In search of poverty predictors: the case of urban and rural Pakistan.

Jamal, Haroon

The main objective of this research is to provide correlates of household consumption or poverty using the latest household survey. The estimated coefficients and their weights may be used to predict poverty incidence from light monitoring survey such as Core Welfare Indicator Questionnaire (CWIQ). The CWIQ survey instrument essentially collects simple welfare indicators from a large segment of population and is not designed to measure income, consumption or expenditure. The paper estimates consumption functions separately for urban and rural areas. These functions are estimated with the help of non-monetary correlates of consumption and applied to predict poverty at provincial and district levels. The paper also provides the latest estimates of poverty in the country using a consistent methodology. Overall, 33 percent people were poor, according to estimates from the latest available household survey of 2001-02. Incidence, depth, and severity of poverty are high in rural areas, as compared to their urban counterpart.

1. INTRODUCTION

The Core Welfare Indicator Questionnaire (CWIQ) survey instrument essentially collects (1) simple welfare indicators and indicators of access, use, and satisfaction of public services. It is not designed to measure income, consumption or expenditure. Nevertheless, to fully analyse the CWIQ data, it is necessary to devise a means for distinguishing poor from non-poor households. Thus, there is a need to identify a set of poverty correlates or predictors and estimate their respective weights to predict household consumption and to rank households for poverty analysis.

This paper provides latest estimates of poverty for the year 2001-02 in the country using consistent (2) methodology. It also computes predicted consumption functions, separately for urban and rural Pakistan. These functions are estimated with the help of non-monetary correlates of consumption and applied to predict poverty at sub-national and sub-provincial levels.

The paper uses Pakistan Integrated Household Survey (PIHS) 2001-02 for the analysis. Household Integrated Economic Survey (HIES) section of the PIHS is mainly used for the estimation of monetary poverty. HIES includes standard and detailed income and consumption modules and is traditionally used to estimate poverty in Pakistan.

The organisation of the paper is as follows. The estimates of the poverty line and poverty during the year 2001-02 are presented in the next section. Section 3 discusses the methodology for modelling predicted consumption function. The estimated poverty correlates are provided in Section 4. Application of the consumption functions to predict poverty at sub-national level is presented in Section 5, while the last section is reserved for some concluding remarks.

2. LATEST POVERTY ESTIMATES

For an inter-temporal comparison of the poverty line and estimates of poverty aggregates (incidence, depth and severity), it is essential to adhere to consistent methodology and norms. The methodology adopted in Jamal (2002) to estimate poverty for the year 1987-88, 1996-97 and 1998-99 is applied to the latest available household survey data (PIHS-HIES, 2001-02). Although the details of various methodological options and recommended steps are provided in that paper, the following is a brief description of the major steps to compute the poverty line and poverty for the year 2001-02.

Poverty can be used to define the poor by total household expenditure falling short of the poverty line by the average dietary pattern the expenditure would translate into fewer calories than required. Therefore to compute the poverty line, calorie norms (cutoff points) and estimated coefficients of the calorie-consumption function (CCF) are required. The idea is to get the estimates of total household expenditure required to obtain the minimum required calories. This paper follows 2550 and 2230 calories per day per adult as calorie norms (3) for rural and urban areas respectively. Household food consumption is translated into calories using Food Consumption Tables for Pakistan [Pakistan (2001)].

Calorie-consumption functions are estimated separately for urban and rural areas. It is argued that consumption behaviour, purchasing patterns, dietary habits, taste, ecology etc. are extremely different for urban and rural groups. Following Jamal (2002), these functions are estimated from the lower quartile of distribution after ranking households by per capita expenditure. Household per adult daily calorie consumption is regressed on total expenditure (excluding taxes). The functional form is chosen on the basis of maximisation of [R.sup.2] criterion. Nonetheless, other statistical tests are also applied before choosing the functional form. The results of these functions are furnished in the Appendix (Table A-1).

Table 1 displays computed poverty lines from these estimated calorie consumption functions. As separate calorie-consumption functions are estimated for urban and rural areas, direct estimation of the national poverty line is not viable. A population weighted average poverty line, however, turns out as Rs 646 per capita per month at the prices of HIES 2001-02 survey. (4)

Table 2 displays various measures (5) of poverty for the year 2001-02. The estimated poverty lines for urban and rural areas are mapped on household per capita total expenditure for computing these measures. Overall, 33 percent of the population was poor, according to the above definition of poverty and the poverty line. The incidence, depth and severity of rural poverty are high as compared with the urban areas.

The trend in poverty incidence is portrayed in Table 3. A few observations emerge. On average, 3 percent annual growth in poverty incidence is estimated between the year 1987-88 and 2001-02. The table indicates a relatively higher increase in urban poverty during 1998-99 and 2001-02. Rural poverty in this period has increased with an annual growth rate of 3.1 percent, while the increase is about 7 percent in the case of urban poverty incidence. (6)

3. MODELLING PREDICTED WELFARE OR POVERTY

It is assumed that the approximating mean function h(x,[theta]), relating to response (welfare) variable to the covariates, x is linear in its parameter [theta]. That is the conditional expectation, E(y|x) of the response y given the covariates is related to the linear predictors by the response link function h(x,[theta]). Some continuous variables with strong predictive capabilities were dichotomised to discriminate between poor and non-poor households. These dummy regressors were constructed and included in the model to capture the effects of qualitative independent variables. The resulting variables were then fitted into a model which contains both continuous and discrete dummy variables. The reduce form of the model is specified by Equation (1) below:

[Y.sub.j] = [X.sub.j][beta] + [[lambda].sub.j1] [[gamma].sub.1] + [[lambda].sub.j2] [[gamma].2] + .... + [[lambda].sub.jk] [[gamma].sub.k] + [[micro].sub.j] ... (1)

where, [Y.sub.i] is the response variable; [X.sub.j] is a matrix of continuous explanatory variables; [lambda]s are the respective explanatory dummies variables; [beta]s are the estimated coefficients relative to the continuous variables; [gamma]s are the estimated coefficients associated with the selected dummy variables; and [[mu].sub.j] is the standard error term. The best poverty predictors were the ones that contributed to a significant marginal increase in the explanatory power of the model.

The response variable may be represented by the total household expenditure. (7) It is a standard multivariate regression analysis and estimates the partial correlation coefficient between expenditure and the explanatory variables. Typically, a logarithmic transformation is applied to the response surface to make the relationship between the y and the x's linear. The transformation stabilises the error variance, reduces asymmetry in the distribution of error terms and improves prediction power. The estimated weighted function is continuous and allows the construction of predicted household expenditure which is used as a basis for poverty analysis in light monitoring survey such as CWIQ.

Alternatively, a dichotomous variable explaining poor/non-poor status may be represented as a response variable. In this case, a logit or probit regression of the binary variable is estimated using the maximum likelihood estimation procedure. Based on the assumptions about the error term of the model, probability is computed to predict the household poor/non-poor status.

The selection of appropriate poverty predictors is the next step in the modelling consumption function. Initially the set of regressors includes a host of explanatory variables both discrete and continuous. These initial regressors are essentially household level variables (8) focusing on: household assets, education level and literacy, employment, household amenities, household structure, and demographic characteristics and geographical location. These variables (9) were constructed from the latest household survey (PIHS-HIES, 2001-02) and only those that strongly correlated with household total expenditure were retained for further testing. A stepwise procedure allows one to calibrate the models by dropping explanatory variables with less predictive power. (10) Optimal poverty predictors are selected using a combination of multiple regression analysis and test for correlation and prediction. Once the poverty predictors were identified, their corresponding weights may be used to predict response (household expenditure) variable.

4. POVERTY CORRELATES

As mentioned above, two alternative methods of specifying the response (dependent) variable are available. A continuous variable (log of household expenditure) or a binary variable may be used to statistically correlate household characteristics with poverty status or consumption behaviour. However, it is argued that poverty status binary variable (poor/non-poor) is computed from household expenditure and by using this variable one may loose much of the information available about the actual relationship between expenditure and its explanatory factors. It is, therefore recommended that the analysis is best carried out with the expenditure variable rather than the poor/non-poor status of households.

Nonetheless, to check the sensitivity of results and relative power of prediction, both methods are applied to estimate the consumption function. To a large extent both alternatives yielded similar prediction power, statistical significance of poverty predictors and goodness of fit. Table 4 portrays a comparative picture of both methods in terms of percentage of correct prediction.

It is evident from the table that both specifications work relatively well in urban areas and estimated 84 percent cases appropriately in the actual category of households. In rural areas, however the prediction power is somewhat less and about 78 percent cases were put in the right category of households. Having reached a conclusion that both specifications are the same in terms of prediction power, further description of results and application are based on a multivariate regression analysis that specifies logarithm of expenditure as the dependent variable. (11)

Tables 5 and 6 present regression results of estimated consumption function for urban and rural areas respectively. The adjusted R-square, which is a measure of goodness of fit, is 0.69 for urban and 0.52 for rural areas. In a cross-section analysis, these magnitudes are considered good enough for acceptability of the model. The magnitudes of Durbin-Watson statistic indicate that the relationship between consumption and poverty predictors is not spurious. Multicollinearity among independent variables, which makes the coefficients statistically less efficient and insignificant, is tested through the condition index. The index value greater than 30 indicates severity of multicolinearity and points to less reliability about the magnitude of the coefficients. The estimated results however, indicate that the value of the condition index is less than 30 for urban as well as rural areas. Having illustrated the summary statistics of estimated consumption functions, some observations regarding poverty correlates are in order.

Family size and dependency are important poverty predictors. The dependency is represented by the proportion of children and members greater than 65. Both determinants are highly correlated with expenditure.

In rural areas, ownership of livestock, poultry, land, non-residential and residential property are all positively correlated with household expenditure. Further, medium and large farmers (ownership of land greater than 12.5 acres) play a dominant role in distinguishing non-poor from poor. In fact, the magnitude of coefficient associated with the variable representing medium and large farmers is the highest. Owner cultivator is also an important determinant of household non-poor status.

One variable that appears to be highly correlated with aggregated household total expenditure with strong predictive capability is the "asset score". This variable is constructed by assigning equal (12) weight to each of the seventeen assets (13) listed in the PIHS questionnaire. A constant 1 is assigned to each of the assets owned by the household, and the assets score is obtained by summing up across all assets at the household level. Of course uniform allocation of score irrespective of the asset characteristics tends to smooth out the distribution of assets across households. To the extent that these assets have different values and all exhibit different rates of depreciation, uniform allocation might even increase the distortion in the distribution of household assets. But, what actually matters in this construction is the ownership of assets by a household and not so much the values of the asset which are difficult to estimate accurately from surveys carried out in a single visit to the household. The maximum asset score is 17 and the minimum is 0 for poorest households which possess none of the assets listed.

The significant and major role of education, especially higher education in urban areas is evident from Table 5. The magnitude of the coefficients associated with higher secondary (intermediate) or tertiary education of the head of a household plays a decisive role in determining the household's consumption/poverty status.

The quality of housing structure in terms of material used is an important determinant of poverty status. Unfortunately, the household survey (PIHS, 2001-02) does not provide the relevant information to capture the quality of housing stock. Therefore, only the housing congestion, represented by number of person per room is included in the consumption function. In housing services, telephone connection appeared as an important determinant of poverty status, both in urban and rural areas.

The magnitude of coefficients associated with domestic and overseas transfers clearly indicates the significance of these variables in determining the household's poverty status. In urban areas, the highest magnitude is associated with households receiving overseas remittances.

Above-mentioned poverty correlates, however estimate medium term expenditure. The methodological framework does not allow to explicitly incorporate some short term transitory factors related to price, and income shocks in the model. Nonetheless, the predicted consumption/expenditure is a good proxy for permanent income barring some short-term fluctuations.

5. PREDICTED POVERTY INCIDENCE AT THE SUB-NATIONAL LEVEL

The estimated non-monetary poverty correlates with the respective weights (14) are applied to determine the provincial and district level poverty incidence in Pakistan. (15) The estimated response on log scale was transformed back and converted into per capita expenditure to remove the effects of the household size. The transformed predicted response was then used to categorise households into poor/non-poor using the poverty lines described above. (16) Table 7 depicts provincial poverty incidences, separately for provincial capitals, large cities, small cities, and towns and rural areas. The urban and rural poverty incidences at district (17) level are presented in Appendix (Table A-4 and Table A-5 respectively).

According to the provincial ranking in terms of lowest poverty incidence, NWFP province ranks second after Punjab province. This may be partly explained with the relatively low rural poverty incidence in NWFP as compared with rural Sindh. The data (not reported here) reveals that overseas and domestic remittances are major contributors towards lowering the poverty incidence in NWFP province. The plight of residents of small cities and towns are also evident from the table. (18) On the average, 40 percent residents of the town are categorised as poor. Balochistan, as expected, ranks the lowest in urban as well as rural poverty levels.

5. CONCLUDING REMARKS

The need to identify a set of poverty predictors and estimate their respective weights arises from the fact that it is expensive to collect detailed household consumption and income data frequently and from a large segment of the population. After devolution of power to the district levels, it is also argued that district-wise poverty estimates should be available to monitor the impact of policies adopted by the district administration. To act in response, the Federal Bureau of Statistics, Government of Pakistan is launching a nationwide survey using the Core Welfare Indicator Questionnaire. This survey instrument essentially collects simple welfare indicators and indicators of access, use and satisfaction of public services. It is not designed to measure income, consumption or expenditure. Nevertheless, to fully analyse the CWIQ data, it is necessary to devise a means for distinguishing the poor from non-poor households.

This paper explored poverty correlates in the context of urban and rural Pakistan. The specificities of developing economies, in particular the dualism between urban and rural areas, motivates one to identify the correlates or determinants of poverty taking into account the clear distinction that must appear either in the analysis of poverty or during the adoption of appropriate economic policies.

At the first step, the poverty incidences for urban and rural areas are estimated using latest available household survey and making use of consistent methodology for poverty estimation. According to the estimates, an overall 33 percent of people were poor during 2001-02. The incidence, depth and severity of poverty is high in rural areas as compared with their urban counterpart. The trend in poverty incidence indicates a relatively high increase in urban poverty during 1998-99 and 2001-02. Rural poverty in this period has increased with an annual growth rate of 3.1 percent, while this percentage is about 7 in the case of urban poverty incidence.

Total household expenditures are then statistically analysed in terms of various household non-monetary (demographic, social, housing etc.) indicators to determine consumption correlates. The results show that in urban areas the main factors which discriminate against poverty include the head of the households education and dependency ratio. In rural areas, asset distribution, especially land and livestock play an important role in distinguishing non-poor from poor. The role of domestic and overseas transfers also appeared significant in discriminating against poverty. Its role is more striking in urban areas.

With the help of these estimated consumption functions, poverty incidences are predicted for provinces and also for selected districts. According to predicted provincial poverty incidence, Punjab ranks first, while Balochistan province ranks forth. Surprisingly, NWFP province ranks second instead of Sindh province. This is, perhaps mainly due to a very low incidence of rural poverty in NWFP. Another important finding, which emerged from this exercise, is that residents of small town and cities are in a vulnerable situation. The poverty incidence in small cities and towns, barring Balochistan rural areas, is the highest in all provinces.

Appendices

Appendix Table A-1
Estimated Calorie-Consumption Functions

 Estimated
 Coefficients t-value

Urban Areas
 Dependent Variable
 Log (Per Adult Calorie Consumption)
 (Constant) 1.290 4.4
 Per Adult Expenditure 0.944 20.4
 Dummy Variable for Sindh -0.384 -1.3
 Dummy Variable for NWFP 0.112 3.3
 Dummy Variable for Balochistan 0.021 0.6
Rural Areas
 Dependent Variable
 Log (Per Adult Calorie Consumption)
 (Constant) 5.977 147.1
 Per Adult Expenditure 0.283 45.5
 Dummy Variable for Sindh -0.302 -3.9
 Dummy Variable for NWFP 0.093 1.1
 Dummy Variable for Balochistan -0.967 -10.8

 [R.sup.2] F-value

Urban Areas 0.25 112.8
 Dependent Variable
 Log (Per Adult Calorie Consumption)
 (Constant)
 Per Adult Expenditure
 Dummy Variable for Sindh
 Dummy Variable for NWFP
 Dummy Variable for Balochistan
Rural Areas 0.20 539.4
 Dependent Variable
 Log (Per Adult Calorie Consumption)
 (Constant)
 Per Adult Expenditure
 Dummy Variable for Sindh
 Dummy Variable for NWFP
 Dummy Variable for Balochistan

Source: Estimates are based on PIHS-HIES, 2001-02.

Appendix Table A-2
Estimates of Logistic Function-Urban Areas (Poor = 1)

 Coefficients Significant
 Level

Demography
 Family Size 0.261 0.000
 Proportion of Children Less than 5 Years 0.012 0.000
 Proportion of Members Greater than 65
 Years 0.015 0.000
 Number of Earners -0.082 0.037
Education
 Proportion of Out of School Children
 (Secondary) 0.589 0.000
 Highest Education Level in Family -0.044 0.000
Head of Household
 Education Level--Illiterate 0.171 0.109
 Education Level--Primary 0.297 0.010
 Education Level--Higher Secondary -0.343 0.179
 Education Level--Tertiary -0.416 0.077
 Occupation--Employer -0.961 0.014
Household Assets
 Asset Score -0.54 0.000
 Ownership of Non-residential Property -0.504 0.039
Housing Quality and Services
 Person per Room 0.220 0.000
 Telephone Connection -1.136 0.000
Transfers
 Overseas Remittances Receiving Household -2.050 0.000
 Domestic Remittances Receiving Household -0.521 0.001
Locational Variables
 Small Cities and Towns 0.771 0.000
 Punjab Province 1.108 0.000
 NWFP Province 0.991 0.000
 Intercept (Constant) -4.383 0.000

Source: Estimates are based on PIHS-HIES, 2001-02.

Appendix Table A-3
Estimates of Logistic Function:-Rural Areas (Poor = 1)

 Coefficients Significant
 Level

Demography
 Adult Equivalent Unit 0.253 0.000
 Proportion of Children Less than 5 Years 0.005 0.019
 Proportion of Members Greater than 65
 Years 0.017 0.000
Education
 Proportion of Out of School Children
 (Secondary) 0.133 0.042
 Proportion of Out of School Children
 (Primary) 0.184 0.028
 Female Highest Education Level in Family -0.054 0.000
Head of Household
 Education Level -0.045 0.000
 Age of Head -0.004 0.092
 Occupation--Own Cultivator -0.437 0.000
 Occupation--Large Farmers -0.943 0.000
 Occupation--Landless 0.068 0.455
Household Assets
 Livestock Ownership -0.24 0.001
 Poultry Ownership -0.731 0.000
 Asset Score -0.648 0.000
 Ownership of Non-agriculture Land -0.602 0.000
 Ownership of Residential House -0.309 0.000
Housing Quality and Services
 Person per Room 0.178 0.000
 Electricity Connection -0.328 0.000
 Telephone Connection -1.333 0.000
 No Toilet in House 0.422 0.000
Transfers
 Overseas Remittances Receiving Household -1.380 0.000
 Domestic Remittances Receiving Household -0.363 0.000
Locational Variables
 Sindh Province -0.572 0.000
 Balochistan Province 0.533 0.000
 Intercept (Constant) -3.107 0.000

Source: Estimates are based on PIHS-HIES, 2001-02.

Appendix Table A-4

Urban Poverty Incidence--District Scenario
(Predicted Percentage of Population below the Poverty Line)

 Overall

Districts Incidence t-value

Islamabad 6.87 (2.6)
Sialkot 13.95 (3.2)
Rawalpindi 24.22 (7.1)
Lahore 24.29 (10.7)
Faisalabad 28.78 (10.1)
Gujranwala 29.30 (8.8)
Multan 35.39 (10.4)
Sargodha 37.27 (9.2)
Bahawalpur 46.59 (10.1)
D. G. Khan 65.79 (10.8)
Karachi 11.38 (7.4)
Sukkur 28.62 (6.7)
Mirpurkhas 32.05 (5.6)
Hyderabad 33.62 (9.0)
Larkana 40.28 (8.0)
Bannu 17.30 (3.2)
Haripur 19.98 (4.5)
Peshawar 31.86 (9.4)
D. I. Khan 37.65 (4.6)
Kohat 44.30 (6.8)
Mardan 46.98 (8.6)
Malakand 53.45 (8.1)
Quetta 23.25 (5.8)
Makran 29.72 (5.2)
Sibi 32.13 (3.7)
Nasirabad 43.23 (5.7)
Zhob 43.83 (4.4)
Kalat 47.67 (8.3)

 Large City Sample

Districts Incidence t-value

Islamabad 6.87 (2.6)
Sialkot 13.95 (3.2)
Rawalpindi 19.39 (4.6)
Lahore 18.62 (7.3)
Faisalabad 23.28 (6.7)
Gujranwala 13.98 (3.6)
Multan 28.65 (5.7)
Sargodha 24.00 (4.4)
Bahawalpur 40.54 (5.2)
D. G. Khan
Karachi 11.38 (7.4)
Sukkur 14.88 (2.1)
Mirpurkhas
Hyderabad 21.21 (4.6)
Larkana
Bannu
Haripur
Peshawar 27.56 (6.2)
D. I. Khan
Kohat
Mardan
Malakand
Quetta 16.49 (3.5)
Makran
Sibi
Nasirabad
Zhob
Kalat

 Small City and
 Town Sample

Districts Incidence t-value

Islamabad
Sialkot
Rawalpindi 31.04 (5.5)
Lahore 41.30 (8.9)
Faisalabad 35.88 (7.6)
Gujranwala 38.94 (8.4)
Multan 39.76 (8.7)
Sargodha 40.80 (7.4)
Bahawalpur 38.04 (8.4)
D. G. Khan 65.79 (10.8)
Karachi
Sukkur 33.46 (5.9)
Mirpurkhas 32.05 (5.6)
Hyderabad 45.22 (7.9)
Larkana 40.28 (8.0)
Bannu 17.30 (3.2)
Haripur 19.98 (4.5)
Peshawar 41.47 (7.7)
D. I. Khan 37.65 (4.6)
Kohat 44.30 (6.8)
Mardan 46.98 (8.6)
Malakand 53.45 (8.1)
Quetta 46.51 (6.1)
Makran 29.72 (5.2)
Sibi 33.13 (3.7)
Nasirabad 43.23 (5.7)
Zhob 43.83 (4.4)
Kalat 47.67 (8.3)

Source: Estimates are based on PIHS-HIES, 2001-02 and the
estimated consumption function.

Appendix Table A-5

Rural Poverty Incidence--District Scenario
(Predicted Percentage of Population below the Poverty Line)

Districts Incidence t-value

Rawalpindi 4.17 (1.7)
Jhelum 7.49 (2.2)
Islamabad 8.96 (2.9)
Gujranwala 10.07 (2.7)
Sargodha 12.64 (4.2)
Mianwali 17.11 (3.2)
Narowal 18.40 (3.5)
Sialkot 19.29 (3.4)
Sahiwal 19.70 (4.0)
T.T. Singh 19.70 (4.1)
Faisalabad 20.51 (6.1)
M. Bahuddin 21.31 (5.2)
Attack 21.75 (4.0)
Bahawal Nagar 22.51 (4.7)
Lahore 24.91 (4.5)
Hafizabad 26.24 (4.7)
Bhakker 27.27 (3.5)
Khushab 27.96 (4.3)
Okara 29.70 (7.3)
Pak Pattan 30.42 (4.1)
Jhang 30.42 (6.6)
Bahawalpur 30.91 (6.5)
Khanewal 32.12 (6.1)
Sheikhupura 32.58 (6.6)
Lodhran 34.06 (4.6)
Kasur 34.68 (7.1)
Multan 37.03 (6.8)
Muzaffargarh 38.39 (7.8)
D.G. Khan 40.24 (5.6)
Layyah 42.24 (4.0)
Vehari 44.60 (7.3)
R.Y. Khan 45.66 (9.6)
Rajanpur 62.09 (6.3)
Karachi 26.29 (3.9)
Tharparkar 28.45 (5.4)
Khairpur 32.71 (7.4)
Badin 33.08 (6.3)
Hyderabad 33.14 (7.2)
Ghotki 33.70 (5.7)
Thatta 34.35 (6.9)
Nawabshah 36.46 (6.1)
Nashero Feroze 38.39 (7.1)
Larkana 42.23 (10.1)
Sanghar 42.53 (8.6)
Dadu 42.92 (9.1)
Mirpur Khas 43.23 (8.4)
Shikarpur 43.87 (6.2)
Sukkur 52.46 (6.0)
Jaccobabad 63.19 (11.3)
Abbottabad 12.15 (2.4)
Mansehra 15.02 (4.0)
Karak 20.24 (2.4)
Swat 21.80 (4.7)
Dir 21.98 (4.7)
Batagram 22.67 (3.3)
Charsadda 24.96 (4.5)
Bonair 25.63 (3.6)
Nowshera 27.11 (3.6)
Shangla 28.88 (3.3)
Mardan 29.64 (5.9)
Swabi 29.77 (4.7)
Kohistan 31.97 (4.0)
Tank 32.81 (4.5)
Bannu 34.47 (5.8)
Hangu 38.73 (3.8)
D.I. Khan 45.98 (8.4)
Peshawar 47.41 (8.2)
Quetta Division 33.48 (10.2)
Mekran Division 37.78 (8.5)
Sibi Division 53.24 (11.1)
Kalat Division 56.20 (16.4)
Zhob Division 58.28 (14.7)
Nasirabad Division 58.73 (15.0)

Source: Estimates are based on PIHS-HIES,
2001-02 and the estimated consumption function.

Note: Due to insignificant t-value (large standard errors),
District Gujrat from Punjab and districts Bonair,
Malakand, Kohat, Haripur and Luki Marrwat of NWPF
province are excluded from the analysis.

Author's Note: The views expressed in this paper are those of the author and do not necessarily represent those of the Social Policy and Development Centre.

REFERENCES

Ercelawn, A. (1992) Absolute Poverty in the 80s--Rural and Urban Trends in Pakistan. Applied Economics Research Centre. (Discussion Paper No. 161.)

Jamal, H. (2002) On the Estimation of an Absolute Poverty Line: An Empirical Appraisal. The Lahore Journal of Economics, July-December.

Pakistan, Government of (2001) Food Consumption Table for Pakistan. Peshawar: Department of Agricultural Chemistry, NWFP Agriculture University.

Pakistan, Government of (2002) Pakistan Integrated Household Survey. Islamabad: Federal Bureau of Statistics, July.

(1) The Federal Bureau of Statistics, Government of Pakistan, is launching a nationwide survey using the Core Welfare Indicator Questionnaire (CWIQ). The survey will provide district level welfare indicators with a sample size of about 77,000 households. The Government has planned to conduct CWIQ survey over alternate years in the entire country for rapid assessment of indicators on health, education, employment, demography, population welfare, household assets and service delivery system.

(2) In Jamal (2002), a consistent methodology is applied to estimate poverty for the years 1987-88, 1996-97 and 1998-99. Similar methodology is applied in this paper for the year 2001-02.

(3) The Poverty line and, hence, poverty incidence is very sensitive to a change in calorie norms or cutoff points. Therefore, it is highly recommended to adhere to a cutoff point, whatever it may be, for inter-temporal comparison of poverty incidence and the poverty line. Same calorie norms are used for 1987-88, 1996-97, and for 1998-99.

(4) The officially national poverty line is Rs 748.56 per capita per month. However, Government does not notify separate poverty for urban and rural areas.

(5) These measures are well known. For detail see Jamal (2002).

(6) According to the official estimates provided in Pakistan (Economic Survey, 2003-04 (page 49), the annual growth in rural poverty between 1998-99 and 2000-01 is 6.2 percent and growth in urban poverty incidence is 4.2 percent. These results are contradictory with our estimates. Although both results are not comparable due to differences in methodology and calorie norms, and important distinction between these two estimates is worth mentioning. This paper uses separate poverty lines for urban and rural areas, while Government uses on poverty line for computing official urban and rural poverty incidences.

(7) The household expenditure is often divided by the poverty line to ensure comparability across regions. Since, in this paper urban and rural welfare predicted functions are estimated separately, it was not felt necessary to divide household expenditure by the poverty line.

(8) The member-level variables such as literacy and enrollment are aggregated at the household level for consistency in the estimation. This aggregation of individual characteristics at the household level produces variables such as proportion of children enrolled in each household, proportion of household members literate etc.

(9) The choice of variable is, however restricted and depends on the availability of data in household survey. For instance, quality of housing stock is an important poverty predictor, but was not included in the initial list of predictors due to non-availability of relevant information in PIHS/HIES. Choice of variables is also depends on the availability of information in CWIQ. All variables that were included are present in the proposed CWIQ survey except some household assets that are not asked in CWIQ.

(10) Various statistical selection criteria are available in selecting best model. These statistics include Akaike Information Criterion, Amemiya Prediction Criterion, Mallows' Prediction Criterion and Schwarz Prediction Criterion. In this paper, Akaike Information Criteria is used to select the best model.

(11) The detailed results of logit estimates are provided in the Appendix, Table A-2 and Table A-3.

(12) One popular method for obtaining weighted score is the Principal Components Analysis (PCA). The weighted Factor Score, which is derived from PCA is also attempted and used as a regressor instead of score computed by assigning equal weight to each assets. However, no improvement and no significant changes in the results are observed. Therefore, simple scoring of assets is preferred.

(13) These assets are; refrigerator, freezer, air-conditioner, air cooler, geyser, washing machine, camera, cooking range, heater, car, motorcycle, VCR, cassette player, compact disk player, vacuum cleaner and computer.

(14) It may be hypothesised that the weights or magnitude of non-monetary correlates would be stable overtime (at least in the short run). However, the phenomenon is not yet empirically proved. An attempt was made to estimate consumption functions from HIES 1996-97 data. Despite some missing variables, the magnitudes of coefficients were not off and quite close. Nonetheless in the absence of any strong empirical support, it is difficult to definitely conclude that coefficients are stable overtime.

(15) The direct estimates of poverty incidence at provincial or district level from household surveys is not recommended due to large standard errors, non-normality and heteroscedasticity in income or consumption variables. The sample design of HIES allows only the computation of the poverty statistics at the national or regional (urban/rural) with an acceptable measure of reliability. Therefore household consumption, which is predicted with the help of non-monetary indictors, is used to estimate poverty statistics for provinces or districts. It is argued that non-monetary variables (demography, education, housing etc.) are less heterogeneous and normally distributed across the sampling stratum. The size of standard error in two-stage estimates depends largely on the degree of disaggreagation sought and the explanatory power of the exogenous variables in the first-stage model.

(16) Poverty incidences (within sample) are indirectly predicted using the information provided by the estimated consumption functions. For out of sample prediction or prediction overtime, the mean consumption level (constant or intercept term) should be used alter adjusting inflation.

(17) District poverty incidences are computed only for those districts which are included in the sampling frame of the Federal Bureau of Statistics.

(18) These findings are consistent with the earlier study by Ercelawn (1992), for poverty incidence during 1980s.

Haroon Jamal is Principal Economist, Social Policy and Development Centre, Karachi.

Table 1
Estimated Poverty Lines (2001-02)

 Urban Rural

Per Day Calorie Requirements--Per
 Adult Equivalent Unit 2230 2550
Per Day Calorie Requirements--Per Capita ** 1889 2104
Poverty Line--Rupees Per Capita Per Month 761 605

Source: Author's estimates based on PIHS-HIES, 2001-02.

** In order to ease in interpretation, minimum calorie requirements
are converted into per capita term using household demographic data
and proportionate minimum requirements. The minimum requirements by
age and sex are available in Food Consumption Table for Pakistan
(2001).

Table 2
Estimates of Poverty Measures, 2001-02 (Percent of Poor Individuals)

 Head Count Index Poverty Gap Index FGT2 Index
 (Incidence) (Depth) (Severity)

Pakistan 33 7.16 2.27
Urban 30 7.10 2.41
Rural 35 7.18 2.21

Source: Author's estimates based on PIHS-HIES, 2001-02.

Table 3
Trends in Poverty Incidence
(Percentage of Population below the Poverty Line)

 1987-88 1996-97 1998-99 2001-02

Pakistan 23 28 30 33
 (2.4%) (3.6%) (3.3%)
Urban 19 25 25 30
 (3.5%) (0%) (6.7%)
Rural 26 30 32 35
 (1.7%) (3.3%) (3.1 %)

Source: Author's estimate fix the year 2001-02 is based on
PIHS-HIES, 2001-02. The poverty incidences for other years
are taken from Jamal (2002). Consistent methodology is
applied for all years.

Note: Annual growth rates from previous period are given
in parenthesis.

Table 4
Predicted Power of Estimated Functions

 Percentage of Correctly Predicted
 Households

 Non-poor Poor Overall
 Correct
 Prediction
Urban Areas
 OLSQ Regression 92.62 58.72 84.02
 Logistic Regression 92.24 59.64 83.97
Rural Areas
 OLSQ Regression 90.04 52.06 78.26
 Logistic Regression 89.19 55.01 78.58

Source: Author's Estimates.

Table 5
Predicted Consumption Function--Urban Areas
(Dependent Variable-Logarithm of Total Household Expenditure)

 Coefficients Significance
 Level

Demography
 Family Size -0.058 0.002
 Proportion of Children
 Less than 5 Years -0.002 0.000
 Proportion of Members
 Greater than 65 Years -0.003 0.000
 Number of Earners in
 Household 0.007 0.005
Education
 Proportion of Out of
 School Children
 (Secondary) -0.077 0.016
 Highest Education Level in
 Family 0.006 0.002
Head of Household
 Education Level-Illiterate -0.046 0.014
 Education Level-Primary -0.044 0.015
 Education Level-Higher
 Secondary 0.093 0.022
 Education Level-Tertiary 0.173 0.019
 Occupation-Employer 0.188 0.035
Household Assets
 Asset Score 0.094 0.003
 Ownership of
 Non-residential Property 0.131 0.024
Housing Quality and Services
 Person per Room -0.051 0.003
 Telephone Connection 0.216 0.015
Transfers:
 Households Receiving
 Overseas Remittances 0.286 0.024
 Households Receiving
 Domestic Remittances 0.094 0.018
Locational Variables
 Small Cities and Towns -0.139 0.011
 Punjab Province -0.218 0.013
 NWFP Province -0.177 0.017
 Balochistan Province -0.053 0.019
 Intercept (Constant) 7.629 0.023
Summary Statistics
 Adjusted R-square 0.69 Condition Index 15.49
 F-value 592.87 Durbin-Watson 1.58

Source: Author's Estimates based on PIHS-HIES, 2001-02

Table 6
Predicted Consumption Function--Rural Areas
(Dependent Variable--Logarithm of Total Household Expenditure)

 Coefficients Significance
 Level
Demography
 Family Size -0.046 0.000
 Proportion of Children Less than 5
 Years -0.002 0.000
 Proportion of Members Greater than
 65 Years -0.003 0.000
Education
 Proportion of Out of School
 Children (Secondary) -0.033 0.000
 Proportion of Out of School
 Children (Primary) -0.030 0.010
 Highest Female Education Level in
 Family 0.005 0.000
Head of Household
 Education Level 0.009 0.000
 Age of Head 0.001 0.000
 Occupation--Own Cultivator 0.057 0.000
 Occupation--Medium and Large
 Farmers 0.181 0.000
 Occupation--Landless -0.059 0.000
Household Assets
 Livestock Ownership 0.055 0.000
 Poultry Ownership 0.075 0.000
 Asset Score 0.107 0.000
 Ownership of Non-agriculture Land 0.085 0.000
 Ownership of Residential House 0.023 0.037
Housing Quality and Services
 Person per Room -0.035 0.000
 Electricity Connection 0.065 0.000
 Telephone Connection 0.164 0.000
 No Toilet in House -0.062 0.000
Transfers
 Households Receiving Overseas
 Remittances 0.182 0.000
 Households Receiving Domestic
 Remittances 0.038 0.000
Locational Variables
 Sindh Province 0.104 0.000
 Balochistan Province -0.138 0.000
 Intercept (Constant) 7.041 0.000
Summary Statistics
Adjusted R-square 0.52 Condition Index 20.62
F-value 413.39 Durbin-Watson 1.54

Source: Author's Estimates based on PIHS-HIES, 2001-02.

Table 7
Predicted Poverty Incidence--2001-U2
(Percentage of Population below the Poverty Line)

 Urban Areas

 Provincial Large
Province Overall Capital Cities

Punjab 26 19 21
 (37.55) (7.34) (13.38)
Sindh 31 11 20
 (30.84) (7.39) (5.26)
NWFP 29 28 --
 (24.72) (6.25)
Balochistan 48 16 --
 (36.13) (3.50)

 Urban
 Areas

 Small
 Cities Rural
Province and Towns Areas

Punjab 42 24
 (23.04) (27.47)
Sindh 38 38
 (13.70) (27.88)
NWFP 41 28
 (17.32) (19.03)
Balochistan 41 52
 (13.80) (32.25)

Note: t-values are given in parenthesis. All estimates are
Statistically significant according to the t-statistics.
Large cities, in Punjab, are Rawalpindi, Islamabad, Faisalabad,
Multan, Gujranwala, Sargodha, Sialkot and Bhawalpur. In Sindh
province, Hyderabad and Sukkur are included in this category.