Estimating inequalities in ownership of insecticide treated nets: does the choice of socio-economic status measure matter?

Abstract

Research on the impact of socio-economic status (SES) on access to health care services and on health status is important for allocating resources and designing pro-poor policies. Socio-economic differences are increasingly assessed using asset indices as proxy measures for SES. For example, several studies use asset indices to estimate inequities in ownership and use of insecticide treated nets as a way of monitoring progress towards meeting the Abuja targets. The validity of different SES measures has only been tested in a limited number of settings, however, and there is little information on how choice of welfare measure influences study findings, conclusions and policy recommendations.

In this paper, we demonstrate that household SES classification can depend on the SES measure selected. Using data from a household survey in coastal Kenya (n = 285 rural and 467 urban households), we first classify households into SES quintiles using both expenditure and asset data. Household SES classification is found to differ when separate rural and urban asset indices, or a combined asset index, are used. We then use data on bednet ownership to compare inequalities in ownership within each setting by the SES measure selected. Results show a weak correlation between asset index and monthly expenditure in both settings: wider inequalities in bednet ownership are observed in the rural sample when expenditure is used as the SES measure [Concentration Index (CI) = 0.1024 expenditure quintiles; 0.005 asset quintiles]; the opposite is observed in the urban sample (CI = 0.0518 expenditure quintiles; 0.126 asset quintiles).

We conclude that the choice of SES measure does matter. Given the practical advantages of asset approaches, we recommend continued refinement of these approaches. In the meantime, careful selection of SES measure is required for every study, depending on the health policy issue of interest, the research context and, inevitably, pragmatic considerations.

Introduction

Socio-economic status (SES) has long been a predictive variable in studies on inequalities in population health status and access to health care services. How to ensure the poor benefit from interventions is a key concern in health and development debates (Grundy and Holt 2001; Gwatkin 2001; Houweling et al.2003; Schellenberg et al.2003; Victora et al.2003). Measures of SES are therefore increasingly incorporated into studies. Given the importance of insecticide treated nets (ITNs) as a public health intervention, estimating inequalities in access to ITNs and how these change over time has been an important aspect of monitoring progress towards the Abuja declaration and the Millennium Development Goals (Nathan et al.2004; Onwujekwe et al.2004; Uzochukwu and Onwujekwe 2004; Grabowsky et al.2005; Wiseman et al.2005; Noor et al.2006, 2007). Findings from these studies have fed into recommendations on how to improve levels of ITN coverage and use among disadvantaged populations.

Various approaches to measuring SES exist, including ‘direct’ measures such as income, expenditure and consumption, and ‘proxy’ measures mainly in the form of asset indices (Falkingham and Namazie 2002). While income and consumption are considered the traditional gold standard measures of SES, asset indices are increasingly being applied in the health and development literature (Filmer and Pritchett 2001; Montgomery et al.2000; Nathan et al.2004; Noor et al.2006, 2007; Onwujekwe et al.2004; Uzochukwu and Onwujekwe 2004). Despite the wide application of asset indices, few studies have shown: the sensitivity of findings to the choice of SES measure; whether the nature of disparities for bednet ownership differs depending on the welfare measure selected; and how the results might differ when rural and urban data are analysed separately or as a combined sample. Ensuring that we can draw similar conclusions from the data regardless of the SES measure used is important to test the robustness of the measures, and to ensure that policies are appropriately designed.

An increasing number of studies apply asset indices to estimate inequalities in coverage of ITNs. We conducted this study to demonstrate the sensitivity of household ranking when two measures of SES—expenditure and an asset index—are applied to the same data. We test whether the degree of inequality in bednet ownership differs depending on how SES is measured, and consider how contextual differences such as infrastructure and livelihoods might influence the findings when rural and urban data are analysed as a single sample. The latter is important because contextual differences between rural and urban settings can influence the value and meaning of individual assets, and thus their contribution to the asset index. While differences between rural and urban settings can be accounted for when expenditure or income data are used through adjusting for costs of living (for example, differences in food purchases, rent and transport), it is not yet clear how the same can be achieved for asset indices. In addition, the degree of inequality in bednet ownership is likely to differ between settings depending on the distribution strategies in place, and whether nets are subsidized or issued for free. Such factors are important to consider in monitoring inequalities in access to bednet ownership. The paper therefore has both a methodological and an empirical goal.

An overview of approaches to measuring socio-economic status

Given concerns regarding how to best measure household welfare, it is important to highlight the potential strengths and limitations of the most commonly used approaches. These include direct measures such as consumption, expenditure or income, and proxy measures, the most widely applied of which are asset indices (O’Donnell et al. 2007).

Direct measures of SES

The main direct measures of living standards are income and consumption. Income is defined as earnings from productive activities and current transfers, while consumption refers to resources actually consumed (Deaton 1997). Consumption is usually measured using household expenditure but the two differ in that expenditure excludes consumption that is not based on market transactions. Consumption for a good or service can be delayed or may have long lasting benefits, for example durable goods. In such cases, consumption should measure the benefits that come from the use of the good rather than the purchase price. Collecting accurate data on either income or consumption is difficult due to measurement errors arising from under-reporting, seasonality, recall bias and imputations.

In many ways, income and consumption data are more prone to measurement errors than expenditure (Deaton 1997; Falkingham and Namazie 2002; Lindelow 2006). First, most agricultural households are producers as well as consumers, and rarely distinguish the two, making it very difficult to disentangle accounts. Second, it is not clear how one should value the consumption of home produced items, especially in economies where markets are not well developed. Even where markets exist, whether one should use buying or selling prices to impute the value of production is a subject of debate. Third, collecting income data in developing countries is difficult because the majority of households rely on subsistence farming. Even where households are self-employed in sectors other than agriculture, collecting data on income is of dubious validity since households are unlikely to differentiate between profit from the business, cost of purchasing items and the amount used for their own consumption. Fourth, income is subject to fluctuations, while expenditure is smoothed over time. The flow of expenditure over a short period of time is therefore more likely to provide a good estimate of SES over a year than that of income. For these reasons, expenditure is the preferred direct measure of welfare in developing countries.

Proxy measures of SES

Due to the difficulties associated with direct measures of welfare, researchers have derived proxy measures of SES,¹ ranging from using single indicators such as education, occupation and area of residence to more complex measures such as asset indices. Asset indices in particular have been applied increasingly in recent years.

Asset indices provide an alternative welfare measure for large household surveys that lack data on income or expenditure (e.g. Demographic and Health Surveys), thus enabling the analysis of socio-economic inequalities in health (Filmer and Prittchet 2001). Asset indices combine a range of household assets and other characteristics such as housing, water and toilet facilities. Proponents of this approach argue that its advantages are: (1) it requires only data that can be easily and quickly collected; (2) assets can provide a convenient way to summarize the living standards of a household; (3) the index is useful as a control variable when estimating effects of variables potentially correlated with household wealth, such as maternal education and; (4) the index may be more representative of the long-term economic status of a household than income or expenditure, as it measures capabilities and the accumulation of resources rather than contemporary cash flow.

Approaches to construct asset indices have included: summing whether or not a household possesses certain assets (Montgomery et al.2000); weighting each asset by its market value (Conteh et al. 2000); factor analysis, which allows the use of a few components to explain the correlation structure between assets (Sahn and Stifel 2001); principal component analysis (PCA) that uses statistical techniques to determine the asset weights used in the index; and the Polychoric PCA, which uses maximum likelihood to capture the quality of assets, by treating assets like ordinal variables, rather than creating binary variables for each category (Kolenikov and Angeles 2004). Of these, the PCA has been one of the most widely applied (Filmer and Pritchett 2001; McKenzie 2003; Schellenberg et al.2003; Nathan et al.2004; Onwujekwe et al.2004; Grabowsky et al.2005; Noor et al.2006, 2007). Briefly, PCA seeks to describe the variation of a set of multivariate data in terms of a set of uncorrelated linear combinations of variables. Each consecutive linear combination is derived so as to explain as much as possible of the variation in the original data, while being uncorrelated with other linear combinations (Filmer and Prittchet 2001; Vyas and Kumaranayake 2006). PCA uses the first component to estimate household well-being. The first principal component gives an index providing discrimination between households, with the assets which vary most across households being given larger weight.

Although asset indices have gained popularity in recent years, there continues to be debate on their validity and reliability (Filmer and Pritchett 2001; Onwujekwe et al.2006; Vyas and Kumaranayake 2006). Concerns raised by these authors are as follows. First, the indices lack theory to motivate either the choice of variables or the appropriateness of the weights. Second, it is not clear how the index should be interpreted since it is more indicative of longer-run wealth than it is of access to resources in the short term, and it does not account for short-run interruptions to or changes in household finances. Where the outcome of interest is the current wealth status of a household, an asset index might not be an appropriate measure. Third, asset indices fail to capture the quantity and value of assets, and fourth, they ignore contextual differences regarding the value assigned to assets when data from different settings are combined to construct a single index (for example, land and livestock might be valued more highly among rural than urban communities). Even when different assets are used for different contexts, it is not clear how appropriate the various assets selected are for each setting.

As a result of such concerns, Lindelow (2006) argues that for some studies it is desirable to estimate SES more comprehensively by supplementing asset indices with other measures such as expenditure or income. In addition, Falkingham and Namazie (2002) point out that there can be problems with using asset indices to monitor changes in SES because there may be significant changes in household ownership or access to some assets, without improvement in household welfare. Despite the on-going debates about the reliability and validity of the asset index in measuring SES, studies have been increasingly adopting this approach rather than direct measures. This calls for a review of the evidence on the validity and reliability of asset indices.

Assets or expenditure: does the choice of SES measure matter?

Few studies have compared asset indices and expenditure as proxies for household SES, and those that have yield differing results. Existing evidence is mainly based on demographic data and education enrollment using data from Living Standards and Measurement Studies (LSMS) in Asia. One of the pioneer works that compared expenditure data and an asset index argued that the two approaches simply provide a different perspective on the same issue, and that the matter is of limited concern (Filmer and Pritchett 2001). Using data from three Asian countries (Pakistan, Indonesia and Nepal), they demonstrate that there is a good correlation between asset indices and expenditure quintiles. The authors therefore argue that the asset index in their setting is at least as reliable as expenditure measures, and sometimes a better measure.

However, they noted differences in household SES classification between the two approaches. For example, classification of households using the Pakistan LSMS revealed that only 60% of the expenditure-poor households are also classified as asset poor, and only 43% of the richest expenditure quintile are also among the richest 20% when asset data are used. In Nepal and Indonesia, 49–56% of households classified rich by expenditure are classified asset rich, while only 10–13% of those classified as expenditure rich are classified as asset poor. The results from Pakistan showed significant differences in household ranking when using the two measures. Such differences highlight the problem of applying a single asset index to compare different communities, especially where environments and livelihoods are diverse. This is likely to be the case when doing cross-country comparisons and within-country comparisons, particularly across rural and urban settings.

Using data from LSMS from five developing countries, Montogomery et al. (2000) found that proxy variables applied in demographic research are very weak predictors of consumption expenditure per capita. They note that although asset indices are often poor predictors of consumption, they may still be useful in testing hypotheses of whether consumption is a significant determinant of health outcomes, particularly with large sample sizes where there is significant variation in consumption. Similarly, Sahn and Stifel (2001) found the correlation of their asset index with household expenditure to be weak, but argue that the asset index was not intended to be a proxy for expenditure but rather an alternative indicator of household wealth.

Lindelow (2006) has shown that using consumption as a welfare measure indicates significant inequality in favour of richer households, while using an asset index indicates inequality in utilization in favour of poorer households. The author argues that in some contexts, the choice of the welfare indicator can have a large and significant impact on socio-economic, inequality data for service use, and therefore on public spending. Stewart and Simelane (2005) analysed similarities of an asset index with income data to measure SES differences in child mortality in South Africa. They show a more mixed relationship: at the household level, the asset index is a proxy for low levels of household income, but the inclusion of an asset index in regression models reduced income coefficients to insignificance. Income-based analyses had a much stronger relationship with child mortality than the asset index. At a community level, the results from the two measures of SES were closely related.

The above review reveals mixed findings when different SES measures are applied to the same outcome. While we recognize that both approaches have strengths and limitations, there is need for more evidence regarding how the two measures of SES compare when applied in developing countries. Of key concern are the extent to which the rankings of households are robust to the choice of SES measure, and how contextual differences alter the findings when data from rural and urban settings are combined to form a single index.

Methods

The study was conducted in a rural and an urban area in Kilifi district in coastal Kenya, following national scientific and ethical approval by the Kenya Medical Research Institute. Mtwapa, the urban area, is located along the main Mombasa–Malindi highway, 42 km south of Kilifi Town. It is a densely populated, low-income area typical of many towns in sub-Saharan Africa. Ganze, the rural area, is located 35 km inland of Kilifi Town, and is a remote setting with limited infrastructure and poor housing. The primary economic activities in Mtwapa are small-scale trading and skilled and unskilled work in the hotel industry, while agriculture is the main source of income in Ganze. The population in Ganze is predominantly Giriama, a sub-group of the Mijikenda ethnic community. Mtwapa houses a wide range of ethnic groups from different parts of the country.

Maps indicating the location and landmarks of every homestead² (rural) and house (urban) were drawn by hand to enable the selection of survey households. A complete list of homesteads or houses was then developed and each allocated a unique number. A total of 285 rural and 467 urban households were selected using simple random sampling. Interviews were conducted by 14 carefully trained field staff, with data collected including monthly expenditure, asset ownership, bednet ownership, housing quality, sources of drinking water and of lighting, cooking fuel, and toilet facilities. Questions on bednet ownership were only posed to households that had at least one child under the age of five, because malaria transmission in the study setting is stable (Snow et al.1993) and children aged under five form the main target group for ITN interventions.

At the time the study took place (2003/2004), the main source of bednets in both settings was the commercial sector.³ There was also a non-governmental organization working in the rural area that issued free bednets to poor households participating in their project. The approximate cost of a bednet at the time of the study was KES 350 (approximately US$ 4.7). Other sources of ITNs have since been introduced in the study areas and in other parts of Kenya, including the sale of highly subsidized ITNs to pregnant women and children under five through government health facilities and mass distribution integrated into the mass measles vaccine campaign (Noor et al.2007).

Measuring socio-economic status

SES categories were estimated separately for each setting to show the degree of inequalities within each area. A similar analysis was conducted using pooled data to demonstrate how contextual differences impact on the asset index and to show how households in both settings fit within the overall sample.

Estimating SES using expenditure data

Households were classified into SES groups using expenditure⁴ data. In both settings, expenditure data were collected for 14 spending categories relevant to the study setting including: food, cooking fuel, cleaning, lighting, rent, transport, remittance, education, debt repayment, contribution to community groups and churches. There were also questions on any other expenditure not captured in the 14 categories. Different recall periods were applied for different items to minimize recall bias. For example, for items that were likely to be purchased on a daily basis (for example, food, kerosene, soap), households were asked how much they spent during the last week and whether or not that time period represented a ‘normal’ week in terms of spending pattern. A monthly recall period was applied for less frequent spending items like education, debt repayment and rent. Expenditure data were converted into per capita estimates weighted for age using adult equivalence scales. Adult equivalence scales require that total monthly expenditure per household be divided by the household size but children are given a lower weight because of the differences in resource demands. Households were divided into quintiles on the basis of their monthly per capita expenditure.

Constructing the asset index using PCA

Data were collected on 15 assets including livestock (goats, cows, chickens), radio, telephone, sewing machines, land, television, cooking stoves (electrical/gas), refrigerators, bicycle, cars, tractors and motorbikes. Other indicators on household characteristics included roofing material, source of energy for cooking and lighting, toilet facilities, type of main wall for the dwelling and source of drinking water.

Using information on the range of assets and household characteristics, we created 42 binary variables. Binary variables are preferred because PCA converts categorical variables into a meaningless quantitative scale (McKenzie 2003). Since there are no standard guidelines on how to select the range of assets to include in an index, we selected variables on the basis of preliminary qualitative data (Chuma et al.2006; 2007), our understanding of the study area, and their use in the Demographic Health Surveys (DHS) and other studies. We also used descriptive statistics to check the mean ownership for each asset. Variables that featured in only a few households (for example, car, motorbike, tractor and refrigerator) were not included in the index. All binary variables created from a categorical variable were included in the index in order to capture complete data for each household. Finally, only 29 variables were used to estimate the asset index. Households that had at least one missing value were excluded from the analysis (n = 8 rural; 15 urban). Households were classified into SES quintiles on the basis of the asset score.

McKenzie (2003) highlights the need to check that the range of assets included in the index is sufficient to avoid clumping and truncation. Clumping occurs when households are grouped into clusters making it difficult to distinguish households from the asset list, while truncation occurs when the distribution of SES is spread over a narrow range that makes it difficult to distinguish the poor and the very poor (McKenzie 2003; Vyas and Kumaranayake 2006). To check for these problems we used the kernel density function of the asset index using the Epanechnikov kernel in STATA. Inequalities in bednet ownership were measured using an equity ratio and the concentration index. An equity ratio compares the values of the lowest and the highest quintile. A ratio of one signifies perfect equity, and a ratio of greater than one indicates that the variable of interest occurs more among the poor than the least poor (Onwujekwe 2005). A concentration index ranges from –1 to 1, with a value of 0 indicating perfect equity. The index takes a negative value when the variable of interest is concentrated among the poorest groups and a positive value when it is concentrated among the richest group (Wagstaff et al. 1991; O’Donnell et al. 2008).

Results

Socio-demographic characteristics of survey households

Mean household size was seven individuals in the rural area and five in the urban. The majority of rural adults had no education (53.5%), while the urban adults were better educated, with only 11.2% reporting no education. Both populations were relatively young, with over 50% of individuals in both areas aged below 18 years. The proportion of children aged below 5 years was 20.5% in the rural setting and 18.0% in the urban setting. There were no significant differences in household size and structure by SES group in both settings. Urban households reported significantly higher monthly per capita expenditure than the rural households. For more information on household SES, see Chuma et al. (2006, 2007).

The asset index

Table 1 shows the summary statistics and scoring factors for the range of assets used to construct an asset index for both settings. The results show that the majority of both rural and urban households owned land (88.9% and 64.9%, respectively). Rural households were more likely to own livestock (goats, cows and chickens), while urban households owned durable assets like radios, television sets and cellular phones. Overall, urban households lived in better quality housing, and had relatively good access to toilet facilities and tapped water. The majority of rural households did not have any toilet facilities (84.9%), and their houses were made of mud walls (91.3%) and grass thatch roofing (81.8%).

The scoring factors indicate different directions of impact on the asset index in both the rural and the urban sample. For example, goats, cows and chickens have a positive sign in the rural sample (indicating a positive impact on SES), while in the urban area only ownership of cows had a positive impact on the index. However, this impact is relatively low (0.007). Unexpectedly, ownership of a bicycle has a negative factor score in the urban sample (-0.031), although it has a positive score in the rural sample (0.072). As expected, low quality housing (e.g. mud walls and makuti/thatched roof) and poor water and sanitation facilities all have negative factor scores in both settings. Housing characteristics scored the highest weights in both settings. For example, having a house with cemented floor recorded the highest score in both settings (0.349 rural; 0.365 urban), followed closely by iron roofing (0.307 rural; 0.341 urban) and brick walls (0.235 rural; 0.359 urban). A similar pattern was observed in the pooled data (cement floor = 0.302; iron roof = 0.282; brick walls = 0.297).

The mean asset ownership of each SES quintile is indicated in Table 2. Ownership of most durable assets increases by SES in both settings, except for those variables that have a negative factor score (Table 1). For example, ownership of a radio in the rural area is 0.0% for the lowest quintile and 75.9% in the highest quintile. In the urban area, ownership of a television set is 7.0% in the lowest quintile compared with 81.2% in the richest quintile. In both settings, households in the richest quintile were more likely to live in better quality houses, to use tap water and to have access to toilet facilities. The combined data show that land was the most commonly owned asset (74.2%), followed closely by radio (68.8%).

The first principal component explains 20% variation in the rural sample and 21% variation in the urban sample. In the combined sample, the first principal component explains 32.4% variation. When households are classified into SES using the combined sample, the results indicate the following:

• The majority of rural households are ranked into the two poorest quintiles (89.5%);

• None of the rural households are in the 5^th (richest) quintile and only one is in the 4^th quintile;

• None of the urban households belong to the 1^st (poorest) quintile and only 8.5% are in the 2^nd quintile;

• The majority of the urban households are in the two richest quintiles (65.6%).

Comparing stability of household ranking using the asset index and expenditure data

The mean monthly per capita expenditure and mean value of the asset index for each quintile are shown in Table 3. The results indicate significantly higher monthly per capita expenditure among urban households for all quintiles, with the urban poor spending five times more than the rural poor. The degree of correlation between the two SES measures was relatively low in both settings (0.29 rural; 0.38 urban).

To show the sensitivity of household ranking to the choice of SES measure, the same households were re-classified into SES using expenditure data. Table 4 shows the number and percentage of households that were ranked into the same SES when both welfare measures were used. For example, in the urban area 43.4% of households classified as expenditure poor (lowest quintile) are also classified as asset poor, and 35.4% of those classified as expenditure rich (highest quintile) are also classified as asset rich. In the 3^rd and 4^th quintiles, the proportion of households that are classified into the same quintiles using both methods is much lower at 27.4% and 13.3%, respectively. In the rural sample, 30.5% of expenditure-poor households are also classified as asset poor and 32.6% of expenditure-rich households are classified as asset rich. However, in both settings the proportion of households classified as expenditure poor but classified as asset rich is low (3.6% urban and 8.3% rural). Overall, the ranking of households between the two SES measures was relatively unstable. Only 26% of households in the rural area and 30.6% in the urban area are classified into the same quintile when either measure is applied.

It is argued that in communities that are relatively poor, classifying households into quintiles might not clearly distinguish them and that it might be more sensible to group households into broader categories. To further examine how stability of SES ranking would compare when broader SES categories are applied, we collapsed the quintiles into three SES categories. The results reveal a similar pattern.

Applying the two SES measures to estimate inequalities in bednet ownership: are there differences in measured inequalities?

An important factor when choosing an SES measure is the exploratory variable of interest. We chose to assess inequalities in bednet ownership: there have been an increasing number of studies examining the same using asset indices, and equity in access to ITNs is a major policy issue for malaria endemic countries, including Kenya. The question of interest is whether the nature of inequalities in bednet ownership varies by type of SES measure. In total, 200 (71.2%) rural and 268 (60.9%) urban households had at least one child under the age of five. Of these, only 73 (36.5%) rural households owned at least one bednet compared with 153 (57.1%) urban households at the time of the study.

The distribution of bednet ownership by SES quintiles is indicated in Table 5. The results show that irrespective of the SES measure, net ownership in both settings is higher among households in the highest quintile, but not necessarily lowest among the poorest households. In both settings, the difference in bednet ownership between the poorest and richest quintile appears to be smaller when households are ranked on the basis of the asset index. For example, when households are ranked by the asset index in the rural sample, 29.3% of households in the lowest quintile own a bednet compared with 54.3% of the highest quintile (Q1/Q5 = 0.54). When households are ranked by expenditure, the equivalent figures are 40.0% and 52.6%, respectively (Q1/Q5 = 0.76). When households are ranked by the asset index in the urban sample, bednet ownership ranges from 42.6% in the lowest quintile to 77.8% in the highest (Q1/Q5 = 0.55). This gap widens when ranking is done by expenditure to 50.8% and 68.4% ownership, respectively (Q1/Q5 = 0.74). When the rural and urban data are combined, the equity ratio is much lower (0.46); ie a narrower gap between the poorest and richest households.

A recognized limitation of the equity ratio is the failure to consider values within the middle quintiles. When the distribution between all quintiles is taken into account, the Concentration Index (CI) shows that inequalities in the rural area are higher using the expenditure quintiles (CI = 0.1024; P = 0.017) rather than the asset quintiles (CI = 0.005; P = 0.172). The opposite pattern was observed in the urban sample; there were wider inequalities using the asset quintiles (CI = 0.126; P < 0.001) than when using the expenditure quintiles (CI = 0.0518; P = 0.232).

Discussion

In this paper we set out to compare how two SES measures (asset index and expenditure) rank households, whether or not results differ when rural and urban data are analysed separately or as a combined sample, and to consider whether the choice of an SES measure matters when assessing inequalities in bednet ownership.

The asset index and contextual differences

The results from the asset index estimated separately for each setting using similar variables explain the same range of variation in each setting (20% rural; 21% urban). These findings may be used to support the application of a single index that combines both the rural and the urban sample. However, an attempt to combine the data revealed that even when similar assets are used to construct an index, urban households appear to be wealthier than their rural counterparts. Although in some cases large differences between settings are related to infrastructure (Filmer and Pritchett 2001), this was not the case in our study setting, except in the case of piped water. It could be that urban households are simply wealthier than rural households. However, urban households can also face harsher livelihood struggles than rural households despite owning the same valuable assets. For example, rural households can use assets such as livestock and land to support daily survival, whereas higher living costs in the urban area might make this more difficult. Such contextual differences highlight the need for computing a separate asset index for rural and urban samples, even when similar data are available for both settings. They also raise concerns regarding cross-country comparisons using an asset index as the sole measure of SES. Nevertheless, given the practical advantages of asset indices, they are particularly useful for such large analyses. We therefore support the need to continue to improve the methodology for estimating SES using asset indices to account for differences in the importance and value of assets between rural and urban settings.

The results also showed a weak correlation between the asset index and monthly expenditure. Similar findings have been reported elsewhere (Montgomery et al.2000; Sahn and Stifel 2001), while others have found a strong correlation (Filmer and Pritchett 2001). As outlined in the literature review, it is argued that asset indices often have a weak relationship with expenditure (in the range of 0.2–0.4) because the two concepts measure different things and because there are no explicit guidelines on how asset variables should be selected (O’Donnell et al.2008). Exactly what is of interest therefore needs careful consideration in selecting indices for studies.

Sensitivity of households’ ranking to the choice of SES measure and implications for findings

A comparison of consistency of ranking of households into SES quintiles revealed that only about a third of households in both settings were allocated the same SES quintile when either approach was applied. Reassuringly, only a minority of households (less than 10% in both settings) shifted from the lowest quintile to the richest quintile and vice-versa. Similar findings were reported in Asia (Filmer and Pritchett 2001). Regarding bednet ownership, the results indicate wider inequalities between the poorest and richest households when expenditure data are used to classify households in both settings compared with asset data. However, when the distribution in bednet ownership is measured across all quintiles, the results are more complicated. There are significant differences by quintile in bednet ownership in the urban area when households are classified using expenditure, and significant differences in the rural area when assets are used.

From these findings it is clear that conclusions about inequalities in bednet ownership depend on the choice of SES measure in our setting. These findings might imply that it is easier to capture SES differences in a rural setting using expenditure rather than asset data, while asset data better capture differences among urban households. However, the nature of bednet distribution in the rural setting might have influenced the direction of this relationship: the poorest households may have received a free bednet from the non-governmental organization working in the area. It is not clear why inequalities in bednet ownership are wider when assets data are used to categorize urban households. Where the main source of bednets is the commercial sector, one might expect wider inequalities when expenditure data are applied. This was not the case in our setting. The data available from our study cannot provide answers to this issue.

Lindelow (2006) observed similar inconsistencies in measured inequalities when different SES measures were applied to the same data, suggesting that caution is required regarding the robustness of health equity analysis to choice of welfare measure. Others have concluded that the choice of welfare measure does not matter (Filmer and Pritchett 2001; Wagstaff and Watanabe 2003). We support Lindelow's argument that in some contexts the choice of welfare indicator does matter and can drive conclusions in important ways.

The choice between SES measures is difficult to make because there are strong arguments for and against each. Expenditure has a more grounded theoretical basis and economists would argue that it is the preferred measure of SES. Asset indices are also important for measuring SES when expenditure data are not available, and are often preferred because collecting data is far less cumbersome and can therefore be considered more reliable. However, a recent study demonstrated that even asset data are not always reliable, and are also prone to measurement errors (Onwujekwe et al.2006).

Given the possibility of accounting for contextual differences through weighting expenditure data, we recommend that expenditure data be collected to classify households into SES, whenever time and survey structure allows. This is particularly the case for studies that aim to compare differences across rural and urban settings, and where the variable of interest is influenced by immediate cash flow. We recognize that estimating SES using expenditure data can underestimate the true SES of households in rural settings compared with their urban counterparts, especially among agricultural communities who are likely to spend relatively low amounts on purchasing food. Given the strong rural-urban links held in our study areas, as elsewhere in sub-Saharan Africa, urban households may benefit from food products produced by their kin in the rural areas (Molyneux et al.2002b). The possibility of underestimating true SES using expenditure data can therefore apply to both rural and urban households; albeit to a lesser extent.

Returning to asset indices, the sensitivities of these approaches should be more widely tested and the methodological concerns addressed where possible. Meanwhile, where data on both assets and expenditure can be made available, we suggest the sensitivity of the SES measures is tested and described before drawing conclusions. Where this is not possible, the potential sensitivity of SES ranking by choice of welfare indicator requires recognition as a potential limitation.

A limitation of our study is that our exploratory variable was based on the ownership of at least one bednet per household, among households that reported having at least one child aged below 5 years. Clearly, the findings would differ if the number of nets or the proportion of children aged under five sleeping under a net were incorporated into the analysis. Nevertheless, the results highlight the need for careful choice of SES measure.

Conclusions

We are not arguing that either asset indices or expenditure are the better approach to measuring SES. Our data do not support this argument and this was not the aim of the study. Both approaches clearly have their strengths and weaknesses. However, the choice of SES measure does matter, potentially leading to different conclusions, with implications for policy recommendations and ultimately for the design of interventions. We have demonstrated this through comparing two SES measures and the implications for equity analyses in bednet ownership. These findings suggest that researchers should carefully consider from the outset whether they are concerned with long-term or short-term wealth, whether the question of interest is related to asset-based inequality or expenditure inequality, the context in which the study is taking place and the nature of polices that they want to inform. Such information can inform the choice of the most appropriate SES measure for a given study or policy question.

Acknowledgements

This study was funded through a grant awarded to CM by The Wellcome Trust, UK and supported by the Kenya Medical Research Institute (KEMRI). We are grateful to the study communities and to the personnel team that worked tirelessly to make fieldwork successful. This paper is published with the permission of the director of KEMRI.

Endnotes

References