Does it matter which poverty measure we use to identify those left behind? Investigating poverty mismatch and overlap for Botswana

This study offers the first attempt in Botswana and adds to the limited literature on poverty mismatch and overlaps in Sub-Saharan Africa. Using the 2015/16 Botswana multi-topic household survey data, the study compares the country’s current official monetary poverty measure with an individual-level multidimensional poverty measure. The results show that multidimensional poverty levels are higher than monetary poverty levels. The results also reveal that significant mismatches and overlaps exist, suggesting that individuals experiencing monetary and multidimensional poverty are not the same. However, the mismatch size and overlaps vary across different subgroups of the populations and place of residence. The econometric estimation results show that age, household size, household head’s education status, household head’s employment status, and location (place of residence) influence poverty mismatch and overlap in Botswana. The findings suggest the need to go beyond traditional monetary poverty measure and complement it with multidimensional poverty measure to identify those left behind. The results are critical for policy interventions, especially for monitoring the trends, understanding poverty dynamics, and targeting social assistance programmes.


Introduction
The 2030 Agenda for sustainable development has reinforced poverty as one of the top priorities (UN 2015). Specifically, SDG 1 (target 1.1) aims to 'eradicate extreme (monetary) poverty for all people,' and SDG 1.2 aims to 'reduce at least by half the proportion of men, women and children of all ages living in poverty in all its dimensions' (UN 2015, 15). These two SDG targets make a clear link between monetary and multidimensional poverty. There is a need for a clear understanding of the links between monetary and multidimensional poverty measures to achieve SDG 1 (Ballón et al. 2018). Therefore, this study tries to provide an in-depth analysis of the link between monetary and multidimensional poverty measures.
The empirical literature on poverty measurement holds that there exists mismatch between monetary and multidimensional poverty (Baulch and Masset 1990;Roelen et al. 2012;Alkire et al. 2015;Kwadzo 2015;Tran et al. 2015;Bader et al. 2016;Suppa 2016;Roelen 2017Roelen , 2018Ballón et al. 2018;Salecker et al. 2020). Notwithstanding this, the research on poverty mismatch and overlap is limited in Sub-Saharan Africa (SSA) (Klasen 2000;Levine 2012;Salecker et al. 2020). However, the SSA region is associated with higher poverty levels (World Bank 2017) and slower poverty reductions (Burchi et al. 2020). Also, most countries in SSA have poor levels of public service provision, infrastructure development, and administrative and financial capacities (Salecker et al. 2020). Therefore, further attention is required to understand the poverty mismatch in SSA. This study is aware of only two in-depth studies on poverty mismatch in SSA by Levine (2012) for the case of Uganda, and most recently Salecker et al. (2020) for the case of Rwanda. Therefore, this study contributes to the limited literature on poverty mismatch in SSA.
Therefore, the purpose of this study is to examine the mismatch between monetary poverty and individual-level multidimensional poverty. This study addresses two primary research questions. First, do the two poverty measures produce similar results in identifying those left behind? Second, what factors contribute to the mismatch and overlaps between the official monetary and multidimensional poverty measures? This study conducts an empirical assessment in Botswana to answer these research questions. Botswana presents a fascinating case. Poverty measurement in Botswana is exclusively based on the monetary approach. Botswana has pursued poverty reduction since independence in 1966. This commitment has led to a decrease in monetary poverty from 59% in 198559% in /86 to 16.3% in 201559% in /16 (SB 201659% in , 2018. Impressive as they are, these figures do not tell the entire story of the country's poverty situation. Despite the significant progress in monetary poverty, Botswana has not had an equally impressive record on other key social indicators such as unemployment, rising inequalities, increasing HIV/AIDS prevalence rates, and child malnutrition. For example, the unemployment rate rose steadily from 10.2% in 1981 to 17.6% in 2015 (SB 2018). The HIV prevalence rate rose from 17.1% in 200417.1% in to 18.5% in 201317.1% in (SB 2014. Malnutrition among children has persisted, with chronic malnutrition estimated at 30% for children under five years (World Bank 2015). These trends indicate that Botswana has not successfully transformed national wealth into improvements in human development. Thus, it is important to examine poverty mismatch when using monetary and multidimensional poverty measures.
Monetary poverty is estimated using the official monetary poverty measure computed by Statistics Botswana (SB, 2018). Multidimensional poverty is calculated using an individual-level multidimensional poverty index that was developed using the theoretical premises of the capability approach (Sen 1985(Sen , 1999 and operationalized using the Alkire-Foster (AF) method (Alkire and Foster 2011a, b). Therefore, the present study compares poverty estimates based on Botswana's official monetary-based poverty measure with findings based on a country-specific individual-level multidimensional poverty index using the 2015/16 Botswana multi-topic household survey collected by Statistics Botswana.
Overall, the results show that multidimensional poverty levels are higher than monetary poverty levels in Botswana. First, consistent with the empirical literature from other countries (Baulch and Masset 1990;Bradshaw and Finch 2003;Levine 2012;Roelen 2017Roelen , 2018Kim 2019), the results reveal limited overlap in findings for monetary and multidimensional poverty. Second, monetary poverty identifies a smaller proportion of the population as poor than multidimensional poverty measures. Third, consistent with other studies from other countries (Bader et al. 2016;Roelen 2017Roelen , 2018Salecker et al. 2020), the results reveal a weak correlation between monetary and multidimensional poverty measures (and various multidimensional poverty indicators). Fourth, the results show that multidimensional poverty levels decline with increasing per capita consumption. However, consistent with other studies, there is a significant proportion of the multidimensional poor in the wealthiest households (Salecker et al. 2020). Last, the econometric estimation results show that age, household size, household head's education status, household head's employment status, and location (place of residence) influence the extent to which individuals are identified as monetary or multidimensionally poor. Therefore, this study concludes that monetary poverty measure alone does not capture the real picture of Botswana's poverty situation. Therefore, it should be complemented with a multidimensional poverty measure.
This study makes several contributions. First, to the best of our knowledge, this study is the first attempt to compare Botswana's monetary and multidimensional poverty measures. Therefore, this study contributes to the limited literature on poverty mismatch and overlaps in Botswana. Exploring poverty mismatch in a country-specific context allows for a deeper understanding from a broader perspective (Roelen 2017). Second, this study also makes a novel contribution to the limited literature on poverty mismatch in SSA (Levine 2012;Salecker et al. 2020) by using a country-specific individual-level multidimensional poverty measure. The study also adds to the debates on poverty mismatches globally (Baulch and Masset 1990;Kwadzo 2015, Bader et al. 2016). Third, the study extends the research on poverty mismatch in SSA by investigating factors influencing poverty mismatches using multinomial logistic regression (Roelen 2018).
The structure of the paper is as follows. The following section ("Literature on the mismatch of monetary and multidimensional poverty measures" section) presents brief literature on poverty mismatch, followed by "Data and methods" section, presenting data and methods. "Results and discussions" section presents results and discussions, and last, "Conclusions and policy implications" section provides conclusions and policy implications.

Literature on the mismatch of monetary and multidimensional poverty measures
Empirical literature points toward evidence of mismatch between monetary and multidimensional poverty (Baulch and Masset 1990;Bradshaw and Finch 2003;Laderchi 1997;Sumarto and De Silva 2014;Tran et al. 2015;Ballón et al. 2018;Roelen 2017Roelen , 2018Kim 2019). These studies' overall finding reveals that poverty measures based on monetary and multidimensional measures identify different groups of individuals or households as poor (e.g., Alkire et al. 2015;Tran et al. 2015;Roelen 2017Roelen , 2018. Notwithstanding rich literature on poverty mismatch, a few studies have investigated poverty mismatch in SSA using country-level analysis (Klasen 2000;Levine 2012;Salecker et al. 2020). In the case of Rwanda, Salecker et al. (2020) found that using a monetary measure alone does not capture the high incidence of multidimensional poverty. Also, the study found that the two measures differ in poverty risk factors. Levine (2012) found large discrepancies between the two measures in the case of Uganda. Klasen (2000) compared a standard expenditurebased poverty measure with a specifically created composite measure of deprivation for the case of South Africa and found that the two measures diverge significantly in identifying the poorest and most deprived sections of the population. Therefore, this study adds to the growing literature of poverty mismatch and overlaps in SSA.
Some studies have argued that individual and household characteristics and structural characteristics (including regional socioeconomic disparities) influence poverty mismatch and overlap (Klasen 2000;Roelen 2018;Tran et al. 2015). Based on this growing evidence of poverty mismatch and overlaps, other studies have examined factors associated with poverty mismatch (Perry 2002;Bradshaw and Finch 2003;Cancian and Meyer, 2004;Alessio et al. 2011;Bader et al. 2016;Ballón et al. 2018;Roelen 2018). Klasen (2000) found that ethnicity, gender and education of the household head to be associated with varying levels of poverty mismatch in South Africa. Tran et al. (2015) identified the same factors in the case of Vietnam. Bader et al. (2016) identified residence, ethnolinguistic families, and access to market as drivers of poverty mismatch for the case of the Lao People's Democratic Republic. Roelen (2018) found that household size, level of education and occupational status of the household head and place of residence significantly influence poverty mismatch in Ethiopia and Vietnam. This study contributes to the limited literature on factors influencing poverty mismatch in SSA.
Some studies also argued that measurement error might influence the mismatch between monetary and multidimensional poverty measures (Hulme and McKay 2008;Roelen 2018;Bradshaw and Finch 2003). The reliability of monetary measure concerning its equivalence scale and indicator for disposable income has been questioned (Brewer et al. 2009). Also, different units of analysis may lead to or compound measurement error since the monetary measure is aggregated at the household level, while multidimensional measure aims to include more individual-level indicators (Roelen et al. 2018). The monetary measure follows an indirect approach of measuring welfare and, therefore, cannot represent the standard of living of a family or individual (Ringen 1988). Therefore, it provides no way to verify the intra-household allocation of resources/income (Alkire and Santos 2014). Others point to the time and lagged effects since monetary indicators are more likely to fluctuate in the short term than non-monetary indicators (Roelen 2018). This study contributes to this literature by investigating factors influencing poverty mismatch and overlaps in SSA.
Other studies have investigated the correlation between income (and thus monetary poverty) and specific dimensions of deprivation (Klasen 2000;Laderchi et al. 1997;Alessio et al. 2011;Singh and Sarkar 2015;Bader et al. 2016;Suppa 2016;Roelen 2017Roelen , 2018. These studies find a weak correlation between the two measures and conclude that one measure cannot serve as a proxy for another (Klasen 2000;Roelen 2018). This study contributes to the literature in this regard. Some studies examined trends in income and multidimensional poverty measures in some countries in SSA (e.g., Alkire et al. 2017;Burchi et al. 2020). Burchi et al. (2020) found a weak correlation between monetary and multidimensional poverty changes, with monetary poverty declining significantly more than multidimensional poverty.
On the other hand, Alkire et al. (2017) found that multidimensional poverty decreased while monetary poverty increased for many countries. These findings show that the two approaches give slightly different results. However, other studies found a high correlation between extreme monetary poverty and multidimensional poverty using comparative cross-country analysis (Burchi et al. 2018). However, these studies did not examine overlaps and risk factors but only focused on aggregate trends.

Data
The analysis of this study utilizes the 2015/16 Botswana multi-topic household survey (2015/16 BMTHS hereafter) collected by Statistics Botswana (SB). This survey is a crosssectional and nationally representative survey, allowing for disaggregation by demographic characteristics, economic variables, and administrative district. The 2015/16 BMTHS collected socioeconomic information, among others, on demographic characteristics, household expenditure and consumption, labor force, health, education, self-assessed well-being and food insecurity, housing, utilities, durable goods and anthropometric measurements (see SB 2018). The dataset contains information on the monetary poverty measure, thereby allowing for the comparison of monetary and multidimensional poverty measures.
The dataset contains information from 24,720 individuals from 7,060 households surveyed in 2015/16. After applying sample weights, this resulted in an estimated 589,909 households and an estimated national population of 2,073,675 individuals (SB 2018). The survey employed a two-stage stratified probability sample design. The first stage was the selection of primary sampling units (PSUs), which were enumeration areas (EAs) using Probability Proportional to Size (PPS), where the measure of size is the number of households in an EA as defined in the 2011 Population and Housing Census. The second stage was the selection of occupied households within the selected EAs. A list of identified occupied households formed the basis of secondary sampling units (SSUs). Thus, the number of occupied households in each selected EA served as a sampling frame for that EA (SB 2018). Stratification was made based on the twenty-six (26) census districts that are heterogeneous and aligned to administrative districts. The districts were further grouped into cities/ towns, urban villages, and rural areas (SB 2018).

Botswana official monetary poverty measure
This study uses the Botswana official monetary poverty measure (BOMPM) to operationalize the monetary approach. The BOMPM is assessed based on consumption expenditure (SB 2018). The measurement of BOMPM relies on the absolute poverty line grounded in a nutritionally based food basket, supplemented by the allowance for non-food needs (World Bank 2015). The PDL is associated with the individual and household composition, considering household sizes, individual gender, age, and region (SB 2018). The computation of the PDL is based on the cost of a basket of goods and services considered necessary to meet household members' basic needs (SB 2018). To calculate the PDL, five components are used: food, clothing, personal items, household goods and shelter; and the cost of each of the five components of the PDL basket was calculated considering household size, individuals' gender and age, and region (SB 2018). Each of the five components' poverty lines is added to obtain each household's poverty line. The poverty line for each household is compared with the reported total consumption. If a household's total consumption 1 3 falls below the corresponding poverty line, then the household and every individual in that household is considered poor (SB 2018). The official poverty measure is reported as a headcount ratio or a percentage of the poor people in the population. It is computed using the Foster, Greer, and Thorbecke (1984) proposed class of poverty measures out of practical demand for decomposable poverty measure.

The individual-level multidimensional poverty measure
This study employs the individual-level multidimensional poverty measure developed by Lekobane (2021) for Botswana. The measure is based on the counting methodology developed by Alkire and Foster (2011a) (henceforth AF). 1 The AF methodology uses a two-step 'dual cut-off' process to identify the poor (Alkire and Foster 2011b). The first step involves identifying individuals as deprived in each indicator according to a dimension specific cutoff. In a second step, the number of weighted deprivations (deprivation scores, c i ) is calculated for each individual, and it is compared to a poverty cutoff (k), which determines the number of weighted deprivations a person must experience to be identified as multidimensionally poor (Alkire et al. 2017). 2 The AF methodology proposes a family of multidimensional poverty measures M α that is based on the FGT class of poverty measures (Foster et al., 1984) to solve the aggregation problem. This study uses the first measure of this family: the adjusted headcount ratio (M 0 ) and contains both multidimensional headcount ratio (incidence of poverty), H and the average deprivation scores, capturing the intensity of poverty, A . This study uses the first component (H) of M 0 to estimate individual-level multidimensional poverty incidences in Botswana.

Proposed dimensions, deprivation indicators and cutoffs
The capability approach, in conjunction with the human rights-based approach, informed the choice of dimensions and indicators. The study also relied on Botswana's policy commitments and development priorities such as Vision 2036, NDP 11, Botswana Poverty Eradication Policy and Strategy (BPEPS), the 2063 Africa Agenda (Agenda 2063), and the SDGs to ensure that the measure is contextually relevant (Lekobane 2021). However, the final list is based on data availability. The following seven dimensions are included in constructing the individual-level multidimensional poverty measure: (i) assets, (ii) housing and living conditions, (iii) water and sanitation, (iv) food security, (v) health, (vi) education, and (vii) security.
The asset dimension measures deprivations related to the possession of household assets. In reference to the capability approach, assets are closely connected with ends (functionings) they facilitate (Alkire and Santos 2014). For instance, having a car or van constitutes the functioning of 'being able to transport oneself.' Four deprivation indicators are considered in this dimension: information, durable goods, transport and tenure. The housing and living conditions dimension captures whether the housing is adequate and is related to health. This dimension captures the functioning of 'being well-sheltered' (Nussbaum 2003). Six deprivation indicators are considered for this dimension: overcrowding, cooking fuel, electricity, floor material, roof material and wall material. This dimension is captured by SDG 11 of the SDGs (UN 2015).
Water and sanitation are linked to health and is considered as a standalone dimension (UN and WHO 2010). The water and sanitation dimension is reflected in SDG 6 (UN 2015) and is also recognized by Agenda 2063 as critical dimensions (AUC 2015). This dimension is captured by two indicators: water supply and toilet facility. The food security dimension is captured using two indicators: food access and nutrition. The food access indicator captures the functioning of 'being free from hunger' (Drèze and Sen 1989;Burchi and De Muro 2016). The nutrition indicator goes beyond the 'access' indicator and captures the functioning of 'being well-nourished' (Sen 1992). Food security is considered a right (UN 1948) and is reflected in SDG 2 of the SDGs (UN 2015) and Agenda 2063 (AUC 2015).
Health is included for its intrinsic as well as instrumental value (Klasen 2000;Alkire and Santos 2014). The health dimension captures deprivations related to access and quality of the nearest health facility and chronic illness. The first indicator is the condition of the nearest health facility capturing the perceived quality of the nearest health facility and problems associated with the health facility. The second indicator is a chronic illness and captures the functioning of 'being healthy' (Rippin 2016). Health is captured by SDG 3 of the SDGs and is considered a basic human right (UN 1948). Education, like health, has intrinsic and instrumental value (Klasen 2000). Not being educated constitutes capability deprivation . Education is a right to which all human beings are entitled (UN 1948) and has a standalone goal in the SDGs (SDG 4). The education dimension is captured using three deprivation indicators: child enrolment, school attainment and literacy.
The security dimension is measured using two indicators (safety and crime). The safety indicator captures the functioning of 'being able to move freely from place to place' while the crime indicator is directly linked to the functioning of 'being secure against crime or violence' (Nussbaum 2000(Nussbaum , 2005. In the SDGs, this dimension is reflected in SDG 16 (UN 2015). Table S3 in the supplementary tables presents the proposed dimensions, deprivations indicators, the deprivation cutoffs, identification level and groups for which the indicators are applicable.

Weighting of dimensions
Different weighting approaches exist in the literature, and these include normative, data-driven and hybrid approaches (Alkire and Santos 2014; Decancq and Lugo 2013). Data-driven and hybrid approaches rely on statistical methods to drive weights, making comparability over time more complex (Decancq and Lugo 2013). Therefore, following other studies in the empirical literature of multidimensional poverty measurement (e.g., Alkire and Santos 2014;Ervin et al. 2018;Burchi et al. 2021), and based on a normative approach, this study adopts an equal weighting scheme across dimensions and equal nested weights within dimensions for each indicator. Setting equal weights is not a choice free from value judgements: it implies assuming that each dimension used reflects their equal importance as constituents of poverty (Burchi et al. 2021). In this case, the choice can be easily justified: the LNOB principle is premised on the human rights approach, and rights are deemed to be equally important. Table S3 in the supplementary tables presents the proposed dimensions, deprivations indicators, the deprivation cutoffs, identification level and groups for which the indicators are applicable. However, actual weights per indicator will differ across age groups as the total number of indicators differs across age groups (as a result of using the individual as a unit of analysis). The constructed MPI is robust for different poverty cutoffs (k values) and changes in weighting structure (w) (see Alkire et al. 2015;Lekobane 2021), meaning our index is stable and can be used for policy analysis.

Analytical strategy
The analytical strategy involves both descriptive and regression analysis. The descriptive analysis consists of three components. Firstly, following other studies (e.g., Tran et al. 2015), this study compares monetary and multidimensional poverty rates across different population sub-groups. Secondly, the study investigates the correlation between consumption and multidimensional poverty (Roelen 2017;Suppa 2016;Salecker et al. 2020). Thirdly, the mismatch and overlap between the two poverty methods are examined. Following other studies analyzing overlap and mismatches (e.g., Baulch and Masset 1990;Laderchi et al. 1997;Suppa 2016;Roelen 2017Roelen , 2018Salecker et al. 2020), a cross-tabulation of monetary and multidimensional poverty incidence is estimated. The cross-tabulation yields a four-cell matrix, representing four different 'poverty categories' (Suppa 2016;Roelen 2017Roelen , 2018. These categories are (i) poverty overlap (both monetary and multidimensionally poor); (ii) positive mismatch (monetary poor but not multidimensionally poor); (iii) negative mismatch (multidimensionally poor but are not monetary poor); and (iv) no poverty overlap (not monetary and not multidimensionally poor) (see Roelen 2018).
The study then employs a multinomial logit model to examine factors contributing to the poverty mismatch. The multinomial logit model is the most commonly applied when examining multiple unordered categorical outcomes. Since the dependent variable ('poverty group status') comprises nominal (no ordering) outcomes, multinomial logit is the appropriate model. Independent variables at the individual level include individual characteristics, household characteristics and community indicators. Individual characteristics include gender, age, disability status and household level variables include gender, age, marital status, educational attainment and employment status, household size and location (see Roelen 2018;Salecker et al. 2020). These selected independent variables are commonly used in the literature as key determinants of poverty (Grootaert 1997;Baulch and McCulloh 2002;Leu et al. 2016;Qi and Wu 2016;Lekobane and Seleka 2017;Salecker et al. 2020).

Incidence of monetary and multidimensional poverty
A comparison between estimates of monetary and multidimensional poverty incidence is presented in Table 1. Overall, the two poverty measures produce significantly different estimates of poverty incidence, with monetary poverty having significantly lower poverty rates. Based on the official monetary poverty measure, the incidence of poverty P 0 is estimated at 16.3% compared to 46.2% of multidimensional poverty incidence (H) . Multidimensional poverty incidence is higher than monetary poverty incidence across all population subgroups, as evidenced by positive differentials between multidimensional and Concerning gender, females exhibited slightly higher poverty incidences than males, regardless of the poverty method used. However, the differences are very minimal. For example, the difference in poverty between males and females is only 0.9 and 1.2 percentage points based on monetary and multidimensional poverty measures, respectively. Regarding age, poverty rates based on the monetary measure exhibit a U-shaped relationship with age with children experiencing the highest poverty level, followed by people aged 65 and over. In contrast, multidimensional poverty findings reveal a positive correlation with age, meaning that the likelihood to be multidimensionally poor increases with age. The rankings of poverty incidences regarding disability status reveal a contrasting picture. Monetary poverty incidence is higher among people with no disability compared to PWDs. The opposite is true for multidimensional poverty, with higher poverty incidence for PWDs than those without a disability. The change in poverty incidence between the two measures is more significant for PWDs (58.7%) than those with no disability (29.2%).
Poverty incidence is consistently higher for citizens than non-citizens regardless of the poverty method used. The change in poverty incidence between the two measures is more than double (30.5%) for citizens compared to non-citizens (13.2%). Across household headship, individuals residing in female-headed households exhibited slightly higher poverty incidences than those in male-headed households regardless of the poverty measure used. Poverty rankings differ concerning the age of the household head. There is a positive correlation between monetary poverty incidence and the household head's age, meaning that households headed by children have lower risks of being monetary poor and vice versa. However, these results should be treated with caution since child-households account for the lowest shares (0.2%) of total households.
In contrast, rankings based on multidimensional poverty measure reveal a U-shaped relationship with the age of household head, meaning that multidimensional poverty incidences decline with an increase in age of household head up to a certain point after which they increase. Poverty rankings showed mixed and different results between the two poverty measures based on the household head's marital status. For example, individuals from households headed by married couples recorded multidimensional poverty rates (32.6%), ranking first (1). In contrast, in the case of monetary poverty, those from households headed by divorced persons recorded the lowest poverty incidences (9.8%), ranking first (1). The rankings also changed for all other marital status categories except those from households headed by widows/widowers and those who never married.
Rankings also differ between the two measures concerning the household size. When using the monetary measure, poverty levels are positively correlated with household size, meaning that an increase in household size will increase monetary poverty levels. However, based on multidimensional poverty measure, the results reveal a U-shaped relationship ∆ Incidence is the difference between H and P 0 (H-P 0 ). H: Multidimensional poverty incidence; P 0 : Poverty incidence based on monetary poverty measure. Number in parentheses represent rankings a All percentages are estimated at population-level using sample weights between poverty rankings and the household size, meaning higher multidimensional poverty incidences decline with an increase in household size to a certain level, after which they increase. Poverty rankings exhibited a negative correlation with household heads' educational attainment, meaning household heads with higher educational attainment have a lower risk of being either monetary or multidimensionally poor. The disparities between the two measures also decline with improvements in educational achievement. Results reveal a mixed and different picture with respect to the employment status of the household head. For example, based on monetary poverty measure, individuals from households headed by the unemployed recorded the highest poverty incidence, ranking last (fifth rank). In contrast, those from households headed by family helpers recorded the highest multidimensional poverty incidences (fifth rank). Figure 1 depicts the incidence of monetary and multidimensional poverty disaggregated by districts. Overall, Fig. 1 depicts varying scenarios based on the two measures across districts. For example, Ngamiland West recorded the highest multidimensional poverty incidence while it ranked fifth based on monetary poverty. The results also show that poverty levels are lower for urban districts (cities and towns) regardless of the poverty measure used. 3 Except for Orapa and Sowa Town, poverty incidence is higher when using multidimensional poverty than monetary poverty measure. This finding is not surprising since these two districts are mining towns with good infrastructure, and most services such as health are provided freely by the mines. In addition, the ranking of poverty levels based on the two measures is examined (Table S1 in supplementary tables). The rankings show diverse differences. Only one district (Sowa Town) maintained its ranking as the least poor district regardless of the poverty measure used. The rankings for all other districts are  different across the two measures, with some districts showing more significant disparities than others. For example, North East ranked second based on monetary poverty and eleventh when using a multidimensional poverty measure. In contrast, Orapa is better off when using multidimensional poverty measures, ranking second than the sixteenth rank when using monetary poverty measures. The differentials in poverty incidences between multidimensional and monetary measures are positive except for Orapa and Sowa, where the opposite is true. These results show that relying on multidimensional poverty alone for policy interventions can be misleading.

Patterns of mismatch between monetary and multidimensional poverty
It is paramount to examine whether the two measures identify the same or different subgroups of the population as poor to shed more light on the differences or similarities. First, the population is divided into four groups: A (monetary poor only, representing negative mismatch); B (multidimensionally poor only, representing positive mismatch); AB (both monetary and multidimensionally poor, representing overlap) and C (non-poor). Table 2 presents the summary results of the observed patterns of mismatch in Botswana. The results reveal that the two measures exhibit significant differences in identifying who is poor. In 2015/16, only 12% of the population was identified as poor by both measures (AB), and roughly 38.5% of the people are either monetary poor (4.3%) or multidimensionally poor (34.2%) (A + B). Of those identified as monetary poor, 73.9% were also identified as multidimensionally poor. About 26% of multidimensionally poor individuals were also monetary poor. Conversely, among individuals not identified as poor by the monetary measure, 40.9% were multidimensionally poor. About 74% of the multidimensionally poor were not poor in monetary terms. The population is divided into income groups (consumption per capita quintiles) to examine further whether monetary and multidimensional poverty measures are related. Figure 2 depicts the results of multidimensional poverty incidence across quintiles based on three poverty cut-offs. 4 The results show that multidimensional poverty levels decrease with increased consumption levels. These results mean that the higher the consumption, the less likely one is, on average, to be multidimensionally poor, regardless of the poverty cutoff chosen. These results suggest that income is essential for avoiding multidimensional poverty in Botswana. The results are consistent with those of Suppa (2016) for Germany  (2020) found only a minimal mitigating effect of consumption on being multidimensionally poor. Estimated multidimensional poverty rates remain high among individuals at the highest consumption levels. However, the wealthiest quintiles exhibited the lowest multidimensionally poverty levels in Botswana.

Correlation between monetary and multidimensional poverty indicators
Following other studies (e.g., Suppa 2016;Bader et al. 2016;Roelen 2017)), the correlation between monetary and multidimensional poverty indicators is examined. This paper used tetrachoric correlation since monetary and multidimensional poverty indicators are dichotomous (Agresti 2010). A point-biserial correlation was used to examine the association between continuous (consumption) and dichotomous variables (MPI indicators). Table 3 presents the results. Overall, the results reveal a weak positive correlation between multidimensional and monetary poverty. The correlation between monetary and multidimensional poverty measures is estimated at 0.439 and is statistically significant, meaning the correlation between the two poverty measures is very limited. Also, multidimensional poverty and per capita consumption exhibited a negative and limited correlation estimated at − 0.137. This finding is consistent with other similar studies in the empirical literature (Roelen 2017). For example, Roelen (2017) found a weak negative correlation between multidimensional child poverty and real per capita consumption in Vietnam.
The results further reveal a weak correlation between monetary poverty and the indicators underpinning the overall measure of multidimensional poverty. Deprivation indicators are positively and weakly related to monetary poverty except for crime, chronic illness and land tenure. The negative correlations between monetary poverty and land tenure, chronic illness and crime are not surprising since monetary poverty incidences are higher for the non-deprived than deprived individuals for these indicators. In contrast, for the rest of the indicators, the opposite holds. The correlation between deprivation indicators and per

Patterns of mismatch across individual and household characteristics
This study examines the extent of the mismatch and overlap between those identified as monetary poor and those identified as multidimensionally poor across different subgroups of the population, such as individual and household characteristics, to gain an indepth understanding of the mismatch patterns. Table 4      Concerning gender, the results reveal that females have slightly higher poverty levels regardless of the poverty method used. However, the differences concerning poverty mismatches (negative and positive) are very minimal. For example, based on the monetary poverty measure alone, poverty incidences for females are only 0.1 percentage points higher than males. In comparison, the difference is only 0.5 percentage points for multidimensional poverty measure alone. Poverty overlaps are slightly higher for females at 12.4% compared to 11.6% for males. Age reveals contrasting results concerning the negative mismatch and positive mismatch. Based on monetary measure alone, age is negatively related to poverty incidences, with children exhibiting the highest poverty incidences of 6.4% compared to 1.4% for older persons. On the other hand, the negative mismatch reveals a positive relationship with age. Older persons exhibit the highest poverty incidences of 62.2% compared to 28% for children, based on multidimensional poverty measure alone. However, poverty overlaps reveal mixed results, with older persons exhibiting higher rates followed by children.
Regarding disability status, the results are mixed. Regarding positive mismatch, people with no disability have higher poverty (4.3%) than PWDs (1.3%) based on monetary measure alone. However, the opposite is true based on multidimensional poverty measure alone, with PWDs having a higher negative mismatch at 60.5% compared to 33.5% of those with no disability. Concerning overlaps, the proportions of PWDs are slightly higher (12.8%) than those without disabilities (12%). With respect to citizenship, citizens exhibited the highest poverty mismatch and overlap, meaning citizens have higher poverty levels than non-citizens regardless of the poverty measure used. Poverty overlap for citizens is more than seven times higher than non-citizens at 12.8% compared to only 1.6% for citizens and non-citizens, respectively.
Regarding the gender of household head, the results reveal slightly higher negative and positive mismatches and limited overlaps for female-headed households than male-headed households. This finding means that no matter which poverty method is used, poverty levels are slightly higher for female-headed households than male-headed households. It should be noted, however, that the differences are very minimal. For example, based on monetary measure alone, the difference between positive mismatch for individuals residing in female-headed households and those in male-headed is only 0.6 percentage points. Based on the multidimensional poverty measure alone, those from female-headed households have a slightly higher negative mismatch at 36.8% compared to 31.5% of those in male-headed households.
In terms of household headship, results reveal that based on monetary poverty measure alone, age of household head is positively related to positive mismatch. This result means poverty incidences based on monetary measure alone increase with an increase in age of household head, with individuals from households headed by older persons exhibiting the highest rates of 5.2%. In contrast, those from households headed by children recorded no HH stands for the household head. A: monetary poor but not multidimensionally poor; B: multidimensionally poor but not monetary poor; AB: overlaps; C: non-poor a All percentages are estimated at population-level using sample weights. Sample size: 24,720 poverty incidence. A similar trend is observed for poverty overlap. However, based on multidimensional poverty measure alone, results reveal a nonlinear relationship between age of household head and negative mismatch. That is, poverty incidence based on multidimensional poverty measure declines with an increase in household head age to a certain point, after which it starts to rise again. Results are mixed concerning the marital status of the household head. For example, based on the monetary poverty measure alone, individuals from households headed by widows/widowers exhibited the highest positive mismatch levels than all other households. However, based on the multidimensional poverty measure alone, individuals from households headed by the never-married recorded the highest negative mismatch, while those from households whose heads are separated exhibited the highest poverty overlaps. Household size positively relates to positive mismatch and overlaps, with larger households exhibiting higher levels than smaller households. Based on monetary measure alone, household size is positively related to positive mismatch. Individuals from smaller households with 1 to 3 members recorded a 1.1% positive mismatch compared to 8% for larger households with more than seven members. The same trend is observed when using both measures, with larger households exhibiting higher overlaps of 23.7% for larger households compared to only 3.1% for smaller households. However, using multidimensional poverty measure alone reveals a U-shaped nonlinear relationship between household size and negative mismatch. This finding means that an increase in household size at lower levels reduces negative mismatch to a certain threshold. A further increase in household size increases negative mismatch.
Overall, poverty incidences decline with improvements in the household head's educational attainment. Based on monetary poverty measure alone, individuals from households whose heads have no educational qualification have the highest positive mismatch of 5.9% compared to 1.5% of those in households whose heads have attained university qualification. The same trend is observed when using multidimensional poverty measure alone, with individuals from households whose heads have no educational qualification recording a negative mismatch of 44.5% compared to 8.4% for households headed by persons with a university qualification. A similar trend is observed when using both measures (overlaps). This finding means that regardless of the poverty method used, families headed by individuals with no educational attainment have higher poverty levels.
Regarding the employment status of the household head, the results are mixed. Among the monetary poor only (positive mismatch), households headed by the unemployed have higher levels. In contrast, households headed by family helpers have the highest poverty levels for negative mismatch and overlaps. This finding is not surprising since individuals from households headed by family helpers have higher multidimensional poverty incidences than those from households headed by unemployed individuals. This descriptive analysis confirmed that poverty mismatch and overlaps exist between monetary and multidimensional poverty measures and differ across different population subgroups.

Patterns of mismatch across administrative districts
The patterns of mismatch and overlap across districts are examined (Fig. 3). Overall, the results reveal diverse patterns of mismatch and overlap across districts and that most of the poor people are over-represented in the negative mismatch (B), implying that most people are multidimensionally poor but not monetary poor. However, in Orapa, the opposite is true, where the positive mismatch is higher than the negative mismatch, implying that many people are considered monetary poor but not multidimensionally poor. The results are not surprising since Orapa is a mining town, and the residents are mostly mining workers and their relatives. Also, most services such as water, electricity, education and health are provided free by the mine, resulting in very low multidimensional poverty incidences among the residents. Another interesting observation is that in Kweneng West, where poverty overlaps (AB) are higher than both positive (A) and negative (B) mismatches, meaning the poorest of the poor are found in Kweneng West. The results for Kweneng West are not surprising. Despite its proximity to the city, the Kweneng West district lacks infrastructural development leading to limited employment opportunities. For example, after 50 years since the country gained its independence, the district has no senior secondary school and low access to health facilities. This district has the lowest literacy rate of 75.3% compared to the national average of 88.7%. The unemployment rate is estimated at 24% compared to the national average of 17.6% (SB 2018). Kgalagadi South also reveal overlaps slightly higher than poverty overlaps. 5 Across strata, rural areas exhibited higher negative poverty mismatch and poverty overlaps (   Table 5 reports the estimated multinomial regression results. The reference category is the non-poor category. This category comprises individuals who are neither monetary nor multidimensionally poor. The log-likelihood ratio (LR) test is significant, showing that there exists a significant relationship between the dependent variables and the explanatory variables included in the estimated model (p < 0.001; R 2 = 0.1918) (Hosmer and Lemeshow 2000;Long and Freese 2006). The significant result of the LR test means that at least one of the regression coefficients in the model is not equal to zero, indicating the overall model is a good fit (Long and Freese 2006). The interpretation of the results is based on relative probabilities (also called relative odds), obtained by exponentiating the estimated coefficients as 100(e − 1) , where τ represents the estimated coefficient for the considered independent variable (Giles 2011; Halvorsen and Palmquist 1980). 6 For example, living in rural areas is associated with 100.7%, 152.6% and 288.9% relative probability of being in positive mismatch, negative mismatch, and poverty overlap (respectively) than being non-poor than living in urban villages. The econometric results confirm the findings based on descriptive analysis. Overall, the regression model results reveal that age, household size, household head's education status, household head's employment status, and location (place of residence) are significant determinants of poverty mismatch and overlap. However, the magnitude of their impacts differs in terms of size and signs. These factors were identified as significant determinants of poverty overlap and mismatch elsewhere in the empirical literature (e.g., Roelen 2018).

Factors affecting poverty mismatch and overlap
The interpretation of the results is limited to only statistically significant variables across all three models. The relative probability of being in a negative mismatch than being nonpoor is higher for males than for females. Age reveals a nonlinear U-shaped effect on negative mismatch and poverty overlap, meaning the relative poverty of being in negative mismatch or poverty overlap than being non-poor declines at lower age levels but increases at higher age levels. People living with disabilities have a higher relative probability of being in negative mismatch than being non-poor compared to those with no disabilities. The relative probability of being in a negative mismatch than being non-poor is higher for citizens than for non-citizens. The same results are observed for poverty overlaps. Household size reveals a nonlinear (inverted U-shaped) effect on poverty overlap and negative mismatch, meaning the relative probability of being in positive mismatch or poverty overlap than being non-poor increases at lower household size levels but declines with an increase in levels of household size. The opposite results are observed for negative mismatch.
Individuals residing in male-headed households have higher probabilities of being in poverty overlap than being non-poor compared to those in female-headed households. Living in households headed by cohabiting couples or those who never married is associated with higher relative probabilities of being in a positive mismatch than being non-poor compared to living in households headed by married couples. The relative probability of being Robust standard errors (SE) clustered at the household level. Significance levels: *p < 0.1; **p < 0.05; ***p < 0.01 in a negative mismatch than being non-poor is lower for individuals living in households headed by married couples than individuals residing in any household type. The relative probability of being in a poverty overlap than being non-poor is higher for individuals whose heads are living together, separated, widowed, or never married than individuals living in households headed by cohabiting couples, married couples.
The relative probability of being in a positive mismatch than being non-poor is higher for individuals residing in households headed by youth than for those in households headed by adults. Results for older persons are mixed. Individuals living in households headed by older persons have lower relative probabilities of being in positive mismatch or poverty overlap than being non-poor compared to those living in households headed by adults. In contrast, the relative probability of being in a negative mismatch than being non-poor is higher for individuals living in households headed by older persons than those headed by adults. The relative probabilities of being in positive mismatch, negative mismatch, or poverty overlap than being non-poor are lower for individuals living in a household with a head having some form of educational attainment than those living in households whose heads have no educational attainment. The magnitude of the impacts increases with increased academic level, with the highest impacts observed for individuals living in households whose heads have a university qualification.
Concerning the employment status of the household head, results are mixed. Living in a household headed by someone engaged in formal paid employment is associated with lower relative probabilities of being in positive mismatch, negative mismatch or poverty overlap than being non-poor compared to living in a household headed by someone unemployed. The relative probability of being in positive mismatch or poverty overlap than being non-poor is higher for individuals living in a household whose head is engaged in self-employment than individuals in a household headed by someone unemployed. In contrast, the relative probability of being in a negative mismatch than being non-poor is higher for individuals living in a household whose head is engaged in self-employment than individuals living in a household headed by someone unemployed. Living in a household whose head is engaged in their own farm is associated with higher relative probabilities of being in positive mismatch, negative mismatch, or poverty overlap than being non-poor compared to living in a household headed by an unemployed person.
The results reveal that individuals residing in rural areas have higher relative probabilities of being in positive mismatch, negative mismatch or poverty overlap than being non-poor compared to living in urban villages. The results are mixed for cities and towns. The relative probability of being in a positive mismatch than being non-poor is higher for individuals living in cities and towns than those in urban villages. The same results are observed for poverty overlap. However, individuals living in cities and towns have lower relative probabilities of being in negative mismatch than being non-poor compared to those in urban villages.

Conclusions and policy implications
This study compares poverty estimates based on the official monetary poverty measure and the individual-level multidimensional poverty measure for Botswana using the 2015/16 BMTHS dataset to investigate poverty mismatch and overlaps. Also, the study investigates factors influencing poverty mismatch and overlaps. Despite overwhelming evidence of poverty mismatches and overlaps, few studies have been carried out in SSA (Klasen 2000;Levine 2012;Salecker et al. 2020). This study fills this gap and contributes to the literature of poverty mismatch in the SSA context.
The results show that multidimensional poverty levels are higher than monetary poverty levels. The results also reveal significant mismatches and overlap. However, the mismatch size and overlaps vary across different subgroups of the population and residence. These results are consistent with findings in the empirical literature (Baulch and Masset 1990;Bradshaw and Finch 2003;Klasen 2000;Laderchi 1997;Levine 2012;Sumarto and De Silva 2014;Roelen 2017Roelen , 2018Kim 2019). This finding confirms that monetary and multidimensional measures identify different groups of individuals as poor Roelen 2017Roelen , 2018. Compared to SSA countries, Botswana recorded lower poverty overlaps, with only 12% of the population identified as poor by both measures. For example, in Uganda's case, 23% of the population was identified as poor by both measures (Levine 2012), and 23.6% of the individuals were identified as poor by both measures in Rwanda in 2013/14 (Salecker et al. 2020). The significant discrepancies in overlaps between Botswana and other SSA countries (Rwanda and Uganda) are attributed to these countries' high monetary poverty, leading to low poverty differentials between the two measures. For example, in Rwanda, the monetary and multidimensional poverty rate difference is 6.7 percentage points compared to 26.3 percentage points for Botswana.
The correlation between monetary and multidimensional poverty measures shows a positive but weak association. Also, per capita consumption shows a weak association with specific dimensions of deprivation and overall multidimensional poverty measure. Similar evidence appears in the empirical literature (Klasen 2000;Laderchi et al. 1997;Alessio et al. 2011;Singh and Sarkar 2015;Bader et al. 2016;Roelen 2017Roelen , 2018Ballón et al. 2018).
Following similar studies (Bradshaw and Finch 2003;Bader et al. 2016;Ballón et al. 2018;Roelen 2018), this study investigated factors influencing poverty mismatches and overlaps and found that individual and household characteristics, as well as regional socioeconomic disparities, influence poverty mismatch and overlap in Botswana. For example, the relative probability of being in positive mismatch, negative mismatch or poverty overlap than being non-poor is lower for individuals residing in households whose heads have higher educational achievements than those whose heads have no educational qualification. Regarding the place of residence, those living in rural areas have higher relative probabilities of being in positive mismatch, negative mismatch or poverty overlap than being non-poor compared to those in urban villages. This finding is consistent with the empirical literature (Klasen 2000;Tran et al. 2015;Bader et al. 2016).
The conclusion that individuals identified as poor by the monetary measure are different from those identified as poor by the multidimensional poverty measure has important policy implications. First, consistent with other findings elsewhere (Roelen 2017(Roelen , 2018Bader et al. 2016;Ballón et al. 2018), this evidence from Botswana suggest that the official monetary poverty measure cannot be used as a proxy for multidimensional poverty measure and vice versa. This finding means that using the official monetary poverty measure alone does not capture the real picture of Botswana's poverty situation. Therefore, there is the need to go beyond traditional monetary poverty measure and complement it with multidimensional poverty measure to identify those left behind.
Second, the weak correlation between monetary and multidimensional poverty means that targeting social assistance programmes based on monetary poverty alone may not be effective. For example, North East districts recorded the lowest monetary poverty levels (ranking second). However, the North East district recorded higher poverty levels and moved to the eleventh rank when using a multidimensional poverty measure. Therefore, relying on official poverty measure alone may send inadequate information to policymakers resulting in weak policy designs, which will yield low impacts on poverty eradication.
Third, policymakers should consider the effects of different factors influencing poverty mismatch and overlaps in designing appropriate policies and programmes for poverty eradication. The use of multidimensional poverty indicators to supplement the monetary measure may assist in monitoring the trends and understanding the multifaceted forms of poverty. Therefore, complementing the current official monetary poverty measure with a country-specific individual-level multidimensional poverty measure would help policymakers better understand the real poverty situation in Botswana and help them put appropriate and specific policy mechanisms in place.