WPS5414 Policy Research Working Paper 5414 Explaining Variation in Child Labor Statistics Andrew Dillon Elena Bardasi Kathleen Beegle Pieter Serneels The World Bank Development Research Group Poverty and Inequality Team September 2010 Policy Research Working Paper 5414 Abstract Child labor statistics are critical for assessing the extent and proxy response versus self-reporting. Use of a short and nature of child labor activities in developing module compared with a more detailed questionnaire countries. In practice, widespread variation exists in how has a statistically significant effect, especially on child child labor is measured. Questionnaire modules vary labor force participation rates, and, to a lesser extent, on across countries and within countries over time along working hours. Proxy reports do not differ significantly several dimensions, including respondent type and the from a child's self-report. Further analysis demonstrates structure of the questionnaire. Little is known about the that survey design choices affect the coefficient estimates effect of these differences on child labor statistics. This of some determinants of child labor in a child labor paper presents the results from a randomized survey supply equation. The results suggest that low-cost experiment in Tanzania focusing on two survey aspects: changes to questionnaire design to clarify the concept of different questionnaire design to classify children work work for respondents can improve the data collected. This paper--a product of the Poverty and Inequality Team, Development Research Group--is part of a larger effort in the department to improve household survey methods. Policy Research Working Papers are also posted on the Web at http:// econ.worldbank.org. The author may be contacted at kbeegle@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Explaining Variation in Child Labor Statistics Andrew Dillon International Food Policy Research Institute Elena Bardasi The World Bank Kathleen Beegle The World Bank Pieter Serneels University of East Anglia Key words: Child Labor, Survey Design, Tanzania JEL code: J21, C81, C93 * We would like to thank Economic Development Initiatives, especially Joachim de Weerdt, the supervisory staff, enumerators and data entry teams for thorough work in the field. We also appreciate the thoughtful comments by three anonymous referees and the suggestions of the Editor, which improved the paper substantially. We thank the seminar participants at IZA Workshop on Child Labor and at the Conference "Survey Design and Measurement in Development Economics", sponsored by the World Bank, the Enterprise Initiative, and Yale University for their comments and feedback. This work was supported by the World Bank Research Support Budget and the Gender Action Plan at PREM-Gender. All views are those of the authors and do not reflect the views of The World Bank or its member countries. Corresponding author is Andrew Dillon, a.dillon@cgiar.org. 1. Introduction and background In the past decade, special attention has been paid to generating empirical evidence on child labor for developing countries. Edmonds (2009) illustrates the boom in studies on child labor and provides an overview of labor force participation rates across a large number of countries. Recognizing the importance of both the definition of child labor and its measurement, the International Labor Organization/IPEC's Statistical Information and Monitoring Programme on Child Labour (SIMPOC) has focused on establishing standardized methods for survey design to measure children's work (see ILO, 2008, and ILO, 2004). Despite these efforts, there is still substantial variation in how child labor is measured. Partly this reflects the practice of measuring child labor as part of a broader survey, a consequence of the limited capacity of statistical offices in low-income countries to field frequent stand-alone child labor surveys. In turn, there can be considerable inconsistency in statistics. Guarcello et al. (2009) carefully document the apparent inconsistency of child labor statistics from large-scale national surveys for several countries. In Ghana, for instance, a comparison between the Core Welfare Indicator Survey (CWIQ) (2003) and the SIMPOC survey (2000) shows a decline in child labor of 27 percentage points from 34 percent of children working in the SIMPOC survey. In Kenya, the Multiple Indicator Cluster Survey 2 (MICS2 2000) and SIMPOC (1998/99) surveys report an increase in child labor of 36 percentage points from 8 percent in the SIMPOC survey. Changes in child labor force participation rates over time certainly could reflect real changes ­ such as rapid economic growth. In this case, we would likely expect to see similar changes in school enrollment, which Guarcello et al. (2009) do not observe. Other explanations for such large fluctuations between two independent surveys administered in close proximity are survey design and sample design. However, Guarcello et al. (2009) find that differences in survey design (including questionnaire type and fieldwork season) explain only some of the variation in child labor estimates across surveys and that samples look otherwise comparable (including age, sex, and urban composition). A sizeable portion of the variation in child labor statistics remains unexplained. There is scant evidence on the impact of survey design on child labor statistics, in contrast with adult labor, where there is more evidence, especially 2 from the United States.1 One exception is Dillon (2010), who compares two different child labor modules within the same survey, one a standard set of labor questions that collects information on participation and hours across various activities posed to parents about their children, and another subjective game played by children that reveals the distribution of their time. Comparisons between the two modules suggest that adults report lower hours of child labor when using a standard labor module relative to children who play the subjective game, but the paper cannot disentangle the proxy effect versus the effect of question type/design. The objective of this paper is to explore further what aspects of survey design affect child labor indicators to assess these tradeoffs. We focus on two main areas: the effect of including screening questions to structure the questionnaire regarding labor market activities and the respondent type. The sequencing of employment questions is posited to have a large influence on labor statistics. This may be particularly relevant in a setting where a significant proportion of individuals are employed in household enterprises or home production and are not directly remunerated in the form of a salary or wage. Classification of activities between those that are considered "work" and those that are not may induce confusion in survey respondents who are not familiar with internationally recognized definitions of labor market activities and may have a very personal concept of "employment." For example, the stand-alone question "Did you work in the last 7 days?" is hypothesized to systematically undercount persons who work in household enterprise activities without direct wage payments, e.g., unpaid family workers or women (Anker, 1983), who may not recognize themselves (or be recognized by other household members) as "employed individuals." This type of employment question may also be flawed for measuring child labor, especially when children participate in economic activities related to the household enterprise or home production, and even more so when such activities are seasonal, occasional, or occupy only a few hours a week. This is especially the case in developing countries. Respondent type may also influence the labor statistics generated. Borgers et al. (2000) illustrate that, given the appropriate question structuring and interview 1 Bardasi et al. (2010) review some of the literature with a focus on evidence from low-income settings. 3 conditions, children older than 10 years of age have sufficient cognitive development to respond accurately to survey questions. However, in practice other household members are often asked to report on the children's activities, rather than the child him or herself. In related work on adults, we find that the effect of proxy response has a large and statistically significant effect on a number of labor statistics, like labor force participation, weekly hours worked, and daily earnings (Bardasi et al., 2010). We also find that the relationship between the proxy and the respondent, with respect to age, education, and gender, influences the estimated adult labor statistics. Focusing on children age 10-15, we assess the implications of survey methods both on average and in relation to the characteristics of the child and his/her household. We draw lessons for measuring child labor force participation, the type and intensity of child work (particularly work that occurs in household enterprises and farms), and the changes in patterns of child work over time. The setting for this work, Tanzania, influences the extent to which these findings might be applicable to other countries. Specifically, we are testing alternative survey designs in a context that we characterize, based on our field experiences, as one where there are not negative perceptions of child labor (see discussion in Bass, 2004, who draws a similar conclusion about perceptions of child labor in Sub-Saharan Africa). Thus, we are not testing whether households try to deny or hide child labor activities and whether specific questionnaire designs can circumvent this problem, but we are assuming-- quite confidently--that this problem is marginal in our setting. In this study, we focus on child labor data from household surveys. Household-based surveys are unlikely to be appropriate sources of data on the most hazardous or worst forms of child labor, which are rare in Tanzania. Such measures should ideally be collected through other methods (see ILO, 2008). Our intent is to measure the extent to which children are engaged in productive activities, which is a first step in the measurement of child labor--not testing how to measure child labor according to the ILO statistical definition. The structure of the paper is as follows. We describe the experimental design and the identification strategy to test differences in questionnaire design and respondent type 4 in the next section. Section 3 provides a description of the data collected; Section 4 presents our results. Section 5 concludes. 2. The survey experiment Whether changes in the measurement method have an effect on the statistics they produce is, ultimately, an empirical question. We designed and implemented a survey experiment in Tanzania focusing on two key dimensions of labor survey design: the inclusion of screening questions in the questionnaire with respect to identifying labor force participants and the type of respondent. In this experiment, we have two different questionnaire designs (which we call "detailed" and "short") and two respondent types (proxy and self-report). Households were randomly selected for the survey from seven districts in Tanzania; we describe the household selection process in the subsequent section. After households were selected within the village, they were randomly assigned to one of four groups defined by the combination of the two experiments, one orthogonal to the other: detailed and self-report, detailed and proxy, short and self-report, and short and proxy. Details on the sampling approach and survey assignment for households and individuals are provided in Section 4. The experiment was conducted for both adults and children--individuals eligible for data collection were all those age 10 and older. In this paper, we focus on the responses from children and the measurement of child labor. We define child labor as the labor force activity of a child ages 10 to 15 who has engaged in at least one hour of labor market activity over the past seven days. An internationally recognized definition of child labor remains an open item in the child labor policy agenda as current international agreements such as the ILO's Convention 138 agree on age limits for child labor, but leave discretion to member countries on hours and activity restrictions in defining child labor. Our definition is consistent with the ILO's Statistical Information and Monitoring Program on Child Labor (SIMPOC) definition of child labor.2 2 In another paper, we focus on the measurement of labor statistics for all the adult population (Bardasi et al., 2010). Note that the "adult labor statistics experiment" and the "child labor statistics experiment" on which we report in this paper are not separate experiments, but rather focus on two different populations in the same survey experiment. 5 For the first dimension of this survey experiment, we have a "detailed" labor module and a "short" labor module. The short module reflects the approach commonly used in short questionnaires, such as the Core Welfare Indicator Survey (CWIQ). This short module is often used to generate statistics with a high frequency, for example with annual regularity, in lieu of multi-topic household surveys that are too demanding to implement on an annual basis. In our survey experiment, the detailed module differs from the short module in two ways: the set of screening questions is longer in the detailed module and the detailed module collects information on second and third jobs. Our objective is to compare the impact of the different screening questions. The detailed module includes several screening questions about labor force participation, specifically, whether the person has worked for someone outside the household (as an employee), whether s/he has worked on the household farm, and whether s/he has worked in a non-farm household enterprise (three separate yes/no questions). These questions are asked with respect to the past 7 days as well as for the past 12 months. In the short module, there is only one question for each of the two reference periods, namely whether s/he has worked in the past 7 days (or past 12 months, respectively). From these screening questions (yes to any of the three in the detailed module; a yes to the one question in the short module), a person is identified working (employed). In the remainder of the paper, we focus on employment statistics with respect to the past 7 days of a worker's main job. Although we expect different results depending on the reference period (seasonal activities are likely to be particularly important in the measurement of child labor), we decided to focus on the past 7 days because this is the standard ILO approach in measuring employment. Moreover, our survey was carried out over a whole year, so seasonality effects should average out between survey assignments. In both the detailed and short versions, the employed are then asked their occupation, sector, employer, hours, and wage. There are too few second and third jobs in the data to analyze those data (6 children and 1 child report a second and third job, respectively). In the second dimension of the experiment, we vary the respondent to whom the questions are asked: directly to the child or to a proxy respondent. Response by proxy rather than self-report is a common practice in household surveys, with the household head often answering all questions. The ILO guidelines for child labor statistics are that these questions be answered by the child, without proxy, and in cases where 6 young children (less than 9 years) have difficulty comprehending or responding to questions, someone else in the family, usually the mother or elder sister, may assist them (ILO, 2004). Although self-reporting (for respondents of some minimum age, typically 10 years or older) is the established standard for multi-topic household surveys (Schaffner, 2000), in practice proxy respondents are often used when individuals are away from the household working or otherwise unavailable to interview in the time allotted in an enumeration area to conduct interviews. In our survey experiment, the proxy respondent is randomly chosen from among the household members who are at least 15 years old. This age threshold reflects common practice in fieldwork as it is unlikely for an enumerator to choose another child (younger than 15) to be a proxy respondent for children (or adults) in the household. The proxy respondent is thus either the head of household, spouse of the head, or an older child or relative living in the household.3 The proxy respondent then reports on two other household members who are at least 10 years old. In practice, proxy respondents are usually not randomly chosen but selected on the basis of availability and knowledge of the person for whom they will respond. Although, in this sense, the experiment does not exactly mimic the actual conditions of proxy respondents, the randomization of proxies allows us to investigate whether proxy characteristics may have an effect on the statistics generated. Tables A1 and A2 in the Annex report the key employment questions in the short and detailed questionnaires and summarize the main features of the two experiments. The full English versions of the labor modules are presented in Bardasi et al. (2010). Combining each of the above two dimensions in our experiment gives rise to a 2 x 2 randomized design that reflects commonly used approaches in practice: a detailed questionnaire with self-respondents, a detailed questionnaire with proxy respondents, a short questionnaire with self respondents, and a short questionnaire with proxy respondents. We use the results from the detailed self-report questionnaire as the benchmark reference for our analysis. This is generally considered to be the "best practice" approach of household surveys. It corresponds to ILO recommendations, which prescribe a detailed questionnaire with children self-reporting, as well as 3 The Tanzanian CWIQ 2006 data indicate that the average Tanzanian household has between two and three adults who could serve as a proxy. 7 recommendations of the World Bank. However, it is not possible to establish with our experiment that the detailed self-report questionnaire (or any other alternative for that matter) is the "gold standard," or the "best" approach to collect child labor statistics. Instead, we will be able to document variations across survey design and identify the most important dimensions along which variations occur. In each of the four designs, in addition to the labor module, the questionnaire also includes five other modules: a household roster, and sections on household assets, dwelling characteristics, land, and consumption expenditures. The questions in these sections follow the same sequence and phrasing, and refer to the same recall periods in the detailed and short questionnaires. The labor module was administered before the consumption module, but after the land module in the questionnaire. Before analyzing the child labor statistics, we first compare household and individual level variables across the assignments to ensure that characteristics are not statistically different on average. From an analytical perspective, we have organized the analysis to address two types of questions: (i) the effects of the change in survey design on child labor statistics, and (ii) whether survey design affects the relationship between child labor and the variables of interest that are typically documented in the empirical literature as being important covariates of child labor. Regardless of whether variations in child employment are found on average, it is possible that the survey design affects reporting on child labor in a non-random way with respect to those characteristics that are generally found to explain (or be correlated with) child labor. To address the first question, we follow two steps. We first estimate differences in mean child labor statistics across assignments.4 We compare the mean outcomes in children's labor force participation, occupation, daily hours worked, and weekly earnings across the four groups for the child's main job. Since the survey assignments 4 In the parlance of randomized control design, to estimate the average treatment effect, we ideally want to estimate = Yt1-Yt0 which is the difference of the outcome variable of interest at time t between two treatments denoted by the superscripts 1 and 0. However, since is unobservable to the econometrician because a household does not receive two treatments simultaneously, one estimates the treatment effect given the observable data, i.e. TE = E (Yt1 | T=1) - E (Yt0 | T=0). Since in a properly implemented randomized design, the treatment and control groups have identical characteristics on average because the groups were composed of randomly allocated households, differing only with respect to the treatment received, the selection bias, E (Yt0 | T=1) - E (Yt0 | T=0), equals zero and the estimate of the treatment effect is unbiased. 8 are randomly allocated, we abstract from unobserved heterogeneity in individual, household, or village characteristics. In a second step, we formally estimate the marginal survey design effects using the following specification: yi = + PPh + SSh + Xi + Dh + h (Eq. 1) where yi are the different labor statistics (such as labor force participation, labor supply, earnings, and occupational choice) for the ith child, Ph is an indicator variable for the proxy treatment of children in household h, Sh is an indicator variable for the short questionnaire treatment of children in household h, Xi is a vector of child and household characteristics for the ith individual, D captures district indicators, and is the stochastic error term, which is randomly distributed across households. Survey data are also used to estimate behavioral equations, for example how the age of the child and other personal and household characteristics impact the probability of the child working. We investigate whether point estimates of key covariates (vector Z) in these equations vary when different survey designs are used, focusing on four important covariates of child work, as identified by seminal papers in the child labor literature: the child's age (Edmonds, 2009), household size (Edmonds, 2005), household assets (Basu and Van, 1998), and household land size (Bhalotra and Heady, 2003). To do this, we interact the survey assignment variables with each of these variables minus its mean value in the sample (Zi ­ mean Z), while controlling for the survey assignment effects, the covariate of interest, household and individual characteristics (Xi, which includes Z variables), as well as district indicators. We estimate the following specification: yi = + PPh + PPh(Zi ­ mean Z) + SSh + PSh(Zi ­ mean Z) + Xi + Dh + h (Eq. 2) 3. The data 9 The survey experiment, the Survey of Household Welfare and Labour in Tanzania (SHWALITA), was implemented in Tanzania. The work was implemented by a well- established data collection enterprise, Economic Development Initiatives (EDI) with the capacity to undertake high-quality field studies. The survey assignments were carefully piloted in a rural and an urban area not part of the sample. A qualitative debriefing with the field supervisors took place at the end of each day during the pilot, in order to solicit their feedback on a range of issues.5 In addition, a subset of households was selected for qualitative interviews with the respondents, in order to see whether wording and structure of the questionnaire could be further improved.6 Training manuals and enumerator instructions were then revised based on these sources of feedback during the pilot. Enumerators were then trained and the survey was implemented.7 SHWALITA was purposively designed and fielded to study the implications of the alternative survey designs for employment indicators and consumption expenditure measures. Here we focus on the component that applies to employment indicators. The field work was conducted from September 2007 for 12 months in villages and urban areas from 7 districts across Tanzania: one district in the regions of Dodoma, Pwani, Dar es Salaam, Manyara, and Shinyanga region, and two districts in the Kagera region. Households were randomly drawn from a listing of all households in the village or urban enumeration area and randomly assigned to one of the four groups defined by the two experiments. The total sample is 1,344 households (with two of 5 The feedback focused on nine areas: 1. General impressions of the respondent's comprehension; 2. Question phrasing; 3. Question sequencing; 4. Completeness of lists of question responses; 5. Clarity of interviewer instructions; 6. Completeness of interviewer manual to resolve field problems encountered; 7. Questions that should be restructured for greater clarity and respondent comprehension; 8. Conceptual or cultural difficulties in translating questions to local language; 9. Areas of emphasis for training enumerators. One of the most important parts of the questionnaire to pilot was the selection of proxy and self-reporting respondents. After a day of training, interviewers spent significant time practicing with examples. 6 During this qualitative interview, respondents were asked open-ended questions to solicit how they thought about the survey questions, why they chose the responses they did, and how they thought about concepts such as work, household production, and their primary activities. 7 The enumerators were trained with the assistance of field supervisors who undertook the questionnaire pre-testing exercise. The training consisted of explaining the research objectives of the survey as well as the "sense" of each question, reinforcing the standards required for correct completion of the household questionnaire and the working relationship between enumerator and supervisor. A field experience to practice administering the questionnaire was part of the training. An interviewer manual was prepared to provide specific guidance during the training period, and to serve as a reference during the implementation. Throughout the training, special emphasis was put on standardization of the manner in which questions are posed and the correct selection of proxy and self- reporting respondents using a random number list. 10 these households being replacement households for refusals to participate), with 336 households assigned to each of the four labor modules. Although the sample of 1,344 is not designed to be nationally representative of Tanzania, the districts were selected to capture variation in Tanzania--both urban/rural as well as along other socio- economic dimensions. The basic characteristics of the SHWALITA households generally match nationally representative data from the Household Budget Survey (2006/07) (results not presented here). Households were interviewed over 12 months, but because of small samples we do not explore the variations across main seasons (such as the harvest season with peak labor demand and the dry seasons with low demand). After the households were randomly selected and randomly assigned to one of the four assignment groups, respondents and proxies were selected according to the following rules. In households assigned to self-report, up to two individuals ages 10 and older were randomly selected to self-report. In households assigned to proxy report, one household member over 15 years was first selected to proxy report; in a second stage, up to two household members age 10 and older (after excluding the individual chosen to be a proxy respondent) were selected to be reported on by the proxy. The proxy also reported for him/herself, and was considered a self-report in this case. Random selection was conducted by first listing eligible individuals (either proxies or self-reports). Then the enumerator examined a random number table pre- printed in the questionnaire that had random numbers generated and listed in columns that corresponded to the potential total number of eligible individuals that could be listed. Each of these tables was generated uniquely for each questionnaire and for each set of listing exercises (either proxies or self-reports) that were required of the enumerator. Because eligible respondents were all those age 10 and older, the sample selected for our analysis included both adults and children. In this paper, we limit our sample to the sub-group of children. Of the total sample of 1,344 households, 494 had at least one child age 10-15 years, resulting in a sample of 566 children. We focus on the subset of households in which these children reside. Some main characteristics of these households are presented in Table 1 by survey assignment. To verify the random nature of the assignment of households to one of the four survey types, we test 11 whether the different household characteristics differ across the four assignment groups. We do this by regressing each characteristic on three indicator variables that reflect the survey assignment (the fourth group being the base category) and test for joint significance of the coefficients using an F test. For most household characteristics, the difference is insignificant, reflecting the random assignment during the field work. Only in three cases do we observe a significant difference between groups, indicating that households assigned to the detailed self-report and short proxy surveys turn out to be slightly larger and own slightly more land than the other two groups. Turning to the 566 children, we classify them on the basis of the survey assignment they receive. This is the combination of the module assigned to the household and sub-household assignment of the child. Children were randomly selected from among all members age 10 and older to self-report or be reported on by proxy. In three cases, children selected to self-report were unavailable and their labor information was collected by proxy respondent. Omitting these children, rather than reclassifying them to their actual assignment as we have done in our analysis, has no effect on the results presented below. To test the random nature of the assignment, we follow the same approach as for households. The results, reported in Table 2, show that we find no difference across the four assignments, except for household size, with the households assigned to the detailed self-report and short proxy surveys slightly larger. Consistent with the design of the survey experiment, there are more households with proxy reported children (Table 1 columns 2 and 4) and more individual children who are proxy reported (Table 2 columns 2 and 4). This is because proxy respondents can only be adults age 15 years and older. Thus, in households selected to the proxy assignment, children age 10-14 have a higher probability of being selected to be proxy subjects than to be self- reports, compared with children in households selected to the self-report assignment. 4. Results We present the results in three parts. In the first part, we examine differences across the children for three key statistics: their labor force participation (LFP) rate, their 12 weekly hours of labor supply, and their main activity in their main job. We also consider time-use statistics focusing on two household chores that are often carried out by children: the collection of firewood and water.8 These time-use questions are identical across the four survey designs and are asked to all children regardless of their employment status, with the only survey design variation arising from the respondent type (self-report or proxy). Throughout we focus on a comparison between the results generated by the short and detailed modules and a comparison between those generated by the proxy and self-reported modules. In the second part, we estimate the average effects of survey type for each of these statistics using standard analysis (probit, OLS, and multinomial logit) where LFP, weekly hours, and main activity are in turn left-hand-side variables, and the survey assignment as well as household and individual characteristics and district effects are right-hand-side variables, as set out in Equation 1. In the third part, we estimate Equation 2 to investigate whether the effects on child labor of the personal and household characteristics commonly analyzed in the child labor literature are sensitive to changes in the survey type. Differences in labor indicators across survey type Table 3 presents differences in LFP, working hours, and time spent on firewood and water collection by questionnaire and respondent type, disaggregated by gender. In each case, we test for a difference in means across survey type groups using a t-test. Row 1 of Table 3, for instance, first reports the mean LFP of boys obtained from the short module (55.4 percent) and compares this with the mean LFP for boys obtained from the detailed module (70.9 percent) and tests whether the difference (-15.4 percentage points) is statistically different from zero. Following the conventional definition, domestic activities (cleaning house and cooking) are not considered economic activities and are not included in LFP. Note that when comparing means 8 Collecting firewood and water are activities that are included in the System of National Accounts definition of economic activities and should in principle be defined as "work," although in practice they are routinely excluded. 13 between two survey designs (e.g., detailed vs. short questionnaire), we are "pooling" statistics with respect to the other experiment (e.g., self-report vs. proxy).9 We find that there are significant differences in reported LFP for boys and girls when using the short module compared with the detailed module. LFP with the short module is 15 percentage points lower for both groups. The difference between the proxy and self-reported statistic, however, is not statistically significant for either boys or girls (the difference is -2.3 and -3.5 percentage points, respectively). One reason why we may observe large differences in LFP between the short and detailed modules is the under-reporting of marginal jobs (i.e., jobs that are especially short, in terms of weekly hours) in the short module. If this is the case, we expect to observe longer average weekly hours conditional on working for the short than the detailed questionnaire, while average weekly hours for the whole sample (i.e., including the zeros) may not differ substantially between the two experiments. This is exactly what we observe, which suggests that when using the short questionnaire, marginal jobs are disproportionately under-reported compared with jobs with longer weekly hours in comparison to what is reported by the detailed questionnaire. Reported time spent on the collection of firewood and water is generally not statistically different across groups, with one exception: boys are reported to spend more time on collecting firewood when reported by proxy. Of particular interest is to assess whether the relationship of the proxy to the child, particularly that of the child to his/her parents, may influence labor statistics. As proxy assignment was random among the eligible respondents in the household who were at least 15 years old, no biases due to the selection of proxy should be present in our estimates. Parents of the child make up 67.5 percent of the proxy responses. Grandparents account for 10.4 percent of proxy responses, while siblings report on their own sibling in 14.2 percent of the cases.10 Restricting the proxy sample to only 9 The intent of the survey experiment was not to generate statistics on child labor for comparison with other surveys in Tanzania, where there will be differences in questionnaire design as well as samples and field supervision. Nonetheless, we note that our LFP rates are higher than the 46 percent LFP reported by Guarcello et al. (2009), perhaps in part driven by a large share of rural households in our sample. 10 Other categories of proxies include nieces/nephews (3.5 percent), other relatives (3.2 percent), and brothers or sisters-in-law (1.2 percent). 14 the sub-sample of parental proxies does not significantly change the estimates for proxy-reported statistics in Table 3 or the regression results in Table 5 discussed below. Fathers as proxy respondents report lower LFP and higher working hours of their children than do mothers, but the difference is not statistically significant. These results are available upon request. In Table 4, we turn to the distribution of children's main activities across broadly defined categories. Participation in domestic duties, while not included in labor force statistics, is commonly collected, particularly in a child labor context. This is usually done by including domestic duties as a possible answer to the questions regarding the individual's main activity. Here we examine how reporting on domestic duties changes when using one overall question about any work (short module) compared with using three screening questions that require the respondent to specify wage work, farm work, and non-farm household enterprise work (detailed module). For the short module, the distribution across main categories is derived from a single question (question 4 in the short module ­ see Table A1, first column); for the detailed module, it is derived from question 9 (see Table A1, second column). The results in Table 4 show that the difference in questionnaire design between the short and detailed modules has a large and statistically significant impact on reports for both boys and girls. The first interesting finding is that the short questionnaire generates lower percentages of "no work" answers than the detailed questionnaire, i.e., higher percentages of individuals who classify themselves in employment.11 The difference is especially large (-20.6 percentage points) and statistically significant for girls but not for boys. However, when asked about the sector of main activity, an extremely large percentage of children who define themselves as "working" in the short questionnaire indicate that they are engaged in domestic duties ­ the difference with the detailed questionnaire is very large and significant for both boys (+21 percentage points) and girls (+35 percentage points). The detailed questionnaire, by contrast, generates higher participation in agriculture for both boys and girls (the difference compared with the short questionnaire is about 15-16 percentage points). As for the 11 In the short questionnaire, "no work" corresponds to those who answer "no" to question 1 (see Table A1, column 1); in the detailed questionnaire `no work' are those who answer `no' to all questions 1,3,5. 15 type of respondent, there is almost no difference between the statistics generated by self and proxy (except for slightly fewer boys working in "other sectors").12 Together this suggests that the additional questions contained in the detailed version work as "screening questions," filtering out at least part of the children that equate domestic duties with employment. It appears that individuals who would classify themselves as "working in domestic duties" if assigned the short questionnaire are "screened out" when using a detailed module and end up classified as "no work." This most frequently happens for girls. At the same time, a non-negligible proportion of children that would classify themselves as mainly engaged in domestic duties in the short questionnaire are classified in agricultural activities in the detailed questionnaire. Regression results: Survey assignment effects To obtain the marginal effect of each survey assignment, we estimate Equation 1 controlling for individual characteristics (age, gender, and education [highest grade attended]), household characteristics (household size, composition, asset holdings,13 and land holdings), and district indicators. In each case, we include separate indicator variables for the short module and the proxy module. Including an additional indicator variable for the short proxy interaction yields very similar results (results not presented). The results for child LFP, obtained by using a probit model, are reported in the first columns of Table 5, and indicate that the short module yields 19 percentage points lower participation rates for boys and 16 percentage points lower for girls (note that this is after re-classifying all domestic duties into "no work," following the ILO definition of employment). The use of proxy respondents also produces underestimation of child labor with respect to self-reporting, but the effects are much smaller (although, again, larger for girls) and not statistically significant for our sample size. These effects are large and their variation is consistent with the widespread differences in child labor statistics noted by Guarcello et al. (2009), who, 12 The non-agricultural sectors were: mining/quarrying/manufacturing/processing, gas/water/electricity, construction, transport, buying and selling, personal services, education/health, and public administration. Only 9 children did work in these sectors in the past 7 days. 13 The household asset index is constructed from a list of 14 durable assets, 7 livestock categories, and 7 housing characteristics. It has mean value 0.11 and standard deviation 0.9. 16 using data for four African countries (Togo, Lesotho, Burkina Faso, and Ghana), find that a CWIQ survey, which is similar to our short questionnaire, generates lower LFP estimates than a more detailed survey. However, since the surveys they compare are implemented two years apart, their results are only indicative. The right-hand-side panel of Table 5 reports the OLS results for the natural log of weekly hours of work in the child's main job conditional on working. The weekly hours of work are significantly higher for boys in the short questionnaire; they are also higher for girls, but the difference is not significant. This is consistent with the hypothesis that in the short questionnaire marginal jobs are disproportionately more likely to be forgotten (or not considered as jobs worth reporting) than jobs with greater weekly hours in comparison with jobs reported in the detailed questionnaire. Next we estimate a multinomial logit to investigate how the survey assignment affects the allocation across three categories ("outcomes"): agriculture and other sectors, domestic work, and the omitted category "no work."14 The results in Table 6a present the marginal effects, while Table 6b presents the predicted probabilities estimated at the mean value of the covariates for the three outcomes and the two experiments. For the pooled sample (Table 6a, panel A), using a short module produces lower participation in agriculture and other non-agricultural sectors with respect to "no work" than a detailed questionnaire produces, with a larger effect for girls. Both girls and boys are more likely to be classified as working in domestic work than identified as not working when given a short module. This effect is also larger for girls than boys. The proxy module is not associated with significant changes in sector classification. The multivariate analysis confirms that the largest difference between the short and the detailed modules is in the allocation of children across the two categories "domestic work" and "no work" (both considered as not in employment, based on the ILO definition). Although the detailed module captures higher participation in employment, the largest and most significant switch is from domestic work (in the short module) to agriculture or no work (in the detailed module). However, the type of 14 We merged "other sectors" with the agricultural sector because there are few observations in the non- agricultural sectors. Alternative categorizations do not change the results. 17 respondent does not appear to produce large impacts on labor force participation and the allocation across employment categories. Using proxy respondents produces similar statistics as when asking the child directly. Regression results: Interaction between survey assignment and covariates To address questions about the effect of survey methods on the estimated coefficients of child labor determinants, we assess the relationships between child labor supply and four variables discussed in Section 3that have been identified in the literature as key covariates.15 All four covariates are expected, and have been observed, to be positively related to child labor. In the subsequent discussion, we assess whether and how the estimated coefficients that reflect these respective relationships are affected by variations in questionnaire and respondent type. Table 7 reports the results of estimating Equation 2 for LFP using a probit regression.16 Columns 1-4 (5-8) present the interactions of survey types and one of the four covariates of interest for boys (girls). The results suggest potentially important impacts of the survey design on the estimated coefficients. For boys there are no differential effects of survey assignment by land holdings or household size. The negative impact of the short module is strongly attenuated for boys in households with higher asset holdings. Conversely, the impact of the short module is greater for boys who are older. For girls, we estimate variations in the short versus detailed impact associated with each of the four variables of interest. The difference in LFP between the short and detailed module is smaller for girls in households with larger land holdings, larger household size, and more assets, and for girls who are older, relative to other girls. We find fewer statistically significant effects of survey design choices in estimates associated with the four covariates we focused on when we consider the effects of those covariates on girls' or boys' conditional hours (Table 8). For girls, we find no 15 The aim of these results is to explore whether survey methods may affect the estimated coefficients. A more detailed analysis would be needed to give a precise meaning to the results. For instance, we limit ourselves to a simple linear relationship and do not explore the interaction with quadratic covariates, which may or may not be more appropriate in some instances. 16 As before, we only report the results for the short and proxy indicators. Including an additional indicator variable for the short proxy module yields similar results. 18 differential effect associated with proxy or short modules with respect to land holdings, household size, assets, and the child's age. For boys, only land holdings are associated with a differential effect of the short questionnaire. Boys from households with larger land size have less of a gap between hours reported by the short vs. detailed modules. Proxy reporting yields greater hours (as in Table 5) but this increase is attenuated for older boys, and is reversed (i.e., proxy respondents generate fewer hours than self-respondents) for boys living in households with a higher level of assets. Although other results are not statistically significant, the small sample size permits only the statistically significant detection of large effects. Results in Tables 7 and 8 indicate that different survey methods may generate employment statistics that are not only different at the mean, but also vary as key covariates vary. That is, the estimation of the economic relationships of interest to the researcher can be affected by the survey method. For both girls' and boys' LFP, we find that the size of the effect of the short versus detailed module varies depending on household assets and the child's age. Given the central importance in the literature of the effect of household wealth and household size on girls' participation in domestic duties--like childcare and food preparation--and economic activities--like the processing of food for market sale--our results raise important questions on the empirical estimation of these effects and point to the need for more research in this direction. Cost implications Alternative survey designs will have cost implications that have to be weighed against the value of "better" data. The difference in length between the detailed and the short module we used in our experiment was small; using the detailed module added only a few minutes to the average duration of the interview, according to field work reports from enumerators and supervisors. The cost implication of using a detailed rather than a short module, therefore, is also small. The additional cost of printing slightly longer questionnaires and the extra data entry requirement are only marginally larger for the detailed questionnaire. 19 By contrast, using proxy instead of self-reports involves substantial savings. The use of self-reports increases the length of field work because more days are spent in each sample village to locate and interview respondents. This survey experiment was carried out in conjunction with a larger consumption expenditure experiment, which required survey teams to spend a full two weeks in a village anyway. We cannot determine the additional field days that would be needed to complete self-report compared to proxy labor modules. However, based on field experience, we can roughly calculate that for two days spent in a village using proxy respondents, the survey team would need at least one more day to track down self-reports. This corresponds to a 33 percent increase in the length of time spent on actual field work. We can assume that all variable costs of field staff (per diems, lodging costs), often the largest category of survey costs, would increase by 33 percent. Transport costs may also raise if field teams used a team vehicle to track down respondents for self- reports. Given that the results of our experiment indicate that using self-reports instead of proxy respondents does not alter significantly the employment statistics collected, we can conclude--even without a rigorous cost-benefit calculation--that using self-reports in this case (for this sample and this type of statistics) would not be worth the extra cost required. 5. Conclusions Child labor has received increasing attention over the past decade and empirical measurement has now become common practice. How child labor is measured does differ across countries and within countries over time, potentially creating problems of comparability. Little is known about whether different survey methods generate different results for child labor statistics or whether the fluctuations we observe in child labor data are explained by other factors. This paper presents a randomized experiment whereby we use two commonly varied survey designs, the level of detail in the questionnaire and the choice of respondent, to estimate the effects of these survey features on the labor statistics they generate. Our findings suggest that using a short employment module generates a much lower incidence of child labor, once the percentage of boys and girls who declare their main occupation was "domestic duties" are correctly classified--as per the ILO definition 20 of employment--as "not in work," and also has some effect on working hours. Both boys and girls are reported to have lower participation in agriculture and more in domestic duties using the short module. Response by proxy seems to have no effect on employment statistics compared with the self-reported response by the child. These observations are confirmed when controlling for a wide range of individual, household, and village characteristics. When we use probit analysis to estimate the marginal effect of the two survey types, we find that the short module yields 17 percentage points lower participation rates for boys and 23 percentage points lower for girls. Using a multinomial logit, we find that both boys and girls are less likely to be reported in agriculture and other sectors than in no work and domestic duties when using the short module. However, response by proxy produces statistics that are not significantly different from self-response. This is in sharp contrast with the effect of survey methods on labor statistics of adults, where response by proxy appears to have the largest impact (see Bardasi et al., 2010). Our finding that there is no significant discrepancy in child labor force participation statistics between proxy and self-reports (that is, between the situation in which questions are asked to adults or the children themselves) is particularly reassuring.17 When discussing the choice of the respondent, in particular the use of household surveys to obtain information on child labor, the ILO guidelines state that "...With regard to respondents, the general practice is to address survey questions to the most knowledgeable adult member of the household (or sometimes the head of household, who is often also the parent or guardian of the working child). However, sections of the questionnaire may be addressed to the children themselves, particularly on hazards at the work place, and the main underlying reason for working." (ILO, 2008, para 49) The ILO document also states the importance of respecting ethical standards to make sure that children are not adversely affected by their participation in the survey, when they are respondents. So, in situations where it is not possible to interview the children directly or it is considered inappropriate, our results indicate that employment statistics should not be significantly affected. 17 An alternative view is that children and parents (and other proxy respondents) are equally disinclined to reveal the actual extent of child labor due to social stigma ­ that it is hidden from surveyors. As discussed in the Introduction, we consider this stigma to be minimal in this setting. This does imply that these results may not be germane in contexts where people would want to hide or deny the extent to which children work. 21 Our results suggest that for measuring child labor, the World Bank and ILO recommendation of using a detailed, self-reported questionnaire has an effect primarily through the appropriate screening of children into reporting their labor market activities. The type of questionnaire has a limited effect on measuring hours correctly for the whole sample, as our results on unconditional hours worked in Table 3 suggest. The screening questions may have an important role in reducing a source of misreporting in labor modules, namely, the respondent's confusion over the economic distinction between labor market activities and domestic activities. The lower LFP but longer hours for those in employment estimated with the short questionnaire compared with the detailed module suggests that more marginal jobs are being under-reported when using the short employment module. This indicates that the survey design may matter more for certain groups of individuals than others, such as in this example for children who combine work with school. We also find evidence that estimated coefficients reflecting the relationship between child labor force participation and economic variables that have been found to be significant explanatory variables, like household size, assets, and land owned by the household, can differ depending on the survey method used. These results provide clear evidence that survey design does matter for measuring and explaining child labor outcomes. Interestingly, the effects are different from those for adults found in our previous work. In the case of children, what appears to be important is a questionnaire design that defines more precisely (through screening questions) what "work" means, while using a proxy or asking the child directly does not seem to affect employment statistics. For adults we came to the opposite conclusions (Bardasi et al., 2010). Although we considered only two dimensions of survey design, our results send a strong signal. In order to compare, monitor, and analyze child labor, more attention should be placed in harmonizing the survey approach that generates the data. Rapid declines or increases in child labor that are solely due to differences in survey approaches may send wrong signals to policymakers. Although shorter, rapid appraisal questionnaires might be advantageous from a policy perspective and for cost 22 reasons--and may be a very acceptable method for adults if they self-report information, based on our research findings--their ability to provide reliable child labor statistics needs to be further considered. These few additional screening questions in the detailed questionnaire--to clarify the meaning of the concept of work--come at very little cost for survey field work. Our results are also an implicit plea for additional, similar survey experiments, as they leave important questions unanswered. Whereas the experiments used in this paper (especially the short vs. detailed questionnaire one) focus on existing survey instruments, future work may want to explore the effects that newly designed instruments would have. In particular, combining survey instruments with direct observation or diary keeping could be especially useful to find out what approach works best, and to help define a "gold standard," on which there is currently no agreement. Another fruitful way forward would be to implement survey experiments to investigate issues related to the System of National Accounts categorization. The experiment used in this paper, while not well suited to address these issues, indicates that how respondents classify children's work may not always be clear. Finally, a more precise way to identify a "pure" proxy effect would involve comparing data on the same person from proxy and self-response. Although this type of experiment could not be implemented in the setting available to us, this is certainly something worth considering for future work under different conditions. 23 References Anker, R. (1983) "Female Labour Force Participation in Developing Countries: A Critique of Current Definitions and Data Collection Methods." International Labour Review 122(6): 709-724. Bardasi, E., K. Beegle, A. Dillon, and P. Serneels (2010) "Do Labor Statistics Depend on How and to Whom the Questions are Asked? Results from a Survey Experiment in Tanzania." World Bank Policy Research Working Paper 5192. Bass, L. (2004) Child Labor in Sub-Saharan Africa. Lynne Rienner Publishers, Boulder, London. Basu, K., and P.H. Van (1998) "The Economics of Child Labor." American Economic Review 88(3): 412­427. Bhalotra, S., and C. Heady (2003) "Child Farm Labor: The Wealth Paradox." The World Bank Economic Review 17(2): 197-227. Borgers, N., E. de Leeuw, and J. Hox (2000) "Children as Respondents in Survey Research: Cognitive Development and Response Quality." Bulletin de Méthodologie Sociologique 66: 60-75. Dillon, A. (2010) "Measuring Child Labor: Comparisons between Hours Data and a Subjective Module." Research in Labor Economics 31: 135-159. Edmonds, E. (2005) "Does Child Labor Decline with Improving Economic Status?" Journal of Human Resources 40(1): 77-99. Edmonds, E.V. (2009) "Child Labor" in T.P Schultz and J. Strauss, Handbook of Development Economics Volume 4, Elsevier Science, Amsterdam, North Holland, forthcoming. Guarcello, L., I. Kovrova, S. Lyon, M. Manacorda, and F.C. Rosati (2009) "Towards Consistency in Child Labour Measurement: Assessing the Comparability of Estimates Generated by Different Survey Instruments." Understanding Children's Work Project Draft Working Paper. ILO (2004) Child Labour Statistics: Manual on methodologies for data collection through surveys. Geneva. ILO (2008) Child Labour Statistics: Report III. Geneva. Schaffner, J. (2000) "Chapter 9 Employment," in P. Glewwe and M. Grosh (eds.), Designing Household Survey Questionnaires for Developing Countries: Lessons from 15 Years of the Living Standards Development Study. Oxford University Press (for the World Bank). 24 Table 1: Household characteristics by survey assignment Households by survey assignment F-test of equality of Detailed Detailed Short Short coefficients across self-report proxy self-report proxy groups Head: female (%) 20.4 22.3 26.7 19.3 0.544 Head: age 48.4 47.3 48.7 48.4 0.882 Head: years of schooling 4.2 4.5 4.6 4.7 0.778 Head: married (%) 74.3 76.2 71.6 81.5 0.277 Household size 6.8 6.0 6.2 6.6 0.046 Adult equivalence household size 5.5 4.9 5.1 5.3 0.082 Share of members less 6 years 16.7 15.5 15.4 16.3 0.876 Share of members 6-15 years 41.2 41.9 42.1 41.0 0.915 Concrete/tile flooring (non-earth) (%) 16.8 17.7 23.3 23.0 0.451 Main source for lighting is 0.199 electricity/generator/solar panels (%) 5.3 4.6 10.3 9.6 Owns a mobile telephone (%) 25.7 24.6 25.0 29.1 0.845 Bicycle (%) 52.2 43.1 45.7 50.4 0.457 Asset index (ln) -0.2 -0.2 -0.1 -0.0 0.206 Owns any land (%) 84.1 87.7 80.2 85.2 0.457 Land size (acres, incld 0s) 4.3 3.2 3.3 4.1 0.082 Month of interview (1=Jan, 12=Dec) 6.4 6.3 6.1 5.9 0.711 Number of households 113 130 116 135 Note: The F-test tests the equality of coefficients across the groups by regressing the group indicators on the household characteristics with clustered household standard errors. 25 Table 2: Children's household and individual characteristics by survey assignment Individual survey assignment Detailed Detailed Short Short F-test of equality of coefficients self-report proxy self-report proxy across groups Female (%) 50.0 56.1 45.5 57.6 0.249 Age 12.3 12.4 12.6 12.4 0.706 Years of schooling 3.1 3.0 3.1 3.0 0.972 Head: female (%) 21.6 22.9 27.6 19.4 0.467 Head: age 48.3 47.7 48.7 48.2 0.937 Head: years of schooling 4.3 4.4 4.8 4.6 0.700 Head: married (%) 74.1 75.2 69.9 81.8 0.128 Household size 6.8 6.2 6.3 6.7 0.079 Adult equivalence hh size 5.4 5.0 5.2 5.3 0.135 Share of members less 6 years 16.8 15.5 14.8 16.5 0.811 Share of members 6-15 years 41.2 43.5 43.4 42.8 0.701 Concrete/tile flooring (%) 17.2 17.2 23.6 21.8 0.555 Main source for lighting is electricity/generator/solar panels (%) 6.0 4.5 9.8 8.8 0.348 Owns a mobile telephone (%) 25.0 23.6 26.0 28.4 0.769 Bicycle (%) 52.6 42.7 47.2 50.0 0.533 Asset index (ln) -0.2 -0.2 -0.1 -0.0 0.264 Owns any land (%) 82.8 88.5 80.5 85.3 0.271 Land size (acres, incld 0s) 4.2 3.3 3.4 4.2 0.184 Month of interview (1=Jan, 12=Dec) 6.4 6.4 6.1 6.0 0.841 Any hours collecting firewood last 24 hours (%) 26.7 31.2 31.7 25.9 0.377 Hours collecting firewood last 24 hours (including 0s) 0.3 0.5 0.4 0.4 0.922 Any hours collecting water last 24 hours (%) 60.3 60.5 70.7 66.5 0.615 Hours collecting water last 24 hours (including 0s) 0.5 0.5 0.5 0.5 0.306 Number of individuals 116 157 123 170 Note: The F-test tests the equality of coefficients across the groups by regressing the group indicators on the household characteristics with clustered household standard errors. Among the sample assigned to self-report, 3 children were unavailable and are re-categorized as a proxy response for the detailed module. 26 Table 3: Child labor statistics of the main job by survey assignment A. B. Short Detailed Diff Proxy Self-rep Diff Labor force participation (%) Boys 55.4 70.9 -15.4*** 61.7 64.0 -2.3 (0.50) (0.46) (0.06) (0.49) (0.48) (0.06) Girls 44.2 58.9 -14.7*** 50.0 53.5 -3.5 (0.50) (0.49) (0.06) (0.50) (0.50) (0.06) Weekly hours last week (unconditional) Boys 12.0 11.2 0.8 11.5 11.7 -0.2 (15.5) (12.4) (1.7) (13.2) (15.0) (1.7) Girls 9.0 9.7 -0.7 9.0 9.9 -0.9 (13.7) (11.6) (1.5) (12.7) (12.7) (1.5) Weekly hours last week (conditional on LFP=1) Boys 21.7 15.7 6.0*** 18.7 18.3 0.3 (14.9) (12.0) (2.1) (12.3) (15.1) (2.1) Girls 20.3 16.5 3.8** 18.0 18.5 0.5 (13.9) (10.8) (2.0) (12.8) (11.8) (2.0) Hours of firewood collection in last 24 hours Boys 0.4 0.5 -0.1 0.6 0.3 0.3** (0.9) (0.9) (0.1) (1.0) (0.7) (0.1) Girls 0.3 0.3 0.0 0.3 0.4 -0.1 (0.6) (0.7) (0.1) (0.6) (0.7) (0.1) Hours of water collection in last 24 hours Boys 0.4 0.4 0.0 0.4 0.4 0.0 (0.7) (0.7) (0.1) (0.5) (0.8) (0.1) Girls 0.6 0.6 0.0 0.6 0.6 0.0 (0.7) (0.7) (0.1) (0.8) (0.6) (0.1) Note: Standard deviation of variables and the standard error of the differences are in parentheses. 27 Table 4: Child's main activity in their main job by survey assignment Boys Girls A. Short or detailed Short Detailed Diff Short Detailed Diff Agriculture 52.5 68.5 -16.0*** 42.9 58.2 -15.4*** Other sectors 2.9 2.4 0.5 1.3 0.7 0.6 Domestic Duties 30.2 9.4 20.8*** 43.5 8.2 35.3*** No work 14.4 19.7 -5.3 12.3 32.9 -20.6*** Number of individuals 139 127 154 146 B. Proxy or self-rep Proxy Self-rep Diff Proxy Self-rep Diff Agriculture 60.3 60.0 0.3 49.5 51.8 -2.3 Other sectors 1.4 4 -2.6* 0.5 1.8 -1.3 Domestic Duties 21.3 19.2 2.1 23.7 20.2 3.5 No work 17.0 16.8 0.2 23.1 20.2 2.9 Number of individuals 141 125 186 114 Note: Other sectors are specifically listed on the questionnaire and include mining/quarrying, manufacturing/processing, gas/water/electricity, construction, transport, trading, personal services, education/health, public administration, and other. *** indicates statistically significant mean differences with the detailed self-report at 1%, ** at 5%, * at 10%. 28 Table 5: Regression analysis of labor statistics by survey assignment Labor force participation in main job Conditional hours worked in main job over the last 7 days over the last 7 days Pooled Boys Girls Pooled Boys Girls Short -0.165*** -0.194*** -0.157** 0.251*** 0.366*** 0.183 (0.047) (0.065) (0.064) (0.087) (0.117) (0.127) Proxy -0.072 -0.039 -0.058 0.052 0.206* -0.059 (0.047) (0.066) (0.066) (0.090) (0.123) (0.130) Observations 566 266 300 321 167 154 Note: Robust standard errors (clustered at the household level) are in parentheses. *** indicates statistical significance at 1%, ** at 5%, * at 10%. Regressions include controls for household size, female headship, share of members under 6 years, share of members 6-15 years, share of members over 60, asset index, acres of land owned, child's age, whether child ever attended school, and district dummy variables. LFP estimates are conducted using a probit model and we report marginal effects. 29 Table 6a: Regression analysis of main activity by survey assignment Pooled Boys Girls Agriculture + Domestic Agriculture + Domestic Agriculture + Domestic other sectors work other sectors work other sectors work Short -0.176*** 0.330*** -0.161** 0.216*** -0.168** 0.423*** (0.048) (0.038) (0.065) (0.046) (0.071) (0.055) Proxy -0.065 0.017 -0.040 0.019 -0.050 -0.012 (0.048) (0.038) (0.061) (0.040) (0.071) (0.062) Observations 566 566 266 266 300 300 Note: The multinomial logit model uses three categories: agriculture and other sectors, domestic work, and the omitted category, no work. Marginal effects are reported. See the note in Table 5 for the list of background characteristics included. See the note in Table 4 for an explanation of other sectors. Robust standard errors (clustered at the household level) are in parentheses. *** indicates statistical significance at 1%, ** at 5%, * at 10%. Table 6b: Predicted distribution from MNL estimates Actual Predicted distribution from MNL regression Short Detailed Proxy Self-report All Ag or other sectors 56.7 49.9 63.2 54.6 59.6 Domestic work 23.5 37.9 8.2 24.0 23.4 No work 19.8 12.2 28.5 21.4 17.3 Boys Ag or other sectors 62.8 54.3 68.1 60.1 63.3 Domestic work 20.3 31.5 8.4 21.0 19.3 No work 16.9 14.2 23.5 19.0 17.4 Girls Ag or other sectors 51.3 45.4 58.1 50.1 53.9 Domestic work 26.3 44.3 8.3 26.6 27.8 No work 22.3 10.4 33.5 23.3 18.2 30 Table 7: Children's labor force participation: Interactions of key covariates and survey assignment Boys Girls (1) (2) (3) (4) (5) (6) (7) (8) Short (S) -0.205*** -0.211*** -0.226*** -0.208*** -0.154** -0.153** -0.182*** -0.159** (0.067) (0.067) (0.069) (0.066) (0.064) (0.065) (0.063) (0.065) Proxy (P) -0.040 -0.028 -0.011 -0.044 -0.085 -0.068 -0.054 -0.050 (0.068) (0.068) (0.070) (0.067) (0.070) (0.067) (0.067) (0.066) Land size 0.022 0.006 0.005 0.007 0.016 0.003 0.001 0.000 (0.014) (0.007) (0.008) (0.008) (0.012) (0.005) (0.006) (0.006) Household size 0.051*** 0.050 0.051*** 0.053*** 0.009 0.005 0.010 0.008 (0.019) (0.030) (0.019) (0.019) (0.015) (0.023) (0.015) (0.015) Asset index 0.204*** 0.207*** 0.183** 0.206*** 0.115* 0.122** 0.032 0.125** (0.060) (0.059) (0.084) (0.060) (0.059) (0.059) (0.111) (0.058) Child age 0.057** 0.060*** 0.064*** 0.121*** 0.050** 0.049** 0.049** 0.020 (0.022) (0.023) (0.023) (0.041) (0.024) (0.024) (0.023) (0.041) S x (land size-mean land size) -0.019 0.018** (0.015) (0.009) P x (land size-mean land size) -0.011 -0.021 (0.015) (0.013) S x (hh size ­ mean hh size) -0.030 0.045* (0.034) (0.026) P x (hh size ­ mean hh size) 0.029 -0.029 (0.034) (0.025) S x (assets ­ mean assets ) 0.215** 0.170** (0.091) (0.086) P x (assets ­ mean assets ) -0.105 0.010 (0.087) (0.098) S x (child age ­ mean child age) -0.077* 0.067* (0.041) (0.039) P x (child age ­ mean child age) -0.039 -0.015 (0.040) (0.040) Number of observations 266 266 266 266 300 300 300 300 2 Adjusted R 0.287 0.285 0.297 0.292 0.237 0.230 0.230 0.230 Note: Robust standard errors (clustered at the household level) are in parentheses. *** indicates statistical significance at 1%, ** at 5%, * at 10%. See the note in Table 5 for the list of background characteristics included. 31 Table 8: Children's hours of work (log): Interaction between key covariates and survey assignment Boys Girls (1) (2) (3) (4) (5) (6) (7) (8) Short (S) 0.379*** 0.345*** 0.354** 0.362*** 0.162 0.206 0.239 0.164 (0.122) (0.126) (0.143) (0.117) (0.132) (0.131) (0.158) (0.124) Proxy (P) 0.208* 0.181 0.366** 0.219* -0.034 -0.052 -0.080 -0.038 (0.122) (0.131) (0.165) (0.124) (0.134) (0.133) (0.152) (0.125) Land size 0.012*** 0.005 0.005 0.004 -0.028 -0.026** -0.025** -0.025** (0.004) (0.005) (0.005) (0.006) (0.019) (0.011) (0.012) (0.012) Household size 0.011 -0.027 -0.002 -0.002 -0.009 0.014 -0.011 -0.010 (0.031) (0.043) (0.030) (0.030) (0.029) (0.039) (0.030) (0.030) Assets -0.179 -0.169 0.098 -0.158 -0.062 -0.052 -0.025 -0.050 (0.169) (0.163) (0.275) (0.169) (0.172) (0.170) (0.257) (0.170) Child age 0.108** 0.102** 0.110** 0.148** 0.041 0.042 0.046 0.096 (0.046) (0.046) (0.046) (0.063) (0.035) (0.037) (0.037) (0.067) S x (land size-mean land size) -0.015** -0.038 (0.007) (0.027) P x (land size-mean land size) 0.007 0.040 (0.017) (0.027) S x (hh size ­ mean hh size) 0.010 -0.042 (0.050) (0.050) P x (hh size ­ mean hh size) 0.039 -0.019 (0.046) (0.052) S x (assets ­ mean assets ) 0.037 -0.172 (0.222) (0.297) P x (assets ­ mean assets ) -0.479* 0.056 (0.264) (0.289) S x (child age ­ mean child age) 0.024 0.023 (0.070) (0.070) P x (child age ­ mean child age) -0.113* -0.092 (0.067) (0.076) Number of observations 167 167 167 167 154 154 154 154 Adjusted R2 0.203 0.193 0.213 0.204 0.013 -0.004 -0.008 0.001 Note: Robust standard errors (clustered at the household level) are in parentheses. *** indicates statistical significance at 1%, ** at 5%, * at 10%. See the note in Table 5 for the list of background characteristics included. 32 Annex Table A1 - Key employment questions in the short and detailed questionnaires Short questionnaire Detailed questionnaire 1. During the past 7 days, has [NAME] worked for someone who is not a member of your household, for example, an enterprise, company, the government or any other individual? YES...1 (»3) NO.....2 (question repeated for the past 12 months) 3. During the past 7 days, has [NAME] worked on a farm owned, borrowed or rented by a member of your household, whether in cultivating crops or in other farm maintenance tasks, or have you cared for livestock belonging to a member of your household? YES...1 (»5) NO.....2 (question repeated for the past 12 months) 5. During the past 7 days, has [NAME] worked on his/her own account or in a business enterprise belonging to he/she or someone in your household, for example, as a trader, shop-keeper, barber, dressmaker, carpenter or taxi driver? YES...1 (»7) NO.....2 (question repeated for the past 12 months) 1. Did [NAME] do any type of work in the last seven 7. CHECK THE ANSWERS TO QUESTIONS 1, 3 days? AND 7. (WORKED IN LAST 7 DAYS) Even if for 1 hour. ANY YES..1 YES...1 (»3) ALL NO.....2 (»37) NO.....2 (question repeated for the past 12 months) 3. What is [NAME]'s primary occupation in 8. What is [NAME]'s primary occupation in [NAME]'s [NAME]'s main job? main job? (MAIN OCCUPATION IN THE LAST 7 DAYS) (MAIN OCCUPATION IN THE LAST 7 DAYS) a. OCCUPATION a. OCCUPATION b. OCCUPATION CODE b. OCCUPATION CODE 4. In what sector is this main activity? 9. In what sector is this main activity? AGRICULTURE. . . . . . . . . . . . . . 1 AGRICULTURE. . . . . . . . . . . . . . 1 MINING/QUARRYING . . . . . . . . . . . 2 MINING/QUARRYING . . . . . . . . . . . 2 MANUFACTURING/ PROCESSING. . . . . . . 3 MANUFACTURING/ PROCESSING. . . . . . . 3 GAS/WATER/ELECTRICITY. . . . . . . . . 4 GAS/WATER/ELECTRICITY. . . . . . . . . 4 CONSTRUCTION . . . . . . . . . . . . . 5 CONSTRUCTION . . . . . . . . . . . . . 5 TRANSPORT. . . . . . . . . . . . . . . 6 TRANSPORT. . . . . . . . . . . . . . . 6 BUYING AND SELLING . . . . . . . . . . 7 BUYING AND SELLING . . . . . . . . . . 7 PERSONAL SERVICES. . . . . . . . . . . 8 PERSONAL SERVICES. . . . . . . . . . . 8 EDUCATION/HEALTH . . . . . . . . . . . 9 EDUCATION/HEALTH . . . . . . . . . . . 9 PUBLIC ADMINISTRATION. . . . . . . . .10 PUBLIC ADMINISTRATION. . . . . . . . .10 DOMESTIC DUTIES. . . . . . . . . . . .11 DOMESTIC DUTIES. . . . . . . . . . . .11 OTHER, SPECIFY . . . . . . . . . . . .12 OTHER, SPECIFY . . . . . . . . . . . .12 33 Table A2. Self-report and proxy selection process Self-report Proxy Eligibility rule: Eligibility rule for selecting proxy respondent: Individuals age 10+ Individuals age 16+ Eligibility rule for selecting subjects for proxy reporting: Individuals age 10+ Selection: Selection: Random selection of 2 respondents 1. Random selection of 1 proxy respondent among among eligible respondents eligible proxy respondents. The proxy respondent self-reports for him/herself 2. Random selection of 2 subjects for proxy reporting (excluding the proxy respondent selected) 34