GCRF: Pseudo-Panels for Long Period Analysis of African Household Surveys

全国住户调查已成为分析发展中国家贫穷情况的标准数据来源。就分析贫困动态的潜力而言，这些非洲调查的一个主要限制是它们不是一个小组- -每一波调查的是不同的家庭，因此它们构成重复的横截面。因此，不可能长期跟踪同一家庭，以调查减贫的驱动因素。这给利用内生变量进行分析带来了挑战，例如家庭规模与贫困或移徙、汇款和家庭收入之间的相互作用。小组的缺失也限制了对长期家庭福利决定因素的分析。我们提出的解决这一数据限制的策略是识别具有代表性的家庭类型，利用重复的横截面家庭调查构建伪面板（见支持案例）。对伪面板的分析使人们能够跟踪类似的家庭，并补充每次调查的家庭层面分析。该项目将制订建造伪面板的方法，在对调查的具体特点作适当修改后，可适用于任何进行三次或更多全国住户调查的国家。原则上，这些方法也适用于人口普查和人口与健康调查数据。虽然该项目的重点是乌干达（1992-2012年使用8个现有调查），但构建和分析伪面板的方法可以应用于其他非洲国家。利用与当地研究伙伴建立的联系，因此在很大程度上是“预算外”的，伪面板方法将应用于加纳（1991-2013年使用6次调查）和坦桑尼亚（1991-2012年使用4次调查）。自上世纪90年代初以来，这三个国家都成功地将贫困人口减少了大约一半。我们使用重复的横断面调查数据，通过将单个家庭（观察单位）分组到基于时不变特征（户主的位置、性别和出生队列）的队列中，形成“代表性”家庭的伪面板。然后，随着时间的推移，这些群体在连续的调查中出现，形成一个带有“滞后值”的伪面板。由于队列固定效应与队列（家庭）特征相关，这些特征由于每次调查中队列成员的变化而无法观察到，并且随着时间的推移而不是恒定的，因此使用变量误差估计器来校正队列均值作为不可观察总体均值的估计值。滞后的因变量由辅助回归和使用时不变工具的增广工具变量估计量构造而成。因此，伪小组允许对家庭福利和减贫的决定因素进行长期（20年或更长时间）分析，有可能为内生变量生成内部工具，并确定政策变化的影响（例如乌干达的普及初等教育）。该项目预计将做出许多贡献。首先，将发展、试验和测试建造和使用伪面板的方法，产生一种可转让的方法，可用于大多数发展中国家，特别是非洲，那里有许多调查，但面板很少。其次，伪小组将用于提供有关乌干达减贫驱动因素的新见解，而在加纳和坦桑尼亚的应用（在本项目的生命周期内）则更为有限。第三，通过培训和研究合作，我们将支持非洲（特别是乌干达）的研究能力建设。为了支持这些目标，在坎帕拉举行的启动讲习班将包括各种利益攸关方和焦点小组，以确定重要的研究和政策问题；将在国内和通过AERC进行传播；还将为当地研究人员提供培训。

National household surveys have become the standard source of data for analysis of poverty in developing countries. A major limitation of these surveys for Africa, in terms of the potential to analyse poverty dynamics, is that they are not a panel - different households are surveyed in each wave so they constitute repeated cross sections. It is therefore not possible to track the same households over time to investigate the drivers of poverty reduction. This creates challenges for analysis with endogenous variables, such as interactions between household size and poverty or migration, remittances and household income. The absence of a panel also limits analysis of determinants of household welfare over long periods.The strategy we propose to address this data restriction is to identify representative household types to construct pseudo panels making use of the repeated cross section household surveys (see the Case for Support). Analysis of the pseudo panel allows one to track similar households and complements household-level analysis for each survey. The project will develop methods for constructing pseudo-panels that can be applied, with suitable modifications for specific features of the surveys, in any country with three or more national household surveys. In principle, the methods are also applicable to census and Demographic and Health Survey data. Although the project focuses on Uganda (1992-2012 using eight existing surveys), the methods for constructing and analysing pseudo-panels can be applied to other African countries. Utilising established links with local research partners, hence largely 'off-budget', the pseudo-panel method will be applied to Ghana (1991-2013 using 6 surveys) and Tanzania (1991-2012 using 4 surveys).These three countries all have managed to roughly halve headcount poverty since the early 1990s.We use the repeated cross-section survey data to form a pseudo panel of 'representative' households by grouping individual households (the observational units) into cohorts on the basis of time invariant characteristics (location, gender and birth cohort of household head). The cohorts are then traced over time as they appear in successive surveys, forming a pseudo panel with 'lagged values'. As the cohort fixed effect is correlated with cohort (household) characteristics that are unobserved and not constant over time due to the changing membership of the cohorts in each survey, an errors-in-variables estimator is used to correct the cohort means as estimates of the unobservable population means. The lagged dependent variable is constructed from an auxiliary regression with an augmented instrumental variables estimator using time-invariant instruments. The pseudo panel therefore permits a long (20 years or more) analysis of determinants of household welfare and poverty reduction, with the potential to generate internal instruments for endogenous variables and to identify effects of policy changes (such as Universal Primary Education in Uganda). The project is expected to make a number of contributions. First, methods for constructing and using pseudo-panels will be developed, tried and tested generating a transferable method of potential use for most developing countries, especially in Africa where there are many surveys but few panels. Second, the pseudo-panel will be used to provide new insights on the drivers of poverty reduction in Uganda, with more limited applications (during the life of this project) to Ghana and Tanzania. Third, through training and research collaboration we will support building research capacity in Africa (specifically Uganda). To support these aims, an inception workshop in Kampala will include a variety of stakeholders with focus groups to identify the important research and policy issues; there will be dissemination in country and through AERC; and training will be provided to local researchers.