Drawing a profile of poverty is a necessary step to identify the characteristics of the population groups that are poor, but it is not sufficient to measure the impact of various household characteristics on poverty. The problem with a poverty profile lies in the fact that it provides information on who are the poor, or on the probability of being poor among various household categories, but cannot be used to assess the correlates of poverty. For instance, the variation of poverty rates across regions is sometimes better accounted for by the differences in households’ characteristics than by the specificities of each region. To sort out the correlates or determinants of poverty and the impact of various variables on the probability of being poor, regressions are thus needed. Also, when estimating such regressions, it is better to rely on linear regressions for the determinants of consumption per equivalent adult than on categorical regressions for the determinants of poverty. This is because using probits or logits implies throwing away valuable information contained in the household consumption information and runs a higher risk of bias In this chapter, as well as in a separate paper by Diallo and Wodon (2007) using the CWIQ surveys, separate regressions are provided for the urban and rural sectors. Apart from a constant, the regressors include (with a few differences depending on the data sets used): (a) geographic location according to key areas or regions; (b) household size variables (number of infants, children, adults and seniors, and their squared value to take into account potential non-linearity in relationships between household size and consumption), whether the household head is a woman, the age of the head, and the marital status of the head (c) characteristics of the household head, including his/her level of education; his/her employment type and sector of activity; (d) the same set of characteristics for the spouse of the household head; and (e) other variables including migration and land ownership status. The regressions are estimated separately in urban and rural areas, with the logarithm of the consumption per equivalent adult as the dependent variable. The specification of the regression has been kept intentionally simple, so as to permit comparisons over time in the determinants of household consumption and thereby implicitly poverty.
Table 16 provides the results from the regression on the determinants of consumption that pertains to the geographic dummy variables as well as the overall constant. Because these variables are not household characteristics, they essentially represent changes in macroeconomic conditions in the country as a whole, as well as in the different regions, for what could be referred to as a typical poor household3. The value of the coefficients in the table can be interpreted as percentage gains in consumption associated with the various explanatory variables, with the caveat that when a coefficient is not statistically significant, it is replaced by the mention “n\s” in the table.
Three comments are in order. First, the values of the constants in the regressions are typically increasing over time, suggesting that for poor households (and more generally the population as a whole) there has been an improvement over time in well-being. Second, there has been a reversal within urban areas in the relative positions of Accra as opposed to other urban areas (Accra households were comparatively better off in 1998/99 while other urban households tend to do better than Accra households in 2005/06, controlling for other variables). As noted earlier,
this may be a result of the fact that in 1998/99, the households that were interviewed in Accra appeared to be better off than the true overall population for Accra, as appearing in the 2001 Census. If we discount the estimates for 1998/99 in urban areas for this reason, what emerges is the fact that in 1991/02, there were no statistically significant impacts of geographic location in urban areas (after controlling for other variables). Third, for the three years of survey data, households in rural Savannah tend to have levels of consumption lower than rural households in other areas, and the gap has been increasing over time (26.7 percent less consumption in 2005/2006 versus 15.8 percent less in 1991/1992).
Table 17: Determinants of real consumption per equivalent adult – economic climate
Urban Rural
2005/2006 1998/1999 1991/1992 2005/2006 1998/1999 1991/1992
Constant
Constant 14.873*** 15.120*** 14.813*** 14.664*** 14.209*** 14.412*** Region
Accra Ref Ref Ref Urban Coastal 0.271*** -0.430*** n\s
Urban Forest 0.160*** -0.164** n\s Urban Savannah n\s -0.508*** n\s
Rural Coastal Ref Ref Ref Rural Forest n\s n\s -0.117** Rural Savannah -0.267*** -0.206** -0.158***
Source: Authors using GLSS data.
The rest of the coefficient estimates from the regressions are provided in table 18. The messages that emerge are similar to those obtained with the poverty profile presented earlier.
• Demographic Characteristics: An additional person in the household tends to reduce consumption per equivalent adult by up to 13 percent to 17 percent, although the impact is lower for elderly individuals. As in a number of other countries, there are few statistically significant differences between male-headed and female-headed households. In terms of marital structure, households whose head is separated, divorced or widowed tend to be slightly poorer (loss in consumption of 6 percent to 13 percentin 2005/2006). • Education Level of the Head and the Spouse: As expected, consumption levels increase
with the education level of the household head, but the effects are statistically significant only as of secondary schooling. The impact of the spouse’s education is smaller than that of the head, probably because spouses are less likely to work and are likely to earn less. There has however been an increase over time in the gains from education at the upper secondary and tertiary level which has probably contributed to the increase in inequality. • Other variables: After controlling for other variables, employment and other variables do
not appear to have large and systematic impacts on consumption. There is weak evidence that households involved in mining have higher levels of consumption than otherwise comparable households whose head works in other sectors. One rather surprising result (given the evidence provided in the section of this paper on cocoa below) is the fact that cocoa producers are at a disadvantage in 2005/06 versus other self-employed heads in
agriculture. It may be that cocoa producers belong to two groups – one group of small land owners with limited production could face substantial deprivation, while the other group, with larger areas cultivated, better equipment and higher production levels would be better off, with their better status picked up by other variables in the regression, but this would have to be confirmed by a more detailed analysis. Finally, households who have not migrated tend to be slightly better off than households who migrated (this does not mean that there were no gains to migration for the households who did migrate).
Table 14: Determinants of logarithm of consumption per equivalent adult, 1991 to 2006 Urban Rural 2005/2006 1998/1999 1991/1992 2005/2006 1998/1999 1991/1992 Age Groups Age 0 to 4 -0.134*** -0.136*** -0.117*** -0.135*** -0.130*** -0.183*** Age 0 to 4 squared n\s 0.013* n\s 0.009*** 0.016*** 0.028*** Age 5 to 14 -0.169*** -0.145*** -0.178*** -0.173*** -0.179*** -0.233*** Age 5 to 14 squared 0.020*** 0.012** 0.015*** 0.023*** 0.017*** 0.034*** Age 15 to 60 -0.134*** -0.203*** -0.182*** -0.145*** -0.202*** -0.151*** Age 15 to 60 squared n\s 0.016*** 0.014*** 0.010*** 0.023*** 0.013***
Age 61 and over n\s n\s n\s n\s n\s -0.181***
Age 61 and over squared n\s -0.104* n\s -0.048** -0.055* 0.064*
Sex of head
Male Ref Ref Ref Ref Ref Ref
Female n\s n\s n\s n\s n\s 0.073**
Education level of head
No education Ref Ref Ref Ref Ref Ref
Primary n\s n\s n\s n\s 0.134*** n\s
Secondary 1 0.170*** 0.180*** 0.145*** 0.146*** 0.190*** 0.101*** Secondary 2 0.265*** 0.340*** 0.387*** 0.293*** 0.224*** n\s
Superior 0.491*** 0.347*** 0.276*** 0.408*** 0.415*** 0.220**
Education level of spouse
No education Ref Ref Ref Ref Ref Ref
Primary n\s n\s n\s n\s n\s n\s
Secondary 1 0.087** n\s 0.117** 0.107*** 0.104*** 0.106** Secondary 2 0.163** 0.295*** 0.413*** 0.426** 0.405** n\s
Superior 0.273*** n\s 0.351** n\s 0.421*** n\s
Marital Status
Married/Informal Ref Ref Ref Ref Ref Ref
Never married n\s n\s n\s n\s 0.214*** n\s
Separated/Divorced/Widowed -0.134*** n\s n\s -0.065* -0.092* -0.105***
Industry of head
Agriculture Ref Ref Ref Ref Ref Ref
Mining/Quarrying 0.195* n\s n\s n\s 0.500*** 0.277** Manufacturing n\s n\s n\s -0.163* n\s n\s Utilities n\s n\s n\s n\s n\s 0.704*** Construction n\s n\s n\s n\s n\s n\s Trading n\s n\s n\s n\s 0.266** 0.188* Transport/Communication n\s n\s n\s n\s 0.375** n\s Financial Services n\s n\s n\s n\s n\s 0.618***
Community & Other Services n\s n\s n\s n\s n\s n\s
Employment status of head
Public n\s n\s -0.145*** n\s n\s n\s
Wage/private/formal n\s -0.109** -0.240*** n\s n\s n\s
Wage/private/informal -0.265*** -0.198** -0.239*** n\s n\s n\s Self-agriculture-export -0.296*** -0.345*** -0.437*** -0.169* n\s n\s
Self-agro-crop Ref Ref Ref Ref Ref Ref
Migration and land ownership
Migration - Yes Ref Ref Ref Ref Ref Ref
Migration - No 0.056 n\s -0.050 0.075** n\s -0.122***
Area of land owned n\s n\s 0.015** n\s 0.001*** 0.002*** Area of land squared n\s n\s n\s n\s -0.000*** -0.000***
By running the same regressions for the three GLSS surveys, it is also feasible to decompose changes in the mean level of consumption per equivalent adult of households over time into changes due to differences in household characteristics and changes due to differences in the returns to these characteristics (using the Oaxaca decomposition). The results are provided in table 15. For brevity, we focus on the discussion of the results obtained for the whole period under review (the changes in levels of consumption observed in table 15 are different from those reported in section 2 because of the logarithmic transformation used in the regressions).
• Impact of general economic conditions: The first line in the table captures the changes in the constant of the regression as well as the geographic dummy variables. These changes do not reflect changes in household characteristics, but rather changes in the general economic conditions in the country, and how these play out in the various parts of the country. In urban areas, for the full period, general economic conditions helped improve household consumption by 20.5 percent in urban areas and 38.9 percent in rural areas.
• Changes in household characteristics: Household characteristics improved in two major ways. First, there was a reduction in household sizes, which accounts for most of the positive impact of the change in demographic characteristics on consumption (gain of 7.9 percent in consumption in urban areas, and 1.4 percent in rural areas). Second, there was an increase in the education level of household heads and spouses, which generated a gain in consumption of 7.8 percent in urban areas, and 2.0 percent in rural areas. However, the fact that the gains from the demographic and education transitions were much larger in urban than in rural areas suggest that additional efforts must be made in rural areas on these issues.
• Changes in the returns to household characteristics: In urban areas, the gains from changes in the returns associated with different types of employment yielded a 12.2 percent increase in consumption over time for the full period. In rural areas, the reverse was observed, with a consumption loss of 8.1 percent. Given that various household characteristics (on the industrial sector of activity as well as the employment status of the head) are combined in this category, one has to be careful about interpretation. But the basic findings that more attractive jobs became available in urban areas, while this was not the case in rural areas, is coherent with the general poverty trend and the fact that at least some categories of rural household are lagging further behind the rest of the country.
• Overall changes in consumption levels: In urban areas, the improvement in general economic conditions accounted for about half of the total gains in consumption, while in rural areas basically all of the gains were due to the improvement in general economic conditions. The fact that in urban areas there were also gains associated with improvements in household characteristics and the returns to these characteristics helps explains why we observe a substantial difference in the total gains in urban as opposed to rural areas. The increase in the average consumption of households between 1991 and 2005 was 46.1 percent in urban areas (21.3 percent between 1991 and 1999 and 24.8
percent between 1999 and 2006). In rural areas, the increase was about eight percentage points lower, at 37.8 percent (22.3 percent between 1991 and 1999 and 15.5 percent between 1999 and 2006). One key message from this analysis is that in rural areas, more efforts could be placed on helping households move faster through the education and demographic transitions.
Table 15: Contributions of key factors to growth in household consumption, 1991-2006
1991 to 1999 1999 to 2006 1991 to 2006 Change in returns Change in characteristics Change in returns Change in characteristics Change in returns Change in characteristics Urban Geography/overall 5.6% 4.7% 8.8% 0.5% 20.5% -1.0% Demographic -3.5% 4.7% 6.6% 2.6% 2.5% 7.9% Education -1.6% 3.9% 1.7% 2.9% -0.9% 7.8% Employment 9.5% 0.9% 2.8% -1.3% 12.2% -0.3% Others -2.4% -0.4% -0.3% 0.4% -3.1% 0.4% Total 7.6% 13.7% 19.6% 5.2% 31.3% 14.8% 21.3% 24.8% 46.1% Rural Geography/overall -2.0% 1.3% 40.7% -1.0% 38.8% 0.2% Demographic 1.3% 2.5% 2.1% -1.0% 2.8% 2.0% Education 4.1% 3.9% -2.8% -1.9% 2.0% 1.4% Employment 10.5% 1.7% -19.6% -1.0% -8.1% -0.3% Others -1.1% 0.0% -0.4% 0.4% -1.1% 0.0% Total 12.8% 9.4% 20.0% -4.5% 34.4% 3.3% 22.2% 15.5% 37.8%
Source: Authors using GLSS data.
The discussion of the determinants or correlates of poverty has focused above on consumption indicators from the GLSS surveys. Before moving to the next section, it is worth mentioning some additional findings from the analysis of the correlates of household wealth carried by Diallo and Wodon (2007). We focus here on differences in findings rather than on similarities. In terms of demographic variables, apart from information on the number of infants, children, adults, and seniors (and their squared values), on whether the head is female, on the age of the head (and its squared value), and on the marital status of the head, the regressors in the wealth analysis also include whether the head is mentally or physically disabled. One key difference in the wealth as opposed to the consumption analysis is that most household size variables have no or fairly small impacts on assets, for the reasons already explained earlier (the authors do not divide assets by household size when measuring well-being). Furthermore, in 1997, but not in 2003, a handicap reduces the assets owned by households by about 6 percent in rural areas and at the national level (in 2003, the coefficient is still negative, but smaller and not statistically significant). This suggests a mild negative impact of handicap on asset-based well-being. In rural areas and at the national level, female heads have slightly higher levels of assets, with gains ranging from 2 percent to 7 percent, but this is not the case in urban areas. Controlling for other characteristics, the age of the head does not have a statistically significant impact on assets in most cases. Finally, heads in a union have slightly higher levels of wealth, probably related to the need for higher accumulation in order to support their wife and children (this is again a
finding that differs from the consumption-based regressions, and the difference is essentially again due to the difference in the treatment of household size).
The impact of education on asset wealth is confirmed. Literacy brings in a gain of about 5 percent to 7 percent versus having a head illiterate, and primary education brings in a bit more (gain of 2 percent to 4 percent in most cases). Completing junior secondary school adds 6 percent to 8 percent in terms of assets versus no education (on top of the gain associated to literacy), while secondary/technical education brings in a larger gain of 13 to 18 percent. At the post-secondary and higher level, the gain in asset wealth versus no education at all varies from 29 to 37 percent (to which one must also add the gain linked to literacy). The impact of employment is lower, and actually in most case not statistically significant once education is controlled for. For example, whether the head is employed or not does not make a large difference, and there is no systematic gain or loss associated to the private or parastatal sector as compared to the public sector, except for rural areas in 1997. What does matter, however, is the sector of activity of the head, with households whose head is not in agriculture doing better, with assets gains of 15 percent to 22 percent versus households whose head is in agriculture. Finally, as is the case for the analysis of consumption, even after controlling for all the above variables, geographic location still maters. In the national regressions, living in urban areas brings in a gain in assets of 31 percent to 37 percent. As for the regional gains or losses, they are also large, which helps explain the relatively high levels of migration observed within the country.