A FRAMEWORK FOR PRIORITY CONTACT OF NON RESPONDENTS

Transcription

1 A FRAMEWORK FOR PRIORITY CONTACT OF NON RESPONDENTS Rchard McKenze, Australan Bureau of Statstcs. 12p36 Exchange Plaza, GPO Box K881, Perth, WA ABSTRACT Busnesses whch have not repled to a mal survey are generally subject to ntensve follow up (IFU), by telephone or other means, to obtan a response. As ths contact s expensve, strateges may be used to determne whch busnesses are gven prorty for contact. Ths paper presents a theoretcal framework for developng and assessng prorty IFU procedures. The stuaton s descrbed by a set of respondents to the ntal mal survey, models for the probablty of response wth and wthout IFU for the set of non respondents (e the remander of the sample) at a partcular pont n tme, and models for the reported and mputed values for ths set of non respondents. A score functon s derved whch, for a fxed number of contacts, maxmses the expected mprovement n accuracy. The framework s then modfed for practcal mplementaton, and a case study presented on ts applcaton to the ABS Survey of Employment and Earnngs. The vews expressed n ths paper are those of the author and do not necessarly reflect those of the Australan Bureau of Statstcs. 1. INTRODUCTION There has been a focus n the lterature n recent years of applyng models based on sound mathematcal prncples to enable the effcent use of survey resources for respondent follow up acton n establshment surveys. Some good examples nclude the work of Latouche and Berthelot (1992), Lawrence and McDavtt (1994) and Granqust and Kovar (1997). Ths work, however, has focused on makng ratonal decsons on whether system generated queres on data receved from respondents (referred to as 'edt' queres) are worth followng up n a cost beneft sense. In ths paper, we present a mathematcal framework to ad n makng ratonal decsons on the extent to whch busnesses, whch are non-respondents to establshment surveys, should be followed up to elct a response. Ths framework s then adapted to enable practcal applcaton through makng a seres of assumptons, and ts effectveness s then evaluated n a case study. The term IFU s used regularly n ths paper, and refers to the process of contnually recontactng a busness (by phone or other means) to elct a response to a mal based establshment survey. 2. ESTABLISHING THE BASIS FOR THE FRAMEWORK Busnesses, whch have not responded to a mal establshment survey, wll be mputed for n some way n the estmaton process. Therefore a realstc am for the IFU process s to mnmse the non-response bas n survey estmates. In ths paper we develop a framework for prortsng IFU such that the non-response bas n survey estmates s mnmsed, assumng the estmated quantty to be a smple level estmate. Our am s to descrbe the beneft from undertakng the IFU process as an objectve functon to be optmsed, subject to a constrant based on the resources avalable for performng IFU. 3. OUTLINE OF A PRIORITY IFU FRAMEWORK FOR THE ONE VARIABLE CASE In ths paper we descrbe a prorty IFU framework for the one varable case, e.g. for a survey varable Y, wth values Y for each unt n the populaton. We assume a true superpopulaton model ξ for Y, whch may depend on auxlary nformaton, denoted as x. In practce ξ s unknown and workng models ζ have to be used. Dfferent classes of unts may have dfferent auxlary nformaton avalable for use leadng to a dfferent workng model beng assocated wth each class A model for prorty IFU We defne the followng notaton for an establshment sample survey: S R NR C NC CR CNR Set of unts chosen n the sample. Set of respondents to a survey pror to performng IFU.e. receved after ntal malout Set of ntal non-respondents to the survey pror to performng IFU. Note that S = R NR. Set of ntal non-respondents (.e unts n NR) that wll be contacted as part of IFU. Set of ntal non-respondents that wll not be contacted as part of IFU. Note that NR = C NC. Set of unts n C that respond as a result of IFU beng performed. Set of unts n C that do not respond as a result of IFU beng performed. Note that C = CR CNR. 473

2 NCR NCNR Set of unts n NC that respond even though no IFU was performed e.g. may be late respondents. Set of unts n NC that do not respond. We assume explct non-response mputaton n the framework whch follows: Let Y # be a random varable whch represents the data we would use n estmaton pror to performng IFU for the th unt and defne t to be: Y : f the unt s a respondent Y {R} : f the unt s a non-respondent (ths notaton has been adopted to ndcate that the mputed value may be dependent on the members of R) Let Y ## be a random varable whch represents the effect on Y # of the IFU process and defne t to be: Y : f the unt responds from IFU, responds late wthout IFU or was a respondent before IFU Y {R CR NCR} : otherwse (the mputed value may depend on the new recevals and orgnal respondents) Defne F = 1 for unts who receve IFU, 0 otherwse. Thus F s a determnstc varable whch we control by our IFU strategy. Also defne I to be the random varable whch ndcates whether unt responds (I = 1), or fals to respond (I = 0). Let I {C} ndcate that unt was an ntal non-respondent ncluded n the IFU process, whch can therefore be descrbed as a Bernoull random varable wth probablty of response parameter p {C}. Let I {NC} ndcate that unt was an ntal non-respondent not ncluded n the IFU process, whch can also be descrbed as a Bernoull random varable wth a dfferent probablty of response parameter p {NC} The effect of IFU n terms of our model The followng algebra expands an expresson for the non-response bas remanng after the IFU process to nclude the quanttes defned n secton 3.1. Ths s necessary to express the framework n a manner whch relates to the outcomes of the IFU process, whch wll facltate assumptons beng made to eventually smplfy the framework to enable practcal mplementaton. The non-response bas remanng after the IFU process s gven by: NR w (Y Y ## ) (w s the estmaton weght for unt ) = NR w F [I {C}(Y Y ) + (1 - I {C})(Y Y {R CR NCR} )] + NR w (1 - F ) [I {NC}(Y Y ) + (1 - I {NC})(Y - Y {R CR NCR} )] = NR w [Y Y {R CR NCR} ] [F (1 - I {C}) + (1 - F )(1 - I {NC})] (1) Introducng the noton of a constrant, we would lke to mnmse (1) subject to F [ m or K F [ K. That s, for a fxed number of unts (m), or a known unt IFU cost (K ) and total resources avalable for IFU (K), mnmse the non-response bas remanng after IFU. Ths s a dffcult 0-1 nteger programmng problem, as Y {R CR NCR} wll depend on all of the F n a nonlnear fashon. If, however, we assume Y {R CR NCR} - Y {R} s small compared to Y - Y {R}, and note that: (Y - Y {R CR NCR} ) = (Y - Y {R} ) + (Y {R} - Y {R CR NCR} ), then (1) NR w [Y Y {R} ] [(1 - I {NC}) - F (I {C} - I {NC})] (2) Ths assumpton essentally means that we don't expect the mputaton error ncurred from mputng for an outstandng unt to change much once addtonal unts have responded as a result of performng IFU. We revst ths assumpton when descrbng the case study of practcal mplementaton n secton Dervng unt level scores and the noton of a beneft functon The am of the followng algebra s to defne a unt level score functon that can be evaluated for all non-respondents n a survey, whch enables non-response bas to be mnmsed through the use of IFU. To facltate ths, we apply 474

3 the trangle nequalty to (2). Ths redefnes the problem to one whch lmts (rather than mnmses) the remanng non-response bas: (2) [ NR w Y Y {R} [(1 - I {NC}) - F (I {C} - I {NC})] = NR w Y Y {R} (1 - I {NC}) - NR w Y Y {R} F (I {C} - I {NC}) (3) Obtanng (2) from (3) s takng a conservatve approach, as t no longer takes nto account the possblty that nonresponse bases from dfferent unts may cancel out. Mnmsng (3) s equvalent to maxmsng: NR w Y Y {R} F (I {C} - I {NC}) (4) subject to the same constrant of F [ m or K F [ K as the frst term of (3) s ndependent of the IFU process. From (4) we can defne a unt level score functon, denoted by S as: S = w Y Y {R} (I {C} - I {NC}) (5) Thus to solve the problem we need to maxmse: B = NR S F Ths expresson can be regarded as a beneft functon (B) that we wsh to maxmse, n order to obtan the maxmum beneft from performng IFU. Ths concept reflects that by maxmsng B through the use of IFU we mnmse the non-response bas n the survey estmate. The parameters I {C} and I {NC} of S are random varables, hence we need to take the expected value of S, whch yelds: E ξ, I{C}, I{NC} {S C, NC, y and R} = w Y - Y {R} [ p {C} - p {NC}] (6) In practce, the components of S would have to be estmated, hence we consder: E { } { }{ S C, NC, x, y and R} = w ŷ y { } [ pˆ { C} pˆ { NC} ] ζ, I C, I NC R (7) We denote ths quantty n (7) as Ŝ and consequently our estmated beneft functon that we wsh to maxmse as: Bˆ = NR Ŝ F Maxmsng Bˆ subject to F [ m leads to choosng the m unts wth the largest values of Ŝ. Maxmsng Bˆ subject to K F [ K leads to a 0-1 nteger programmng problem (not easy to solve). 4. ISSUES TO CONSIDER IN THE PRACTICAL APPLICATION OF THE FRAMEWORK 4.1. Intal score functon In (7) above we have defned the followng estmated unt level score functon: Ŝ = w ŷ y { } [ pˆ { C} pˆ { NC} ], R whch we want to determne for every unt n the sample whch s a non respondent pror to IFU. Ths rases a number of ssues n regards to applyng ths theory n practce. 475

4 4.2. Multple varables In the majorty of surveys, more than one data tem s collected, and the unt of measurement for dfferent data tems collected wll often be dfferent e.g. for a survey collectng, say, employment and gross wages and salares. We would lke to assess the beneft of IFU on each data tem to derve an overall score for a non respondng unt. Ths would requre some form of scalng to ensure scores for each varable are comparable, and then combnng the scores for each varable n some way to form an overall score for the unt Controllng for dfferent levels of estmates It s unlkely that the users of survey output are only nterested n estmates at one aggregate level, e.g. for a varety of ABS surveys, users may be equally nterested n Australan level estmates and Industry or State level estmates. If the scorng functon n (7) was appled to all unts, t s lkely that the unts wth the hghest weghts (e.g. those n the most populous States or Industres) wll attract the largest scores (assumed wth all other thngs beng equal). Prortsng IFU on ths bass may therefore brng the greatest beneft to estmates at the aggregate level and the largest sub populatons, but ths could be at the expense of the resultant qualty of estmates for smaller sub populatons, whch could be of equal mportance to users. Therefore some scalng of scores to take account of ths ssue may also be necessary Maxmsng beneft subject to fxed costs In developng the framework t was mentoned that maxmsng F S subject to K F [ K leads to a 0-1 nteger programmng problem that s dffcult to solve. In practce, estmatng the K for ndvdual unts would be dffcult, as separate unts may requre a substantally dfferent amount of IFU effort to elct a response, and ths s unlkely to be constant across survey cycles for the same unt. It may be possble to estmate an average K (say k) over all unts, however f ths was used the nequalty: K F [ K would break down to F [ K / k whch s equvalent to F [ m. Hence the focus n the practcal applcaton of the framework wll be to choose the m unts wth the hghest S scores, after some form of scalng has been performed to address the ssues rased n 4.2. and Probabltes of response A key component of the scorng functon s [ pˆ { C} pˆ { NC} ]. The pˆ { C} depends on the nature of the IFU contact. As we are attemptng to ensure that all unts regarded as mportant by our scorng functon should be followed up untl a response s acheved (whch may even nvolve vstng the respondent), then unless a busness refuses to be part of the survey a response s most lkely to be obtaned. Experence wth ABS surveys s that only a small proporton of busnesses wll refuse to be nvolved n the survey, under whch case the ABS then has the power to pˆ C = 1 n our make the survey ther legal oblgaton. Consequently t would be reasonable to use a value of { } model. The pˆ { NC} depends on when IFU s performed. The longer ths s after the specfed due date for the orgnally maled out form the more lkely t s that ths value wll tend to 0. Consequently assumng a value of pˆ { NC} = 0 s also reasonable for our model. These assumptons smplfy our scorng functon to the expresson below, whch wll be used n the practcal applcaton of the framework presented n ths paper. Ŝ = w ŷ y { R} (8) The score functon n (8) can be nterpreted as the expected absolute mputaton bas of not recevng a response for the th unt whch s mnmsed over the set of non respondents by achevng a response through IFU from the m unts wth the largest Ŝ scores. 476

5 5. APPLICATION OF THE PRIORITY IFU FRAMEWORK TO THE AUSTRALIAN SURVEY OF EMPLOYMENT AND EARNINGS 5.1. The Australan Survey of Employment and Earnngs The Australan Survey of Employment and Earnngs (SEE) s a quarterly sample survey collectng data from approxmately 13,500 busnesses n Australa across all ndustres excludng Agrculture, Forestry and Fshng. The data tems collected are total employment (splt by full tme / part tme) for the pay date closest to the mddle of the quarter and gross quarterly earnngs (consstng of gross wages and salares, severance termnaton and redundancy payments, and drectors fees) pad to all employees of the busness n the quarter. The proporton of busnesses new to the SEE sample each quarter s approxmately 8%. Estmates are frst produced n the SEE approxmately fve and a half weeks after the reference quarter to feed nto the prelmnary Natonal Accounts release. Fnal estmates for publcaton purposes are then produced approxmately 12 weeks after the reference quarter. The key estmates produced from the SEE are total employment (TE) and gross quarterly earnngs (GQE) at a varety of estmaton levels, the fnest beng at the State (8 States) by Industry (16 Industres) level. The followng collecton strategy s currently n use for the SEE: 1) Due date for the form s the second Thursday of the month followng the reference quarter, at whch tme approxmately 35-40% of busnesses have responded. Mal remnders are sent to outstandng busnesses the day after the form due date. 2) Phone IFU acton commences for busnesses outstandng three weeks after the reference quarter, at whch tme approxmately 55-60% of busnesses have responded. All outstandng busnesses are contacted at least once wthn a two and a half week perod. 3) Prelmnary estmaton s performed at fve and a half weeks after the reference quarter, at whch tme approxmately 80% of busnesses have responded. 4) Phone IFU acton contnues for all unts, wth the am of achevng 100% response pror to close off for fnal estmaton at 12 weeks after the reference quarter, n general approxmately 99% of busnesses respond each quarter. It has been estmated that roughly the same amount of resources are expended on phone IFU n the SEE to acheve a response rate of 90% and to get the response rate from 90% to 99%. Consequently the applcaton of a prorty IFU technque to SEE could result n large effcency gans Method used for applyng the prorty IFU framework to the SEE The model adopted for scorng a unt We need to make use of auxlary nformaton (.e. the x ) to determne a relevant value for ŷ from the scorng functon n equaton (8), whch s reproduced below. Ŝ = w ŷ y { R} Also n dervng equaton (2) from equaton (1) n secton 3.2. t was assumed that y {R} was unlkely to change much as a result of performng IFU (.e. y {R} s assumed not to vary much wth the survey response rate), or more mportantly that any change n a unt s mputed value as a result of IFU s lkely to be small relatve to the value of ŷ pror to performng IFU. ( ) y { R} 477

6 In applyng the framework of ths model to the SEE the relance on ths assumpton s removed, as we estmate the actual ( ŷ y { }) from the prevous quarter s survey data fle (SDF). That s, our mputaton model ζ s R ndependent of the set of respondents R at any stage of the current survey cycle. Ths s acheved by usng the knowledge of what unt reported for the data tem y last quarter as our value of ŷ, and what the mpute of data tem y for unt would have been last quarter f t had not responded as our value of y {R}. The assumpton we are ntroducng however s that the expected absolute mputaton bas (.e. the quantty ( ŷ y { }) ) for a busness s R relatvely constant from one quarter to the next. Dfferent mputaton strateges are used n the SEE for unts n sampled strata and completely enumerated (CE) strata (.e. all unts n the strata are selected). The above calculaton cannot be made for busnesses that were nonrespondents n the prevous quarter. In ths case the relevant data from 2 quarters ago s used. If a busness has been a non-respondent for the past two quarters, or s new to the sample, then a default score s gven to ensure the busness receves IFU (ths s explaned n greater detal n secton ) Scalng for dfferent varables and levels of estmaton Users of the SEE are most nterested n estmates of total employment (TE) and gross quarterly earnngs (GQE) at the State (8 groups) by ndustry (16 groups) level and relevant aggregates of these. Therefore scores for prorty IFU need to be scaled to account for the dfferent magntudes of the man data tems and dfferent levels of estmates requred as dscussed n 4.2. and The procedures below descrbe how ths was done for the SEE n our case study. The basc dea was to derve a smple nteger score for each unt between 0 and 10 whch s then easy for survey processng staff to nterpret. 1) Group together ndustres domnated by a partcular sector (.e. publc vs prvate) and / or havng smlar total employment. Ths enabled the 16 ndustry groups to be reduced to 7, requrng prorty IFU to be controlled for 8 (for the 8 States) x 7 = 56 separate groups. 2) Usng hstorcal SDF's, calculate the values of Ŝ by the method descrbed n for all unts n each hstorcal SDF where possble (.e. all unts except for chronc non respondents or unts new to the survey for the partcular quarters' SDF). 3) At the 56 State by ndustry group levels, determne decles for the dstrbuton of Ŝ values for the partcular quarters SDF for each varable and then medan smooth these decle values over the quarters analysed. 4) To determne the scaled IFU score for an outstandng unt n a future quarter, calculate the unt s Ŝ value for both TE and GQE, and determne whch decle (where decles are numbered from 0-9) t belongs to for ts partcular State by broad ndustry groupng based on the smoothed decle values obtaned by the process descrbed n pont 3 above. Take the maxmum decle (.e. the maxmum between TE and GQE) the unts Ŝ belongs to as the unts prorty IFU score e.g. f an outstandng unt s Ŝ value for TE fell n the th percentle, but for GQE fell n the th percentle, then the unt s prorty IFU score would be 3. The maxmum prorty IFU score for a unt s therefore 9. We denote ths scaled Ŝ score as Z to ndcate the fnal scaled IFU score for unt. 5) For the purpose of ths analyss, outstandng unts who were new to the survey for the partcular quarter or had been a non respondent for the prevous two quarters were gven a default prorty IFU score of 10, and thus can be separately dentfed. The decson to group a unt s Ŝ scores for TE and GQE nto decles was arbtrary. However t s a smple method to scale dfferent varables and produce a meanngful dscrete score that can be used to effectvely dstngush between those unts for whch recevng a response s mportant and other unts for whch obtanng a response through IFU s less mportant. 478

7 Determnng cutoff prorty IFU scores The framework refers to the optmal soluton as choosng the m unts wth the hghest Z scores to be subject to IFU, where m s based on some knowledge of the unt cost for IFU and the total resources avalable to undertake IFU n the partcular survey. In applyng the theory to the SEE we have taken a slghtly dfferent approach, where we are nterested n a cutoff Z score for whch the beneft n obtanng a response through performng IFU for unts below ths score s margnal n relaton to the expected mpact on the survey estmates relatve to the samplng error of the estmate. It s expected that choosng ths cutoff score wll equate to a relatvely stable value of m each quarter, or equvalently a stable number of unts that wll not receve IFU Results from applyng the prorty IFU framework to the SEE for a test quarter Descrpton of data fles used and resultng scores Data from the SEE for all four quarters of 1998 and the frst quarter of 1999 were used to determne the approprate medan smoothed decles by the procedure outlned n From performng ths process t was found that the absolute error of mputaton for a unt was reasonably consstent from quarter to quarter, whch supported the assumpton requred for ths scorng technque to be effectve as dscussed n secton Wth the medan smoothed decles calculated, t was possble to score outstandng unts for the second quarter of 1999, referred to as 2/99. Scorng for unts n the 2/99 survey was performed for all unts outstandng at prelmnary estmaton tme, whch equated to approxmately 17.5% of the survey sample. Note that ths analyss has been performed retrospectvely, and approprate fles dd not exst to allow scorng at the commencement of IFU or at any other tme n the survey cycle. The followng table lsts the dstrbuton of resultng scores. Table 1: Scores set for outstandng unts at prelmnary estmaton n 2/99 Scores Number of unts assgned ths score Fnal response rate (%) f unts less than ths score are not receved after prelmnary estmaton All It s nterestng to note that the dstrbuton of scores across outstandng unts s not even, and ncreases wth the score value. Ths ndcates that non-response s not random. Hstorcal analyss of data from the SEE shows that large unts, whch contrbute sgnfcantly to estmates, tend to be later respondents Assessment of the mpact on estmates The fnal SEE estmaton fle for 2/99 contaned responses from approxmately 99% of unts. All unts, whch faled to respond pror to fnal estmaton, have been excluded from ths analyss, and were also excluded from table 1 above. Consequently the 100% n table 1 actually refers to all unts whch had responded by the fnal estmaton close off tme. The stated purpose of the prorty IFU technque was to choose a cutoff Z value for whch the beneft n obtanng a response through performng IFU for unts below ths score s margnal n relaton to the expected mpact on the survey estmates relatve to the samplng error of the estmate. Wth knowledge of each unts fnal reported values, 479

8 we can assess the mpact on estmates of not havng performed IFU for unts wth a Z below a range of cut-off values, by re-runnng fnal estmaton wth mputes substtuted for unts whose Z value was below the cutoff level. Ths of course assumes that the unt would not have responded f IFU had not been performed. The graphs below present these results for both TE and GQE at the State and Australan levels. Graph 1: Percentage dfference from actual TE estmate for each maxmum score cutoff Percentage dfference from actual estmate Percentage dfference by maxmum score cutoff AUS NSW VIC QLD SA WA TAS NT ACT Maxmum cutoff score Graph 2: Percentage dfference from actual GQE estmate for each maxmum score cutoff Percentage dfference by maxmum score cutoff Percentage dfference from actual estmate AUS NSW VIC QLD SA WA TAS NT ACT Maxmum cutoff score The graphs show that the estmates converge to the actual fnal estmate qute rapdly after responses have been obtaned for the hghest scorng unts, wth the estmate of GQE showng somewhat more varablty than the 480

9 estmate of TE. A logcal choce for the cutoff score based on the above graphs would be a value of 7, whch equates to a survey response rate of 93.2%. Table 2 shows the expected devance from the fnal estmate ths would cause compared to the survey relatve standard errors for the State level estmates. Table 2: Impact on estmates of usng a cutoff score of seven compared to survey standard errors Varable Total Employment Gross Quarterly Earnngs State % absolute devance from fnal estmate Relatve standard error of estmate (%) % absolute devance from fnal estmate Relatve standard error of estmate (%) New South Wales 0.07 % 2.2 % 0.17 % 2.4 % Vctora 0.01 % 2.2 % 0.13 % 2.7 % Queensland 0.04 % 2.5 % 0.13 % 2.7 % South Australa 0.04 % 2.7 % 0.12 % 2.9 % Western Australa 0.04 % 3.4 % 0.04 % 2.7 % Tasmana 0.00 % 3.4 % 0.43 % 2.8 % Northern Terrtory 0.14 % 4.2 % 0.20 % 5.8 % Aust Capt Terrtory 0.07 % 3.1 % 0.23 % 3.7 % Australa 0.02 % 1.1 % 0.13 % 1.3% The table shows that the mpact of not obtanng a response for unts wth a Z value of less than 7 has had neglgble mpact on the survey estmates relatve to the survey's standard errors, and as such any nferences beng made on the survey data would not have been affected had ths prorty IFU polcy been employed for the 2/99 quarter. Testng of ths prorty IFU strategy on more than one quarters worth of data s necessary to ensure that the choce of an approprate cutoff score s robust and can expect to yeld consstent results over tme. To assess ths, data from the SEE for the thrd quarter of 1999 was also analysed n the same manner as descrbed n ths paper and yelded very smlar results Restrctons of ths analyss and related ssues Scorng for IFU at an earler stage of the processng cycle The method of scorng descrbed n ths paper reles completely on hstorcal data thus a unts Z value s known pror to the form beng dspatched. Ths s advantageous, as t s therefore possble to prortse IFU at a very early stage n the survey cycle. For the purpose of ths paper the model was appled to the SEE to ascertan whether total IFU costs could be cut wthout notceably affectng the qualty of the survey estmates. By retrospectvely decdng whch unts would contnue to receve IFU after prelmnary estmaton occurred, t appears we could have drastcally cut the large IFU costs assocated wth obtanng a response from the fnal 10% of outstandng unts (whch as stated prevously equates to approxmately half of the total IFU cost for the survey), wthout notceably affectng the qualty of the estmates produced. In applyng ths framework n general, dfferent survey areas may wsh to cease IFU acton for unts wth a score below the cutoff much earler n the survey cycle dependng on ther objectves for applyng the framework. The ultmate am s to ensure the most strategc use of the survey resources avalable for the IFU process Smplfcaton of the framework by use of assumptons Several assumptons have been made n ths paper to smplfy the mathematcal framework developed for ease of practcal applcaton. The results of the case study ndcate that the smplfed model should work well n practce, however further gans could be possble by revstng some of the smplfcatons. For example the role of the [ pˆ { C} pˆ { NC} ] n (8) suggests we should collect and analyse ndvdual busnesses response hstores. Effectvely estmatng these quanttes should lead to more effectve applcaton of the framework. 481

10 Scorng at hgher aggregate levels or controllng estmate qualty at a lower level In ths example we appled the scorng functon at the State by broad ndustry level but decded on the cutoff score based on the mpact on estmates at the State level. Decdng on a cutoff score from consderng the potental mpact on all state by broad ndustry estmates would be cumbersome, and the fact that the actual scores were determned at ths level s enough to exert some control on the mpact of the process on estmates at ths level. If we were not nterested n exertng some control over the mpact on State by ndustry estmates by performng prorty IFU we would have appled the scorng functon at a hgher level e.g. just at the state level. Ths may have resulted n a hgher cutoff score beng chosen (.e. requrng less IFU to be performed) to acheve the same expected level of accuracy as the scorng functon would be drectly amed at ths level Collecton strategy for new unts to the survey and long term non respondents In ths example we gave all new unts to the survey and those whch had not responded for the prevous two quarters a default prorty IFU score of 10. For applcaton of the prorty IFU framework to be successful t s mportant that specal attenton s pad to these two groups of unts. New unts to the survey should undergo some form of nducton process amed at maxmsng the lkelhood of response pror to IFU beng undertaken (ths s the practce used n the SEE). Unts, whch have not responded for the last two quarters, should be dentfed pror to dspatch and be gven specal attenton to ensure ther early response. Ths may nvolve havng to negotate wth the provder as there may well be good reasons why they have not responded n the past (e.g. lack of data avalablty, not understandng the survey requrements etc). If specfc procedures to address long term non respondents are not developed there s a danger that the number of unts wth default scores wll ncrease over tme and thus the savngs n processng costs from applyng the prorty IFU framework would be eroded. 6. CONCLUSION Ths paper has developed a mathematcal framework to mnmse the non-response bas n survey estmates from performng IFU for a fxed number of ntal non-respondents to a sample survey. In consderng the ssues nvolved n the practcal applcaton of ths framework, smplfcatons to the mathematcal model based on a number of assumptons and approxmatons have been requred. Testng of ths prorty IFU model n the Australan Survey of Employment and Earnngs has shown that substantal reductons n survey IFU costs are lkely to be possble wthout notceably affectng the qualty of estmates produced from the survey. Consequently the applcaton of ths technque to establshment surveys n general s lkely to enable effcences n the survey process to be realsed. References Carlton, S (1998). A framework for prortsng IFU. Unpublshed report, Methodology Dvson, ABS. Granqust, L. and Kovar J.G (1997). Edtng of Survey Data: How Much s Enough? n Survey Measurement and Process Qualty. New York. Wley, Latouche, M. and Berthelot, J.M (1992). Use of a Score Functon to Prortze and Lmt Recontacts n Edtng Busness Surveys. Journal of Offcal Statstcs, 8, Lawrence, D. and McDavtt, C. (1994). Sgnfcance edtng n the Australan Survey of Average Weekly Earnngs. Journal of Offcal Statstcs, 10,