory and two in statisticalmethods), and that already seems inadequate toprovide the foundation needed to cover the newertechniques in survey sampling. To prepare for thefuture, more courses are likely to be needed.

Survey sampling is now a highly developed fieldwith a rich literature and a wide range of methods.A practicing survey statistician needs at least threebasic courses in this area (e.g., applied samplingmethods, basic theory, and variance estimation),and then there are many special topics that shouldideally be covered in some fashion. The following isa partial list of important topics that cannot beadequately covered, if mentioned at all, in the basiccourses:

! Nonresponse and calibration weightingadjustments

! Edit and imputation methods! Design and estimation for surveys over time! Measurement error models

! Sampling rare populations! Small area estimation! Sampling for establishment surveys! Design and analysis of observational studies! Disclosure avoidance! Treatment of outliers! Sample coordination over timeA selection of these topics can appear in a fourthsampling course, but not all of them can becovered, and those that are covered cannot betreated in depth. Several of these topics warrantone (or even more) three-credit courses for a fulltreatment, and others warrant one or two credits.Additional courses could be considered, but theyare likely to run up against the constraints on theacceptable maximum number of credit hours for anMS program. With only one course on these topics,a difficult choice has to be made between whetherto attempt to make the students aware of all theareas of survey sampling methods that they mayneed in their work, or whether to let them chose asmall number that they will study in somewhatmore depth.

The justification for the need for all surveyresearchers to acquire a broad knowledge of thewhole survey process has been made in theprevious section. That justification clearly applies tosurvey statisticians, who must design samples thatmeld with the other survey design features in aneffective manner. A broad knowledge of the surveyprocess requires at least two or three 3-creditcourses (e.g., data collection methods, sources oferror in surveys), and one or two more courses if asurvey practicums is included. Carefully chosenapprenticeships that give exposure to the wholesurvey process may also serve the importantrequirement of giving students an appreciation ofthe overall design considerations that apply insurvey practice.

The need for computing skills, particularly the useof statistical software packages, is clear.Computing features in most modern courses instatistical methods, and it can also be included inrelevant sampling courses. That should in generalsuffice, provided that attention is given tocomputing in course development.

A widely acknowledged weakness of muchstatistical training is a lack of attention to oral andwritten communication. Strong communicationskills are essential for applied statisticians ingeneral, and this is certainly so for surveystatisticians. In a paper on preparing statisticiansfor careers in the Federal Government, an ASA


Committee cited the need for communication skills:�The need for writing skills cannot beoveremphasized. As statisticians move intomanagerial positions, communication�both writtenand oral�becomes a vital ingredient of asuccessful performance.� (Eldridge et al., 1982, p.76). A paper entitled �Preparing Statisticians for aCareer at Statistics Canada,� by Denis et al. (2002)contains recommendations for universities inpreparing students to work in a statistical agencythat are in line with the criteria proposed earlier,including again a call for writing skills:

! �Provide students with a solid grounding instatistics�;

! �Give students the opportunity to work on �real�problems�;

! �Generate interest in learning to use statisticalsoftware�; and

! �Put more emphasis on writing skills�.Training in oral and written communication can bestbe achieved by requiring presentations and reportsin many courses. Assignments of 15-minutepresentations supported by well-prepared slides inseveral classes can help students develop theiroral communication skills and prepare them formaking presentations at professional meetings.Training in report writing can be obtained throughcourse assignments, but to be effective, instructorsor writing experts must spend adequate time on thewriting rather than on the technical aspects of thereports. In some cases, a separate course on theprinciples of report writing may be advisable.

The challenge in developing an MS program forsurvey statisticians is to accommodate all theseobjectives within a reasonable course load. Giventhe unavoidable constraints on the length of an MSprogram, some priorities need to be established.The initial set of priorities noted earlier is a solidgrounding in statistical theory and methods (at leastfour to six courses), an extensive coverage ofsampling methods (at least four courses), and abasic overview of the whole survey process (atleast two courses). After that, I favor flexibility toallow choices of additional topics in any of theseareas or in other relevant areas, such as asubstantive area of application. As a rule, studentswishing to embark on careers as specialistsamplers and those wishing to pursue PhDdegrees in survey sampling would be advised toelect more courses in statistical theory andmethods and in topics in survey sampling.

However an MS course is structured, it cannot beexpected to provide a detailed coverage of all thesurvey sampling topics that a survey statistician will

encounter at work. Also, as the field of surveysampling continues to advance, new methods willemerge. As a consequence, I see a strong need forcontinuing education programs for surveystatisticians to develop their range of skills.Advanced short courses are helpful in exposingparticipants to specialized techniques, but they lackthe means to ensure that the participants undertakethe supplementary study needed to fully master thematerial. I believe that longer term formal courses,with assignments and examinations, are required,together with some form of certification forsuccessful performance. This kind of continuingeducation, which is required in some otherdisciplines, is now needed for survey sampling.Hopefully employers will recognize the benefits ofsuch continuing education, encourage staff toparticipate, and reward those who complete thecourses successfully. It is of interest to note thatStatistics Canada invests a good deal inprofessional training. It has established its ownTraining Institute to offer courses in-house, and itexpects supervisors and employees to discuss atraining plan for the coming year during the annualperformance meeting (Denis et al., 2002).

4. Training Courses for Survey Statisticiansfrom Developing Countries

This section addresses the training needs forsurvey statisticians from developing countries. TheMS programs described earlier can serve theeducational needs for survey statisticians fromdeveloping countries well. However, it is expensiveand time-consuming to send persons fromdeveloping countries to MS programs abroad.Alternative shorter training programs are, therefore,much needed.

One model for such a training program is the SPSSthat Kish developed over 40 years ago, asmentioned earlier. Another long-term program fortraining staff from national statistical offices is thatrun by the International Program Center at the U.S.Census Bureau. The center now runs workshopsfrom two to six weeks in length on a variety oftopics, including survey sampling(http://www.census.gov/ipc).

The IASS has successfully run short courses insurvey sampling and other aspects of surveyresearch in association with the biennial sessionsof the International Statistical Institute for manyyears. Kish initiated short half-day meetings in1973, and these were expanded into longerworkshops in 1987. At the ISI session in Seoul in2001, the IASS offered the following courses:Workshop on Survey Sampling, VarianceEstimation in Complex Surveys, Introduction to


Small Area Estimation, Nonsampling ErrorResearch, and Edit and Imputation of Survey Data.The workshop on survey sampling is specificallydesigned for survey statisticians from developingcountries. Together with the course on varianceestimation, it provides an overall introduction to thesubject.

Courses like those offered by the IASS serve avaluable role both by providing training in surveysampling and also by bringing together surveystatisticians from different countries to share theirexperiences. Such courses should emphasizesurvey sampling practice and introduce basictheory only to the extent necessary. They shouldcover basic sampling techniques (e.g., systematicsampling, stratification, multistage sampling, PPSselection) and also nonresponse, noncoverage,weighting, and the use of software for varianceestimation. They should also include illustrationsfrom developing countries. Simple exercises can behelpful in reinforcing key points being made. Wherepossible, hands-on use of computer softwareshould be included.

The key problem with providing training courses forsurvey statisticians from developing countries,whether longer courses like the SPSS or shortercourses like those of the IASS, is that of funding.The SPSS was started with funds from the FordFoundation and now is supported in part by theLeslie Kish International Fellows Fund. Manystatisticians from developing countries have beenable to attend the IASS courses with fellowshipsupport from the UN Statistical Division. In bothcases these sources of funding support haveprovided the core basis for maintaining thecourses, with other support coming from variousother sources.

When training persons from different countries,language becomes an important issue. The abovecourses are generally given in English and requireparticipants to understand English. However, atleast when there is only one other major languagein addition to the language of instruction, atranslation of presentation slides�or at least keyterms�into the other language can be valuable.The use of two projectors for the two languagescan then help those who are not so fluent in thelanguage of instruction.

A type of training program that seems particularlydesirable is one that is given in a developingcountry for statisticians in that country andneighboring countries. A one- or two-week regional

program could be a very effective use of limitedresources. In his invited discussion of this paper,Hermann Habermann suggested an enhancementto this scheme that would extend the program to aseries of training activities over a period of severalyears. I thoroughly endorse this extension, which Isee as a way to develop the internal trainingresources in the region, with local instructors takingover in later years.

In earlier years the World Fertility Survey directedby Sir Maurice Kendall for the InternationalStatistical Institute and the UN Statistical Division�sNational Household Survey Capability Programmade major contributions to establishing surveyinfrastructures in many developing countries. Sincethat time, a sizable number of international surveyprograms have begun, but unfortunately none ofthem has a training mission in survey statistics orsurvey methodology. Much would be gained if theycould be persuaded to devote small proportions oftheir budgets to training and capacity building.

The programs listed in this section play a veryvaluable role in training survey statisticians indeveloping countries. Yet much more needs to bedone. Might ASA and IASS collaborate in fosteringsuch training programs? Both associations areconcerned with improving the use of statistics, andefforts in this area would serve that objective well.

ReferencesDenis, J., Dolson, D., Dufour, J., and Whitridge, P.(2002). Preparing statisticians for a career atStatistics Canada, (Working Paper No. HSMD-2001-006E/F). Ottawa: Statistics Canada.

Eldridge, M.D., Wallman, K.K., Wulfsberg, R.M.,Bailar, B.A., Bishop, Y.M., Kibler, W.E., Orleans,B.S., Rice, D.P., Schaible, W., Selig, S.M., andSirken, M.G. (1982). Preparing statisticians forcareers in the Federal Government: Report of theASA Section on Statistical Education Committee onTraining of Statisticians for Government. AmericanStatistician, 36, 2, 69-89.

Kish, L. (1965). Survey sampling. New York: Wiley.

Kish, L. (1996). Developing samplers fordeveloping countries. International StatisticalReview, 64, 143-162.

Wolf, F. E. (Ed.). (1992). Center for SurveyMethods: Summary of a workshop. WashingtonDC: National Academy Press.


Special Articles: Censuses Conducted Around the World

The Possible Impact of QuestionChanges on Data and Its Usage:

A Case Study of Two South AfricanCensuses (1996 and 2001)

Maurius CronjéStatistics South Africa

Disclaimer: The views expressed in this paperrepresent those of the author and do notnecessarily reflect the views of Statistics SouthAfrica.

The supply of accurate data is very important in afast moving society. Policy makers, planners,business people, and many others make use ofdata provided to them, on a daily basis. Theapplication of this data directly impacts our dailylives. It is for this reason that all people who makeuse of any sort of data should have a clearunderstanding of the data and its limitations. Thisprocess starts right at the beginning, when theinstrument for data gathering is developed.

Bradburn (2000) states that �since the beginning ofscientific surveys, researchers have been sensitiveto the fact that changes in the wording of questionscould produce dramatic changes in the distributionof responses.� This article sets out to explore brieflythe impact of differences in questionnaire wording,order, and topics on the final data that will beproduced. It also explores other factors that mightplay a role in eliciting differential responses to aquestionnaire. These issues are discussed in thecontext of the recent South African Census (2001),as compared to the South African Census of 1996.

BackgroundData that are gathered over time will be used todraw a comparison to determine change, establishtrends, and predict the future. For this reason, thecompatibility of different data sets is very important.Researchers need to be made aware of thedifferences between two data sets and the possibleeffect of these differences on the quality of thedata. Differences in methodologies used to gatherthe data might have an impact on the measurementof a certain variable or the application of the data tosmall areas. Differences in the measurement of thevariables might also be due to differences ininstrument design. It is therefore essential that astudy be conducted where these differences arehighlighted and the implication shown. This type of

research would lay the foundation for the work ofother researchers utilizing the data. If majorproblems exist with the data, prospective usersshould be made aware so that they can decide forthemselves whether they would like to use it or not(Jaffe, 1951).

Statistics South Africa (Stats SA) is the officialagency responsible for conducting the NationalPopulation Censuses in South Africa. Stats SA isrequired, under the terms of the Statistics Act, No.6 of 1999, to conduct a census every five years. InOctober 2001, Stats SA conducted the secondpopulation census since the first democraticelections of 1994. The first census was conductedin October 1996. Before any person can start touse the two data sets, a careful analysis should bemade to establish differences between them. Thereare many factors that might affect this comparison.This paper will, however, focus on issues relating tothe questionnaires used.

Questionnaire Wording and OrderMcLaughlin (1999) argues at length about how itemcontext can affect answers. She mentions that thecontext of an item, including format, wording,response option, and order of questions, can havean unintended (or intended) effect on the answers.

One of the most important factors in solicitingproper responses from respondents is the wordingof the questions. For instance, the questions �Howold are you?� and �In what year were you born?�have only one word in common but are understoodas calling for the same information. The questions�When were you born?� and �Where were youborn?� share all the same words except for one.They will, however, result in two totally differentresponses. The wording of the questions is whatthe respondent must process in order tounderstand what information is sought. In general,wording should use simple language, be clear, andavoid phrasing in the negative.

One needs to establish that the respondents willunderstand the wording that is used. Questionsusing slang, culture-specific, or technical wordsmight not measure what the researcher intended atthe beginning of the study. The wording might leadthe respondent to answer in a certain manner. Theuse of �and� in a question must be avoided so thata double question is not asked. The use of theword �not� should also be avoided, especially if therespondent is to answer �yes� or �no.�


Survey processing is often easier with close-endedquestions, where respondents are given optionsfrom which they can choose. Such questions canbe answered more quickly, but, on the other hand,respondents may rush through the questions tooquickly and make mistakes. The answers providedmight also not be what the respondents preferred.For close-ended questions, the choices mustexhaust the entire range of possible answers andmust be mutually exclusive. For open-endedquestions, no options are provided for therespondents. It is left to them to think and fill in theirown responses in their own words. This might leadto more meaningful information and to dissimilarbut intelligent opinions. The processing of thesequestions is often more difficult, with a lot ofproblems experienced in the coding of the answers.

In addition to questionnaire wording, the positioningof the actual questions can affect responses.People are often influenced by visual and otherstimuli, and it is therefore important to place thequestions in an order that will result in a positiveresponse. It has also been established that peopleremember the first and the last items on a list muchbetter than the middle items. This can lead to afalse impression when it comes to issues such asagree/disagree or like/dislike. Bradburn (2000)discusses the Tourangeau and Rasinski model,which conceptualises the question/answer processas one of stages that require different processingand may be susceptible to different types of contexteffects. The focus in this model is on interpretation,retrieval, judgement, and response. An interestingexample that is mentioned is the question �Howmany children do you have?� when asked ofteachers. In the first case, it was preceded byquestions related to their families, and they theninterpreted this question as relating to their families.In the second case, the question was preceded byquestions relating to the school, and theyinterpreted it as referring to the number of childrenthey had in their class. The question got aresponse of 0-4 in the first instance and 25-40 inthe second instance, although the question wasexactly the same.

Two types of order effect can be identified, namelyassimilation and contrast effects. The first happenswhen respondents adjust their responses to laterquestions so that their responses to a set ofquestions are consistent. Contrast effect happenswhen respondents discount or subtract informationfrom a previous question when answering asubsequent question (Gendall, Carmichael, &Hoek, 1997).

Design and layout were traditionally determined bywhether a questionnaire would be self-administeredor interviewer administered. Although these

aspects still guide the design, the design andquestions are now also influenced by thetechnology used for the processing of the results,specifically whether one might use scanning andoptical character recognition, optical mark reading,and image recognition software. These moderntechniques sometimes guide the development ofthe questionnaire in terms of layout and order.Careful consideration should be given to thewording and the order in conjunction with the use ofmodern technology to ensure that the questionsmeasure what the researchers set out to measureat the initiation of the research.

Topics CoveredSometimes the topics covered can have an effecton responses. In censuses, for instance, questionson income are usually associated with a lot ofcontroversy. The data related to these questionsare also usually suspect. The questions are placednear the end of the questionnaire so that they donot influence the frame of mind of the respondent.Debates take place on a regular basis as towhether questions of this nature should beincluded. Another issue usually hotly debated isquestions about ethnicity or population group. Hurst(2000) has described at length the whole debatethat Statistics Canada went through on thequestion of ethnicity. Similar problems wereexperienced in the United States, which promptedthat country to change the question after the 1990census; this resulted in a situation wherestraightforward comparison between censuses isnot possible.

Topics are covered by making use of several typesof questions. These are filtering questions (alsoknown as skip questions), where respondents aredirected towards relevant questions; binaryquestions, in which respondents choose one of twoanswers; and categorical questions, in which a listof options is given to respondents. Ordinalquestions consist of scaled responses such as the5-point Likert scale, while linked questions have thesame style of question and use a matrix to recordresponses.

Other Factors Influencing ResponsesBradburn (2000), in collaboration with Sudman,designed a model for understanding the sources ofresponse effects. The model is based on the viewthat the survey interview is a special kind ofconversation between individuals who speak thesame language. It is seen as a micro-social systemconsisting of three roles held together by acommon task reflected in the questionnaire. Oneperson (interviewer) has the role of asking thequestions, a second person (respondent) has therole of answering the questions, and a third person(researcher) defines the tasks. All three of these


role players have an effect on the responsesreceived.

The process of answering a questionnaire itemevokes a number of cognitive processes(comprehension, information retrieval, judgmentand evaluation of retrieved information, and theselection of an appropriate response), each ofwhich can lead to variation and erroneousresponses by the respondent. The manner in whichthese various processes operate will depend on thenature of the response task, the mode ofadministration (self or interviewer), andcharacteristics of the respondent.

Variation and errors in response to questionnaireitems can be due to a number of factors, includingthe respondent�s confusion, ignorance,carelessness, or need for social desirability;ambiguous wording of the items and responsecategories; the order or context in whichquestionnaire items are presented; the choice andpresentation order of response categories; and themethod of administration (Meadows, Abrams, &Greene, 1998).

The social context in which the interview takesplace also plays a vital role in interview outcomeand the accuracy of the data. Social context isrelated on one level to the physical environment inwhich the interview is conducted. For instance,respondents might feel more comfortable if aquestionnaire is completed in their home asopposed to completing it in a work environment.

Another factor influencing responses might be thecharacteristics of the respondent and theinterviewer. Aspects such as race, gender, and agemay influence the willingness to express attitudes.The way in which the interview is conducted mightalso have an influence. For instance, the interviewmight be conducted by telephone, might be self-administered, or might be performed by usingvarious forms of computer assistance (Bradburn,2000).

The South African Perspective QuestionnairesIt is not unusual for the number and types ofquestions included in the census questionnaire tochange from census to census in response to newpolicy initiatives, a review of redundant questions,other user needs, society in general, and publicperception. The look and feel of the censusquestionnaire has also changed in response todesign techniques.

For both South African censuses, the questionnairewas developed with active participation by usersand the census advisory committee. Items to beconsidered were suggested by various

stakeholders. International consultants alsoprovided expert assistance in the design and layoutof the questionnaires. In both cases, behind-the-glass and field testing were done to ensure that auser-friendly instrument was developed. Testresults facilitated the development of appropriatewording for questions and provided qualitative andquantitative feedback on the reliability of responsesand the acceptance of questions.

Other factors that played a role in the developmentof the questions were decisions on processingmethodologies. For instance, technologicaladvances in scanning and image recognition, aswell as automatic and computer-assisted coding,guided the actual layout and design of thequestionnaire.

For both censuses, the questionnaire wastranslated into all 11 official languages. Theapplication and availability of the translationsdiffered between the two censuses. For Census 96,all 11 languages were available in the field. ForCensus 2001, every enumerator carried atranslation booklet that could be used by therespondent to read the questions in his or herpreferred language. The answers would then betranscribed onto the English version of thequestionnaire. However, for Census 2001 thetranslated questionnaire was also available onrequest. For Census 96, every language had adifferent colour. For Census 2001, all the differenttranslations came out in the same colour.

In Census 96, the following questionnaires wereused:

Questionnaire 1 (Household questionnaire): Thisquestionnaire was used in private households andin hostels that provided family accommodation. Itcontained 50 questions for each person and 15 foreach household. A total of nine people could beincluded on the questionnaire.

Questionnaire 2 (Summary book for hostels): Thisquestionnaire was used to list allpersons/households in the hostel and included 9questions about the hostel.

Questionnaire 3 (Personal questionnaire): Thisquestionnaire was used for individuals withinhostels and contained 49 questions (the same ason the household questionnaire, with the exceptionof relationship).

Questionnaire 4 (Enumerator�s book for specialenumeration): This questionnaire was used toobtain very basic information for individuals withininstitutions such as hotels, prisons, and hospitals,as well as for homeless persons. Only 6 questions


were asked of these people. The questionnairealso included 9 questions about the institution.

For Census 2001, three different questionnaireswere used. The first questionnaire (questionnaireA) was used for all households, including single-person households, and in workers� hostels,university student residences, and residentialhotels. This questionnaire was designed to collectinformation about each person in a household (41questions) and information about the household asa whole (1 question), including services (13questions).

The other two questionnaires were used forinstitutions, tourist hotels, and the homeless.Institutions included hospitals, childcare institutions,boarding schools, convents, defence forcebarracks, prisons, community and church halls, andrefugee camps. One questionnaire (questionnaireB) was used to capture personal information aboutindividuals (40 questions), whilst the other(questionnaire C) contained questions that appliedto the institution as a whole (8 questions). Again,the personal questionnaire omitted the questionson relationship and the household.

One of the major changes from Census 96 toCensus 2001 was the introduction of barcodes onthe Census 2001 questionnaire. Everyquestionnaire for Census 2001 was uniquelyidentified by a barcode printed on every page of thequestionnaire.

Topics CoveredBoth censuses covered basic demographic, social,and economic topics. Some questions applied toeach individual in the household; for example,respondents were asked to indicate the age,gender, population group, home language,education level (people over age 5 years),occupation (age 5 years and older), and income ofeach person present in the household on censusnight. Other questions dealt with the circumstancesof the household as a whole, for example, whetheror not the household had access to electricity andpiped water.

For Census 2001, there was a coordinated effort toharmonize census questions across the SouthernAfrican Development Community member states. Anumber of meetings and workshops were held todiscuss a set of core questions. The main aim wasto have compatible data across the region. StatsSA included these core questions in thequestionnaire and supplemented them with somespecific questions pertaining to South African dataneeds.

The 2001 questionnaire included a total of 14questions that were not asked in the 1996 census.There were also 9 questions that were asked in1996 but not repeated in 2001. This means thatthere are 23 questions that cannot be used forcomparing the data between the two censuses.Issues to be debated are whether the inclusion orexclusion of certain questions adds value or takesvalue away from the census data set. Some of thereasons for the exclusions might also beinvestigated. Sometimes questions are redundant,and sometimes new research shows that a changeof focus might be needed. The usage of the datamight also result in a question being dropped orincluded in the next census.

The Way ForwardMuch research must be conducted to determine theimpact on the data of the differences describedabove. Some of these issues might be resolvedwith proper data editing. This can be done as soonas the data sets become available. In addition,studies are needed on the effect of other factors,such as cultural differences, responses, and,ultimately, on data.

All of the issues discussed above lead to adirection in which the selection of the topics and thedesign of the questionnaire will become a full blowninterdisciplinary exercise, with inputs from fields asfar apart as psychology, linguistics, informationtechnology, graphic design, and statistics.

It is especially important that societies with a largemixture of cultures give special attention to aspectsof questionnaire design and question wording. Thisis to ensure that the data from different culturalgroups are consistent and comparable.

ReferencesBradburn, N.M. (2000). Questionnaire design: fromart to science. Paper delivered at the FifthInternational Conference on Social ScienceMethodology, Cologne, Germany.

Gendall, P., Carmichael, V., and Hoek, J. (1997). Atest of the conversational logic analysis model ofquestion order effects. Marketing Bulleting, 8, pp.41-52.

Hurst,L. (22, January 2000). The politics of people-counting. Question on ethnicity haunts census-takers. Toronto Star.

Jaffe, A.J. (1951). Handbook of statistical methodsfor demographers. Washington, DC;U.S.Government Printing Office.


McLaughlin, M.E. (1999). Controlling methodeffects in self-report instruments. ResearchMethods Forum, Vol. 4.

Meadows, K. A., Abrams, C., and Greene, T.(1998). The application of cognitive aspects ofsurvey methodology (CASM) in the development ofHRQOL measures. Quality of Life Newsletter,[Online]. Available: http://www.mapi-research-inst.com/index02.htm

To Contact the Author:

Maurius CronjéStatistics South [email protected]

http://www.mapi-research-inst.com/index02.htm

http://www.mapi-research-inst.com/index02.htm


Discussion Corner

Editor’s Note: Two discussants have provided their points of view on the topic of substitution. We encourageother members to get involved in this discussion. Send a brief description of your approach in dealing withsubstitution. We would like to know how different organizations in countries around the world deal with thechallenges that nonresponse imposes on surveys.

To Substitute or Not to Substitute—That is the QuestionDavid W. Chapman

Nonresponse in surveys is generally classified asone of two types: unit or item nonresponse. Withunit nonresponse, no usable survey information isobtained from an eligible sample unit. With itemnonresponse, some of the survey items are notobtained from the respondent (which could includeedit failures), but enough items are obtained toinclude the sample participant as a unit respondent.

To account for unit nonresponse, there are threemethods that I am aware of that are used by surveypractitioners:

! Adjusting the weights of the respondents;! Imputation of entire records; and! Substitution (or field substitution).Based on my experience, weight adjustment is byfar the method used most often to account for unitnonresponse, at least for U.S. surveys. With thismethod, the respondents in a weighting class(which may be a stratum or cluster) are �weightedup� to the full sample of eligible units in theweighting class.

I am not aware of many instances where imputationis used, though I believe that this is done for theshort form component of the U.S. DecennialCensus (essentially �hot decking� entire records).Although this method may be acceptable foraddressing nonresponse bias, it would increase thevariance of survey estimates because, in effect, theweights of the �donor� respondents are at leastdoubled.

By substitution, or field substitution, I am referringto the method of substituting for each samplenonrespondent a population unit that was notoriginally selected for the sample, and treating thesubstitute as though it had been selected for theoriginal sample. The goal is to select substitutesthat have survey characteristics that are similar tothose of the nonrespondents. My experience hasbeen that most statisticians view substitution as an

inferior (or even unacceptable) method ofaccounting for nonresponse.

The purpose of this article is to describe the basictypes of substitution, and to discuss theadvantages and disadvantages of the use ofsubstitution. Much of the material here has beenabstracted from an earlier document I wrote as achapter in the book, Incomplete Data in SampleSurveys (Chapman, 1983). I hope that this articlegenerates some interest and discussion ofsubstitution as a method for accounting for unitnonresponse.

I believe that, because it has been misused in thepast, substitution is an underrated method. I am notsuggesting that it is the best method to use for allsurveys, but one that may be appropriate in manyapplications. My impression is that substitution isnot used often to account for unit nonresponse forU.S. surveys. However, it may be used morefrequently in other countries. In his excellent articleon substitution, Vehovar (1999, p. 377) reports that��although it is rarely recommended intextbooks�,� ��we can observe a relativelywidespread use of this procedure in manyprobability sample surveys.�

There are two basic types of substitution methodsthat are used in applications:

! Selection of a specially designated substitute;and

! Selection of a random substituteThe first method involves a purposive selection of asubstitute, one that that matches closely thesampling frame characteristics of thenonrespondent. For example, for an establishmentsurvey, the substitute could be defined as theestablishment (of those not selected for thesample) that comes closest to matching thenonresponding establishment on framecharacteristics such as type of business,geographic location, degree of urbanization, publicvs. private, for profit vs. nonprofit, number ofemployees, and revenues or profits.


The second method uses a random selection toobtain a substitute from among thoseestablishments not selected for the sample.Typically, the substitute is selected randomly froma subpopulation (e.g., the same stratum or clusteras the nonrespondent) to improve the chances ofselecting a substitute that has surveycharacteristics that are similar to those of thenonrespondent. For the case of randomly selectedsubstitutes, Vehovar (1999) derives expressions forthe impact on bias and variance of the use ofsubstitution.

Of these two substitution methods, randomselection of substitutes is the method probablyused most often in practice, at least for probabilitysampling (which is the context of this article). Thisis because random selection would appear to bemore scientific and avoids additional biases thatmight occur with purposive selection. However, interms of obtaining replacement data for anonrespondent, a specially selected substitute thatmatches the nonrespondent on all the relevantframe characteristics may be superior to arandomly selected replacement from asubpopulation defined by some of the framecharacteristics.

Developing statistical expressions that characterizethe impact on bias and variance of substitution canonly be done (without creative modeling) for themethod of selecting substitutes randomly. Forrandom substitution, Vehovar (1999, p. 238) pointsout that any additional bias associated withsubstitution would arise from differences in theprocedures used to contact and recruitrespondents. For example, there may be more callattempts to reach initial sample units than forsubstitute units. He concludes that, although thereare likely to be such difference in the datacollection procedures, the bias is likely to be"relatively small."

Regarding variance comparisons, Vehovar (1999,Section 2.2) develops an expression for theincrease in variance for weight adjustmentcompared to substitution for the case of two-stagecluster sampling. He concludes that "�we canbenefit from substitution only in very specificcircumstances which are not encountered veryoften." His conclusions are similar for multi-stagecluster designs and for stratified sampling.

In terms of empirical studies of the impact ofsubstitution, I summarized the results of fourstudies that I identified in my earlier research(Chapman, 1983, Section 3). Also, while I was atthe U.S. Census Bureau, I participated in anempirical study of substitution for an RDD (randomdigit dialing) survey (Chapman and Roman, 1985).

In addition, Vehojar (1999, Section 2.3) describedan empirical investigation that he had conducted.For these empirical studies, there were no decisiveresults in terms of the value of the use ofsubstitution as a method of accounting for unitnonresponse.

So, the decision as to whether to use substitutionmay rely mostly on judgment and practicalconsiderations relating to the specific survey. Forexample, in a survey of elementary and secondaryschools I worked on while at Westat in 1971-72, animportant goal of the study was to developequivalence relationships between several nationalachievement tests. To do this, the analysts neededa precise number of schools participating in thestudy, with exactly two schools from each stratum.Substitution was the only approach to achieve thetype of sample needed for the analysis. In myearlier research on substitution (Chapman, 1983, p.47), I described the substitution procedure used forthat study, which was a combination of using aspecially designated substitutes (first priority) andrandomly selected substitutes.

There are two major advantages to the use ofsubstitution that I want to mention. First, to theextent that substitutes can be made for all of thenonrespondents, the original sample design andsample size will be preserved. Thus, if the designwas one with uniform selection probabilities (andtherefore "self-weighting"), the final sample will stillhave those properties. This will simplify theanalysis of the survey results. In addition, thevariances of survey estimates are likely to be aboutthe same as would have been realized had theresponse rate been 100 percent.

The second advantage is the potential for biasreduction in a case where there are a lot of framecharacteristics available, as there are for manyestablishment surveys. In this type of situation, ifsubstitutes can be identified that match (or nearlymatch) those of the survey nonrespondent onseveral frame characteristics, the nonresponse biascould be reduced, compared to that associated withweight adjustments, where the missing surveyvalues are effectively imputed by a blend of thesurvey values of respondents in the sameweighting class. Of course, the frame variablesmust be highly correlated to the survey variablesfor such benefits to be realized.

There are several potential disadvantages to theuse of substitution that need to be considered.First, there is the potential to introduce bias insurvey estimates from the way substitutes areselected, especially if interviewers are allowed tochoose substitutes. To avoid such biases, well-defined criteria for identifying substitutes must be


developed. Ideally, substitutes would be selectedas an office procedure, and the interviewer wouldnot know whether he/she was attempting tointerview an initial sample unit or a substitute unit.This type of control should be easy to establish withrandom digit dialing telephone surveys, using CATI(computer-assisted telephone interviewing).

Second, there may be a tendency for less effort tobe made to contact and recruit initial sample units,when substitutes are available. To some extent,substitutes may be viewed by interviewers, andperhaps by some survey practitioners, to be nearlyas good as the original sample units. This willreduce the response rate and increase the potentialfor nonresponse bias.

A related difficulty with the use of substitution is thelimitation associated with it because of thesequential aspect of the implementation process(i.e., a substitute would not be called until attemptsto recruit the initial sample unit failed). For a fixeddata collection period, the time allowed to contactand recruit initial sample units would have to bereduced to allow adequate time to recruitsubstitutes. More likely, a longer data collectionperiod would be used to allow sufficient time to tryto recruit original sample units and substitutes. Thiswould probably increase data collection costs (ascompared to a weight adjustment approach), evenfor comparable numbers of completed interviews.

An approach to partially resolve the problem of thesequential nature of the use of substitutes is tobegin attempts to contact a substitute before allplanned attempts have been made to recruit theoriginal sample unit. Then, simultaneous attemptswould be made to contact both the original andsubstitute units. Of course, a weakness with thisapproach is that you may waste interviewer timecontacting the substitute unit, and may even end upwith interviews with both the original and substitutesample units, in which case the data from thesubstitute unit should not be used.

Another disadvantage of substitution is the problemof identifying the substitute respondents sufficientlyin the final database so that the true response ratecan be calculated. There may be a tendency tocompute the response rate as though thesubstitutes were initial sample units.

A final weakness with the use of substitution that Iwant to mention is that a substitute may not berecruited for every nonrespondent. This willcertainly be true if only one substitute is selectedfor each nonrespondent. (For this reason, Irecommend that multiple substitutes be selected for

each nonrespondent, preferably as part of thesample selection process.) In those cases wheresubstitutes are not obtained for everynonrespondent, some other method of accountingfor the residual nonresponse (like weightadjustment) would still have to be applied.

In conclusion, the fundamental question is whetherit is better to use a substitute unit for a surveynonrespondent, rather than imputingnonrespondent data from a blend of the datareported by respondents in the same weightingclass. I think that in many applications it would bebetter to use a substitute unit, as long as thedisadvantages of implementing a substitutionprocedure discussed above can be sufficientlyaddressed.

I would appreciate any feedback from readers thathave used substitution in surveys, including thespecific procedure(s) used, the success rate interms of obtaining cooperating substitutes for unitnonrespondents, and any observations regardinghow well the procedure worked, or didn't work. Ofcourse, any views on whether substitution shouldbe used at all to account for unit nonresponse insurveys are also welcome.

ReferencesChapman, D. (1983). The impact of substitution onsurvey estimates. In W. Madow, I. Olkin, and D.Rubin (Eds.). Incomplete data in sample surveys,vol. II, Theories and bibliographies, (pp. 45-61),Academic Press: New York.

Chapman, D. and Roman, A. (1985). Aninvestigation of substitution for an RDD survey.Proceedings of the Survey Research MethodsSection, ASA, 269-274.

Vehovar V. (1999). Field substitution and unitnonresponse. Journal of Official Statistics, 15, (2),335-350.

To Contact the Author:David W. Chapman, Chief StatisticianFederal Deposit Insurance CorporationRoom 2022, 550 17th St. N.W.Washington, DC 20429USA(202) [email protected]


Field Substitutions Redefined

Vasja Vehovar

IntroductionThe term �field substitution� usually refers to thepopular practice of replacing the nonrespondingunit with another unit during the fieldwork stage ofthe survey. As mentioned by Chapman (2003) thispractice was �misused� in past decades andobtained a bad reputation. Overview in Vehovar(1999) shows that, in general, substitutions gainonly small improvements in sampling variance, thatis outweighed by numerous practical drawbacks,particularly the lack of control over the interviewersand the delays due to the sequential nature of thesubstitution (as the status of replaced units must beclarified first before the substitute units can beused). It is also true that none of the examples fromthe literature properly demonstrated that thispractice was justified. We could conclude thatsubstitutions are often applied only due to theignorance of the practitioners. The relatively strictrejection of this practice in probability samples inmany face-to-face surveys in the U.S. thus seemsto be the proper attitude.

The (false) justification of this procedure usuallycomes from comparisons with the plain alternativeof taking no action to deal with nonresponse. Moresophisticated justification also arises from the(never proven) belief that substitutions outperformalternative weighting adjustments. This is, in fact, avery tricky intuitive misconception, because it issomehow strange that by obtaining optimalstructure of the sample we get almost no gains inprecision compared to the unbalanced sample ofresponding units without substitutions. In the caseof two-stage cluster sample with n = 100×10 units itthus seems much better to get the designedstructure of respondents (i.e., 100 clusters of 10units each) with substitutions, than the same 1,000units with 100 clusters of varying sizes (e.g., 5-15units) without substitutions. The reason that thedifference is rather small is the fact that the well-known Kish formula for the increase in variancedue to weighting does not hold for nonresponseweights in cluster samples. This fact is alsosomehow surprising, because it confronts anotherintuitive conviction of the practitioners�the beliefthat the Kish�s formula is extremely robust.

Defining SubstitutionsHowever, despite the above theoretical objections,a much more profound criticism relates to theinterviewer�s influence on the selection of thesubstitutes. This also includes the situation wherethe interviewer may declare a certain difficult-to-

deal unit as nonresponding and then obtain somerandom substitute from the management.Nevertheless, the interviewer�s influence�althoughone of the key disadvantages�is not a necessarypart of the procedure. In Vehovar (1999) this basicdistinction was not recognised, so some confusionremained throughout the paper about the exacttype of substitutions being discussed. Let us drawthis distinction here. To start improving thedefinition, we should separate two components, (a)the interviewer influence on selection and (b) thesequential nature of replacing the missing unit withthe substitute one. Only the latter is the necessaryand essential component of the substitutions.There are also no reasons to restrict substitutionsto nonresponse, as we may sequentially substitutealso for noneligible units (e.g., noneligible units).

Figure 1. Types of substitution procedure

Definition: Substitutions relate to sequentialreplacing of the missing unit with another unitduring the field stage of the survey. Two typesshould be distinguished:

! Interview influenced substitutions. Here, theinterviewer may impact whether substitute unitwill be applied. It is true that in face-to-facesurveys it is impossible to completely eliminatethis impact; we can only approximate it withstrict field controls. With respect to how thesubstitution is selected, we have two furtheroptions:

• randomised selection, where the substituteunits are selected from some pre-preparedlist based on a strictly controlled probabilitysampling mechanism;

• nonrandomised selection, where thesubstitution are selected either with somepurposive selection criteria, or, with somefreedom of the interviewer.

A. Substitution procedure in probability sample survey

B. Interviewer influenced

substitutions

C. Management generated

substitutions

D. Randomised selection

E. Nonrandomized selection

F. Substitution procedure in

nonprobability sample surveys


! Management-generated substitutions. Here,the interviewer has no impact on whether orhow to select the substitute. Even when hedeclares a unit as nonresponding, he doesn�tknow, whether this will lead to substitution. Thissituation is very close to what Kish (1965) calls�supplement sample��a list of pre-preparedunits applied when survey managementdecides to do so. However, with supplements,no sequential one-to-one replacement isperformed, which is the essence ofsubstitutions. The most typical example for thistype is the full CATI environment, where thetelephone interviewer has no awareness thatmanagement actually applies substitutions. Onthe other hand, when interviewer knows thesituation within his cluster (i.e., how far he istowards meeting the criteria of completing the�take� b of eligible or responding units) or heeven decides whether and how to select thesubstitute unit, this becomes the interviewer-influenced type of substitutions. Obviously, wesubstitute also in the standard Waksberg-Mitofsky procedure, however, only thenoneligible units are substituted there, as it wassomehow agreed in this procedure not toreplace for nonresponse, despite someattempts to do so.

Figure 2 below illustrates some typical examples ofsubstitutions in probability samples. If we do notexplicitly distinguish among them, we easily end upwith a communication mismatch, which also existsin Vehovar (1999), where the bulk of the treatmentis devoted to type D (Figure 1), with occasionalcomments that actually refer to types E and C. Asthese are all different types, much different criticismrefers to each one of them.

Classic Substitution ExampleThere exists a top substitution-preferred example ofa face-to-face survey with 2 units (e.g., schools,household) per stratum, or per cluster. With thiscase, both, Chapman (1983) and Vehovar (1999)basically agree that substitutions are the preferredoption. Should we really allow the interviewer totake the advantage of being in a certain locationand let him take a substitute, so that he brings aneeded piece of information from this specificgeographical segment included into the sample?Or, should we apply the alternative:

! let the initial sample size be inflated withanticipated nonresponse. If it is 20 percent letall or a portion of the clusters/strata be of theincreased size (3 instead of 2),

! let there be supplement samples ready in casethat nonresponse exceeded our expectations.In cluster samples these may be additionalunits within existing clusters, or in the case of

specific costs structures, the fresh clusters maybe prepared,

! the problem of one or no unit per strata/clustercan be solved with some collapsed samplingtechniques or with imputation/weighting.

In the above case, substitutions seem very�natural� and advantageous to common sense andalso to statistical intuition. However, only profound(and complicated research) can find the properanswer to this problem. The same is true for otherapparently advantageous situations, particularly forthe Waksberg-Mitofsky procedure, where the pastresearch is not yet exhaustive (Chapman andRoman, 1985). This substitution procedure is stillslightly overestimated compared to the modifiedone without substitutions (Brick and Waksberg,1992). Of course, it is also true that the Waksberg-Mitofsky procedure was frequently abandoned inrecent years in the United States because of goodlist assisted samples and also due to thedrawbacks of its sequential nature. However, withmobile phone surveys, at least in Europe, this maydramatically change.

Figure 2. Examples of substitution procedure inprobability samples

ModeSelection Telephone Face-to-faceManagementgeneratedsubstitutions(C)

Waksberg-Mitofsky in fullyautomated CATIprocedure

Supplements(onlyconditionally)

Interviewerinfluencedsubstitutions(B"D)

Standardrandomisedsubstitutionprocedures intelephonesurveys

Standardrandomisedsubstitutionprocedures inface to facesurveys

Substitutions in Nonprobability SamplesSo far, the discussion concentrated only onprobability samples. However, today it is becomingincreasingly important to acknowledge theproblems of nonprobability samples. We may keepsaying that there cannot be any modification in thenotion of scientific sample, nevertheless, we arefaced with a sharp increase of nonprobabilitycomponents in all samples. Is it thus final thatnothing can be done with these samples, exceptthat each practitioner develops his own local artsand tricks for drawing and adjusting hisnonprobability sample?

Two directions are worth mentioning here. One isthe use of substitutions with additional applicationof multiple imputations (Rubin and Zanutto, 2002).The basic idea is to obtain information from the


substitutes at certain locations (cluster, strata,family) while the interviewer is there and use itwithin the multiple imputation strategy. The strictrules of probability samples are slightly abandoned,however, cost-error advantages are obtained inexchange.

Another direction relates to quota samples and tothe contemporary flow of nonprobability samples onthe Web (Vehovar et al, 2002), where the demandfor some theory guidance is particularly strong. Thequestion is whether nothing can be added to Kish(1965, 1999), who declared these samples simplyas the object of art with no scientific substance. Onthe other hand, for the majority of marketresearchers, particularly those working onhousehold or Internet panels, the notion of�sampling�s dead� has already been true fordecades. Careful pre-selection together withsophisticated post-survey adjustments oftenprovide data suitable for the majority of their needs.Yet, there exists a feeling that we are still not ableto take the whole potential of the informationhidden in these nonprobability samples. Propensityscore weighting is an example of an attempt toobtain better inference from such data. Here, thesubstitutions (type F and E) share the same questwith other nonprobability samples for moreresearch efforts elaborating the correspondingcost-error strategies (Vehovar et al, 2001).

ConclusionsThe very essence of substitutions is the sequentialreplacements of missing units. On the other hand,the interviewer�s impact on selection of thesubstitutes exists only in some specific types ofsubstitutions. There, the substitutions cannot bejustified. The only possible exception is the extremecase with only few units per strata/cluster, wherethe situation is still unclear and more research isneeded. However, none of this criticism hits themanagement-generated substitutions, including theWaksberg-Mitofsky procedure in telephonesurveys. There, the substitutions face�besidesgeneral disadvantage arising from their sequentialnature�another problem: they provide a very smallgain in precision, which makes them much lessfavourable than they seem. When substitutions areapplied in nonprobability samples they share thegeneral destiny of this growing family ofnonscientific surveys: they outperform probabilitysamples very often (if judged from overall cost-errorconsiderations), yet little research has been

performed to elaborate on this advantage and helpthe practitioners.

ReferencesBrick, J. M. and Waksberg, J. (1991). Avoidingsequential sampling with RDD. SurveyMethodology, 17(1), pp. 27-41.

Chapman, D. W. (2003). To substitute or not tosubstitute-that is the question. Survey Statistician,48, 32-34.

Chapman, D. W. and Roman, A. (1985). Aninvestigation of substitution for an RDD survey.Proceedings of the Survey Research MethodologySection, ASA, 269-274.

Chapman, D. W. (1983). The impact ofsubstitutions on survey estimates. In Meadow etal.: Incomplete data in sample surveys, vol. II,National Academy of Sciences and AcademicPress.

Kish, L. (1965). Survey sampling. New York:Wiley.

Kish, L. (1999): Quota sampling: Old plus newthought. Available on http://ris.org/vasja/kish.doc.

Rubin, D. B. and Zanutto, E. (2002): Usingmatched substitutes to adjust for nonignorablenonresponse through multiple imputation. InGroves et al (Eds): Survey nonresponse. NewYork:Wiley.

Vehovar, V. (1999). Field substitution and unitnonresponse. Journal of Official Statistics.

Vehovar V., Lozar, Batagelj & Zaletel (2002).Nonresponse in web surveys. In Groves et all(Eds): Survey nonresponse. New York:Wiley.

Vehovar, V., Lozar, K. and Batagelj, Z. (2001).Sensitivity of e-commerce measurement to surveyinstrument. International Journal of ElectronicCommerce, 2001, 6/1.

To Contact the Author:Vasja VehovarUniversity of Ljubljana, [email protected]


IASS General Assembly

The next General Assembly of IASS members willtake place during the ISI meetings in Berlin onThursday, August 14 from 11:15 a.m. to 1:00p.m. All IASS individual members, as well as therepresentatives of institutional members, arewarmly welcomed to participate in this meeting. Foradditional information, please see the web site:http://www.isi-2003.de/.

IASS Elections

In February, ballots were distributed to members ofthe IASS for the election of a President-Elect, twoVice-Presidents, a Scientific Secretary, and sixCouncil Members. Results of the election will beannounced and the candidates terms will begin atthe end of the General Assembly.

Joint IMS-SRMS Mini Meeting onCurrent Trends in Survey Sampling and

Official Statistics, Calcutta, India,January 2-3, 2004

The Institute of Mathematical Statistics (IMS) andthe Survey Research Methods Section (SRMS) willjointly sponsor a mini meeting on Current Trends inSurvey Sampling and Official Statistics to be heldnear Calcutta, India, January 2-3, 2004. This will beco-sponsored by the U.S. Census Bureau, GallupResearch Center at the University of Nebraska-Lincoln and the Department of Statistics, Universityof Calcutta. The mini meeting is intended to serveas a bridge between mathematical statisticians andpractitioners working on sample surveys and officialstatistics either in government or private agencies.

The following researchers have tentatively agreedto present their papers:

Arijit Chaudhuri (India), William Bell (USA), HelenoBolfarine (Brazil), Jay Breidt (USA), Daniela Cocchi(Italy), Stephen E. Fienberg (USA), David Findley(USA), Marco Fortini (Italy), Wayne Fuller (USA),Malay Ghosh (USA), J.K. Ghosh (India/USA), PilarIglesias (Chile), Michael Larsen (USA), BruneroLiseo (Italy), Viviana Lencina (Argentina), MichelMouchart (Belgium), S. James Press (USA), J.N.K.Rao (Canada), Stuart Scott (USA).

The meeting web page http://www.jpsm.umd.edu/ims contains relevant information concerningregistration, the program, and other topics. If youare interested in presenting your paper at themeeting, please contact Partha Lahiri [email protected]. Local information can beobtained from Tathagata Banerjee [email protected].

IASS Program Committee for theSydney ISI Session (2005) Invites

Suggestions

By the time you read this, our committee will be inthe process of completing the short-listing of topicsfor the invited paper sessions to be held during theSydney ISI for the IASS part of the program. Thediscussions to reach a final decision will take placeduring the Berlin ISI Session in August. If youwould like to propose a topic, please do so bysending a title and short abstract (up to 100 words)as well as any other comments and suggestions tothe chair of this committee at the address below.Last minute suggestions are still welcome, providedthey reach our committee before July 31.

Pedro SilvaRio de Janeiro � BrazilPhone: 55-21-25144548Fax: 55-21-25140039E-mail: [email protected]

Announcements


Exercises on Survey Methods withSolutions Editions Ellipses, 2003

There are very few works containing surveyexercises with solutions. In an effort to foster solidunderstanding of the discipline, Pascal Ardilly(INSEE, France) and Yves Tillé (Université deNeuchâtel, Switzerland) offer a selection of 116exercises, each accompanied by a detailedsolution. These exercises are grouped in chaptersand cover the main aspects of survey theory. Theinitial chapters are devoted to sampling methods:simple random sampling, stratified sampling,unequal probability sampling, multi-stage sampling,balanced sampling, and quota sampling. The nextchapters deal with adjustment methods: ratioestimator, regression estimator, poststratifiedestimator, generalized calibration. One chapter is

devoted to variance estimation. The bookconcludes with exercises on nonresponse.This work (300 pages, 25 Euros) containsexercises at all levels, some developing theoreticalaspects, others dealing with more applied problemswith numeric applications. Each chapter beginswith a brief course reminder specifying thenotations and the main findings.

Pascal Ardilly is the head of Sampling andStatistical Data Processing at INSEE in France.Yves Tillé holds a Ph.D. in statistics from BrusselsUniversité Libre and is currently a professor at theUniversité de Neuchâtel in Switzerland. Bothauthors have already published textbooks inFrench:

! Les techniques de sondage (SamplingTechniques) P. Ardilly, (1994). Technip;

! Théorie des sondages, échantillonnage etestimation en population finie (SamplingTheory, Finite Population Sampling andEstimation) Y. Tillé, (2001). Dunod.



Visit the new and improved IASS web site andread The Survey Statistician on line!

www.isi-iass.org

Important Notices

! A PDF file of the newsletter is available on the IASS web site. Very few members no longer wish toreceive the hard copy, but simply to be notified when a new issue is posted. Currently, we do not have aprocess in place to support this option. A process will be developed when an adequate number ofmembers choose the above. Until that time, all members will continue to receive hard copies of thenewsletter. Please send an e-mail to [email protected] if you would like to take advantage ofthe option mentioned above.

! Members are encouraged to view the IASS website (www.isi-iass.org) and provide comments orsuggestions to Eric Rancourt: [email protected].


In Other Journals

Survey Methodology

A Journal Publishedby Statistics Canada

Contents Volume 28, Number 2, December 2002

In This Issue.................................................................................................................................................... 115

GORDON J. BRACKSTONEStrategies and Approaches for Small Area Statistics ..................................................................................... 117

DENNIS TREWINThe Importance of a Quality Culture ............................................................................................................... 125

YVES THIBAUDEAUModel Explicit Item Imputation for Demographic Categories.......................................................................... 135

BALGOBIN NANDRAM, GEUNSHIK HAN and JAI WON CHOIA Hierarchical Bayesian Nonignorable Nonresponse Model for Multinomial Data from Small Areas ............ 145

JAY STEWARTAssessing the Bias Associated with Alternative Contact Strategies in Telephone Time-Use Surveys .......... 157

ROBERT M. BELL and DANIEL F. MCCAFFREYBias Reduction in Standard Errors for Linear Regression with Multi-Stage Samples .................................... 169

MONROE G. SIRKENDesign Effects of Sampling Frames in Establishments Survey ...................................................................... 183

LOUIS-PAUL RIVESTA Generalization of the Lavallée and Hidiroglou Algorithm for Stratification in Business Surveys................. 191

WILSON LU and RANDY R. SITTERMulti-way Stratification by Linear Programming Made Practical..................................................................... 199

ROBBERT H. RENSSEN and GERARD H. MARTINUSOn the Use of Generalized Inverse Matrices in Sampling Theory.................................................................. 209

Acknowledgements ......................................................................................................................................... 213


Journal of Official Statistics

An International Review Publishedby Statistics Sweden

Contents Volume 18, Number 4, 2002

JOS is a scholarly quarterly that specializes in statistical methodology and applications. Survey methodologyand other issues pertinent to the production of statistics at national offices and other statistical organizationsare emphasized. All manuscripts are rigorously reviewed by independent referees and members of theEditorial Board.

Measures to Evaluate the Discrepancy Between Direct and Indirect Model-basedSeasonal AdjustmentEduardo Otranto and Umberto Triacca........................................................................................................... 511

Satisfying Disclosure Restrictions with Synthetic Data SetsJerome P. Reiter ............................................................................................................................................. 531

Item Nonresponse as a Predictor of Unit Nonresponse in a Panel SurveyGeert Loosveldt, Jan Pickery, and Jaak Billiet................................................................................................ 545

Using Administrative Records to Improve Small Area Estimation: An Examplefrom the U.S. Decennial CensusElaine Zanutto and Alan M. Zaslavsky............................................................................................................ 559

Asymptotically Efficient Generalised Regression EstimatorsGiorgio E. Montanari and M. G. Ranalli .......................................................................................................... 577

Editorial Collaborators..................................................................................................................................... 591

Index to Volume 18, 2002 ............................................................................................................................... 595


Journal of Official Statistics

An International Review Publishedby Statistics Sweden

Contents Volume 19, Number 1, 2003

Multiple Imputation for Statistical Disclosure LimitationT. E. Raghunathan, J.P. Reiter, and D.B. Rubin............................................................................................. 1

Confidence Interval Coverage Properties for Regression Estimators in Uni-Phase andTwo-Phase SamplingJ. N. K. Rao, W. Jocelyn, and M. A. Hidiroglou .............................................................................................. 17

Sequential Weight Adjustments for Location and Cooperation Propensity for the 1995 NationalSurvey of Family GrowthVincent G. Iannaccione ................................................................................................................................... 31

Constrained Inverse Adaptive Cluster SamplingEmilia Rocco ................................................................................................................................................... 45

Poisson Mixture Sampling Combined with Order SamplingHannu Kröger, Carl-Erik Särndal, and Ismo Teikari ....................................................................................... 59

Ischemic Heart Disease and Acute Myocardial Infarction as Cause of Death in aCause-of-Death RegisterBoo Svartbo, Gösta Bucht, Lars Olov Bygren, Anders Eriksson, and Lars Age Johansson .......................... 71

All inquires about submissions and subscriptions should be directed to [email protected]


Statistics in Transition

Journal of the Polish Statistical Association

Contents Volume 5, Number 5, October 2002

From the Editor .............................................................................................................................................. 721

Classification and Data Analysis

K. JAJUGAForeword ......................................................................................................................................................... 723

H. H. BOCKClustering Methods: From Classical Models to New Approaches.................................................................. 725

W. J. KRZANOWSKIOrthogonal Components for Grouped Data: Review and Applications .......................................................... 759

Cz. DOMAŃSKISome Remarks on the Tasks of Statistics on the Verge of the Twenty First Century.................................... 779

A. ZELIAŚSome Notes on the Selection of Normalisation of Diagnostic Variables ........................................................ 787

D. BANKS and R. T. OLSZEWSKICombinatorial Search in Multivariate Statistics............................................................................................... 803

M. GREENACRE and J. G. CLAVELSimultaneous Visualization of Two Transition Tables .................................................................................... 823

E. GATNARWhat is Data Mining? ...................................................................................................................................... 837

M. D. LARSENImpact of Latent Class Clustering of NSF Doctoral Survey Data on Adjusted Rand Index Values................ 843

Other Articles

A. MŁODAKAn Approach to the Problem of Spatial Differentiation of Multi-feature Objects UsingMethods of Game Theory ............................................................................................................................... 857

H. P. SINGH and N. MATHURAn Alternative to an Improved Randomized Response Strategy.................................................................... 873

Reports

International Federation of Classification Societies Conference�IFCS-2002, Cracov, July 16-19, 2002(K. Jajuga)....................................................................................................................................................... 887

The Ninetieth Anniversary of the Foundation of the Polish Statistical Association�the ScientificConference,Krakow, Poland, July 14-15, 2002 (A. Zeliaś) ............................................................................. 891




Contents Volume 5, Number 6, December 2002

From the Editor .............................................................................................................................................. 897

Classification and Data Analysis

D. THORBURNForeword ......................................................................................................................................................... 899

M. CARLSONAssessing Microdata Disclosure Risk Using the Poisson-Inverse Gaussian Distribution .............................. 901

M. DAVIDSEN and M. KJØLLERThe Danish Health and Morbidity Survey 2000�Design and Analysis.......................................................... 927

D. HEDLINEstimating Totals in Some U.K. Business Surveys......................................................................................... 943

A. HOLMBERGA Multiparameter Perspective on the Choice of Sampling Design in Surveys ............................................... 969

V. KIVINIEMI and R. LEHTONENWeb Tools in Teaching and Learning of Survey Sampling: The VliSS Application ........................................ 995

S. LAAKSONENTraditional and New Techniques for Imputation ............................................................................................. 1013

K. MEISTERAsymptotic Considerations Concerning Real Time Sampling Methods ......................................................... 1037

Other Articles

B. PRASAD, R. S. SINGH, and H. P. SINGHModified Chain Ratio Estimators for Finite Population Mean Using Two Auxiliary Variables in DoubleSampling ......................................................................................................................................................... 1051

G. N. SINGHEstimation of Population Ratio in Two Phase Sampling................................................................................. 1067

T. P. TRIPATHI, H. P. SINGH, and L. N. UPADHYAYAA General Method of Estimation and Its Application to the Estimation of Coefficient of Variation ................. 1081

Book Reviews

W. Charemza and K. Strzala (Eds), East European Transition and EU Enlargement:A Quantitative Approach, (by Subrata Ghatak)............................................................................................... 1103




Contents Volume 5, Number 6, December 2002 (cont)

Economic and Social Trends in Transition Countries

Basic Economic Trends in Countries of Central and Eastern Europe and CIS Countries(by M. Bieńkowska, E. Czumaj, J. Gniadzik) .................................................................................................. 1103

Obituary

Tore E. Dalenius (1917-2002)......................................................................................................................... 1117

Acknowledgements

Referees of Volume 5 ..................................................................................................................................... 1121


The Allgemeines Statistisches Archiv (AStA)

www.uni-koeln.de/wiso-fak/wisostatsem/asta/

The Allgemeines Statistisches Archiv (AStA), Journal of the German Statistical Society is edited by Prof. Dr.Karl Mosler, University of Cologne. It provides an international forum for researchers and users from allbranches of statistics. The first part (Articles) contains contributions to statistical theory, methods, andapplications. A focus is on statistical problems that arise in the analysis of economic and social phenomena.The second part (News & Reports) consists of news, reports, and other material that relates to the activities ofthe Society. Book reviews are published in the third part (Books). Invited papers which have been presentedto the Annual Congress of the Society are regularly published in the second issue of a volume. All papers inthe first part are anonymously refereed. In order to be acceptable, a paper must either present a novelmethodological approach or a result, obtained by a substantial use of statistical method, which has asignificant scientific or societal impact. Occasional review papers are also encouraged. Authors ofcontributions are requested to mail three copies to the editor for refereeing. The journal publishes originalcontributions in English and German language, articles preferably in English. After acceptance the author hasto provide a LaTeX- or ASCII file of the final manuscript that follows the style of the Allgemeines StatistischesArchiv in citation and layout. For submissions, please contact:

Petra SablotnyAllgemeines Statistisches Archiv (Prof. Mosler)Universität zu KölnAlbertus-Magnus-PlatzD-50923 Kölne-mail: [email protected]

Contents Volume 87, Issue 1, 2003

Articles

Core Inflation in the Euro Area: Evidence from the Structural Linebreak VAR ApproachElke Hahn....................................................................................................................................................... 1-23

On the Optimal Design in Stratified Regression EstimationRalf Münnich ................................................................................................................................................ 25-38

Systematische Ausfälle im Mikrozensus-Panel: Ausmass und Auswirkungen auf die Qualitätvon ArbeitsmarktanalysenSylvia Zühlke................................................................................................................................................ 39-58

Heaping and Its Consequences for Duration Analysis: A Simulation StudyJoachim Wolff, Thomas Augustin ................................................................................................................ 59-86

News and Reports

Die Statistische Erfassung von Gründungen in Deutschland�Ein Vergleich von Beschäftigtenstatistik,Gewerbeanzeigenstatistik und den Mannheimer GründungspanelsMichael Fritsch, Reinhold Grotz, Udo Brixy, Michael Niese, Anne Otto ...................................................... 87-96

Rolf Krengel In MemoriamReiner Stäglin............................................................................................................................................... 97-99



Contents Volume 87, Issue 1, 2003, (cont.)

Books

Buchbesprechungen ................................................................................................................................ 101-105

Kucerá, T., Kucerá, O., Opara, O., Schaich , E.New Demographic Faces of Europe. The Changing Population Dynamics in Countries of Centraland Eastern EuropeReviewed by G. Uebe

Shumway, R. H., Stoffer, D. S.Time Series Analysis and Its ApplicationsReviewed by W. Schmid

Wilcox, R. R.Fundamentals of Modern Statistical Methods, Substantially Improving Power and AccuracyReviewed by G. Uebe

Härdle, W., Rönz, B.MM* Stat, Eine Interaktive Einführung in die Welt der StatistikReviewed by P. von der Lippe

Buchinformation ...............................................................................................................................................106

Contents Volume 87, Issue 2, 2003

Articles

Volkswirtschaftliche Gesamtrechnungen im Zentrum der WirtschaftspolitischenDiskussion�Zur EinführungRuth Meier, Reiner Stäglin ....................................................................................................................... 107-112

Stability Criteria and Convergence: The Role of the System of National Accounts for FiscalPolicy in EuropeTilmann Brück, Andreas Cors, Klaus F. Zimmermann, Rudolf Zwiener .................................................. 113-131

Das System der Volkswirtschaftlichen Gesamtrechnungen als Informationsquelle für dieGeldpolitik in der EWUHermann Remsperger.............................................................................................................................. 133-146

Gemeinschaftliches Europa�Nationale Volkswirtschaftliche GesamtrechnungenJosef Richter ............................................................................................................................................ 147-163

Wachstumsrückstand in Deutschland?�Probleme der DeflationierungHeinrich Lützel ......................................................................................................................................... 165-175

Volkswirtschaftliche Gesamtrechnungen als die Kunst des MöglichenWolfgang Strohm ..................................................................................................................................... 177-182

Qualität und Genauigkeit der Volkswirtschaftlichen GesamtrechnungenAlbert Braakmann .................................................................................................................................... 183-199



Contents Volume 87, Issue 2, 2003, (cont.)

An Approach for Timely Estimations of the German GDPAndreas Cors, Vladimir Kouzine .............................................................................................................. 201-220

Gibt es Eine Theorie der VGR?Utz-Peter Reich........................................................................................................................................ 221-237

INTERNATIONAL ASSOCIATIONOF SURVEY STATISTICIANS

CHANGE OF ADDRESS FORM

If your home or business address has changed, please copy,complete, and mail this form to:

IASS Secretariatc/o INSEE-CEFILAtt. Ms. Claude Olivier3, rue de la Cité33500 Libourne — France

Last name: Mr./Mrs./Miss/Ms. _______________ First name: __________________________

E-mail address (please just indicate one): ___________________________________________

May we list your e-mail address on the IASS web site?

Yes No

Home address

Street: ____________________________________________________________________________

City: ______________________________________________________________________________

State/Province: ___________________________ ZIP/Postal code: _____________________

Country: __________________________________________________________________________

Telephone number: ________________________________________________________________

Fax number: ______________________________________________________________________

Business address

Company: _________________________________________________________________________

Street: ____________________________________________________________________________

City: ______________________________________________________________________________

State/Province: ___________________________ ZIP/Postal code: _____________________

Country: __________________________________________________________________________

Telephone number: ________________________________________________________________

Fax number: ______________________________________________________________________

Please specify address to which your IASS correspondence should be sent:

Home Business

IASS Officers and Council Members

President (2001-2003): Xavier Charoy (France), [email protected] Elect (2003-2005): Luigi Biggeri (Italy), [email protected] Presidents (2001-2003): David Binder (Canada), [email protected]

Anders Christianson (Sweden): [email protected] Secretary (2001-2003): Seppo Laaksonen (Finland): [email protected] Members (1999-2003): Florentina Alvarez (Spain), [email protected]

Cynthia Clark (USA), [email protected] Do Nascimento Silva (Brazil), [email protected] Martin (UK), [email protected] Muba (Tanzania), [email protected] Tikkiwal (India), [email protected]

Council Members (2001-2005): Kari Djerf (Finland), [email protected] Fitch (USA-Guatemala), [email protected] Gligorova (Croatia), [email protected] Huang (China), [email protected] Sicron (Israel), [email protected] Thiongane (Senegal-Ethiopia), [email protected]

Committees Chairs:Berlin 2003 Program Committee: Danny Pfeffermann (Israel), [email protected] 2005 Program Committee: Pedro Do Nascimento Silva (Brazil), [email protected] Election Nominations Committee: Oladejo Ajayi (Nigeria), [email protected] of 2003 Cochran-Hansen Prize: Chris Skinner (UK), [email protected]

The Secretariat:Executive Director: Alain Charraud (France), [email protected]: François Fabre (France), [email protected] Secretary: Anna Maria Vespa-Leyder (France), [email protected]: Claude Olivier (France), [email protected]

Institutional Members

4 international organizations

AFRISTATCICRED

EUROSTATUNITED NATIONS STATISTICAL DIVISION

31 bureaus of statistics

ARGENTINA - INSTITUTO NACIONAL DE ESTADISTICA Y CENSOS - INDECAUSTRALIA - AUSTRALIAN BUREAU OF STATISTICS

BELGIUM - INSTITUT NATIONAL DE STATISTIQUEBRAZIL - INSTITUTO BRASILEIRO DE GEOGRAFIA E ESTATISTICA - IBGE

CANADA - STATISTICS CANADACHINA - GOVERNO DE MACAU

CÔTE D'IVOIRE - INSTITUT NATIONAL DE LA STATISTIQUECZECH REPUBLIC - CZECH STATISTICAL OFFICE

DENMARK - DANMARKS STATISTIKFINLAND - STATISTICS FINLAND

FRANCE - INSTITUT NATIONAL DE STATISTIQUE ET D'ÉTUDES ÉCONOMIQUES - INSEEGAMBIA - CENTRAL STATISTICS DEPARTMENT

GERMANY - STATISTICHE BUNDESAMTGREECE - NATIONAL STATISTICAL SERVICE OF GREECE

IRAN - STATISTICAL CENTER OF IRANITALY - INSTITUTO CENTRALE DI STATISTICA - ISTAT

MEXICO - INSTITUTO NACIONAL DE ESTADISTICA, GEOGRAFIA E INFORMATICA - INEGINETHERLANDS - CENTRAL BUREAU OF STATISTICS

NEW ZEALAND - STATISTICS NEW ZEALANDNIGERIA - FEDERAL OFFICE OF STATISTICS

NORWAY - CENTRAL BUREAU OF STATISTICSPORTUGAL - INSTITUTO NACIONAL DE ESTATISTICA - INE

REPUBLIC OF KOREA - NATIONAL STATISTICAL OFFICE - NSOSPAIN - INSTITUTO NACIONAL DE ESTADISTICA

SWEDEN - STATISTICS SWEDENSWITZERLAND - OFFICE FÉDÉRAL DE STATISTIQUE

TANZANIA - BUREAU OF STATISTICSUNITED KINGDOM - OFFICE FOR NATIONAL STATISTICS

USA - BUREAU OF THE CENSUSUSA - DEPARTMENT OF EDUCATION

USA - DEPARTMENT OF HEALTH AND HUMAN SERVICES

6 universities, research centers, private firms of statistics

ARGENTINA - UNIVERSITAD NACIONAL DE TRES DE FEBRERODENMARK - SFI

EUROPE - A.C. NIELSEN MANAGEMENT SERVICEFRANCE - INSTITUT NATIONAL D'ÉTUDES DÉMOGRAPHIQUES - INED

USA - RESEARCH TRIANGLE INSTITUTEUSA - WESTAT

Documents

In This Issue · newsletter. This issue is the last under her watch. Our next editor will be Steve Heeringa ([email protected]). I thank him for agreeing to keep the Survey Statistician