Essays on nonsampling errors in household panel surveys

Abstract

Household surveys represent the predominant form of data collection in low- and middle-income countries and function as crucial substitutes to constrained administrative data. In recent years, following an increasing demand for data, researchers and policymakers alike have addressed the continued issue of low-quality data. While much progress has been made, many sources of data, including household surveys, have been identified as being insufficiently accurate and reliable, thus constraining informed decision-making on behalf of policymakers. Indeed, the importance of obtaining high-quality outputs has been recognised in the Sustainable Development Goals, which emphasise that to date, data is key to informing policy, monitoring progress, and ultimately achieving formulated goals. This thesis aims to provide a better understanding of survey methodological issues in low- and middle-income countries and provide an outlook on the future of panel survey applications. Thereby, the first two essays deal with identification of nonsampling errors in household survey datasets, factors influencing their prevalence, and their impact. Conversely, the third essay examines the continued role of agriculture in rural development. The first essay investigates the prevalence of nonsampling errors in the seventh survey wave of a long-term household panel survey conducted in Thailand and Vietnam, which encompasses 3,812 households. An analysis of the distribution of nonsampling errors is undertaken in order to ascertain which type of error is most prevalent in the underlying computerised survey instrument. These findings are then compared with those of an earlier study, which examined the prevalence of nonsampling errors in a paper-based survey instrument. Thereafter, a negative binomial model is applied to analyse factors influencing nonsampling errors, which simultaneously assesses the influence of the interviewer, respondent, and interview and survey environment. The second essay utilises data from the same panel, albeit making use of the longitudinal nature of data. Using seven waves of panel survey data from Thailand, which were collected between 2007 and 2019, interviews of 1,542 identical households were examined with a focus on the consistency of reported employments. A three-stage approach is developed to identify inconsistent reporting thereof between pairs of consecutive survey waves. Additionally, a two-stage multilevel logistic model is applied in order to analyse interviewer and employment characteristics that influence inconsistent reporting. Further, the impact of inconsistent reporting on policy pertaining to household welfare is examined. The third essay utilises three waves of household survey data from Thailand, which were conducted in 2007, 2013, and 2019, and considers 1,160 identical households. A descriptive analysis is undertaken in which changes in livelihoods of rural households in Northeast Thailand are examined. Further, a logit regression is applied to identify factors influencing poverty incidence, which differentiates by the typology of household based on the importance of agriculture. The first essay finds that computerised survey instruments have a substantially lower count of missing data, whereas measurement errors remain a pressing issue. The findings of the negative binomial regression model highlight the importance of interviewer training and indicate that more outgoing and sympathetic interviewers produce interviews of higher quality. Additionally, conditions of the interview and survey are shown to influence the prevalence of nonsampling errors. Notably, the results suggest that measurement errors are most likely to occur in initial survey weeks, whereas the likelihood of refusal increases as the survey progresses. In Vietnam, incongruence of ethnicity between interviewers and respondents indicated a substantial increase in nonsampling errors. Further, survey providers in endeavours to collect high-quality data must account for differences in survey implementation. The second essay identifies substantial cases of underreporting of employments throughout pairs of consecutive survey waves. Notably, informal employments are less likely to be consistently reported and more complex household compositions are positively correlated with inconsistency. The impact of omitted employments on welfare indicators is demonstrated to be substantial with poverty headcounts being overestimated by, on average 6.7 percentage points at the provincial level. The third essay highlights that while income has been observed to increase over a 12-year period, which has coincided with an increasing proportion of agriculture-based households being classified as non-poor, little has changed in rural livelihoods in rural Northeast Thailand. Despite substantial out-migration of working-aged household members, most households remain engaged in agriculture and can be described as part-time, small-scale farmers. Further, those households mainly engaged in agriculture are observed to become increasingly dependent on government interventions due to the region’s propensity to droughts. In conclusion, the essays examining data quality of household surveys in Thailand and Vietnam provide new perspectives regarding factors that survey providers must consider in conducting surveys. Further, shortcomings of labour modules that are typically used in household surveys in developing countries are identified and provide an entry point to a debate on possible approaches to more accurately collecting employment data. The third essay highlights that rural populations remain highly reliant on agriculture and that the role of agriculture in development cannot be understated.