14 Multidimensional Assessment of Child Welfare for Tanzania

Identifying trends in living standards in Tanzania has been a subject of considerable interest. Analysis of a household budget survey conducted in 2007 revealed consumption poverty rates approximately similar to the rates calculated from a comparable survey conducted in 2001 (Government of Tanzania 2009). This stagnation in consumption poverty occurred despite relatively high published rates of economic growth over the same period and little change in measured inequality. Price inflation over the same period as measured by the household budget survey also differed drastically from inflation rates derived from the published consumer price index (CPI) and the GDP deflator (Adam et al. 2012). The growth–poverty–inequality conundrum alongside the wide divergences in measured inflation provoked a great deal of analysis.¹

More recently, the World Bank (2015) published a poverty assessment based on a household budget survey conducted in 2011/12. This recent assessment focused heavily on comparisons of the results from 2011/12 with the data available from the 2007 survey and found a reduction in consumption poverty of about six percentage points. In the companion volume to this book, Arndt et al. (2016a) draw upon this and other analyses to assess growth and poverty for Tanzania, and Arndt et al. (2017) conduct a macroeconomic assessment of the growth–poverty relationship using a structural model. They find that the six percentage point reduction in poverty from 2007 to 2011/12 lies at the optimistic end of a reasonable range.

The assessment of consumption poverty trends in Tanzania over this most recent period (2007–11/12) has been substantially complicated by changes in the data collection methods employed in 2011/12 compared with all earlier surveys. In their poverty assessment, the World Bank (2015) also took the opportunity to apply a series of methodological changes to the computation of the nominal consumption aggregate and the poverty lines. These differentials rendered the analyses of the 2011/12 survey non-comparable with published analyses from 2007 and earlier. In order to account for these differences, the World Bank (2015) took a series of steps to revise 2007 data and calculations.

The revisions to the 2007 data were considerable. World Bank (2015: 2) reports that ‘consumption per adult rose by almost one-third’. The poverty line was also adjusted upward substantially, leaving the measured poverty rate at the national level essentially at the same value as reported in previously published assessments. Nevertheless, the issue of achieving comparability in data and methods clearly dominates any analysis of consumption poverty trends over the 2007 to 2011/12 period.

Rather than enter this fray, the work presented here seeks to analyse welfare trends from a multidimensional perspective, relying on data from four Demographic and Health Surveys conducted over the period 1991/2 to 2010. The chapter is structured as follows. Section 14.2 provides a brief review of multidimensional poverty measures. Both the first-order dominance (FOD) method and the Alkire–Foster (AF) approach are considered. Section 14.3 presents the datasets employed and the choices made to derive a set of comparable indicators. Section 14.4 presents results, including a comparison across the FOD and AF approaches. A final section concludes by highlighting the need for a collection of poverty tools to fully capture the complex nature of poverty dynamics.

The FOD methodology and implementation are described in detail in Chapters 3 and 4. They highlight that FOD analysis is an approach to comparing populations using multiple, binary welfare indicators without imposing weighting schemes or making assumptions about preferences for each indicator. Briefly, multidimensional welfare comparisons are based on the simple criterion that it is better to be not deprived than deprived in any indicator. FOD comparisons of population A and B result in one of three outcomes: population A dominates population B; population B dominates population A; dominance is indeterminate. Indeterminate outcomes occur when two populations are too similar or too different for definitive comparisons to be made (without further information or assumptions). For example, when comparing two individuals using three binary indicators with outcomes (0,1,0) and (1,0,1), dominance cannot be established because we do not assume it is better to be not deprived in any given dimension. The same logic can be extended to populations.²

Dominant outcomes are binary and thus provide no information about the extent of domination. To mitigate this shortcoming, we draw bootstrap samples from the surveys considered and conduct FOD analysis for each sample.³ The share of dominant outcomes for each pair of populations across all bootstrap samples can be interpreted as a probability of domination. Thus, while the welfare indicators are ordinal in nature, the application of bootstrap sampling produces probabilities of one population performing better than another. Probability of net domination across all bootstraps is used to rank areas. The probability of net spatial domination of area i is defined as the sum of the probability that i dominates each other area minus the sum of the probability that each other area dominates i. This probability of net spatial domination can be linearly transformed into an index that falls in the interval [−1,1] where higher values indicate that an area is better off. Analogously, bootstrap samples can be employed to calculate temporal net domination of a given area in time period t relative to time period s.

Next, we consider an alternative approach to multidimensional analysis, the Alkire–Foster (AF) approach developed by Alkire and Foster (2007). The method is well known for its application to the Multidimensional Poverty Index of the United Nations Development Programme (UNDP) which assesses welfare in over one hundred countries (see for example, Alkire and Santos 2010). This section provides a brief overview of the methodology. Alkire et al. (2015) provide a recent and comprehensive discussion of an array of multidimensional poverty measures.

The AF approach to multidimensional analysis aggregates information obtained from a set of binary welfare indicators into a single index that captures both the incidence and intensity of multidimensional poverty. The process of defining this index can be described in two steps: identification and aggregation. Identification is achieved in what Alkire and Foster (2007) refer to as a dual cut-off method. First, as with FOD, the approach begins with a set of binary welfare indicators, where in each dimension an individual is deemed to be deprived or not deprived according to a dimension-specific threshold. Second, an across dimension cut-off must be specified to distinguish the poor from the non-poor. In this context, the cut-off (k) identifies the poor as those with a weighted deprivation count greater than a cutoff level k. This provides a poverty headcount (H). When weights are equal across dimensions, k can be expressed as a number of dimensions such that individuals who are poor in k or more dimensions are considered poor.

Identification of the poor (via H) provides no information about the intensity of poverty. If an individual with a weighted deprivation count greater than k (i.e. one who is defined as poor in the multidimensional sense) becomes poor in an additional dimension, the multidimensional headcount ratio would not reflect this increase in the intensity of poverty. Therefore, an additional aspect of poverty is introduced to reflect the intensity of poverty. Intensity is measured by the average weighted deprivation count among those who are identified as poor. The final AF poverty index is referred to as the adjusted headcount ratio (M₀) and is expressed as the product of the multidimensional headcount ratio (H) and the average deprivation count among the poor (A),

Thus, a change in M₀ cannot be understood without considering both H and A. Though the method is sensitive to thresholds within and across dimensions as well as dimensional weights, the adjusted headcount ratio is simple to compute and convenient for comparisons across time and space.

Two important differences between the FOD and AF methodologies could lead to dissimilar results. First, FOD results use information from the full distribution of outcomes whereas M₀ is the product of two averages: H and A. For FOD, indeterminacy may result between two populations B and C when B outperforms C for all but a small segment of population B. In the same situation, AF is likely to clearly establish that population B outperforms C. Second, the use of weights allows the AF method to result in clear outcomes that may be indeterminate with FOD. As noted, because no assumptions are made about the relative importance of each dimension, FOD dominance cannot be established between pairs of welfare outcomes such as (0,1,0) and (1,0,1). However, with the AF method, the comparison is dependent upon how weights are assigned. For instance, with equal weighting the second pair is clearly superior to the first. On the other hand, with a weighting scheme (0.2, 0.6, 0.2) the first outcome is associated with greater welfare.

Results derived from FOD rely on few assumptions and strict criteria for establishing dominance. Thus, when dominance is established, the result is quite robust. AF, on the other hand, applies a weighting scheme and cut-off levels that may influence results. Despite this potential for different conclusions, in a comparison across thirty-eight countries Permanyer and Hussain (2015) find that the methodologies align closely with a correlation coefficient of 0.95. Arndt et al. (2016b) similarly find high correlations using census data for Mozambique. For these analyses, the indicators and thresholds determining deprived and not deprived in each dimension, which both FOD and AF are obliged to specify, were the same.

In this analysis, data from four Tanzania Demographic and Health Surveys (TDHS) are used to define five binary welfare indicators that allow multidimensional welfare to be estimated using both the FOD and AF methodologies in two subpopulations of children.

The 1991/2, 1996, 2004/5, and 2010 TDHS provide the data used in this analysis (National Bureau of Statistics and Macro 1993, 1997, 2005, 2011). The TDHS aims to provide estimates for the entire country, for urban and rural areas, and regions. The 1991/2 and 1996 TDHS samples were drawn in a three-stage design, with the goal of selecting 500 households each in Dar es Salaam and Zanzibar, and 300 households in the remaining regions. Using the 1988 census sampling frame, 357 enumeration areas (EAs) were first selected from wards/branches and then within wards/branches such that rural and urban EAs were selected proportionally within each region. In the third sampling stage, households were selected from complete household listings in each EA. The sampling design for the 2004/5 and 2010 TDHS involved two stages where in the first stage 475 clusters were selected from a list of EAs based on the 2002 census with eighteen clusters selected in each region except Dar es Salaam where twenty-five clusters were selected. In the second stage, households were then systematically selected from complete household listings in each EA.

From this micro data, we capture the non-monetary multidimensional nature of poverty by defining two population groups: school-age children at least seven and less than eighteen years old and young children less than five years old. The 7–17 sample includes 13,608, 11,472, 14,357, and 14,687 children and the under-five sample includes 7287, 6080, 7461, and 7526 children for 1991/2, 1996, 2004/5, and 2010, respectively. In each population group, children’s welfare is examined over time and across regions. Spatial areas include the nation, urban/rural areas, and geographical zones. Larger sample sizes for the 7–17 population group also allow analysis of administrative regions.⁴

For each population group we identify a set of five binary welfare indicators based on the Bristol Indicators (Gordon et al. 2003); the indicators are presented in Table 14.1.

Ideally, the sanitation threshold would be specified such that children using unimproved sanitation (e.g. uncovered latrines or no facilities) would be considered deprived. However, in 1992, 1996, and 2004 the TDHS does not distinguish between covered and uncovered latrines. In 2010, 73 per cent of school-age children used latrines and of these children 89 per cent used uncovered latrines. It is logical then to classify all latrines to be a deprivation. In section 14.4, we examine the sensitivity to the sanitation indicator choice by considering an alternative sanitation threshold where the use of any kind of latrine is not deemed to be a deprivation.

Browsing household surveys, the possibilities of examining a rich variety of deprivations appear to be great. However, both the FOD and the AF methodologies require that all indicators be non-missing for every individual or household in the sample. Care must be taken in constructing indicators that apply to the full population being examined. For instance, immunization histories seem to provide a useful measure of the health of children under five. Yet, children under the age of one would not be fully immunized and therefore should not be deemed deprived based on incomplete immunization records. Consequently, the sample would need to be restricted to children aged one to five rather than zero to five.

Women’s health indicators present similar difficulties. The Demographic and Health Surveys offer information on a wide range of family planning, fertility, and maternal health topics. However, these questions tend to be posed to a narrow range of women for whom these issues apply, and thus care must be taken to restrict the sample to the relevant population. For instance, maternal health issues would limit the population to not only women of childbearing age, but also women who were pregnant in the recent past. Depending on sample sizes or analytical goals, necessary restrictions may render the inclusion of certain indicators impractical due to the concomitant restrictions on the sample.

Figure 14.1 presents mean deprivation trends for children aged 7–17 at the national and urban/rural level. Table 14.2 also reports deprivation at the zonal level for all indicators including the alternative sanitation indicator. Overall, Figure 14.1 exhibits positive signs of advancement in most indicators. School-age children make considerable progress in access to education and information with national deprivation in education reduced by more than half between 1992 and 2010. Similar trends are observed in rural and urban areas and in all zones.

Access to safe water follows the most variable pattern. Urban water deprivation is relatively low but increases over time from 9 to 14 per cent. While national and rural areas achieve gains over the entire period, welfare backslides somewhat between 2004 and 2010 to 29 per cent and 33 per cent, respectively. In the zones, only Western makes progress between each survey while Central, Eastern, and Southern Highlands deteriorate over the eighteen-year period.

Urban areas progressed in terms of the housing and the primary sanitation indicator. Though access to urban sanitation improved by thirty-two percentage points over the study period, deprivation remained high at 60 per cent. In contrast, rural areas achieved little gains in either indicator with deprivations in sanitation and housing of 97 per cent and 83 per cent in 2010. Within the zones, Zanzibar, and Eastern zone follow urban patterns while the remaining zones generally mirror rural areas.

The vast majority of the population uses covered or uncovered pit latrines (83 per cent in 1992 and 73 per cent in 2010). The primary sanitation indicator classifies children using any pit latrines as deprived while the alternative sanitation indicator shifts this large percentage of children to being not deprived. As a result, deprivation in the alternative sanitation indicator (children with no sanitation facility) is extremely low. In contrast to the primary sanitation indicator, the percentage of children deprived in the alternative indicator increased at the national, rural, and urban areas, with more substantial increases in Central, Southern Highlands, and Western zones. Zanzibar is the only area to significantly reduce alternative sanitation deprivation.

Table 14.3 presents mean deprivation levels for children under five. Deprivations in water, sanitation, and housing closely follow the levels and trends seen with school-age children. Deprivation in education for under-fives measures whether children’s mothers have completed primary school. Though declining in every year, under-five education deprivation is greater than that of school-age children. Under-five nutrition, as evidenced by anthropometric measures, improved over the eighteen-year period. However, these figures remained high with 31 per cent and 41 per cent of urban and rural children nutritionally deprived in 2010. Though improvement occurred in all zones, as many as 51 per cent of children in Central were still nutritionally deprived in 2010.

We begin by examining whether child welfare, as defined by our set of five indicators, improved between 1992 and 2010. FOD temporal analysis compares the performance of a given area between survey years and is reported as the average probability of net domination across 100 bootstrap iterations. Net probability of domination measures the probability that the welfare of an area improves between two years minus any probability of regression.

Table 14.4 reports the temporal FOD outcomes for school-age children. Both the static results and bootstrap probabilities provide strong evidence of welfare progress at the national level and in rural areas from 1992 or 1996 to 2004 or 2010. In contrast, urban areas advance between 1992 and 1996 and then stagnate in the remaining years. National and rural stagnation between 1992 and 1996 is consistent with very little to no change in the percentage of children who are deprived in sanitation, housing, and education. Urban stagnation across most years and national and rural stagnation between 2004 and 2010 are directly associated with decreasing welfare in the water indicator. Among the zones, only the Central zone shows little to no signs of advancement during the study period. In line with substantial improvements in all indicators, Zanzibar exhibits the greatest probability of advancement among the zones.

To evaluate the sensitivity of temporal outcomes to the sanitation threshold, FOD comparisons were re-estimated using the alternative sanitation indicator and reported in Table 14.5. Consistent with alternative sanitation deprivation increasing over time (Table 14.2), evidence of temporal advancement is drastically reduced. Notably, static advancement in national and urban areas disappears and only moderate bootstrap probabilities of welfare gains in national and rural areas remain between 1992 or 1994 and 2004. However, Zanzibar, where the alternative sanitation indicator improved in all years, exhibits strong probabilities of advancement.

Finally, the temporal FOD results for children under five are reported in Table 14.6. Though the indicator trends for school-age children are generally comparable to under-fives deprived in water, sanitation, housing, and education, the under-five temporal results demonstrate the strict nature of the FOD criteria. For example, children under five and school-age children advance in all indicators between 1996 and 2004 nationally and between 1996 and 2010 nationally and in rural areas. However, unlike outcomes for school-age children, in the under-five static case, 2004 does not dominate 1996 for the nation or rural areas. In both periods, the probability of domination is lower than that of school-age children. This example demonstrates that the FOD criteria demand progress not only on average, but throughout the distribution (see Chapter 11 for more a more detailed discussion of indeterminacy).

In each year, FOD comparisons are made between all areas to determine the degree of domination of each area and zone. Values in the inner table represent the probability that the row area dominates the corresponding column area.⁵ Row averages measure the probability that the row population dominates all other populations, and column averages measure the probability that the column population is dominated by all other populations. In interpreting a population’s relative wellbeing, both row and column averages should be considered.

The 1992 and 2010 spatial comparisons for school-age children are presented in Tables 14.7 and 14.8.⁶ Within the tables, all domination in the static case (bold values) and significant bootstrap probabilities occur when urban areas, Eastern, Northern (1992), and Zanzibar dominate or when rural areas, Lake (1992), and Central (2010) are dominated. Column averages indicate that Southern Highlands, Southern, and Western zones also have moderate probabilities of being dominated in both years. Between the remaining areas, FOD is indeterminate or the probabilities of domination are quite low. Column averages for urban, Eastern, and Zanzibar and row averages for rural areas and Central increase considerably between 1992 and 2010, indicating a greater disparity between the welfare of the better-off and worst-off areas. In both years, the nation is nearly as likely to dominate other areas as it is to be dominated.