Accuracy of Poor County Designation

К оглавлению
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 
34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 
136 137 138 139 140 141 142 

We have noted above that many counties were designated as poor on the

basis of broadly political criteria. Initial evidence on targeting accuracy

can be found in the frequency distributions of poor and non-poor counties

across income levels. In 1986, only half of the counties in the lowest income

decile were designated as poor, even though there were even more counties

designated as poor in the next income group (see Figure 4.1). By 1993, far

fewer counties in the lowest income groups were being excluded, implying

better coverage, but there were many more counties designated as poor in

the middle-income groups, implying greater leakage (see Figure 4.2).

Overall targeting effectiveness can be evaluated more formally by defi ning

‘targeting gaps’ and ‘targeting errors’ (Park et al., 2002). Targeting gaps

describe mis-targeting in the full sample with respect to a reference poverty

line, while targeting errors describe mis-targeting given a set number of

targeted benefi ciaries. Similar to poverty measures, gaps and errors can be

aggregated using different weights.

Two types of targeting gaps can be calculated: the targeting count gap

(TCGt) and the targeting income gap (TIGt). The targeting count gap is

defi ned as

0

100

200

300

400

500

600

<10 10–25 25–50 50–75 75–90 >90

Income percentile group

Frequency

Non-poor

Poor

Figure 4.1 Income per capita distribution in poor and non-poor counties,

1986

Non-poor

Poor

0

50

100

150

200

250

300

350

400

450

<10 10–25 25–50 50–75 75–90 >90

Income percentile group

Frequency

Figure 4.2 Income per capita distribution in poor and non-poor counties,

1993

TCG

N

I P Y Z I P t it it it t

i

N

it it 

1

0 1 1

1

2 { ( , ) ( ,Y Z it t )}

Here N is the total sample of counties, indexed by i. Iit1 is an indicator

variable for type one error (or incomplete coverage) that equals one if a

county is not designated as poor (Pit = 0), but its income per capita (Yit) is

below the poverty line (Zt). Iit2 is an indicator variable for type two error

(or leakage) that equals one if a county is designated as poor (Pit = 1), but

its income per capita is above the poverty line.4 TCGt can be interpreted as

the percentage of counties that are mis-targeted and is easily disaggregated

into type one and type two error.

The targeting income gap is defi ned as

TIG

N

Z Y I Y Z I t t it it it t

i

N

it 

1

1

1

2 {( ) ( ) }

where the indicator variables are as defi ned above. TIG is similar to TCG

except that mis-targeting is weighted by the magnitude of mis-targeting,

measured as the difference between county income and the poverty line.5

Yearly TCG and TIG measures for poor county designation are presented

in Tables 4.6 and 4.7. Both measures are sensitive to the chosen poverty

line; as the line is increased, type one error increases and type two error

decreases. The tables give the TCG and TIG for each year from 1986–1995

for two different lines, the offi cial poverty line and a relative poverty line

equal to 60 per cent of mean income per capita.

The results show that targeting effectiveness has deteriorated steadily over

time, that incomplete coverage or omission of the poor has fallen, while

leakage has increased, and that using the offi cial poverty line, targeting gaps

jumped noticeably after the new poor county designations in 1993. As seen

in Table 4.6, the percentage of counties that were mis-targeted increased

from 14 to 22 per cent using the offi cial poverty line and from 15 to 19 per

cent using the relative poverty line. While failure to designate a poor county

as poor was nearly twice as likely as designating a non-poor county as

poor in 1986 (using either the offi cial or relative poverty lines), by 1995 the

opposite was true using the relative line, and using the offi cial line virtually

all mis-targeting was due to leakage. Considering that about one fi fth of

counties are mis-targeted, the TIG of 77 yuan in 1995 for the offi cial line

implies that the average magnitude of mis-targeting in mis-targeted counties

is about 385 yuan, or nearly two thirds of the poverty line.6 Only part of the

targeting gaps can be explained by preferential treatment towards minority

and revolutionary base counties. In 1986, 25 per cent of leakage in the TCG

(using the offi cial poverty line) was due to minority counties and 35 per

cent to revolutionary base counties. By 1995, the comparable fi gures were

35 per cent and 19 per cent, respectively.

Table 4.6 Targeting count gap, 1986 to 1995

Offi cial poverty line Relative poverty line

(60% of average income per capita)

Type I Type II Total Line Type I Type II Total

1986 0.094 0.050 0.144 598 0.099 0.050 0.149

1987 0.082 0.065 0.146 611 0.097 0.061 0.158

1988 0.044 0.101 0.144 586 0.086 0.073 0.159

1989 0.056 0.096 0.152 538 0.096 0.079 0.175

1990 0.078 0.093 0.171 570 0.093 0.085 0.178

1991 0.058 0.101 0.158 590 0.093 0.084 0.177

1992 0.038 0.107 0.145 628 0.087 0.083 0.171

1993 – – – 655 0.028 0.150 0.178

1994 0.005 0.232 0.237 703 0.047 0.137 0.185

1995 0.004 0.218 0.222 793 0.065 0.120 0.185

Note: Calculations based on sample of 1837 counties with complete data for all years.

Source: Park et al. (2002).

One problem with the targeting gap measure is that it is sensitive to the

number of poor counties designated. If the number of designations is less

than the number of truly poor counties, type one error is unavoidable,

and if designations exceed the number of poor counties, type two error

is unavoidable, even when targeting is perfect in that designations go to

the poorest counties. Another way to assess targeting, then, is to compare

outcomes with the perfect targeting case given the number of poor

county designations. The targeting count error (TCE) is the percentage

of designations not given to counties that would be targeted under this

defi nition of perfect targeting, or

TCE

D

I Y Z P t it it t it

i

N

1

0

1

( *, )

Here Zt

* is the income level of the marginal, or threshold, county when

targeting is perfect given the number of available designations (D). Similar

to targeting gaps, the indicator functions can be weighted by income

differences with counties that were mistakenly targeted to calculate targeting

income error (TIEt) or by rank differences to calculate targeting rank error

(TREt).7 These statistics are reported in Table 4.8, and show that by any

measure, targeting error was substantial in the original designations (in

fact a majority of designations were mis-targeted), increased steadily over

time, fell dramatically after new designations in 1993 to levels even below

that of the original designations, and then began increasing once again.

Thus, the 1993 designations reduced targeting error, but through a strategy

of expanded coverage benefi cial to counties above the absolute or relative

poverty thresholds.

Table 4.7 Targeting income gap, 1986 to 1995

Offi cial poverty line Relative poverty line

(60% of average income per capita)

Type I Type II Total Line Type I Type II Total

1986 9.6 6.2 15.8 598 11.6 6.1 17.7

1987 8.2 9.1 17.3 611 11.1 7.5 18.6

1988 3.3 16.4 19.7 586 8.5 9.6 18.1

1989 4.3 17.3 21.7 538 9.6 11.1 20.7

1990 6.5 16.2 22.7 570 9.7 13.0 22.7

1991 4.5 21.9 26.5 590 9.9 15.9 25.8

1992 2.9 29.9 32.9 628 9.7 19.1 28.8

1993 – – – 655 1.9 26.0 27.8

1994 0.3 65.8 66.1 703 4.6 29.6 34.2

1995 0.2 76.4 76.5 793 7.8 31.2 39.0

Note: Calculations based on sample of 1837 counties with complete data for all years.

Source: Park et al. (2002).

However, even if poor county designation was perfect, there would still

be mis-targeting due to the existence of the non-poor in poor counties and

of the poor in non-poor counties. Both of these phenomena have been

important. In 2002 the total rural population in the 592 poor counties was

around 200 million, while the poor population was only 28 million by the

offi cial poverty line or a little less than 100 million measured by the US

$1 a day purchasing power parity standard. Even if all of the poor were

resident in poor counties, the majority of households in poor counties

would still not be poor.

Table 4.8 National targeting error, 1986 to 1995

Targeting count Targeting rank Targeting

error error income error

1986 0.524 363 242

1987 0.504 381 265

1988 0.574 447 264

1989 0.625 532 302

1990 0.649 564 332

1991 0.629 621 378

1992 0.618 682 422

1993 0.280 260 153

1994 0.319 313 212

1995 0.334 323 267

Note: Calculations based on sample of 1837 counties with complete data for all years.

Source: Park et al. (2002).

Data from the National Bureau of Statistics indicate that of the 80

million rural poor in 1992, only 23 million lived in non-poor counties,

accounting for 29 per cent. However, this proportion has increased. An

estimate from the same source suggests that the poor living in non-poor

counties accounted for 38 per cent of the total poor population in 2001.

Rural household data provide evidence that an even larger percentage of the

poor live in non-poor counties. For example, one study indicated that about

half of the poor in four southern provinces did not live in poor counties

(Ravallion and Jalan, 1999).