Individual Wealth Inequality: Measurement and Evidence from Low- and Middle-Income Countries

The accumulation of personal wealth, stemming from ownership and control of assets, plays a critical role in advancing women’s and men’s economic opportunities. Yet, it is an understudied dimension of inequality across the developing world. To study individual-level wealth inequality and gender differences in wealth, this paper leverages unique data from nationally-representative multi-topic household surveys that were conducted in Cambodia, Ethiopia, Malawi and Tanzania and that interviewed men and women in private regarding their personal ownership and valuation of physical and financial assets. The analysis documents substantial gender inequalities in asset ownership and wealth, overall and for specific asset classes. Individual-level wealth inequality measures are substantially higher vis-à-vis comparators based on per capita household consumption expenditures and per capita household wealth — and intra-household wealth inequality has a substantial role in explaining overall wealth inequality. While land is a key contributor to wealth inequality across countries, there is cross-country heterogeneity in the relative contributions of asset classes. Self-reporting on asset ownership and valuation, the internationallyrecommended best practice, is also shown to lead to higher inequality estimates compared to the businessas-usual survey practice of interviewing a single, most-knowledgeable household member to identify intrahousehold asset owners and values. The discussion expands on the implications of the findings for future surveys and methodological research. JEL Codes: C81, C83, D31, J16, O12.


Introduction
The accumulation of wealth-stemming from ownership and control of land, durables, and other assetshas critical roles in advancing economic opportunities for men and women. These include improved productivity and investment, raising one's ability to withstand economic and other shocks, as well as improving economic security for future generations. Land and equipment are necessary for expanding farm and nonfarm enterprise operations, for example; access to credit and other financial services can also be eased through the use of land as collateral and ownership of bank accounts (Cheng et al., 2021;Twumasi et al., 2020). Ownership of vehicles, and other durables such as mobile devices and computers, can improve access to markets and lower transaction costs (Hasanbasri et al., 2021a). Across all these areas, assets can be passed down or transferred to others -improving the economic mobility of younger generations, for example (Quisumbing et al., 2004). All of these channels can also help contribute to individuals' economic status and bargaining power within households (Doss, 2013).
Understanding wealth inequality along individual dimensions such as gender, age, and sector of work, are important for shaping effective, distributionally-sensitive policies around asset ownership. These include designing programs to expand individual property rights (Wang, 2014;Kilic et al., 2021); access to finance (Dupas et al., 2017;Field et al., 2021); as well as understanding the implications of property taxation (Komatsu et al., 2021;Stephens and Ward-Batts, 2004). Given the critical roles assets play in income generation, investment, and intergenerational economic mobility, individual-level data collection on asset ownership can better inform how these initiatives should be targeted, particularly amid surging wealth inequality across countries, and rising costs of land and capital (Zucman, 2019).
Multi-topic household surveys are an ideal source of information on these different dimensions, and to better highlight the socioeconomic and demographic profiles of asset owners for accurate policy targeting.
When included in surveys, however, variables on wealth (as well as other welfare aggregates such as consumption, on which national measures of inequality are based), are often collected at the household level (De Vryer and Lambert, 2021;OECD, 2013), or by a single "most knowledgeable" member reporting for others, as opposed to individuals themselves (Kilic, Moylan and Koolwal, 2021;Hasanbasri et al., 2021b).
In this vein, self-reported, individual-level data collection on asset ownership in these surveys can yield two contributions in developing a more nuanced, policy-relevant understanding of wealth: first, to reveal greater insights on wealth across sub-groups within the population, and second, to shed greater light on intra-household disparities in wealth. A large literature challenging the unitary model of the household has underscored how measures of inequality based on household-level data can inherently mask unequal resource allocation and welfare outcomes among individuals (Chiappori and Meghir, 2015;Haddad and Kanbur, 1990). Relying solely on household-level data collection on asset ownership to inform program targeting, can therefore miss substantial intra-household inequalities in how returns to assets are shared across members -limiting the efficacy of such programs in reducing broader inequalities across different sub-groups. While it can be argued that individuals who do not own assets within the household may still benefit from other members' ownership (including the possibility to inherit later on) -as well as benefiting from greater household wealth overall, even if they never own assets themselves -this may not always be the case, particularly in contexts where social and cultural restrictions on the ownership and transfer of property and other assets to certain groups, including women and individuals facing shocks such as divorce and widowhood (Djuikom and van de Walle, 2022). In contexts where women face greater legal and cultural restrictions on ownership of physical and financial assets, for example, studies have shown that expanding their personal ownership has been shown to improve their socioeconomic outcomes along different dimensions (see Wang, 2014, andDoss, 2006, on the links between women's individual property ownership and changes in household expenditure patterns; and Dupas et al., 2017 on the links between individual financial account ownership and women's savings). Understanding features of ownership, including exclusive as well as joint ownership in the household, as well as rights over assets to sell, bequeath, and make investments, can shed further light in this area.
In this paper, we use new data from nationally-representative household surveys that were conducted in Cambodia, Ethiopia, Malawi and Tanzania over the period of 2016-2020 with support from the World Bank Living Standards and Measurement Study-Plus (LSMS+) program; the surveys included cross-country comparable questionnaire modules based on intra-household, private interviews with adult household members regarding their personal ownership, control and valuation of a comprehensive set of physical and financial assets, including whether assets were exclusively or jointly owned. The LSMS+ supported questionnaire modules were based on recently developed international guidelines for collecting individual disaggregated survey data on asset ownership and rights (UN Statistical Division (UNSD), 2019). The 2019 UNSD guidelines, in turn, were based on several country pilots supported by the Evidence and Data for Gender Equality (EDGE) initiative, that recommends self-reported data on individuals' asset ownership and rights. 2 Having self-reported information on asset ownership and values allows us to create individual-level wealth aggregates, whose analysis reveals a number of findings that have important implications for both survey design and understanding of wealth inequalities within populations.
This paper is primarily structured as a measurement study. We discuss how the LSMS+ data were collected across countries, associated measurement issues with reporting data on asset ownership and valuation, 2 The EDGE pilots were conducted between 2014-2016, and included Uganda, Mexico, Georgia, Philippines, Mongolia, Maldives, and South Africa. These pilots and the resulting guidelines were also informed by the ongoing work of several survey initiatives that using different approaches have included some individual-level data on ownership and control of assets (also see Doss et al., 2014, for a discussion of the Gender Asset Gap project). and how individual-level data on wealth sheds greater insights on inequalities within populations, vis-à-vis standard household per capita measures. There are several headline findings from our research: • First, across countries and different asset classes, there are substantial gender differences in the incidence of asset ownership and values of asset holdings, underscoring the importance of disaggregated data collection (i.e. elicited from individuals, and at the asset-level) for characterizing gender inequities in wealth in low-and middle-income countries.
• Second, inequality as measured by household per capita wealth far outpaces more standard measures based on household per capita consumption expenditure, and individual-level variation in wealth leads to even higher measures of inequality across countries, similar to recent findings on intra-household consumption poverty (Brown et al., 2018;Bose-Duker et al., 2021). Intrahousehold wealth inequality has a substantial role in explaining overall wealth inequality, and land, in particular, is an important driver of inequality, although there is variation by country as to whether residential or non-residential land plays a greater role. Data on asset values are also not without their own measurement issues, including missing responses and variation in reported values among joint owners, raising questions on how to treat jointly owned assets. We use these features of the data as an opportunity to understand how constructing different specifications of individual-level wealth to account for these sources of measurement error affect our findings, and continue to find that our results are robust to these sensitivity checks.
• Third, within the context of understanding individual wealth inequality, self-reporting on asset valuation matters. In Malawi, we find that measures of individual wealth inequality are higher using self-reported data from the LSMS+ are higher than estimates from a concurrent, national survey that covered similar survey strata, and which had the same identical household and agricultural questionnaires and field teams, with the exception of using the more standard business-as-usual survey practice of interviewing a single, most-knowledgeable household member to identify asset owners and values within the household. 3 Overall, a key contribution of this paper is highlighting the importance of disaggregated data collectionacross individuals and asset classes -as well as carefully considering aspects of respondent selection (i.e. self versus proxy reporting) in understanding and targeting economic inequalities. Little is understood as to how specific assets -across land and durables, for example -drive wealth inequality within countries, given data gaps on individual ownership of different asset classes, particularly in national surveys that could allow for a deeper understanding of inequality within populations. And even when these topics are covered in surveys, household members' asset ownership and control are typically asked only of one person considered by survey enumerators to be knowledgeable about these issues. The significance and consistency of our findings across countries, and robustness to sensitivity checks addressing measurement errors around missing values and jointly owned assets, point to the need to broaden the scope of surveys and methodological research to better understand longer-run inequalities among men and women stemming from differences in personal asset ownership and wealth.
The paper is structured as follows. Section 2 discusses the LSMS+ methodology and survey data. Section 3 provides descriptive statistics on within-and across-household variation in asset ownership and valuation, including how asset valuation are constructed, accounting for different scenarios on how missing data as well as joint ownership of assets/multiple valuations of the same asset are treated. Section 4 discusses the 3 The same surveys were used by Kilic, Moylan and Koolwal (2021) in a separate comparison of how self-reporting affected estimates of men's and women's agricultural land ownership; their study found that asking one household member to report on others' ownership, compared to having all adult members reporting themselves, led to lower reporting of women's land ownership and higher reporting for men. findings on inequality measurement using the individual-level data on wealth. Section 5 presents a discussion on implications for how self-versus proxy reporting affects estimates of inequality. Section 6 concludes.

Background
Since 2016, and as part of the 18 th replenishment of the International Development Association (IDA18) that committed to launch pilot data collection in at least six IDA countries to "gather direct respondent, intra-household level information on employment and assets," 4 the LSMS+ program has coordinated with country national statistical offices to include self-reported, individual-level modules on asset ownership and rights, as well as labor, as part of ongoing, multi-topic, nationally-representative surveys. The LSMS+ modules follow 2019 United Nations guidelines for producing statistics on asset ownership from a gender perspective, prepared under the Evidence and Data for Gender Equality (EDGE) project. 5 With an emphasis on self-reported data to fully capture intra-household gender inequalities in asset ownership, LSMS+ aims to address widespread data gaps on individual wealth and improve the accuracy in estimates of wealth. This is in contrast to more common survey approaches that -when data on wealth are collected -tend 4 World Bank, 2016. 5 See UN (2019). EDGE was launched in 2013 as part of a collaboration between the US Government, the United Nations Statistics Division and UN Women, along with key regional and international agencies including the OECD and the World Bank. The UN-EDGE guidelines were also informed by country pilots including the Methodological Experiment on Measuring Asset Ownership from a Gender Perspective (MEXA) in Uganda (Kilic and Moylan, 2016).
to either collect this information at the household level, or by one respondent per household that answers for others (see Kilic, Moylan and Koolwal, 2021 for a discussion).
In this paper, we use publicly available data from LSMS+ supported national surveys that were conducted in Malawi (2016 Integrated Household Panel Survey), Tanzania (2019/20 National Panel Survey), Ethiopia (2018, and Cambodia (2019 LSMS+ Survey). 6 These countries all have different civil and customary laws around men's and women's land ownership and rights, as well as broader norms around the pursuit of economic opportunities, affecting individuals' ability to accumulate and manage assets (see Hasanbasri et al., 2021d. Table 1 presents the main features of these surveys. Individual wealth is constructed in our analysis using data on individuals' self-reported ownership and valuation of different assets in the LSMS+ surveys. The aforementioned UN guidelines recommend that countries collect information, at a minimum, on the principal dwelling, agricultural land, and other real estate as this "core" set of assets has been found to comprise the majority of personal wealth. Additionally, the guidelines advised countries to determine additional assets (including non-agricultural enterprise assets, livestock, large and small agricultural equipment, financial assets and liabilities, valuables, and consumer durables such as vehicles) to collect data on based on the policy needs and prevalence of assets within the country.
In line with these recommendations, LSMS+ supported national surveys have focused on the following asset classes on a cross-country basis: residential and non-residential (which, in these countries, was primarily agricultural) land; mobile phones; and financial accounts. A module on livestock ownership was also included covered in Ethiopia and Cambodia, and the Cambodia LSMS+ Survey additionally included a module on individual-level ownership of durables, such as computers as well as motorized and nonmotorized vehicles, including bicycles, motorcycles, cars, tractors, boats and tuk tuks (rickshaws). Similar to other national multi-topic surveys under the LSMS program, the LSMS+ supported surveys collected detailed data on the same range of individual and household socioeconomic and demographic outcomes critical for understanding poverty and welfare, with the same questionnaire structure, flow of modules and question wording. This allows for cross-country comparisons in the paper alongside the within-country analysis.

LSMS+ methodological approach to collecting individual-level data on assets
Survey modules on ownership of and rights to assets should be premised on the basic idea that assets can often be owned by multiple individuals, and in other cases only one individual. To account for these scenarios, a key component of the LSMS+ approach is to conduct intra-household and private interviews with all household members aged 18 and older on their personal ownership of and rights to assets (see Hasanbasri et al., 2021b, for a detailed discussion of the survey methodology as well). 7 Following the UN guidelines, this involved first obtaining a roster of assets, in each asset class, belonging to household members, and then separately interviewing all adult household members, feeding a common roster of assets into each individual-level interview. 8 Across all LSMS+ supported surveys, information on individual asset ownership and valuation was entirely self-reported.
The LSMS+ assets modules also examine nuances of ownership that are not well-covered in surveys on wealth. This includes differentiating between respondents' exclusive versus joint reported ownership of 7 As part of the Malawi Integrated Household Panel Survey 2016, the first LSMS+ supported survey, up to 4 adults were randomly selected for interview, following the approach used in MEXA in Uganda. This covered 99 percent of adults in the survey sample.
Interviews prioritized the selection of the head of household, and his or her spouse, if working in a couple household, and the rest of the respondents were selected at random. By the second LSMS+ supported survey in Tanzania each asset in a given asset class. 9 In the case of joint ownership, respondents were asked to identify who shared the asset within the household (through those individuals' unique household roster identifiers), as well as the number of men and women outside the household who shared ownership of the asset. For residential and non-residential land, different ownership and rights constructs were covered as well. For residential and non-residential land, for example, individuals were asked about overall reported ownership, economic ownership (being able to control money over the proceeds from hypothetical land sales), as well as documented ownership/whether the respondent's name was on a title or certificate to the land.
Whether landowning respondents had rights to sell, bequeath, rent, use as collateral, and invest were also asked. Additional questions were asked on whether these nuances of ownership and rights were held jointly or exclusively. This level of detail highlights important patterns of ownership and decision-making that can inform policy efforts to expand access to financial services, land, and property rights in general.
Our results also show, under the individual-level survey approach, a substantial share of agreement among couples on ownership and rights to dwelling and non-dwelling land -around 70 and 75 percent of land parcels in Tanzania and Malawi, respectively, 90 percent in Ethiopia, and 94 percent in Cambodia (Appendix Table A1). 10 On valuation specifically, the UN guidelines recommend that in estimating wealth, each asset should be valued individually at its current market price (the potential sales value or realization value, namely how much the respondent would expect to receive if they sold the asset today). This recommendation had been based on the experience under the World Bank Living Standards Measurement Study-Integrated Surveys 9 The one exception was for consumer durables in Cambodia (computers, as well as motorized and non-motorized vehicles), where respondents were only asked about ownership overall, and not asked to identify joint owners. This stems from the protocol and questionnaire used in MEXA, also in line with the sample durable module in the UN-EDGE guidelines.
on Agriculture (LSMS-ISA), the Gender Asset Gap Project, and the surveys that were supported by the EDGE project, which found that the potential sales value provides a more consistent measure of asset prices as compared to the construction value or rental value of the asset, as well as other approaches to sales valuation such as the "quick sale" price and "reservation" price. 11 The LSMS+ assets modules followed the UN guidelines and asked respondents that of each asset they owned, "if you were to sell <THIS ASSET> today at current market prices, how much would you receive for it (in local currency)?" 12 The approach was item-by-item and was asked of respondents for a point estimate valuation for each asset they owned (each land parcel, for example), as opposed to an aggregate estimate for all assets within a specific asset class.
For financial accounts, respondents were asked about the current value of each account.

Descriptive statistics: individual-level variation in wealth
Wealth is a function of accumulated asset value but can also be viewed through the lens of ownership overall. Understanding inequalities in wealth rely on a careful examination of both, particularly in contexts and among groups where ownership of specific assets may be very limited. In this section, we first examine statistics on ownership of different asset classes, and then discuss aggregate measures of wealth we construct based on asset values reported by owners. 11 The "quick sale" price is what the owner could expect to receive if he/she sold right away, i.e. not necessarily at current market prices, and which could lead to an undervaluation of the asset. The "reservation" price is the price that would induce the respondent to sell the asset, even if he/she was not intending to sell; this might lead to an overvaluation. For more discussion see UN (2019).
12 In Ethiopia, since land is owned by the state, respondents were asked to provide the rental as opposed to sales value. In practice, however, land has been bought and sold through side channels. Table 2 presents the share of men and women owning assets covered in each survey -and shows that depending on the country, gender differences in ownership can vary substantially. Women are not always less likely to own assets -for residential and non-residential land, for example, women are more likely in

Asset ownership
Malawi and Cambodia to have reported as well as economic ownership, with higher shares of ownership for residential as compared to non-residential land. These differences are much greater in Malawi, where exclusive ownership of land among women tends to be substantially higher than men's, attributable to higher share of women-headed households and matrilineal communities where land is inherited through the matrilineal line (Kilic, Moylan and Koolwal, 2021). 13 In Cambodia, joint ownership of land is greater overall compared to other countries, likely due to civillaw designating most assets acquired during marriage as joint. In Ethiopia and Tanzania, on the other hand, we find lower shares of property ownership overall, with women being less likely to own land in both countries. Table 2 reveals statistically significant gender inequalities in reported and economic ownership in Ethiopia, as well as economic ownership in Tanzania. Women are also significantly less likely in Ethiopia to own livestock.
13 Ownership, however, may not necessarily correlate with rights -in separate estimates among landowners, gender disparities widen substantially in rights to sell and bequeath land (Hasanbasri et al., 2021b(Hasanbasri et al., , 2021c. (1) Estimates weighted with the household sampling weights. Significant gender differences are indicated by asterisks; ***p<0.01 **p<0.05 *p<0.10 (2) Because of some missing observations for specific assets, the share of men and women with any asset can be lower than the shares of individual assets (Hasanbasri et al., 2021b(Hasanbasri et al., , 2021c. (3) In Cambodia, joint versus exclusive ownership of durables was not asked explicitly.
Furthermore, the results point to the importance of collecting data on different asset classes, when trying to better understand differences in ownership across groups, including by gender. In Cambodia, for example, women's ownership of livestock and durables like bicycles was greater than for men. Men, however, were more likely to own mobile phones, financial accounts, computers and different classes of motorized vehicles. In Ethiopia, Tanzania, and Malawi, gender differences in mobile phone ownership (where women were less likely to own) widened considerably compared to other asset classes; the same was true for financial account ownership in Ethiopia and Tanzania. In addition, nuances of ownership can vary by country as well as asset class. The share of men and women in Cambodia with joint ownership of land and mobile phones is much higher compared to countries in Sub-Saharan Africa, for example, but not for financial accounts or livestock.
In separate OLS regression estimates controlling for socioeconomic and demographic variables including age, marital status and employment, along with household consumption quintiles and enumeration-area fixed effects (Appendix Tables A2 and A3), we find a set of cross-country consistent correlates of men's and women's asset ownership. As would be expected, ownership of different assets is significantly higher among respondents who are household heads and/or older. Being married has, in fact, a stronger positive association with women's ownership of different assets than for men, and particularly for residential and non-residential land. Ownership of specific assets is also often tied to specific employment and education outcomes -land ownership, for example, is higher among respondents working in agriculture (with stronger effects in Cambodia and Tanzania), while financial asset and mobile phone ownership are higher among those working for a salary or in a non-farm enterprise, and with higher schooling. Table A4 also highlights a substantial degree of within-household heterogeneity in asset ownership, among households with both adult men and women. For example, in the countries from Sub-Saharan Africa,15-20 percent of households have men owning land, while women do not; in general around 10 percent or more of households in these countries, and in each asset class, have women not owning any assets, while men report exclusive or joint ownership with other household members. In many households as well, women and men are each exclusive owners of different assets within the same asset class, including livestock in Cambodia and Ethiopia.
3.2 Constructing individual-level wealth measures, and addressing measurement issues around missing valuations and joint ownership of assets Following the UN guidelines that were discussed in Section 2.2, we use respondents' self-reported valuations of each asset they own, based on the potential sales value, to construct aggregate valuations of their wealth. Data on asset valuation, however, can be subject to different forms of measurement error, that might affect an overall understanding of inequality within populations. There are two measurement issues with the valuation data that we address in the paper, specifically related to missing valuations in respondents' self-reports, as well as how valuations around joint ownership are treated.

Missing valuations: multiple imputation approach
On the first issue, a key measurement challenge is that in Ethiopia and Tanzania, a substantial number of respondents not report values for specific assets, such as land and livestock (see Appendix Table A5). In Ethiopia, for example, 43 percent of men and 56 percent of women who were residential landowners did not report a potential sales value. In Tanzania these shares were 19 percent of men and 48 percent of women. Appendix Table A5 shows that among those with missing valuations, women were generally more likely to have missing data, and the most common reason among men and women was not knowing the value of the asset. In Malawi and Cambodia, on the other hand, the shares of respondents with missing valuations across asset classes was much lower -5 percent or less among landowners in Malawi, and on average 15 percent in Cambodia. In Cambodia and Ethiopia as well, there were no missing observations on valuation for financial accounts, mobile phones and durables. Aggregating missing data will lead to zeros within specific asset classes and could lead to an underreport of individuals' total wealth.
Missing data are not uncommon in household surveys, and in this case we used the presence of missing data as an opportunity to understand how alternate approaches to treating this issue could lead to overall differences in findings on individual wealth. Specifically, we compared individual wealth when leaving missing values as is (and hence treated as zeroes when aggregating values across asset categories), with a multiple imputation model (Rubin, 1987). Multiple imputation is commonly used to address issues related to missing data or the need to replace sensitive data in large-scale surveys (see Reiter, 2012 for a discussion) and Kilic et al. (2017) also rely on multiple imputation to derive predicted GPS-based area measures for up to 1/3 of agricultural plots that are not visited for GPS-based area measurement in the context of LSMS-ISA-supported national surveys in Tanzania and Uganda.
The multiple imputation approach uses information on observed values of variables that are predictive of the missing values -for example, individual, household, enumeration-area and interviewer variablesand conducts a predictive mean matching model based on repeated imputations. In our analysis, following the approach used by Kilic et al. (2017), on the individual-level data, the imputation is based on running the model 50 times, yielding 50 complete datasets, the distribution of which is intended to reflect the sampling variability. The average of these 50 predictions is taken to be the estimated value of a specific asset whose value was not reported by its owner.
Overall, the LSMS+ supported surveys have a rich set of individual, household and community characteristics -along with data on features of asset ownership -that were used explanatory variables in our multiple imputation model that relies on predictive mean matching with 5 neighbors. The covariates that are used in the multiple imputation model are detailed in Appendix Table A2. These variables are strongly associated with asset valuation and factors related to the missing data generation process, and include survey design features in accordance with the guidance provided by Rubin (1996) and van Buuren et al. (1999). We provide a set of estimates that are obtained with no imputation and another set that is based on multiple imputation to recover missing asset values.

Approaches to treating multiple valuations on jointly-owned assets
Among joint owners of the same asset (for example, the same land parcel), while there was a substantial share of agreement over ownership (as discussed in Section 2.2), there was some variation in asset valuation. If some individuals or groups tend to systematically underreport their asset valuation relative to others, their overall wealth will be lower even among the same set of assets owned. Within each of the no-imputation and multiple imputation approaches, we examined how results would vary by either (a) retaining joint owners' self-reports (hence ignoring the intra-household discrepancy) or (b) assigning the maximum asset value reported among any of the joint owners.
In total, then, we have four different versions of the individual wealth variable that allows us to understand how the distribution of wealth and trends in wealth inequality change depending on differences sources of measurement error: (A) No imputation -using self-reports for joint owners,  Table 3 presents the valuation of assets, converting local currency estimates to USD, for men and women across the four specifications above. Because of the focus of the analysis on overall wealth in the population, and since there are many men and women in the country samples that do not own any assets, Table 3 examines the entire sample including non-owners -Appendix Table A6, however, includes the same estimates conditional on ownership. Significant differences are also indicated across men and women, within each specification.

Descriptive statistics on asset valuation and individual wealth
In general, the results show that as one moves from specification (A) to (D), the average value of landed assets tends to increase for both men and women, with greater increases for residential as opposed to non-residential land. Comparing (A) and (C), however, the magnitude of gender differences does not change substantially. The results also again highlight importance of examining different asset classes separately, when considering gender differences as well as the overall contribution to wealth. While the overall value of assets is not significantly different across men and women in Malawi, Ethiopia and Cambodia, for example, women's value of livestock assets is significantly lower than men's in Ethiopia and Cambodia, as well as for mobile phones, computers and motorized vehicles in Cambodia. Despite variations in ownership that were discussed earlier, as well as some differences in magnitude of land values (these disappear among the sample of owners in Appendix Table A6) the value of non-residential and residential land was not significantly different across men and women in these countries. In Tanzania, however, women's value of non-residential and residential land is significantly lower than men's. Figure 1, which relies on the asset value specification (C) (multiple imputation with self-reported values for joint owners) 14 , shows that important intra-household differences can be masked when only considering the distribution of household wealth. Even among the top 20 th quintile of household wealth, for example, around 20-30 percent of men and women in these households have either no assets or are in the bottom two quintiles of the individual wealth distribution. Appendix Table A7, which presents results from OLS household fixed effects regressions, shows that women in this group are more likely to be young (30 years of age or less), with less schooling, and less likely to be married (for men in this group, on the other hand, only age had a significant positive association). Figure 1 also shows that in Sub-Saharan Africa, among households in the top two quintiles (above 60 th percentile) of household wealth, the share of women in the top two quintiles of individual wealth is smaller than for men.
Since it is feasible that the factors affecting gender differences in wealth may vary along the distribution of wealth, Appendix Figure A1 presents predicted means of recentered influence function (RIF) regressions of men's and women's total wealth, estimated at 10 percentage point increments from 10 th to 90 th percentile (see Firpo, Fortin, and Lemieux, 2009, for a discussion). The recentered means at each percentile are presented separately for specifications (A) and (C), separately for men and women and by different age groups, given the strong links between age and overall asset values that were highlighted earlier in Table   5. The RIF regressions control for the same individual and household variables as in Appendix Table A2 along with geographic fixed effects, and the selected results from the estimations of RIF regressions at selected percentiles of men's and women's total wealth are presented in Appendix Table A8 to better illustrate distributional gender differences in relationships between covariates and total wealth. 15 Importantly, the RIF results show that gender inequalities in wealth widen for older (51+) age groups, compared to younger cohorts, although men and women in this group are more likely to own assets overall.

Comparing Measures of inequality using the individual wealth data
Using the total value of assets across different asset classes, Figure 2 presents estimates of competing measures of inequality among individuals as measured by Lorenz curves based on (i) household per capita expenditure, (ii) household per capita wealth (scaling individual wealth up to the household level), and (iii) individual wealth as measured by specification (C). A striking factor is how high inequality is in wealth overall, as compared to household per capita consumption expenditure (see Boyer, 2020, for a discussion of measurements of wealth and consumption inequality across countries, as well as underlying data issues).
The results underscore the importance of individual-level data collection on wealth: across countries, as   Table 4 compares additional inequality indicators, apart from the Gini, across different specifications of individual wealth: the 75/25 ratio, which compares how many times greater wealth is at the 75 th compared to the 25 th percentile of the distribution; as well as the Generalized Entropy measures GE(0) (mean log deviation, or Theil L index) and GE(1) (Theil T index). 16 Bootstrapped standard errors (250 repetitions) are also estimated for the different measures, and t-tests that were conducted revealed statistically significant differences across household per capita wealth and individual wealth for every wealth specification and inequality measure. The magnitudes of these differences do vary, with larger differences between different household per capita wealth and individual wealth inequality measures in Cambodia, Tanzania and Ethiopia.
We also find, for the most part, that inequality measures do not vary substantially by individual wealth specification (A)-(D), and particularly among the Gini and Theil indices. The 75/25 ratio, on the other hand, is more sensitive to specification, particularly in Ethiopia and Tanzania where there was a greater share of missing data on land valuation. 17 The general trend, regardless of which specification (A)-(D) of individual wealth we examine, however, is that individual-level data on wealth leads to greater variation and higher inequality estimates than household-level approaches. Results also did not change when dropping respondents for whom there was disagreement within the household over whether they owned specific assets.
16 Theil indices are based on ratios of incomes to the mean, and while they increase with greater inequality, they are not capped at one, like the Gini. 17 The ratio might have increased substantially in Ethiopia, for example, under the multiple imputation approach, where a sizeable share of asset values is missing would have been valued at zero under specification (A) but are then imputed with market prices that might in some cases be quite large/at the higher end of the distribution, depending on land size. In Tanzania, on the other hand, based on the distribution of land values, the multiple imputation model allocated land values that shrank the ratio. In Cambodia and Malawi, the ratio is less sensitive to specification.     Indiv. wealth specification: (1) (1) "Other" in the case of Cambodia includes computers, tuk tuks, boats and bicycles.

How can respondent selection affect the measurement of inequality?
Given the emphasis on self-reported asset ownership in the LSMS+, a natural question might arise as to how different individual-level data on wealth might be if collected instead by one representative per household (i.e., who would report as a proxy for others). Malawi provides a unique setting to make this comparison, where the LSMS+ supported Malawi IHPS 2016 was conducted concurrently with another national survey, the Malawi Fourth Integrated Household Survey (IHS4) 2016/17. 18 Both surveys were implemented by the Malawi National Statistical Office, relying identical household and agricultural questionnaires and field teams that covered both IHPS and IHS4 EAs. 19 Unlike the IHPS, the IHS4 did not include an individual questionnaire that included LSMS+ assets modules, and a "most knowledgeable" member was to report on other members' land ownership and valuation (see Kilic, Moylan and Koolwal, 2021, who compare statistics on men's and women's agricultural land ownership across the two surveys). 20 Both surveys identical household and agriculture questionnaires and following the same protocol for administering interview questions. The two national survey samples were also similarly distributed across interview months and survey strata (see Kilic, van den Broeck, Koolwal and Moylan, 2020, for a detailed discussion).  3 below presents the same Lorenz curve for the IHS4, which shows that even asking only a "most knowledgeable" household member about others' land ownership and valuation leads to a higher Gini coefficient than household per capita wealth. Table 10, however, shows that when comparing individuallevel wealth inequality estimates across the IHPS and IHS4, inequality across different measures is still significantly higher in the IHPS (and substantially greater in magnitude as well) compared to the IHS4. There were other differences as well; the Gini decomposition revealed that non-residential land contributed more to the Gini coefficient (factor share of 59 percent), compared to residential land (41 percent). Apart from collecting individual-level data, therefore, the mode of reporting (self-versus proxy) also matters significantly in estimating inequality, and can vary by asset class. Lower inequality in individuallevel wealth in the IHS4 may be due to the key respondent reporting the same valuation among all joint owners of a given asset. Even among those not owning jointly, however, the intra-household variation in reporting (as measured by the standard deviation in land values) for the most knowledgeable household member was also significantly lower in the IHS4 than in the IHPS (results available upon request). Relying strictly on one individual per household to report on asset valuations, therefore clearly leads to less

Conclusions and looking ahead
Using nationally-representative, self-reported survey data elicited from men and women in Sub-Saharan Africa (Malawi, Tanzania and Ethiopia) and Southeast Asia (Cambodia), this paper highlights how intrahousehold, individual-disaggregated survey data collection on ownership and valuation of physical and financial assets can case a brighter light on economic inequality among men and women. The data from the surveys supported by the World Bank LSMS+ program reveal significant (i) intra-household differences in asset ownership and wealth aggregates among men and women, and (ii)more pronounced estimates of overall inequality based on individual asset wealth vis-à-vis household per capita expenditures and per capita asset wealth. Land, in particular, is an important component of individual wealth inequality, and the contribution of residential versus non-residential land varies by country context. Compared to the survey approach of interviewing a single, most knowledgeable household member to report asset owners and values, self-reported information on asset ownership and valuation leads to greater estimates of wealth inequality. Additional sensitivity checks to address missing asset values and variation in asset valuation among joint asset owners yield similar results.
Overall, this paper makes an important contribution towards a more nuanced of understanding longerterm economic inequalities among men and women that are often disguised in existing survey data. The demonstrated utility of individual-disaggregated survey data on asset ownership and values further contributes to the broader momentum, including as part of the World Bank LSMS program, to improve the scope of and approach to data collection in household surveys for better capturing the extent and drivers of gender inequities in development outcomes. In the context of future large-scale household surveys in which the target variables include "wealth," the estimated parameters informing sampling design simulations (e.g. intra-cluster correlation) are expected to be different for individual wealth vis-à-vis per capita consumption expenditures. Our findings regarding the asset classes that contribute the most to the measures of wealth inequality among individuals could influence the approach to sampling design simulations and questionnaire design (particularly when assets modules cannot be as extensive as those implemented as part of LSMS+ supported surveys). Future methodological research can better document the drivers of measurement error in asset valuation, including missingness, variation in asset valuation among joint owners, and possibly with respect to third-party, more objective valuation approaches (highly dependent on the country context and asset class). (1) In Malawi, there was an issue in the data on how respondents identified other joint members for residential land ownership, so discrepancies could not be constructed.   Notes: (1) All estimates are weighted. Statistically significant differences between men and women, within each survey, are indicated by asterisks (***p<0.01, ***p<0.05, * p<0.10). * Excluded category is 35-44. †Indicates dependency ratio of children and elderly. Table A3. Socioeconomic correlates of asset ownership across countries -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000 -0.000***       (1) and (3)   Notes: (1) All estimates weighted with the household sampling weights, and standard errors clustered by enumeration area. ***p<0.01 **p<0.05 *p<0.10 (2) Estimates based on specification (C), i.e. imputing missing data on valuation, and using self-reports among joint owners. Results were similar for the other specifications (A), (B) and (D) as described in Table 4.