«2 Chapter Measuring Poverty Summary The first step in measuring poverty is defining an indicator of welfare such as income or consumption per capita. ...»
The first step in measuring poverty is defining an indicator of welfare such as income
or consumption per capita. Information on welfare is derived from survey data.
Good survey design is important. Although some surveys use simple random sampling, most use stratified random sampling. This requires the use of sampling
weights in the subsequent analysis. Multistage cluster sampling is also standard; it is
cost-effective and unbiased, but it lowers the precision of the results, which calls for some adjustments when analyzing the data.
The World Bank-inspired Living Standards Measurement Surveys (LSMS) feature multitopic questionnaires and strict quality control. The flexible LSMS template is widely used.
Income, defined in principle as consumption + change in net worth, is generally used as a measure of welfare in developed countries, but it tends to be seriously understated in less-developed countries. Consumption is less understated and comes closer to measuring permanent income. However, it requires one to value durable goods (by assessing the implicit rental cost) and housing (by estimating what it would have cost to rent).
While consumption per capita is the most commonly used measure of welfare, some analysts use consumption per adult equivalent, in order to capture differences in need by age, and economies of scale in consumption. The Organisation for Economic Co-operation and Development (OECD) scale (= 1 + 0.7 × (NA − 1) + 0.5 × NC) is popular, but such scales are controversial and cannot be estimated satisfactorily.
Other popular measures of welfare include calorie consumption per person per day, food consumption as a proportion of total expenditure, and nutritional status as measured by stunting or wasting. However, there is no ideal measure of well-being, and analysts need to be aware of the strengths and limitations of any measure they use.
Haughton and Khandker Learning Objectives After completing the chapter on Measuring Poverty, you should be able to
1. Summarize the three steps required to measure poverty.
2. Recognize the strengths and limitations arising from the need to use survey data in poverty analysis, including the choice of sample frame, unit of observation, time period, and choice of welfare indicators.
3. Describe the main problems that arise with survey data, including:
• survey design (sampling frame/coverage, response bias)
• multistage cluster sampling.
4. Explain why weighting is needed when surveys use stratified random sampling.
5. Describe and evaluate the use of equivalence scales, including the OECD scale.
6. Define consumption and income as measures of welfare, and evaluate the desirability of each in the context of measuring well-being in less-developed countries.
7. Summarize the problems that arise in measuring income and consumption, and explain how to value durable goods and housing services.
8. Identify measures of household welfare other than consumption and income, including calorie consumption per capita, nutritional status, health status, and food consumption, as a proportion of total expenditure.
9. Argue the case that there is no ideal measure of welfare.
Introduction: Steps in Measuring Poverty
The goal of this chapter is to set out a method for measuring poverty. Given the enormous literature available on the subject, we simply set out the main practical issues, with suggestions for further reading for those interested in pursuing the subject more.
Three steps need to be taken in measuring poverty (for further discussion, see
• Defining an indicator of welfare
• Establishing a minimum acceptable standard of that indicator to separate the poor from the nonpoor (the poverty line)
• Generating a summary statistic to aggregate the information from the distribution of this welfare indicator relative to the poverty line.
CHAPTER 2: Measuring Poverty This chapter defines an indicator of welfare; chapter 3 discusses the issues involved in setting a poverty line; chapter 4 deals with measuring aggregate welfare and its distribution.
Household Surveys All measures of poverty rely on household survey data, so it is important to recognize the strengths and limitations of such data and to set up and interpret them with care.
Key Survey Issues Ravallion (1992) lists a number of issues related to surveys that require attention
before one even attempts to measure or analyze poverty:
• The sample frame: The survey may represent a whole country’s population, or some more narrowly defined subset, such as workers or residents of one region.
The appropriateness of a survey’s particular sample frame will depend on the inferences one wants to draw from it. Thus, a survey of urban households would allow one to measure urban poverty, but not poverty in the country as a whole.
• The unit of observation: This is typically the household or occasionally the individuals within the household. A household is usually defined as a group of persons eating and living together.
• The number of observations over time: Most surveys are single cross-sections, covering a sample of households just once. Longitudinal surveys, in which the same households or individuals are resurveyed one or more times (also called panel data sets) are more difficult to do, but these have been undertaken in a several countries (for example, the Vietnam Living Standards Surveys of 1992–93 and 1997–98, or parts of the Thailand Socioeconomic Surveys of 2002 and 2004).
Several common problems arise when using and interpreting household survey data. We review these, organizing our thoughts largely along the lines set out in Ravallion (1992).
Survey Design. If the sample on which a survey is based is not random, then the resulting estimates of poverty are almost impossible to interpret. They are likely to be biased, but we do not know by how much.
A simple national random sample would create a list of everyone in the country and then randomly choose subjects to be interviewed, with each person having an equal chance of being selected. In practice, sampling always falls short of this ideal for three reasons. First, some people or households may be hard to find; for instance, most surveys interview people at their homes, but this completely overlooks homeless persons, a group that is likely to be poor.
Second, some of the surveys that have been used to measure poverty were not designed for this purpose in that their sample frames were not intended to span the entire population.
Examples: This is true of labor force surveys, which have been widely used for poverty assessments in Latin America; the sample frame is typically restricted to the “economically active population,” which precludes certain subgroups of the poor. To take another example, household surveys in the Republic of Korea have typically excluded one-person households from the sample frame, which renders the results unrepresentative.
Key questions to ask about any survey are the following:
• Does the sample frame (the initial listing of the population from which the sample was drawn) span the entire population?
• Is there likely to be a response bias? This may take one of two forms: unit nonresponse, which occurs when some households do not participate in the survey, and item nonresponse, which occurs when some households do not respond fully to all the questions in the survey.
Third, it is very often cost-effective deliberately to oversample some small groups (for example, minority households in remote areas) and to undersample large and homogeneous groups. Such stratified random sampling—whereby different subgroups of the population have different (but known) chances of being selected but all have an equal chance in any given subgroup—can increase the precision in poverty measurement obtainable with a given number of interviews. When done, it is necessary to use 12 weights when analyzing the data, as explained more fully in the following section.
CHAPTER 2: Measuring Poverty Sampling. Two important implications flow from the fact that measures of poverty and inequality are always based on survey data.
First, it means that actual measures of poverty and inequality are sample statistics, and so estimate the true population parameters with some error. Although it is standard practice to say that, for instance, “the poverty rate is 15.2 percent,” it would be more accurate to say something like “We are 99 percent confident that the true poverty rate is between 13.5 percent and 16.9 percent; our best point estimate is that it is 15.2 percent.” Outside of academic publications, such caution is, unfortunately, rather rare.
The second implication is that it is essential to know how the sampling was done, because the survey data may need to be weighted in order to get the right estimates of such measures as mean income or poverty rates. In practice, most household surveys oversample some areas (such as low-density mountainous areas, or regions with small populations), to get adequately large samples to compute tolerably accurate statistics for those areas. Conversely, areas with dense, homogeneous populations tend to be undersampled. For instance, the Vietnam Living Standards Survey of 1997/98 (VLSS98) oversampled the sparsely populated central highlands and undersampled the dense and populous Red River delta (Vietnam 2000).
In cases such as this, it is not legitimate to compute simple averages of the sample observations such as per capita income to make inferences about the whole population. Instead, weights must be used, as the following example shows.
Example: Consider the case of a country with 10 million people who have a mean annual per capita income of $1,200. Region A is mountainous and has 2 million people with average per capita incomes of $500; region B is lowland and fertile and has 8 million people with an average per capita income of $1,375.
Now suppose that a household survey samples 2,000 households, picked randomly from throughout the country. The mean income per capita of this sample is the best available estimator of the per capita income of the population, and so we may calculate this and other statistics using the simplest available formulae (which are generally the ones shown in this Handbook). For example, the Vietnam Living Standards Survey of 1992–93 (VLSS93) essentially chose households using a simple random sample, using the census data from 1989 to determine where people lived; thus, the data from the VLSS93 are easy to work with, because no special weighting procedure is required.
Further details are set out in table 2.1. If 400 households are surveyed in Region A (one household per 5,000 people) and 1,600 in Region B (one household per 5,000 people), then each household surveyed effectively “represents” 5,000 people; a simple average of per capita income ($1,215.60), based on the survey data, would then generally serve as the best estimator of per capita income in the population at large, as shown in the “case 1” panel in table 2.1. 13 Haughton and Khandker Table 2.1 Illustration of Why Weights Are Needed to Compute Statistics Based on Stratified Samples
Source: Example created by the authors.
* Estimated income per capita is likely to differ from true income per capita, due both to sampling error (only a moderate number of households were surveyed) and nonsampling error (for example, underreporting, poorly worded questions, and the like).
In picking a sample, most surveys use the most recent population census numbers as the sample frame. Typically, the country is divided into regions, and a sample is picked from each region (referred to as a stratum in the sampling context). Within each region, subregional units such as towns, counties, districts, and communes are CHAPTER 2: Measuring Poverty usually chosen randomly, with the probability of being picked being in proportion to population size. Such multistage sampling may even break down the units further, for example, to villages within a district.
At the basic level (the primary sampling unit such as a village, hamlet, or city ward), it is standard to sample households in clusters. Rather than picking individual households randomly throughout a whole district, the procedure is typically to pick several villages and then randomly sample 15 to 20 households within each chosen village.
The reason for doing cluster sampling, instead of simple random sampling, is that it is far cheaper to survey several households in a small area than to have to find households scattered widely over a potentially very large area.
But the use of cluster sampling, which is now almost ubiquitous, has an important corollary: The information provided by sampling clusters is less reliable as a guide to conditions in the overall area than pure random sampling would be. To see this, compare figure 2.1.a (simple random sampling) with figure 2.1.b (cluster sampling).
Although, on average, cluster sampling will give the correct results (for per capita income, for instance), so the expected mean values are unaffected—it is less reliable because we might, by chance, have chosen two particularly poor clusters, or two rich ones. Thus, cluster sampling produces larger standard errors for the estimates of population parameters. This needs to be taken into account when programming the statistical results of sample surveys. Not all statistical packages handle clustering; however, Stata deals with it well using the svyset commands (see appendix 3 for details).
Most living standards surveys sample households rather than individuals. If the variable of interest is household-based—for instance, the value of land owned per household or the educational level of the household head—then the statistics should be computed using household weights.
But a survey that samples households will give too little weight to individuals in large households. To see this, consider the realistic case of a survey that, at the village
Figure 2.1a Simple Random Sample Figure 2.1b Cluster Sampling