Integration into the South African core economy: household‐level covariates

Abstract

In this analysis of household survey data, households' main income sources are used as indicators of integration into the South African core economy. The allocation of main income sources is studied as the outcome of households' demographic composition, geographic location and earners' characteristics. The emerging picture of household income generation is one that disputes the common perception of African households as raising their incomes from a multitude of sources. The majority of surveyed households rely to a large extent on a single source of income and a single income earner. Separate multinomial logit models are estimated for urban and non‐urban households where, in addition to the considerable association with non‐urban residence, prominent earner covariates of low‐integration income sources are female gender, old or young working age, and low levels of education. Both provincial location and within‐provincial, subregional locations display strong impacts. The study also finds associations between main income sources and households' demographic compositions that are compatible with findings both in studies on private transfer behaviour and in the growing literature on endogenous household formation in South Africa.

1 INTRODUCTION

The relevance of income sources to the welfare of the South African household can be illustrated by two recurrent findings in research on poverty and inequality in the country. First, it is widely recognised that poor households derive a substantial share of their income from either remittances sent by migrant family members or public pensions. Secondly, access to wage income has been attributed considerable importance in avoiding poverty and is a crucial explanatory factor in income inequality (e.g. Carter & May, 1999; Leibbrandt & Woolard, 1999; Leibbrandt et al., 2000; Van der Berg, 2000). There exist several historical reasons to expect households' access to wage income or dependence on transfer incomes to be associated with other microeconomic factors. This study therefore analyses variation in households' main income sources of various origins, as a process associated with households' demographic composition, location and earner characteristics.

The South African setting for household income generation can be summarised by four characteristics: very high urbanisation rates; peasant agriculture being mostly absent among the rural non‐white population; labour markets often being inaccessible in rural areas; and, typically, very high unemployment rates in both rural and urban areas. However, the articulation of a complete model for the allocation of income sources in this setting is not a goal of this study. Our investigation focuses on the differences in main income sources among South African households, and on whether any empirical regularities can be identified in characteristics of households that would explain these differences. The study is not entirely dissimilar in scope from other investigations (e.g. Lipton et al., 1996; Carter & May, 1999; Posel, 2001; Leibbrandt & Woolard, 2001), but it augments previous research in several ways.

Conceptually, the quantitative analyses draw on the finding that a considerable proportion of households derive the bulk of their income from a single main income source, which is very often earned by one or a few individuals in the household. Categories of main income sources are further designed to reflect households' degree of integration into the core economy of South Africa. In terms of methodology, the concentration of household earnings around one main income source warrants an approach whereby households' allocation to income source categories is analysed, rather than the more common investigation into households' average shares of income from various origins. The allocation process is analysed through the estimation of the probability for households of holding each specific type of main income source, as associated with a group of explanatory variables. Probability models are estimated separately for the rural and urban areas using two multinomial logit regression frameworks.

The article proceeds in the following manner. Section 2 introduces the 1995 October Household Survey data (StatsSA, 1996) and explains how this sample is constructed based on the study's definition of a main income source. Section 3 provides the information about the context in which household‐level characteristics would have assumed their impact on main income variation in South Africa. A discussion based on descriptive statistics in Section 4 provides an informal assessment of the representativity of the main income source concept and provides an income level and labour market context. In Section 5, the empirical modelling and variables are introduced. The results from regression analyses are presented in Section 6 and conclusions are drawn in Section 7.

2 THE DATA, SAMPLE DELIMITATION AND DEFINITION OF MAIN INCOME SOURCE

In October 1995, Statistics South Africa (StatsSA) conducted questionnaire‐based interviews on a wide range of living standards issues with almost 30 000 households, representing all households in the country and containing nearly 131 000 inhabitants. Two months later, most of the households were revisited in a more detailed investigation of their incomes and expenditures. These two surveys are often referred to as the 1995 October Household Survey (OHS) and the 1995 Income and Expenditure Survey (IES). In the two surveys, a household is defined as ‘a person or a group of people dependent on a common pool of income who normally occupy a dwelling unit or a portion thereof and who provide themselves with food or the necessary supplies or arranged for such provision’. A member resides at least four nights a week in the household.

The sample for the two surveys was stratified by province, urban and non‐urban area, and population group. Altogether 3 000 enumerator areas were drawn as primary sampling units, and within each ten households were visited. The data concerning households were weighted by the estimated number of households in each stratum and, in accordance with instructions from Statistics South Africa, the set of weights with the IES are applied here, as the two surveys are being linked (StatsSA, 1996, 1997a, 1997b). In order to preserve as much as possible of the relative weights among the households in the retained sample, the household weights supplied by Statistics South Africa are applied renormalised to sum to zero, as suggested by Deaton (1997) in the context of missing or unusable observations.

For the multivariate analyses, a subsample was selected consisting of 14 621 households that met three specific criteria. As a first criterion, only African and coloured households would be under study, as these groups are overrepresented among low‐income households and would face similar historical legacies. In accordance with the ‘racial’ categories (African, coloured, Asian/Indian and white) applied in official statistics in South Africa, this study uses the same categories, sometimes referring to them as ‘population groups’.

As the information on individuals' labour market characteristics in the OHS was of a better quality than in the IES, it was deemed desirable to extract that information from the former survey. Households in the two data sets are easily matched, but individuals are not. The second criterion therefore requires that all earners in a household must be identified in both surveys. As a consequence, 5 per cent of the households that met the first criterion were dropped from the analyses.

Finally, the focus in this study will be on the households that have a main income source, which can be defined by the proportion of total income originating from that source. Here a main income source will be defined by a cut‐off contribution set at 67 per cent. Table 1 shows the implications from where the defining cut‐off contribution is placed on the proportion of sample households considered to have a main income source and on the numbers of earners involved.

Households with numbers of main income earners by main income definitions and by various cut‐off contribution levels (n = 21 032)

Main income's contribution to total household income	Households fitting the definition	Number of contributors to main income
		1	2	3	4	5	Sum
50%	84,5%	71,7	23,6	3,6	1,0	0,1	100,0
67%	70,7%	70,8	24,2	3,9	1,0	0,1	100,0
75%	63,5%	70,3	24,5	3,9	1,1	0,2	100,0
90%	45,8%	71,0	24,2	3,7	1,0	0,1	100,0
100%	25,1%	73,0	23,0	3,2	0,7	0,1	100,0
Source: StatsSA OHS/IES95, own computations, weighted figures.

The second row of the table shows that 70 per cent of the households have a main income source if the defining contribution is drawn at 67 per cent. In more than 70 per cent of these households the main income is earned by one member, and in roughly one‐quarter of the households the main income is jointly raised by two earners. As can be seen in the other rows, among the households that meet the various definitions, the very high reliance on one or two earners is remarkably persistent to where the cut‐off line is drawn.

Two other observations are especially noteworthy. First, the figures in the second column show that 45 per cent of the households raise 90 per cent of income from one source category, whereas one‐quarter of the households rely on one source for all their income. Thus, almost regardless of which contribution one uses to define a main income source, the vast majority of households seem highly dependent on one or two earners and on a single source of income.

3 CONTEXTUAL INFORMATION AND INCOME SOURCES

Compared with the rest of the continent, perhaps the most divergent features of household income generation among African and coloured households in South Africa are the generally very small contributions from agricultural income and the historically entrenched and widespread dependence on transfer incomes among rural African families (Jooma, 1991; Reardon, 1997). These two features of income generation in the country cannot be explained outside the context of the legacies of racial segregation, dispossession of land rights forced removals and the migrant labour system (Nattrass, 1981; Wilson & Ramphele, 1989; Lester, 2000).

This study does not, however, attempt to give an account of the historical developments that have led up to the current setting of complex interlinkages between geographical locations, institutional legacies and households' demographic, financial and physical asset endowments. Nevertheless, as will be shown, several aspects of the extent to which South African households were integrated into the core economy in 1995 originate presumably in the legacies from the migrant labour system, as interwoven with the apartheid programme's influx control into the urban areas. Also, batteries of other laws gravely undermined the position of the underprivileged population groups in the labour market.

Two of the most relevant consequences of the migration of working‐age males from the few employment opportunities in the former ‘homelands’ or ‘tribal areas’ – to which apartheid attributed the citizenship of the majority of the African population – is first a ‘peculiar (and quite unnatural) household structure’ (Wilson & Ramphele, 1989) in sending areas, in which children, the elderly and women were vastly overrepresented. Simultaneously, with many of the migrants spending most of their earnings in the core areas of the economy or on the majority of goods that were produced there, the process became one of increasing spatially uneven economic development, with a highly inequitable distribution of employment opportunities (Wilson & Ramphele, 1989).

In the wake of the apartheid era, microdata from the early 1990s accordingly attest not only to high poverty rates but also to very high unemployment rates in both urban and rural areas (SALDRU, 1994; World Bank, 1995; Standing et al., 1996). The high unemployment rates are partly due to the economic stagnation and consequent need for structural transformation instigated by the first oil shock in 1973, but unemployment was also augmented by increasingly distorted relative costs of often subsided capital and labour. Such subsidisation practices also contributed to the increased mechanisation of agriculture, the consequences of which were particularly grave for rural African wage employment (Wilson & Ramphele, 1989; Bhorat et al., 2001, 1998).

In this context the South African literature usually distinguishes, by one set of labels or another, between four broad groups of household income sources: private transfers, public transfers, self‐employment and wage income (Carter & May, 1999; Leibbrandt & Woolard, 1999; Leibbrandt et al., 2000; Bhorat et al., 2001).

Bhorat et al. (2001) break down the South African labour force into three groups according to access to the ‘modern consumer economy’. As will be reflected in the descriptive statistics in the next section, a similar classification of main income sources allows the latter to serve as indicators of households' integration into the core economy. The main income sources are therefore classified by their origin, as either from the core sectors, from the marginal sectors, or of a peripheral nature:

1.	The core sector main income includes salaries and wages from all sectors except the primary sectors and domestic services. In the ‘salaries and wages’ concept is included bonuses and income from overtime, commissions and director's fees, part‐time work and cash allowances in respect of transport, housing and clothing. Self‐employment in the form of net profit from business or professional practice or activities conducted on a full‐time basis is also considered core, as is capital income from the letting of fixed property, royalties, interests, dividends and annuities.
2.	The primary sectors and domestic services constitute separate subcategories under ‘marginal sectors’.
3.	The two subcategories of ‘peripheral’ main income sources are finally private transfers and public transfers. Private transfers are considered to be alimonies, maintenance and similar allowances from divorced spouses or family members living elsewhere and regular allowances from family members living elsewhere. Public transfers consist of pensions resulting from own employment, old‐age and war pensions, social pensions or allowances in terms of disability grants, family and other allowances, or from sources such as the Workmen's Compensation Act, Unemployment Insurance Fund and the Pneumoconiosis and Silicosis Funds.

4 MAIN INCOME SOURCES IN AN EARNINGS AND LABOUR MARKET CONTEXT

This section discusses four aspects of the social relevance and suitability of the main income concept as an indicator of integration. First, it is shown how the distribution of main income sources differs in urban and non‐urban areas. Secondly, the relationship between households' main income sources and the income distribution is discussed. Thereafter, some perspective is provided on the extent to which a main income source is representative of households' total income generation activities. Finally, individuals' labour market statuses are related to their households' main income source.

4.1 Urban and non‐urban main income sources

For the historical reasons referred to in the previous section, one would expect core sector access and main income sources to differ in the rural and urban samples, but not in the same fashion as elsewhere, where agricultural activities are more prominent in rural areas. Leibbrandt et al. (2000) have noted that the IES data do not capture agricultural activities for own consumption well. In this study's sample, 8,4 per cent of all households were recorded as having either slaughtered domestic animals or harvested crops in the year preceding the interview. Profit from agricultural activities had to be registered under ‘self‐employment’ in the IES questionnaire, but only 1,1 per cent of the households that had slaughtered or harvested had records of any self‐employment profits at all. These figures presumably understate the importance of agriculture which, according to May (1996), assumes several important functions as, for example, a supplementary source of nutrition and a safety net for vulnerable house holds in South Africa. However, left with little choice other than taking the data at face value, agricultural production is not treated as a separate source of income.

As the term ‘rural’ has an intuitive connotation of pastoral activities, which is thus misleading in this context, the term ‘non‐urban’ will henceforth be applied. Also, the sometimes very high population densities found in ‘rural’ areas of South Africa raise doubts as to the appropriateness of the terminology. For instance, Mabin (1989) defines ‘rural slums’ as the many areas that were ‘urban’ in respect of their population densities but ‘rural’ in respect of the absence of proper urban infrastructure or service.

Table 2 shows the distribution of main income sources in the urban and non‐urban subsamples. As can be seen, core sector income is much more prevalent in the urban than in the non‐urban sample, with shares of 77,7 and 42,2 per cent of the households in each sample, respectively. Further, urban main income sources are considerably more concentrated around either core sector or public transfer main incomes, which together account for more than 90 per cent of the households. At shares of 27 and 14 per cent in non‐urban areas, households that rely on private and public transfers constitute proportions, respectively, twice and four times as large as those of their urban counterparts. A valid objection to the use of a ‘dominant source of income’ for the analysis of rural livelihoods is raised by Ardington & Lund (1996), who emphasise that such sources may be of a temporary nature.

Distribution of main income sources in the sample by location

Main income source	Urban	Non‐urban	Total
Core sectors	77,7	42,2	57,6
Primary sectors	1,4	14,1	8,8
Domestic services	3,1	2,9	3,0
Public transfers	14,3	27,0	21,9
Private transfers	3,5	13,9	8,7
Total	100,0	100,0	100,0
N	7·394	7·272	14·621
Source: StatsSA OHS/IES95, own computations, weighted figures.

4.2 Main income sources and the income distribution

Tables 3 and 4 portray the distributions of non‐urban and urban households across household income deciles and main income source categories. In order to show where this study's sample of households belongs in the population‐wide income distribution, the deciles are computed based on the whole, original OHS/IES sample. Before turning to the distribution of main income sources, it needs to be noted that the proportion of households in the four lower deciles in the non‐urban areas is nearly twice that of the urban areas. A common trend in both areas is that roughly 60 per cent of the households with core sector main income sources are found in the fifth to eighth deciles, whereas proportions to the magnitudes of 70–80 per cent of households with other main income sources are found in the first to fourth deciles. Moreover, the largest proportions of households at the lower end of the income scale are found among households relying on peripheral main income sources.

Distribution of main income sources among non‐urban households by household income deciles (n = 7 227)

			Income source category
		Marginal		Peripheral
Non‐urban decile	Core sectors	Primary sectors	Domestic Service	Public transfers	Private transfers	Total
1	3,1	22,4	35,2	35,8	32,0	19,6
2	4,3	24,6	18,7	25,4	26,5	16,4
3	6,6	19,3	15,4	20,1	17,4	13,8
4	12,5	15,7	13,1	11,9	10,1	12,5
5	18,5	7,9	9,4	4,0	7,5	11,3
6	16,7	6,2	4,7	1,5	4,1	9,0
7	15,9	2,8	1,4	0,6	2,0	7,6
8	12,4	0,9	1,7	0,3	0,3	5,5
9	6,8	0,2	0,0	0,3	0,3	3,0
10	3,2	0,0	0,4	0,0	0,0	1,4
Sum	100,0	100,0	100,0	100,0	100,0	100,0
Source: StatsSA OHS/IES95, own computations, weighted figures.

Distribution of main income sources among urban households by household income deciles (n = 7 394)

	Income source category
Urban Decile	Core sectors	Primary Sectors	Domestic service	Pubilc transfers	Private transfers	Total
1	1,6	20,9	20,3	28,2	33,8	7,4
2	2,4	20,2	19,2	21,8	23,0	6,7
3	4,7	10,8	22,3	23,3	13,4	8,3
4	8,6	17,2	17,1	13,1	12,8	9,8
5	13,0	11,0	10,2	6,4	8,6	11,8
6	15,9	10,4	7,5	2,6	3,8	13,2
7	17,5	7,3	2,9	2,4	2,3	14,2
8	17,4	0,2	0,5	0,7	0,8	13,7
9	13,1	1,0	0,0	1,1	1,0	10,4
10	5,7	1,0	0,0	0,3	0,4	4,5
Sum	100,0	100,0	100,0	100,0	100,0	100,0
Source: StatsSA OHS/IES95, own computations, weighted figures.

4.3 Main income sources as representative of sample households' income generation

Table 5 shows the distribution of the number of additional, non‐main income sources in the final sample. As can be seen, the vast majority of households in the sample do not have another source of regular income, with the only noteworthy deviations found among households in the core sector and domestic service categories, where additional income is found in 13 and 10 per cent of households, respectively. It should be noted that in the data, ‘total income’ is defined as the sum of ‘direct’ and ‘indirect’ income, where the latter is of a non‐regular nature and thus in many cases accounts for the difference between the main income and total income.

Number of non‐main income sources by main income source (n = 14 772)

	Main income source
Number of non‐main incomes	Core sector	Primary Sector	Domestic services	Public transfers	Private transfers	Sum
0	86,6	97,0	89,3	95,0	96,9	90,2
1	11,3	2,7	7,4	4,4	2,4	8,3
2	1,7	0,4	2,4	0,5	0,2	1,2
3 or more	0,4	0,0	0,9	0,2	0,6	0,3
Sum	100,0	100,0	100,0	100,0	100,0	100,0
Source: StatsSA OHS/IES95, own computations, weighted figures.

The distribution of the number of contributors to individual households' main income earners in the sample is shown in Table 6. In just over 70 per cent of the households in this sample the main income is earned by one individual, but deviations from the one‐earner pattern are found in the domestic services and private transfer categories, where the corresponding figures are 82 and 92 per cent, respectively.

Number of contributors to main income by main income source (n = 14 772)

	Main income source
Number of earners	Core sector	Primary Sector	Domestic services	Public transfers	Private transfers	Sum
1	68,0	69,5	82,4	68,3	92,5	70,8
2	25,7	24,7	15,2	28,7	6,0	24,2
3 or more	6,3	5,8	2,4	3,0	1,5	5,0
Total	100,0	100,0	100,0	100,0	100,0	100,0
Source: StatsSA OHS/IES95, own computations, weighted figures.

4.4 Labour force participation and the main income source categories

Table 7 shows the labour market statuses of adults in the sample categorised according to their households’ main source of income. The left‐hand side of the table focuses on the non‐participants, whereas the right‐hand side shows the distribution of participants across the statuses ‘unemployed’, ‘self‐employed’ and ‘employed’. This study follows the official definitions (StatsSA, 1997b) of expanded unemployment, including ‘discouraged seekers’, and economically non‐active (henceforth ‘inactive’). A child is defined as being 14 years old or younger. The term ‘working age’ refers to females below 60 years of age and males below 65, corresponding to the gender‐specific retirement ages. A ‘retired’ individual is above working age and has been captured in that status of labour force activity in the OHS data.

Adults' labour force status by households' main income source (n = 66 841)

	Non‐participants in labour force				Labour force status among participants
Main income source	Enrolled	Retired	Other	Total	Unemployed	Self‐employed	Employed	Sum	Share of all adults
Core	21,2	1,1	4,4	28,7	14,9	4,7	80,3	100	60,4
Primary	11,7	0,2	2,4	17,2	6,4	0,1	93,5	100	6,7
Domestic	28,3	0,9	6,2	39,3	45,1	6,6	48,3	100	2,5
Public transfers	31,7	10,5	22,6	69,7	93,5	1,3	5,2	100	22,9
Private transfers	47,1	0,9	11,3	65,2	95,5	0,9	3,6	100	7,4
Total	23,8	2,6	7,7	34,1	24,5	3,9	71,6	100	100
Source: StatsSA OHS/IES95, own computations, weighted figures.

As can be seen, around two‐thirds of adult members in households with either type of peripheral main income source are non‐participants, but their distributions within the non‐participation status differ quite dramatically. Non‐participation is also high in the domestic services category at nearly 40 per cent, which makes the rate 10 and 20 percentage points higher than in the core sector and primary sector categories, respectively. The right‐hand side of the table shows evidence that the small fraction of labour force participants in the peripheral income households are unemployed to a dramatically higher extent – at an immense 95 per cent – than the participants from households with core sector main income. Also, the very high unemployment rate in domestic services households and the relatively low unemployment rate in primary sector households are noteworthy.

Using households' main sources of income as indicators of integration thus presents the latter as a highly spatially driven phenomenon, with low integrations associated with low household incomes, low labour force participation and high unemployment rates. Further, while one‐third of the households that met the first two criteria (population group and identification) did not have a main income source, the main income source is of considerable relevance to income generation among the approximately 70 per cent of households that do have one. Few of those households have other income sources or other members with a regular income, which naturally also makes the households extremely vulnerable to the loss of main income earners or incomes.

5 MODELLING MAIN INCOME SOURCES

The identification of characteristics of households with different sources of main income proceeds through the use of two five‐way multinomial logistic models, the explanatory variables of which are introduced below. However, at least two sets of analytical complications are associated with these specifications, pertaining to the potential statistical endogeneity and to the predetermined nature of certain explanatory variables. These are addressed in separate subsections below.

5.1 Empirical modelling

The application of the multinomial logistic model as a probability model in this context is based on the assumption that the probability of a given household i of holding a specific income source m is a function of its endowment vector of S explanatory variables X_i and a vector of income source‐specific parameters, â_m, according to: (-1)

where n is the sample size. Long (1997) shows how this model may be derived either as a probability model or a discrete choice problem. In order for the expression to be uniquely defined, one set of âs (for the core sector category in this case) is normalised to zero. By the vector of explanatory variables, the ensuing probabilities are thus functions of the characteristics that influence households' access to various types of income. Long (1997) refers to Amemiya (1985), who has shown that ‘under conditionswhich are likely to apply in practice the implied likelihood function is globally concave, ensuring the uniqueness of ML estimates’.

The earner‐specific explanatory variables are all recognised determinants in the literature on access to the various income sources, and encompass geographical variables, earners' age, gender and educational characteristics. In addition, the study implicitly assumes that dependence on non‐labour income can be explained by the number and composition of households' non‐earners in shares of children out of total household size and the proportions of adults who are unemployed, inactive and retired out of the household. The use of household members' unemployment status as an explanatory factor for main income sources implicitly suggests that unemployment is considered involuntary (and, as will be discussed below, exogenous). In view of the extremely high unemployment rates among households with transfer main income sources and the high concentration of those income sources at the very bottom of the income distribution, the assumption of involuntary unemployment appears reasonable.

Households where several individuals contribute to the main income are assimilated by the use of proportions of earners in each age, gender and education category. Furthermore, all non‐binary explanatory variables are measured in terms of deviations from the subsample median values, as it was deemed more informative to express impacts as originating from divergences from modal living arrangements. Summary statistics of these variables in non‐normalised format, but including median values, are given in Table 8.

Summary statistics of explanatory variables

	Non‐urban subsample			Urban subsample
Variable	N	Mean	Std dev.	Median	Min.	Max.	N	Mean	Std dev.	Median	Min.	Max.
African	7227	0,93787	0,24140	1	0	1	7394	0,71788	0,45006	1	0	1
Number of non‐earners	7227	3,54103	2,65037	3	0	28	7394	2,89099	2,27835	3	0	24
Western Cape	7227	0,03999	0,19595	0	0	1	7394	0,15810	0,36486	0	0	1
Eastern Cape	7227	0,26346	0,44054	0	0	1	7394	0,17758	0,38218	0	0	1
Northern Cape	7227	0,02767	0,16405	0	0	1	7394	0,05937	0,23634	0	0	1
Free State	7227	0,05618	0,23028	0	0	1	7394	0,12794	0,33405	0	0	1
KwaZulu‐Natal	7227	0,18915	0,39166	0	0	1	7394	0,11901	0,32383	0	0	1
North West	7227	0,09852	0,29804	0	0	1	7394	0,08480	0,27860	0	0	1
Gauteng	7227	0,02242	0,14804	0	0	1	7394	0,17568	0,38058	0	0	1
Mpumalanga	7227	0,13754	0,34444	0	0	1	7394	0,05613	0,23018	0	0	1
Limpopo	7227	0,16508	0,37127	0	0	1	7394	0,04138	0,19919	0	0	1
Subregion 1 non‐urban: ‘semi‐Town’/urban: ‘informal dwellings’	7227	0,05189	0,22182	0	0	1	7394	0,08791	0,28318	0	0	1
Subregion 2 non‐urban: ‘town without local authority/urban: ‘hostels’	7227	0,05936	0,23632	0	0	1	7394	0,04274	0,20228	0	0	1
Subregion 3 non‐urban: ‘tribal area’	7227	0,56441	0,49587	1	0	1	Not defined for urban sample
Subregion 4 non‐urban: ‘agricultural or amenities area’	7227	0,22707	0,41896	0	0	1
Share of earners female	7227	0,42777	0,44530	0,5	0	1	7394	0,42041	0,41892	0,5	0	1
Share of earners in education	7227	0,27615	0,43380	0	0	1	7394	0,11458	0,30608	0	0	1
Category: none
Share of earners in education	7227	0,45702	0,47909	0	0	1	7394	0,38386	0,45893	0	0	1
Category: primary
Share of earners in education Category: secondary	7227	0,14667	0,33760	0	0	1	7394	0,23388	0,39031	0	0	1
Share of earners in education Category: matriculated	7227	0,06111	0,22508	0	0	1	*
Share of earners in education Category: tertiary			*				7394	0,11248	0,29892	0	0	1
Share of earners in age category: ≤ 19 years	7227	0,00830	0,09074	0	0	1	*
Share of earners in age category: ≤ 24 years			*				7394	0,02637	0,16025	0	0	1
Share of earners in age category: 20–24 years7227	7227	0,03003	0,17067	0	0	1	*
Share of earners in age category: 25–34 years	7227	0,21295	0,38875	0	0	1	7394	0,27130	0,41311	0	0	1
Share of earners in age category: 35–59 years	7227	0,48871	0,47703	0,5	0	1	7394	0,52334	0,465214	0,5	0	1
Share of earners in age category: 60–64 years	7227	0,06741	0,23292	0	0	1	7394	0,04761	0,19823	0	0	1
Share of earners in age category: ≥ 65 years	7227	0,15533	0,34845	0	0	1	7394	0,09157	0,27676	0	0	1
Share of non‐earners children	7227	0,41124	0,33924	0,44444	0	1,5	7394	0,38215	0,34897	0,4	0	3
Share of adult non‐earners	7227	0,17754	0,32347	0	0	1	7394	0,21071	0,35339	0	0	1
Unemployed
Share of adult non‐earners inactive	7227	0,51112	0,45124	0,5	0	1	7394	0,43263	0,44312	0,33333	0	1
Share of adult non‐earners retired	7227	0,02861	0,12838	0	0	1	7394	0,03974	0,16016	0	0	1
Notes: * The two specifications differ in category variables left out to avoid multicollinearity. In both samples the earner age category 60–64 years and the fraction of non‐earners in employment were left out. However, in the non‐urban sample the matriculated education category was left out whereas in the urban sample tertiary education was omitted. For similar reasons the two youngest earner age categories in the urban subsample were collapsed into one category.

A rich literature exists in which both macroeconomic and microeconomic determinants of labour force participation, employment and earnings have been identified (e.g. Willis, 1986). Several studies of related areas have also been conducted on South African data, which attest to determinants of employment being found among age, experience, gender, education, marital status and race (e.g. Mwabu & Shultz, 2000; Naudé & Serumaga‐Zake, 2001). The channels through which individual characteristics influence the allocation to economic sectors are individual expected earnings and reservation earnings (Wambugu, 2003), the determinants of which are similar to those of labour force participation, employment and earnings. A dense review of determinants' remittance behaviour is provided by, for instance, De la Brière et al. (2002). (The interested reader is also referred to Stark, 1995.) Posel (2001) tests several hypotheses about South African remittance behaviour and estimates the impact on remitted amounts in sole migrant households.

With respect to the South African social security system, there are social support programmes to cover several circumstances but not unemployment benefits. The old‐age pensions system encompasses some 60 per cent of the total social security budget (Budlender, 2000). Women are entitled to pension at the age of 60 and men at 65, and the system is financed by general government revenue. While a means test does apply in practice, it seems to have little effect on or not be binding to African households (Ardington & Lund, 1995; Case & Deaton, 1998; Jensen, 2001; Bertrand et al., 2003).

It follows from equation (1) that the marginal effect of explanatory variable s on the probability that household i has main income source m is given by: (--1)

The marginal impact thus depends not only on the change in variable s and the coefficient for that variable, but on the level of variable s and of all other variables, as well as all the other slope parameters. Consequentially, marginal effects will vary with the variable values at which they are estimated and the sign of the marginal effect need not match that of the slope parameter. Hence, the individual slope parameters convey little information per se. The regression output is therefore presented by their exponential value or in ‘relative risk ratio’ format. See Long (1997) or Borooah (2001) for a discussion of various modes of presenting output from the multinomial logit approach.

5.2 Endogenous household formation

A growing body of literature suggests that the living arrangements of South African households alter in response to the economic circumstances of individual members, such as access to certain sources of income (Klasen & Woolard, 2001; Bertrand et al., 2003; Edmonds et al., 2003). While no attempts are made here to draw inferences as to the nature of such intrahousehold processes, the case may yet be that two‐way causality applies between income source and demographic composition. This would render the latter group of explanatory variables statistically endogenous.

Edmonds et al. (2003) provide a number of findings that suggest impacts from income sources on household structures. The first finding relates to migration where, quite naturally, absent members constitute a defining characteristic of households that rely on private transfers (see also Wilson & Ramphele, 1989). However, if younger members are encouraged to migrate due to successful outcomes of the households' previous migration histories, the demographic characteristics of migrants' households are transplanted and augmented among consecutive generations in the household. Secondly, Edmonds et al. (2003) have also found that the old‐age pension received by a member of the household may serve to finance younger members' migration.

In analysing household formation and unemployment in South Africa, Klasen & Woolard (2001) use two‐stage least‐squares regression techniques in order to control for causality running from unemployment to household formation around a non‐labour income source. Note that a growing body of international literature exists on patterns of household formation and unemployment most studies of which take household formation as exogenous (e.g. Atkinson & Micklewright, 1991; Arulampulam & Stewart, 1995; Gregg & Wadsworth, 1996; OECD, 1998). Klasen & Woolard (2001) find that access to state transfers increases the likelihood of attracting unemployed people to a household and that unemployed adults reside with their parents longer than do the employed.

Consistent with findings by Bertrand et al. (2003), Klasen & Woolard also find that households' collection of remittance income, pensions and other non‐wage private income is correlated with lower shares of working age adults in labour force participation and employment. The authors do not apply the main income source concept, but find that 60 per cent of the unemployed in their study live in households where someone is employed and 20 per cent live in households receiving remittances.

On the same note, Edmonds et al. (2003) find that female, pensions‐eligible household heads are more likely to reside with their adult children than with certain other relations. In summary, it thus appears advisable to investigate into simultaneity between income sources and each of the proportions of the unemployed, children, and inactive members.

5.3 Predetermined variables

Arguably, variables such as those reflecting household wealth, unemployment or non‐participation are predetermined outcomes of past choices. As stated by Glewwe (1991) in the context of investigating determinants of household welfare, in the absence of an identification of the processes and determinants that led up to such past choices the analysis is incomplete. Further, parameter estimates for predetermined explanatory variables must be perceived as explaining the variation in household main income sources conditional on the past decisions and events through which they have taken on their current values.

6 EMPIRICAL RESULTS AND SIMULATIONS

While the relative risk ratio format is useful in terms of the direction and magnitude of the impact from changes in explanatory variables on relative probabilities of given pairs of outcomes, the coefficients provide little information on how the absolute probabilities of each of the individual outcomes are ultimately affected. After a brief overview of the general fit of the two regression models and the significance of the various categories of explanatory variables, a set of simulation exercises follows which show impacts from changes in hypothetical household characteristics on main income source probabilities.

6.1 Regression results

Tables 9 and 10 display the regression output from the non‐urban and urban subsamples. The coefficients in the two models are not directly comparable, however, as variable definitions differ in certain cases. Table 2 in the descriptive section showed a considerably larger variation in main income sources in the non‐urban areas, and several findings also suggest that the approach is more warranted as applied to non‐urban households than to urban ones.

Multinomial logit estimates of main income category for non‐urban households

Number of observations 7227
Wald χ^2 (132) 8481,75
Prob. > χ^2 0,0000
Pseudo‐R^2 0,46530
Log pseudo‐likelihood −5292,4811
Explanatory variables	Primary sectors	Domestic services	Public transfers	Private Transfers
Number of non‐earners	0,40269***	0,63699**	0,69266***	0,64107***
African	1,95325	0,61823	0,99315	6,20648**
Western Cape (‡‡)	0,66125	0,28560*	0,41992*	0,22745*
Eastern Cape	1,35777	0,78995	2,04318***	2,92144***
Northern Cape	2,56191**	0,59466	0,46335*	1,04375
Free State (‡‡)	1,54488	1,08210	0,84746	1,88100
North West	1,12546	0,27480***	0,50335***	1,74649**
Gauteng	0,42690**	1,10011	0,35069**	0,00000***
Mpumalanga	2,16795**	1,26062	0,36272***	0,48242***
Limpopo	2,48950**	0,69581	0,86055	1,21457
‘Semi‐town’ (‡‡)	0,30623*	0,49497	1,80941	1,51507
‘Town’	0,86912	1,88772	2,91000***	2,18790*
‘Tribal area’	0,16712***	0,84164	1,95046**	2,42624**
Agricultural/amenities area	9,40751***	4,39276***	1,95114**	0,55195
Share of main income earners	0,13286***	20,71139***	2,38631***	4,37819***
(Sh. MIE) female
Sh. MIE no education	161,46310**	83,64364***	70,94666***	72,08230***
Sh. MIE primary education	101,36020***	75,11304***	30,36159***	35,88247***
Sh. MIE secondary education	29,89462***	25,72992***	10,33534***	19,82048***
Sh. MIE matric education	9,07762***	3,94063	1,53559	6,52037***
Sh. MIE≤19 years	1,61322	2,76234	1,81192	35,24345***
Sh. MIE 20–24 years	1,07565	2,08541	0,12926***	3,92955***
Sh. MIE 25–34 years	0,19222***	0,64582	0,01194***	0,48177*
Sh. MIE 35–59 years	0,30040***	0,68831	0,06612***	0,61308*
Sh. MIE≥65 years	0,42824*	0,60990	7,99522***	0,79091
Share of non‐main income earners (Sh. NMIE)≤14 (§)	369169,90***	29,29530	109,61080**	791,6230***
Sh. NMIE unemployed (§)	9,49793**	5,04592	5,19879*	0,53000
Sh. NMIE inactive (§)	13,98190***	2,35612	2,77849*	4,26535**
Sh. NMIE retired	0,38287**	0,40104	0,41879**	0,03984***
χ^2 value Hausman test of IIA	−11,684	−48,962	19,289	−29,622
when category is omitted	Supported	Supported	Supported	Supported
Notes: Symbols: (†)/(‡) Wald test H(0): variables all coefficients = 0, not rejected at 5 or 10 per cent. (§) Endogenous, observed variable replaced by prediction. Significance levels are based on standard errors that take the clustered nature of the sample into account. Source: StatsSA OHS/IES95, own computations, weighted data.

Multinomial logit estimates of main income category for urban households

Number of observations 7394
Wald χ^2 (132) 7456,21
Prob. > χ^2 0,0000
Pseudo‐R^2 0,4151
Log pseudo‐likelihood −3287,0091
Explanatory variables	Primary sectors	Domestic services	Public transfers	Private Transfers
Number of non‐earners	0,91432	0,925926	0,95905	0,89340**
African	1,04992	1,632152**	0,78297	2,96949***
Western Cape (†)	5,95991***	0,855529	1,16850	0,71978
Eastern Cape	2,12973	0,873331	2,37955***	2,85653***
Northern Cape	7,56284***	2,168811*	3,34436***	3,81096***
Free State	1,31239	1,416153	2,03931***	2,85198***
North West (‡‡)	4,78987*	0,600125	1,08411	1,10442
Gauteng	0,47306	1,068179	0,51492**	0,25044***
Mpumalanga	7,32566***	0,841719	0,71261	0,53572
Limpopo	35,25985***	0,105925**	0,95059	1,06688
‘Informal dwellings’	3,08046**	1,704814**	1,38963*	0,99245
‘Many hostels’	0,03670***	0,054554***	0,06080***	0,27563
Share of main income earners	0,96613	17,53389***	4,34963***	6,17050***
(Sh. MIE) female
Sh. MIE no education	19,12191***	20,78237***	31,98307***	5,44395***
Sh. MIE primary education	11,02114***	12,75822***	10,58078***	3,96337***
Sh. MIE secondary education	5,58107***	3,882472***	4,89480***	2,25146***
Sh. MIE matric education	0,00000***	1,61E‐01	1,16208	0,16442***
Sh. MIE≤24	1,94605	5,108697**	0,25153***	7,95197***
Sh. MIE 25–34	0,74982	2,87273**	0,04808***	1,12336
Sh. MIE 35–58	0,42178**	2,299924*	0,12671***	0,95745
Sh. MIE≥65	0,52508	2,329331	23,83244***	3,13204*
Share of non‐main income earners (Sh. NMIE)≤14	0,90525	0,344943***	0,57190***	0,86880
Sh. NMIE unemployed (‡‡)	0,45602*	0,623586	1,03930	0,93292
Sh. NMIE inactive	0,81788	0,538909**	0,56167***	0,72891*
Sh NMIE retired	0,05840**	0,163018**	0,15467***	0,20707*
χ^2 value Hausman test of IIA	−16,664	14,161	1203,949	−47,451
when category is omitted	Supported	Supported	Not supported	Supported
Notes: Significance levels are based on standard errors that take the clustered nature of the sample into account. Symbols: (†)/(‡) Wald test H(0): variables all coefficients = 0, not rejected at 5 or 10 per cent. (§) Endogenous, observed variable not replaced by prediction. Source: StatsSA OHS/IES95, own computations, weighted data.

First, Hausman tests of the independence of irrelevant alternatives (IIA) support this assumption for non‐urban households but not for the urban ones. Secondly, the larger value of the pseudo‐coefficient of determination of 0,465 for the non‐urban subsample vs 0,415 in urban areas indicates that the non‐urban model explains more variation than the urban one. The R² values may, to some extent, exaggerate explanatory powers, as the null hypothesis that a variable may have no effect on the outcome cannot be rejected for certain variables. This applies at the 10 per cent level for three variables in the non‐urban and two in the urban categories. However, the pseudo‐R² value must be seen in the perspective of estimates' significance. A general impression can be derived by studying the fractions of estimates significant at the 10 per cent level or higher in bold type in Tables 9 and 10, which reveals that roughly three‐fifths of the estimates in each sample are significant by this measure.

Further, the variables in the output are divided into three sections in the vertical dimension, where the upper section includes household‐level variables, with dashed lines serving to separate the African, provincial and subregional dummy variables. The middle section constitutes the earner characteristics, while the lower section contains the non‐earner composition characteristics. In both sets of output, the middle section has the highest prevalence of significant estimates, which attests to the high relevance of earner characteristics and indeed of education levels.

In both subsamples main income sources appear subject to both interprovincial and regional variation, while the African population dummy has a significant estimate for only one outcome category in the urban output and two in the non‐urban output. Further, three‐quarters of the estimates for non‐earner characteristics are significant in the non‐urban subsample and over half in the non‐urban one. It should be noted, however, that in urban areas all the coefficients for all the non‐earner fraction variables, including that of children, are less than unity, indicating diminishing impacts on probabilities of non‐core income sources. The opposite is true for non‐urban areas and especially for the fraction of children.

However, in the non‐urban subsample the joint exogeneity of the variables for the proportions of children out of total household size, and for the unemployed and the inactive out of total adult members, was not supported. Consequently, these variables were replaced with predictions through a non‐simultaneous two‐step procedure. The variables were tested for endogeneity by the method suggested by Rivers & Voung (1988) using four variables for the proportions of adult non‐earners in each education category as additional exogenous variables. (Test results with more detailed account of the procedures are available from the author.) When testing for endogeneity in the urban sample, none of the estimates for the residuals was individually significant and jointly the three residuals' estimates were significant only at the 10 per cent level. The relevant cells in Table 9 are therefore shaded in grey to caution the reader as to the invalid significance levels. While the direction of causality from these variables may be open to discussion, the high and prevalent significance of the predictors is consistent with some interaction between household non‐earner composition and income sources.

6.2 Simulations

The direction and strength of impact, as well as the absolute probabilities associated with some of the key explanatory variables, will now be illustrated through four sets of simulations in Tables 11–14, three of which are based on the non‐urban estimates and one on the urban estimates. Cells shaded in grey indicate that estimated probabilities are based in part on results that were not significant at the 90 per cent level or higher in the regression analyses. The first simulation in Table 11 illustrates the impact of a single main income earner's gender and age in a household of five, with three children and an inactive member. The household is assumed to reside in a ‘tribal area’ in KwaZulu‐Natal and the earner has primary education only.

Simulation of impact from main income earner's gender and age

	Probabilities of main income categories
Earner	Primary sector	Domestic services	Public transfers	Private transfers	Core sector
Male aged 35–59	10,4	0,6	9,9	22,6	56,5
Female aged 35–59	0,7	6,1	12,3	51,4	29,4
Female aged 25–34	0,6	7,3	2,8	51,7	37,6
Female aged 60–64	0,8	2,8	60,0	27,0	9,4
Male aged 25–34	8,0	0,6	2,2	21,3	67,9
Male aged 60–64	12,4	0,3	53,9	13,2	20,3

With a male main income earner in the age category 35–59 years (in the first row), the probability that the household has a core sector main income source is over 55 per cent. However, if the earner is a female of the same age (as illustrated in the second row), the probability of having a core sector source is reduced to below one‐third and a probability just over 50 per cent exists that the household relies on private transfers instead. If the female earner belongs to the younger age category of 25–34 years, the same probability prevails for having a private transfer main income, but probability for the household of having core sector main income has increased slightly to almost 38 per cent. If the female earner is older and aged 60–64 years, the probability of private transfer income is reduced to below 30 per cent while the likelihood of public transfers is now 60 per cent. If the same age alterations are applied to the male earner, the probability of core sector income increases considerably to over two‐thirds in the younger case, but is reduced to 20 per cent in the older case. Thus, variations in the gender and age of the main income earner seem strongly associated with changes in probabilities of different types of main income sources.

The second simulation in Table 12 illustrates the impact of education and subregional location for the same household under the assumption that the main income earner is a female in the age category 35–59. In the first case, the household still resides in a ‘tribal area’ but the main earner has no education. The probability that the household depends on private transfers is just over 60 per cent and for having core sector main income the probability is below 20 per cent. Raising the educational level to primary schooling (equivalent to the second case above) yields the corresponding probabilities at just over 50 per cent and slightly below 30 per cent, respectively. A marked increase in the probability of having core sector main income to 45 per cent follows an increase in the earner's education to secondary schooling, but the probability of private transfers is almost as high.

Simulation of impact from main income earner's education and non‐urban households' subregional location

		Probabilities of main income categories
Education	Location	Primary Sector	Domestic services	Public transfers	Private transfers	Core sector
None	‘Tribal area’	0,7	4,0	17,0	61,0	17,4
Primary	‘Tribal area’	0,7	6,1	12,3	51,4	29,4
Secondary	‘Tribal area’	0,3	3,2	6,5	44,2	45,7
None	‘Agricultural/amenities area’	35,5	19,5	15,9	12,9	16,2
Primary	‘Agricultural/amenities area’	32,2	25,3	9,8	9,3	23,4
Secondary	‘Agricultural/amenities area’	19,0	17,3	6,7	10,3	46,8

If the same household had resided in an ‘agricultural or amenities area’, the trend with respect to changes in the probability of core sector main income is similar and ends at the same level. A most notable difference, however, is that probabilities of marginal sector incomes are dramatically greater; initially their added probabilities are 55 per cent but decrease with education to 36 per cent. Thus, while increased levels of education seem to vastly improve chances of households having core sector main income regardless of location, the probabilities of its alternatives seem strongly affected by location and also affected by the earners' level of education.

The importance of provincial location and the number of non‐earners in the household is illustrated in Table 13. The default urban household resides in KwaZulu‐Natal and consists of four members, including two children and one earner in the age category 60–64 years. With a male earner the model returns a probability of having core sector main income of over 80 per cent. If the earner is female, the probability of the same sector income drops to 50 per cent and to two‐fifths for public transfer incomes.

Simulated impact from urban households' provincial location

	Probabilities of main income categories
Province	African	Earner	Primary Sector	Domestic services	Public transfers	Private transfers	Core sector
KwaZulu‐Natal	Yes	M	0,8	0,2	15,1	1,6	82,3
KwaZulu‐Natal	Yes	F	0,5	2,3	40,5	6,0	50,7
Eastern Cape	Yes	F	0,6	1,2	57,6	10,3	30,3
Northern Cape	No	F	1,4	1,3	72,8	3,3	21,3
Gauteng	No	F	0,3	1,9	33,5	0,6	63,7
Gauteng	No	M+F	0,2	0,2	9,0	0,2	90,4

Moving the last household constellation to the Eastern Cape increases the public transfer probability to over 55 per cent and reduces the likelihood of core income to less than one‐third. Shifting location to the Northern Cape and the population group to coloured raises the probability of having public transfers to over 70 per cent and reduces the chances for core income to almost one‐fifth. If the household resided in Gauteng, however, the probability of having a core sector main income increases dramatically to almost 65 per cent. Finally, if the latter household rather has two earners – one of each gender – the likelihood that it has a core sector main income peaks at over 90 per cent.

A final simulation in Table 14 illustrates the associations between main income sources and household composition factors, represented by the impact from the labour force status of non‐earners and the presence of children in the households. The initial household again resides in a non‐urban, ‘tribal area’ in KwaZulu‐Natal and has a female earner aged 20–24 years with primary education. The other household members initially encompass two children and two inactive non‐earners. The default household has a probability of just below 70 per cent of being dependent on private transfers. If an inactive adult were replaced by an additional child, the probability of relying on private transfers increases – as would be expected from theory – to over 80 per cent. By shifting the remaining inactive member into retired status, the same probability drops to below 65 per cent. This is still very high, but the impact suggests a negative association between the presence of retirement‐aged members and dependence on private transfers.

Simulated impact from non‐earner household characteristics

			Probabilities for main income categories
Age of female earner	No. of children	Adults	Primary sector	Domestic services	Public transfers	Private transfers	Core sector
20–24	2	2 inactive	0,2	6,5	6,8	69,9	16,5
20–24	3	1 inactive	0,6	4,6	6,0	81,6	7,3
20–24	3	1 retired	0,7	7,4	9,4	64,1	18,4
60–64	3	1 inactive	0,8	2,8	60,0	27,0	9,4
60–64	3	1 unemployed	0,7	3,4	70,3	16,0	9,5
60–64	2	2 unemployed	0,2	4,6	72,4	6,3	16,5

By turning the adults' roles around and letting the retirement‐aged member be a female earner and the younger female be an inactive member, the chances of dependence on private transfers drops below 30 per cent and yields a 60 per cent probability of the household relying on a public transfer. If the younger female's labour force status is then altered to unemployed, the probabilities shift by one‐tenth in favour of having public transfers, which almost halves the likelihood of the other transfer main income to 16 per cent. This prediction is compatible with earlier findings of a positive association between unemployed household members and pension‐collecting households (Klasen & Woolard, 2001; Edmonds et al., 2003; Keller, 2003). Finally, by replacing one child with an additional unemployed adult member, the probability of having public transfers increases by only two percentage points, whereas the probability of private transfers drops drastically. In conjunction, these shifts illustrate a strong, positive association between the presence of children and dependence on either type of transfers in rural areas.

7 CONCLUSIONS

This study has shown that, among the majority of coloured and African households captured by Statistics South Africa's 1995 October Household Survey, income generation is highly concentrated to one or two members. The study is based on a classification of main income sources according to core economy integration, which – with the exception of households relying on primary sector income – shows low integration to be associated with the lower end of the household income distribution and very high non‐participation and unemployment rates. This categorisation also reveals a high concentration of household income around a single source and warrants a perspective on household income generation in the sample as revolving around one main income source.

While inference to the total South African population is prohibited by the intentional selection of households with a main source of income, the study proceeds to identify statistical regularities that account for a considerable fraction of the variation in the sample's main income sources. Most prominently, the spread across main income source categories is much greater in non‐urban areas than in the urban areas, where core economy sources account for over three‐quarters of the households. Within these two types of geographical areas, variations in main income sources are associated to a considerable extent with differing characteristics of main income earners and with the households' provincial as well as subregional location. However, also non‐earner characteristics contribute to explaining differences, indicating that different household structures are associated with the various main income sources. Results are also consistent with other findings, which suggest that household formation may be endogenous to income sources (Klasen & Woolard, 2001; Edmonds et al., 2003; Bertrand et al., 2003).

Results imply that the gender, education and age of main income earners all have a considerable impact on integration by main income sources. With small variation across non‐core main income sources, the likelihood of low integration increases if the main income is earned by women, by the elderly or earners of young working age, and by individuals with low levels of education. Within the urban and non‐urban subsamples, main income sources are also subject to interprovincial variation. Of particular concern regarding low core sector integration among both urban and non‐urban households are the Eastern and Northern Cape provinces. Within both subsamples the probabilities of holding specific main income sources also vary across subregions. In non‐urban areas, higher probabilities of reliance on private or public transfer incomes were found to be associated with residence in the former ‘tribal areas’. Residence in agricultural or otherwise commercialised non‐urban areas raises the probabilities of main income sources from the primary sector or domestic services.

Another trait particular to non‐urban households is the many and strong associations between main income source probabilities and the characteristics of household members that do not raise that income. After controls for endogeneity and in consistence with previous findings, higher‐than‐modal fractions of unemployed household members are strongly and positively associated with public transfers, but not with private transfers. The high unemployment rates found among adults in households with private transfer main income can thus only be assimilated if such households have large numbers of adults, taking into account that the earners may also be unemployed. Having larger‐than‐modal proportions of economically non‐active members is positively associated with reliance on either type of transfer income sources, but stronger for private transfers. Furthermore, but only in non‐urban areas, higher‐than‐modal fractions of children are very strongly and positively associated with probabilities of all non‐core main income sources.

It has been noted by authors such as Keller (2003) that poor households differ from the non‐poor in terms of generation structure. The results here support an explanation of that phenomenon based on public transfer incomes largely being age driven and strongly associated with low‐income households, unemployed or inactive members as well as young children. As elderly individuals receive pensions and younger women often have young children, multigeneration households arise when receivers of public pensions support their children and grandchildren (Klasen & Woolard, 2001; Edmonds et al., 2003). The negative association between retired non‐contributors and non‐core main income sources, however, indicates that outside the core economy, children or grandchildren may not share incomes with older generations.

It is questionable whether the now eight‐year‐old patterns of living arrangements and income sources depicted in these data still prevail and whether derived policy implications apply. However, not unlike in a number of other studies, results from this investigation would promote ambitions towards ‘employment creation’. If such polices could somehow be invented, they would be especially useful if spatially targeted and if assimilated to the very different patterns of non‐integration that exist between both urban and non‐urban areas, between provinces and across subregions. The vast variability in means of income generation across space also warrants more spatially targeted research efforts.

The considerable impact even from very low levels of education on core sector access suggests that adult literacy programmes may promote the integration of marginalised or peripheral households. Finally, one can hope that the collection of both old‐age pensions and child support grants has increased since 1995. In particular, the latter policies are supported for poverty alleviation purposes, as targeting transfers at children and young mothers would target the low‐income, transfer‐dependent households and may also be to the benefit of the elderly. However, to the extent that household formation is endogenous to such transfers, households may reshape along with increased collection of such transfers. Answers to that question, and to whether the high concentration around a single income source still applies, can only be answered if reliable, recent income data are made available.