Effect of recall duration on reporting of household expenditures : an experimental study in Ghana

~~~~~~~8905 SOCIAL DIMENSIONS OF ADJUSTMENT IN SUB-SAHARAN AFRICA Effect of Recall Duration on Reporting of Household Expenditures An Experimental Study in Ghana Chris Scott and Ben Amenuvegbe  SDA Working Paper Series Editorial Board Chairman Ismail Serageldin Director, Occidental and Central Africa Department World Bank Members Ramesh Chander, Statistical Adviser, World Bank Dennis de Tray, Research Administrator, World Bank Yves Franchet, Director General, Statistical Office of the European Communities Ravi Kanbur, Editor of World Bank Economic Review and World Bank Research Observer and Senior Adviser, SDA Unit, World Bank Gabriel Kariisa, Chief Economist, African Development Bank F. J. C. Klinkenberg, Director, Directorate General for Development, Commission of European Communities Jacques Loup, Coordinator of Assistance to Developing Countries, United Nations Development Programme E. M. Morris-Hughes, Chief, Nutrition Planning, Assessment and Evaluation Service, Food and Agriculture Organization F. Stephen O'Brien, Chief Economist, Africa Region, World Bank Graham Pyatt, Professor of Economics, University of Warwick Paul P. Streeten, Director, World Development Institute, Boston University Victor E. Tokman, Director, Employment and Development Department, International Labour Office R. van der Hoeven, Senior Adviser, United Nations Children's Fund Editor Michel Noel Chief, SDA Unit World Bank Managing Editors Marco Ferroni Christiaan Grootaert Senior Economist, SDA Unit Senior Economist, SDA Unit World Bank World Bank  SOCIAL DIMENSIONS OF ADJUSTMENT IN SUB-SAHARAN AFRICA WORKING PAPER NO. 6 Surveys and Statistics Effect of Recall Duration on Reporting of Household Expenditures An Experimental Study in Ghana Chris Scott and Ben Amenuvegbe The World Bank Washington, D.C.  Copyright e 1990 The World Bank 1818 H Street, N.W. Washington, D.C. 20433, IJ.S.A. All rights reserved Manufactured in the United States of America First printing May 1990 The findings, interpretations, and conclusions expressed in this paper are entirely those of the author and should not be attributed in any manner to the World Bank, to its affiliated organizations, or to members of its Board of Executive Directors or the countries they represent. The World Bank does not guarantee the accuracy of the data included in this publication and accepts no responsibility whatsoever for any consequence of their use. In order to present the results of research with the least possible delay, the manuscript has not been edited in accordance with the procedures appropriate to formal printed texts, and the World Bank accepts no responsibility for errors. The material in this publication is copyrighted. Requests for permission to reproduce portions of it should be sent to Director, Publications Department, at the address shown in the copyright notice above. The World Bank encourages dissemination of its work and will normally give permission promptly and, when the reproduction is for noncommercial purposes, without asking a fee. Permnission to photocopy portions for classroom use is not required, though notification of such use having been made will be appreciated. The complete backlist of publications from the World Bank is shown in the annual Index of Publications, which contains an alphabetical title list (with full ordering information) and indexes of subjects, authors, and countries and regions. The latest edition is available free of charge from Publications Sales Unit, Department F, The World Bank, 1818 H Street, N.W., Washington, D.C. 20433, U.S.A., or from Publications, The World Bank, 66 avenue d'1ena, 75116 Paris, France. ISSN 1014-739X Christopher Scott is a private statistical consultant based in London, United Kingdom, and Ben Amenuvegbe is a statistical advisor for the 'Vitamin A Supplementation Trial in Northern Ghana. Library of Congress Cataloging-in-Publication Data Scott, Christopher, 1927- Effect of recall duration on reporting of household expenditures: an experimental study in Ghana / Chris Scott and Ben Amenuvegbe. p. cm. - (Social dimensions of adjustmnent in Sub-Saharan Africa, ISSN 1014-739X) Includes bibliographical references. ISBN 0-8213-1572-2 1. Cost and standard of living-Ghana-Statistical methods. 2. Household surveys-Ghana-Statistical methods. I. Amenuvegbe, Ben. II. Title. III. Series. HD7067.8.S37 1990 339.4'7'09667-dc2O 90-12403 CIP  SDA Working Paper Series Foreword Integration of social and poverty concerns in the struc- economic crisis in Africa on the one hand and the tural adjustment process in Sub-Saharan Africa is a adjustment response on the other hand affect the liv- major driving force behind the design of the World ing conditions of people. Empirically, major improve- Bank's adjustment lending program in the Region. To ments are needed in our knowledge of the social further the goal, the Social Dimensions of Adjustment dimensions of life in Africa, how they change, and (SDA) Project was launched in 1987, with the United whether all groups in society participate effectively in Nations Development Programme and the African the process of economic development. Gaining this Development Bank as partners. Since then many other knowledge will demand new efforts in data collection multilateral and bilateral agencies have supported the and policy oriented analysis of these data. Most im- project financially as well as with advice. The task portantly, policy actions are needed in the short term presents a formidable challenge because of the sever- to absorb undesirable side-shocks stemming from the ity of economic and social constraints in Africa and the adjustment process so that the poor and disadvan- intrinsic difficulty of tracing the links between eco- taged are not unduly hurt, and in the long term to nomic policies and social conditions and poverty. It is ensure that these groups fully participate in the newly essential to have a continuous professional dialogue generated growth. The SDA Project's mandate is to between all concerned parties, so that the best ideas operate, in a concerted way, in all three domains: get discussed by the best minds, and become, as concepts, data, actions. This working paper series will quickly as possible, available for implementation by report progress and experience in all three areas. I policymakers. This is the aim of the SDA working encourage every reader's active participation in the paper series. series and the work it reports on. It is meant to be a To fulfill its mission, the SDA Project operates on forum not only for exchange of ideas but even more different levels. Conceptually, contributions need to importantly to advance the cause of sustainable and be made which advance our understanding of how the equitable growth in Africa. Edward V.K. Jaycox Vice President, Africa Region 111   The Social Dimensions of Adjustment (SDA) Project Working Paper Series The SDA Project has been launched by the U1NDP The Surveys and Statistics subseries focuses on the Regional Programme for Africa, the African Develop- data collection efforts undertaken by the SDA Project. ment Bank, and the World Bank in collaboration with As such, it will report on experiences gained and other multilateral and bilateral agencies. The objective methodological advances made in the undertaking of is to strengthen the capacity of governments in the household and community surveys in the participat- Sub-Saharan African Region to integrate social dimen- ing countries to ensure an effective cross-fertilization sions in the design of their structural adjustment pro- in the participating countries. The subseries would grams. The World Bank is the executing agency for also include "model" working documents to aid in the the Project. Since the Project was launched in July implementation of surveys, such as manuals for inter- 1987, 30 countries have formally requested to partici- viewers, supervisors, data processors, and the like, as pate in the Project. well as guidelines for the production of statistical The Project aims to respond to the dual concern in abstracts and reports. countries for immediate action and for long-term in- The Policy Analysis subseries will report on the stitutional development. In particular, priority action analytical studies undertaken on the basis of both programs are being implemented in parallel with ef- existing and newly collected data, on topics such as forts to strengthen the capacity of participating gov- poverty, the labor market, health, education, nutrition ernments (a) to develop and maintain statistical data and food security, the position of women, and other bases on the social dimensions of adjustment, (b) to issues that are relevant for assessing the social dimen- carry out policy studies on the social dimensions of sions of adjustment. The subseries will also contain adjustment, and (c) to design and follow up social papers that develop analytical methodologies suitable policies and poverty alleviation programs and pro- for use in African countries. jects in conjunction with future structural adjustment Another subseries, Program Design and Implemen- operations. tation, will report on the development of the concep- The working paper series "Social Dimensions of tual framework and the policy agenda for the project. Adjustment in Sub-Saharan Africa" aims to dissemi- It will contain papers on issues pertaining to policy nate in a quick and informal way the results and actions designed and undertaken in the context of the findings from the Project to policymakers in the coun- SDA Project in order to integrate the social dimensions tries and the international academic community of into structural adjustment programs. This includes economists, statisticians, and planners, as well as the the priority action programs implemented in partici- staff of the international agencies and donors associ- pating countries, as well as medium- and long-term ated with the Project. In the light of the three terrains poverty alleviation programs and efforts to integrate of action of the Project, the working paper series con- disadvantaged groups into the growth process. The sists of three subseries dealing with (a) surveys and focus will be on those design issues and experiences statistics, (b) policy analysis, and (c) program design which have a wide relevance for other countries as and implementation. well, such as issues of cost-effectiveness and ability to reach target groups. v   Table of Contents Executive Summary 1 1. Introduction and Background 2 2. Literature Review 3 3. Methodology of Investigation 5 A. Study design 5 B. Field work 7 C. Data preparation and processing 8 4. Analysis of Results 9 A. Main results 9 B. Weekly and monthly cycles 10 C. Seasonal variation 11 5. Factors Affecting Recall 13 A. Timne trends within interviewers and within respondents 13 B. Alternatives to the recall lapse hypothesis 14 C. Change of respondent 14 D. Expenditure characteristics 15 E. Respondent characteristics 16 6. Conclusions 17 Appendix: Schedule of field visits 18 References 19 vii   Executive Summary An experimental study was carried out in Ghana with also ruled out, and in principle the recall periods were a sample of 135 households each of which was inter- always bounded. In these circumstances the associa- viewed 11 times at varying time intervals. Recall of tion of recall lapse with recall duration became even expenditures on 13 frequently purchased items was clearer with an average loss of 2.9% for every day compared according to the length of the recall period. added to the recall duration. Between different expen- Clear evidence of recall loss was seen, with average diture items the extent of the decline was associated daily expenditures falling consistently in association positively with mean frequency of purchase of the with increasing length of recall period from 1 day to 1 item. week. There was little or no further fall in going to 2 The authors favor the interpretation that the 1-day weeks. For a I year recall period data were erratic. recall data are the most accurate and argue the urgent Examining recall in the 1 day to 1 week range, need for further studies to explore the limits of the weekly and monthly cyclic variations in expenditure recall effect and possible means of palliating it. were eliminated by re-weighting. Start-up bias was 1  1. Introduction and Background A crucial issue in the design of household budget This study arose from the needs of the Ghana Living surveys is the length of the recall period for reporting Standards Survey (GLSS), a multi-purpose household transactions. It is known that the longer the recall survey sponsored by the World Bank and imple- period the greater the likely recall error, but on the mented by the Ghana Statistical Service, in which a key other side a longer period makes it possible to cover a element is the estimation of detailed household expen- larger sample of transactions for a given number of ditures. The GLSS design, worked out by the LSMS interviews, and hence for a given field cost. Every unit at the World Bank, called for a 2-week recall survey designer has to face this conflict and choose a period for each household, a period chosen essentially solution believed to be optimal. The problem is partic- for logistic motives rather than in response to any ularly acute in third world surveys, where widespread evidence of recall efficiency. The World Bank's Social illiteracy seems to rule out the only alternative, the Dimensions of Adjustment Project is also currently diary method. proposing widespread use of the same design in Afri- Inmost African surveys food expenditures, at least, can countries. The question of the viability of the have been recorded by interviewers visiting the 2-week recall thus becomes an urgent issue. household either every day or every second day.' In The present authors, already associated with the Asia, recall periods of one week or one month have GLSS as consultants, prepared a minimal research been more common (UNESCAP, 1985). The World proposal to test the issue and were able to obtain a Bank's Living Standards Measurement Survey research grant from the UK Overseas Development (LSMS) project, already applied in 8 developing coun- Administration to conduct the experiment in Ghana.2 tries, uses a fortnightly recall period, together with a The design calls for an experimental survey, inde- 1-year estimate based on normative reporting. pendent of the GLSS although sponsored and sup- These wide variations reflect the almost total ab- ported by the Ghana Statistical Service. Its essential sence of evidence for developing countries on the level purpose was to investigate recall efficiency as a func- of recall error and its relation to recall duration. More- tion of the length of the recall period, for expenditures over, in balancing recall error against sampling error on common articles by African households. there is a similar gap in knowledge of the main factor affecting the latter: the regularity of consumer pur- chases in third world countries. The range of variation found in the solutions Notes adopted, from one day to one month, implies of course an approximately corresponding cost range, hence 1. In an exhaustive bibliography covering all household budget 30-fold. The practical importance of this issue is thus surveys conducted in Africa up to 1970, Scott (1971) found 146 out of 177 surveys in which daily recording was used for food expendi- so extreme, in relation to costs of third world survey tures at least. work, that the need to settle the question on an empir- 2. We wish to thank the ODA for financial support and the ical basis seems almost self-evident. It can reasonably Ghana Statistical Service for sponsoring the study and providing be argued that this is the most urgent design issue logistic support. The authors take sole responsibility for the conclu- facing third world survey workers today. sions expressed in this report. 2  2. Literature Review There is fairly extensive literature on recall lapse in bered. in addition Scott reports a recall loss of 12% of surveys. However the number of experimental studies transactions for the average of his 6 most frequently in which recall efficiency is compared for different purchased items and 38% for the 4 least frequently recall durations is not very large. Obviously the nature purchased. In a similar experiment in Botswana, Scott of the information to be recalled is a crucial factor. (1971) reports a loss of 71% in going from daily to Confining attention to household expenditure/in- weekly recall; however the implementation of this come/consumption, we have identified 15 studies in experiment was acknowledged to be unsatisfactory which recall periods of different durations have been and it can be argued that the results should be set compared experimentally, none of them recent.3 Most, aside. Finally, Plewis (1972), reporting an experiment though not all, of these report substantial memory loss in Malawi, found no difference between daily and with increasing duration, in some cases as much as weekly reporting. However, in this experiment the 50% or more. A striking finding reported by Politz weekly reporting period covered the same days as the (1958) is worth citing as an example of the obstinacy 7 daily reporting periods for the same respondent, so that of the recall problem: Politz arranged for interviewers the two types of response are not independent; in to observe shoppers in a supermarket and record their addition the sample was excessively small, with an purchases, then follow them home for an immediate 88% relative standard error for the estimate for daily interview. The number of items reported as purchased reporting. fell below the observation by 9% and the expenditure In the studies by Ghosh and Mahalanobis & Sen an fell below the observation by 6%. attempt was made to check the errors against objective Among the studies mentioned above, just 6 relate to methods. Ghosh placed an observer in the local shops household expenditure/consumption/income in de- who had to record all purchases for a year. veloping countries. Ghosh (1953) examined frequency Mahalanobis & Sen tested recall of consumption of of purchase for 9 common expenditure items and certain selected foods for 1-week and 1-month recall found a steadily falling average as the recall period and checked responses by weighing the household's went from I day to 1 week to I year. Mahalanobis & stocks before and after. When this was found imprac- Sen (1954) compared reported expenditure for a tical because stocks were too large, the household was weekly and a monthly recall period. The latter gave asked to set aside a much smaller stock from which all consistently lower results, with a greater discrepancy consumption was to be taken. Ghosh's observational for items purchased less frequently (up to 30% for data were consistently lower, sometimes much lower, sugar). In the Ghana study (Ghana, 1961) daily record- even than the 1-year recall reports. But considering the ing was compared with weekly, the latter yielding way in which they were obtained, it is difficult to have 20% less by value. Dutta Roy & Mabey (1968), also in confidence in them as a reliable criterion. Mahalanobis Ghana but using a more formal experimental design, and Sen found their check data agreeing on the whole found a 30% fall in expenditure in going from daily to more closely with the monthly (lower) figures than the weekly reporting for 4 commonly purchased food weekly but there are anomalies in the figures and one items. Scott (1971, 1972), using this same survey and can question whether households always obeyed the comparing daily with weekly recall but analyzing fre- instruction to take their consumption only from the quency of purchase rather than amount of expendi- special stock. These writers, as well as (apparently) ture, found a decline of 15% for the same 4 items and Zarkovich, appear to favor long recall periods and a similar figure for the pooled group of 10 items ana- attribute the apparent recall decline to over-reporting lyzed. Note that this discrepancy between the effects on short periods. This may account for the rather for frequency and amount implies that heavier expen- consistent use of longer periods in Asian surveys, ditures (within a given item) were less well remem- where daily interviewing is practically unknown. 3  4 However the evidence for this interpretation of the Note decline is decidedly weak. The Ghana 1968 study and the Botswana studywere 3. Cole & Utting (1956), Dutta Roy and Mabey (1968), Ghana poorly supervised and implemented and the Malawi Central Bureau of Statistics (1961), Ghosh (1953), Goldberg (1957), Lamale(1959),Lewis (1948),Mahalanobis&Sen (1954),Metz (1956), study is inconclusive. The need for a more rigorous Neter & Waksberg (1964), Olson (1980), Plewis (1972), Scott (1971 African study was clear. and 1972), Sudman & Bradburn (1973), Zarkovich (1966).  3. Methodology of Investigation A. Study design to be implemented. Before detailing the schedule of interviews, we need to give an overview of the exper- In most survey experiments comparing different imental design. methods of questioning, the different treatments have The backbone of the study is a series of 9 interviews been applied to distinct samples of respondents. In the with each household spaced at intervals of 1 day, 2 case of household expenditures this is an inefficient days, 3 days and 7 days in such a way that the different design because the recall effect has to be detected interval lengths were presented, twice each, in ap- against the background of between-household variance, proximately random order. For example, a given almost inevitably a high level of variation. household might be allocated interviews on days 1, 4, The Dutta Roy-Mabey experiment in Ghana was 11, 12,14,17,24,25,27. The intervals here are, in days: based on a proposal by one of the present authors for 3,7,1,2,3,7,1, 2. (The second half of the series always a design which does not seem to have been used before repeated the order of the first half.) At each interview in research on recall. Scott suggested that all the recall except the first, the respondent was asked about ex- period lengths be applied to each household: a succes- penditures "since my last visit". The first interview sion of interviews should be conducted at varying was introduced mainly in order to provide a starting intervals with the same household, with recall always point for the first interval, hence ensuring bounded based on the period "since my last visit", and with the recall. This series of 9 interviews was called the X-in- different intervals presented in a different, and ran- terviews. domized, order to different households. The starting An additional objective was to test the various recall dates should also be staggered across days of the week approaches used in the World Bank's LSMS/SDA and days of the month to eliminate any bias from surveys, namely: a 2-week bounded recall question, a cyclic variations in household expenditure patterns. 1-year unbounded recall question, and a synthetic Later calculations based on the results of the experi- 1-year estimate based on three normative questions. ment suggested that, for comparable sensitivity in Before presenting these questions, we introduce an detecting the main recall effect, a design using inde- important distinction. pendent treatment groups would have required about Everyday experience suggests that when we are 10 times as many households in the sample. A further asked to recall repeatable events over a short period advantage of the proposed design was that, since the we attempt to answer factually, but if the period be- recall periods ran consecutively for each household, comes too long we switch to normative reporting, that all recall periods were bounded: that is, the start of each is, we report our usual experience. A distinction can period was signalled by a previous interview and the thus be made between factual and normative response. end by the current interview. There is clear evidence A similar distinction can be made between questions from several studies (for example, Neter & Waksberg, asked in a factual or normative mode-though it is 1965; Plewis, 1972) that unbounded recall can lead to clear that a factual question may lead to a normative serious telescoping error (mis-dating), with conse- response. Thus one may distinguish between ostensi- quent over-reporting. bly factual (OF) questions and explicitly normative (EN) This design, already tried out in the Ghana 1968 and questions, independent of the nature of the response Botswana 1967-68 experiments (but in the Ghana case actually obtained.4 without the staggered starting dates), was adopted for The World Bank sponsored questionnaire uses two the present study. In order to meet the above con- OF questions, covering 2 weeks and 12 months, and a straints and achieve a constant interviewer workload set of three EN questions which together lead to a each day, a very tightly controlled interview plan had 12-month normative estimate. They are as follows: 5  6 2-week OF question to avoid contamination: we did not want the respon- dent, or the interviewer, to use the factual responses The World Bank's recommended question reads: as a source of norms on which to base normative " mhs.since long-term estimates.5 For this reason it was decided to my last visit?" place the long-period (12 month and 2 week) questions before the short period ones. A smaller risk, but still For ease of interviewing, this is preceded by the filter one to be avoided, is that the normative data obtained question: in the long-period estimate mnight influence the factual visit have the members of responses. For example, if the interviewer is having "Since my last youtrhave the four difficulty obtaining a response for a 7-day period (per- household spent mnything on the following haps the person who made the purchases is absent and information is coming from another household mem- These two questions were transferred almost exactly ber) he might refer to previous responses given to the into our experiment. two normative questions: "How often do you buy?" and "How much do you usually spend each time?" 12-month OF question and work out a suitable response for 1 week. To pre- vent this, the best solution seemed to be to have the The World Bank's recommended 12-month OF normative questions asked by a different person and question reads: the answers recorded on a questionnaire not made "How much did the household members spend available to the short-term factual questioner. "Hon mucinthe path12 montehs?" These considerations led to the following design: on ..mi the past 12 months?" This question follows the two mentioned above in the The 1st visit to each household is made by the World Bank questionnaire. For our experiment we supervisor, who explains the survey, obtains the World Bank questionnaire. Fo or grexterimenaw household's firm commitment to cooperate, es- mistodif this question slv r tablishes the list of household members, records some background information about the house- "I want you to look back 1 year to-day. From that hold, and asks the 12-month OF question for each day to this day, how many cedis have you and item. This was called the S1 interview. your household spent on ...?" TThe 2nd visit is made by the supervisor again, 2 It should be mentioned that in the World Bank ques- weeks after the 1st. This time he asks the filter tionnaire this form of question was used only for question, the 2-week OF question and the three purchases of clothing and domestic supplies. For all EN questions. This is the S2 interview. other items, including all of those chosen for our ex- The 3rd visit comes I day later. This is the first of periment, only the EN question set below was used for the X-interviews. The supervisor accompanies the 1 -year response. the interviewer and introduces him to the house- EN questions hold. The interviewer then asks about purchases since the supervisor's visit of the day before. The World Bank's recommended set of three nor- These data are not used in the main analysis, mathve Wons rec n e t though they are of some interest in investigating mative questions reads: satu is start-up bias. "During which month-is in the past 12 months did The 4th to 11th visits. Here the interviewer carries the members of your household buy...?" [WRITE out the planned schedule of visits which involve each month cited. CODE number of months.] two of each of the intervals: 1 day, 2 days, 3 days, "During these months, how often did they usu- 7 days. At each visit he asks about purchases ally buy ... ?" [WRITE number of times and time "since my last visit". unit. CODE number of times per month.] Thus the household is visited 11 times in all, twice "How much did they usually spend each time?" by the supervisor and 9 times by the interviewer. The These questions were used without modification in total duration covered by the factual bounded recall our experiment. The annual synthetic estimate is ob- questions in each household comes to 40 days. tained by multiplying the three responses as coded. Calculations based on the similar Ghana experi- As the experiment called for both factual and nor- ment of 1968 suggested that a sample of between 100 mative responses for the same items, it was important and 150 households would suffice to yield significant evidence of a recall trend in the 1-7 day experiment  7 for the most frequently purchased items. It is clear, batch after an interval of 1, 2, 3 or 7 days. The batches however, that the method would become much less for one interviewer were started off one after another efficient for rarer purchases. Moreover there was a at 1-day intervals and the 3 interviewers of each team need to limit the length of interviews both to reduce started one after another at 3-day intervals. Thus the the burden on respondents and to keep down overall starting points for the 9 batches allocated to one team costs. For these reasons the study was deliberately were staggered day by day over a 9-day period. How- limited to the 13 most frequently purchased items, ever, all 4 teams started this sequence on the same based on GLSS results. Only two of these were non- day.6 food items. Table 1 shows the frequency of purchase The visits for any one interviewer were interleaved and mean daily expenditure per household in our in such a way that the interviewer always had 4 house- sample, for each of the selected items, based on the holds (1 batch) to interview each day over the whole X-interviews. The items are listed in the order of their 27 days of his program, with the exception of two rest appearance in the questionnaires. Together these 13 days. A specimen schedule is shown in the Appendix. items account for about one third of the total annual The success of the experiment depended on very cash expenditures by households in Ghana (estimate strict adherence to the time table. Interviewers were from Ghana Living Standards Survey). offered a 100% salary bonus payable at the end of the project if they completed all interviews on schedule. B. Field work Absence on compassionate or health grounds was allowable only on condition that enough advance A sample of 144 households was identified and allo- warning was given to the supervisor to allow him to cated among 4 field teams, each consisting of 1 super- arrange for substitution by a reserve interviewer. Four visor and 3 interviewers. The 12 households for each interviewers and one supervisor were trained and interviewer were divided into 3 batches of 4 house- held in reserve for this purpose. Interviewers were holds. Each batch represented a day's interviewing by also required to visit any given household at the same one interviewer, who would then return to the same time of day on every visit, to the nearest hour. How- ever, provision for failure to achieve this was made by requiring the interviewer to enter the time and date on Table 1. Mean Frequency and Value of Daily each record page and allowing a space for separate Expenditure per Household recording of any purchases occurring in the over-run Frequency of Mean expenditure period, that is, that part (if any) of the actual interval purchase per day per day per which fell later than the planned interval by more than Item per household household (cedis) one hour. Amounts so recorded were transferred to (1) (2) (3) the proper period prior to data processing. Finally, the Cassava 0.456 51 interviewer's presence in the household at the right Beans 0.237 20 time and on the right day was checked through a Garden egg 0.463 16 clocking-in sheet kept always by the household. All Plantains 0.470 31 field workers had to enter on this sheet the time and Tomatoes 0.848 32 date at every visit to the household and this could be Kieshkey 0.9383 17863 checked by the supervisor. Supervisors were required Palm oil 0.370 26 to be present in their area every day and at least one Bread 0.656 53 senior officer visited each area every day. Soap 0.493 34 These very strict control measures were rather suc- Charcoal/firewood 0.467 33 cessful: only one interviewer-day was missed through Onion 0.733 20 field worker error (out of 324). The 4 households af- Maize 0.170 71 fected were dropped from the sample, as were 5 others Column (2): Proportion of non-zero responses among the which went absent or failed to cooperate. In all, 135 1-day recall responses. This measure does not take households provided the complete required record, a account of the possibility of two purchases of the same response rate of 93.7%. Over-run intervals were re- item on the same day by one household. corded in 4.7% of the sample interviews, but in more Column (3): Daily mean expenditure per respondent than half of these (2.4% of the sample) the over-rn household over the 26 days covered by the 2nd-9th X-interviews. No adjustments made for possible recall was 2 hours or less. error. The research method is essentially model based and Garden egg: Eggplant no attempt was made to provide a representative sam- Kenkey: Crushed maize paste ple of the population. Two areas were selected purpo- Maize: Covers maize in grain form only sively, an outer suburb of Accra and a rural area some  8 20 km from Accra. Two teams were allocated to each The next step was to transfer all transactions re- of these; two separate blocks were chosen in the for- ported in over-run periods into the interval where mer and two neighboring villages in the latter. The 36 they would have fallen had the interview taken place households required for a given team and location at the scheduled time. This procedure corrects the were selected in advance by the supervisor using a analytic basis of the data (an interval classified as d determinate procedure designed to spread the sample days in length maintains an expected total expendi- through the area and to eliminate subjective choice. ture proportional to d) at the expense of a slight dis- Selected households were numbered with a chalk tortion in the recall analysis (a few expenditures are mark on the wall, or census numbers were used where classified with the wrong recall duration). they survived. All dates of visits were pre-entered on Some of the questions required coding. This was the appropriate questionnaires with the selected carried out, followed by data entry. household number at the end of the training session. Preliminary visual examination of the data listings Each interviewer was also provided with a complete revealed two anomalous observations. Two house- daily schedule of visits and each supervisor was given holds had reported a bulk purchase of maize from 50 the complete schedule for all of his team. to 100 times the daily mean. One fell in a 1-day recall Since the X questionnaire consisted essentially of interval, the other in a 2-day interval. Their inclusion only one question (repeated for 13 items), a single would dominate the recall effect for maize and distort day's training was sufficient for interviewers. Empha- that for the all-items total. They have been replaced by sis was placed on the need for discipline in maintain- zero (the modal value) in all the main analyses (they ing the planned schedule. Supervisors were trained are retained only in Table 1 and in the evaluation of over a 2-day period, again with primary emphasis on regularity of purchase of maize). No other observa- rigid adherence to the time table. tions come anywhere near this degree of eccentricity Interviewers were locally recruited teachers. Super- and no other such adjustments have been made. visors were regular staff of the Ghana Statistical Ser- Computer work was undertaken on the main-frame vice. A day's work for an interviewer would be about installation of the Regional Institute for Population one hour; on this basis the interviewers' pay could be Studies at the University of Ghana. considered lavish. These conditions, together with the very close supervision and the brevity of the inter- view, must be considered exceptionally favorable to Notes the production of high quality data. It would be rea- sonable to expect a lower error rate than in a full 4. The word "factual" is used above to refer to the semantic survey. intention, not to the accuracy of the response. 5. There is evidence from an analysis of the LSMS survey con- C. Data preparation and processing ducted in Cte d'Ivoire that such contamination occurs in the LSMS interview. (Grootaert & Scott, 1986.) 6. The inadequate randomization over days of the week noted The first step was the elimination of 9 households in Section 8 eow could have been overcome most easily had this which did not supply the complete data set. This left latter feature been modified, for example by staggering the teams' 135 households for analysis. starting dates at 1-day intervals.  4. Analysis of Results A. Main results validity of this assumption in later sections. Mean- while we remind the reader that each of the 135 re- Analysis of the main recall effect is based on the simple sponding households contributes the same amount of principle that, for any expenditure item, in the absence data, namely 2 values for each of the recall periods 1 of recall error average daily expenditure should be day, 2 days, 3 days and 7 days, pIus 1 value for each independent of the length of the recall period-or of the other estimates: 14 days OF, 12 months OF and equivalently, expenditure over a recall period of 12 months EN-composite estimate. One of the 13 ex- length d should be proportional to d. Thus, recall error penditure items, plantain, was omitted from the ques- will be indicated by the observation of a trend in tionnaire S2 through a printing error. This affects some average daily expenditure with increasing d. of the comparisons possible at the all-items level. Common experience leads us to expect that, where Table 2 shows the basic findings. recall error occurs, this trend will be downwards, and For a variety of reasons these results should not be the error will be least for smallest d. However it is not taken at their face value. In the following sections we our intention to take the latter expectation for granted discuss these reasons and present adjusted results. and the question of other interpretations will be dis- Before doing so we examine the statistical signifi- cussed in a latersection. Nevertheless, for convenience cance of some of the differences observed in Table 2. of presentation we shall take d = 1 as a baseline, After converting the raw data into daily mean ex- corresponding to a recall index 100. penditures we carried out an analysis of variance be- The principle cited above assumes that the observa- tween the 4 recall periods 1 day, 2 days, 3 days, 7 days, tions have been properly randomized over all natu- regarded as 4 treatments, with 2 replicates for each rally occurring trends or cycles. We examine the observation. (The 1st and 2nd replicates differ in over- Table 2. Mean Reported Daily Expenditure per Household for Various Recall Durations, Indexed to 1 day= 100 14 days I year I year Base 100 Item I day 2 days 3 days 7 days OF OF EN in cedis Cassava 100 120.00 116.00 107.00 91.00 145.00 95.00 50.52 Beans 100 94.00 100.00 98.00 68.00 148.00 108.00 18.96 Garden egg 100 108.00 95.00 86.00 80.00 118.00 98.00 15.52 Plantains 100 91.00 77.00 91.00 - 138.00 - 30.59 Tomatoes 100 93.00 81.00 78.00 74.00 93.00 84.00 31.67 Kenkey 100 95.00 84.00 85.00 76.00 113.00 80.00 72.96 Fish 100 92.00 81.00 85.00 81.00 101.00 95.00 183.87 Palm oil 100 74.00 86.00 76.00 81.00 98.00 75.00 26.44 Bread 100 90.00 83.00 72.00 80.00 108.00 80.00 53.44 Soap 100 139.00 115.00 104.00 96.00 147.00 95.00 33.39 Charcoal/firewood 100 103.00 80.00 85.00 98.00 115.00 98.00 33.07 Onion 100 87.00 82.00 75.00 70.00 74.00 76.00 20.24 Maize 100 108.00 108.00 96.00 83.00 123.00 112.00 42.43 Totalalitems 100 9. 8920 87.10 - 113.76 - 613.10 Total witloutplantain 100 9&58 68981 86.89 82.05 112.51 91.38 582.51 OF: Ostensibly factual answer form EN: Explicitly normative question form, composite estimate 9  10 all mean by only 0.2%.). Using the between-replicates significant at the level P < 0.005. Thus a consistent estimate of variance in the denominator we tested the day-of-week variation is present, with high expendi- between-treatments variance with an F-ratio for each tures on Wednesdays and the weekend. of the 13 items and for all items pooled. Table 3 gives The authors had expected that their rough random- the results. ization of the interviewers' work schedules together The variances within treatments differ substan- with the staggered starting dates would suffice to tially. We therefore repeated the analysis at the all- produce a reasonably even distribution of days of the item level after introducing weights inversely week between the 4 treatment categories. In the event proportional to the standard deviation within treat- however the randomization proved excessively ments, as estimated from the 2 replications. The lumpy: Table 5 shows the distribution of achieved weights were: interviews by day of week. Duration Weight The most serious distortion here is the under-repre- I 1.000 sentation of Wednesday and Saturday (two high- spending days) among the 1-day recall interviews. 2 1.242 Note that the day of the week dassification in Table 3 1.407 5 refers to the day of interview, not the day of expendi- 7 2.007 tUre. If we consider the days of expenditure which the Re-runn-ting the ANOVA we obtained an increased respondent is trying to recall, obviously all 7 days value of F: 7.882 compared with the unweighted value appear with equal frequency in the 7-day column, 7.184. while in the 1-day column they occur with widely These results leave no reasonable doubt that a recall varying frequencies. This effect vitiates comparability effect is pesen,,afectng mstfnoafteitms. between the recall results for the four durations. To effect is present, affecting most if not all of the items. u l ordrtoso neulfoigw edt The overall trend is clearly downwards, with increas- put all four durations on an equal footing we need to The verll tendis darl donwars, ith ncras- re-weight the data so that the 7 days of the week are ing recall loss at increasing recall duration, at least up to 2 weeks. B. Weekly and monthly cycles Table 4. Variation in Expenditure by Day of Week (Data from 1-day Recall Period Only) Marked variations in expenditure are found according Mean per household to day of the week. These show up most clearly, of Day Dummy interview Main survey course, in the 1-day recall data. Table 4 shows these Monday 472 571 data, separating out the 1-day recall dummy intermvew Tuesday 578 558 which preceded the main experimental series. A sim- Wednesday 815 788 ple way of testing the significance of the day-of-week Thursday 526 557 variation is to observe the correlation between the two Friday 559 528 columns of the table. This yields r - 0.886, which is Saturday 860 702 Sunday 711 697 Table 3. Significance Test of Differences Between First 4 Columns of Table 2. Variance Ratio F. Item F(3; 942) Probability Cassava 1.24 > 0.1 Beans 0.07 > 0.1 Table 5. Distribution of Achieved Interviews by Garden egg 2.66 < 0.05 Day of Week and Recall Duration Plantains 2.72 < 0.05 Recall duration Tomatoes 3.56 < 0.025 Kenkey 4.44 < 0.01 Day I day 2 days 3 days 7 days Fish '7.81 < 0.001 Monday 59 30 29 29 Paln oil 3.07 < 0.05 Tuesday 45 30 46 61 Bread 9.17 < 0.001 Wednesday 30 46 45 32 Soap 6.78 < 0.001 Thursday 47 45 28 13 Charcoal/firewood 2.17 > 0.1 Friday 31 30 45 58 Onion 4.78 < 0.005 Saturday 15 45 46 61 Maize 0.42 > 0.1 Sunday 43 44 31 16 A IV 7.18yy 0.001 All 270 270 ;;y5y;00C00000SS0C00005000-Ng0000i0000000000 t0-0 X tj0i