Pakistan panel household survey: sample size and attrition.

Durr-e-Nayab ; Arif, G.M.

1. INTRODUCTION AND BACKGROUND

The socio-economic databases in Pakistan, as in most countries, can be classified into three broad categories, namely registration-based statistics, data produced by different population censuses and household survey-based data. The registration system of births and deaths in Pakistan has historically been inadequate [Afzal and Ahmed (1974)] and the population censuses have not been carried out regularly. The household surveys such as Pakistan Demographic Survey (PDS), Labour Force Survey (LFS) and Household Income Expenditure Survey (HIES) have been periodically conducted since the 1960s. These surveys have filled the data gaps created by the weak registration system and the irregularity in conducting censuses. The data generated by the household surveys have also enabled social scientists to examine a wide range of issues, including natural increase in population, education, employment, poverty, health, nutrition, and housing. All these surveys are, however, cross-sectional in nature so it is not possible to gauge the dynamics of these social and economic processes, for example the transition from school to labour market, movement into or out of poverty, movement of labour from one state of employment to another. A proper understanding of such dynamics requires longitudinal or panel datasets where the same households are visited over time. Since panel surveys are complex and expensive to carry out, they are not as commonly conducted as the cross-sectional surveys anywhere in the world and in Pakistan they are even rarer.

One of the available panel surveys in Pakistan has been conducted by International Food Policy Research Institute (IFPRI) over a period of five years from 1986 to 1991 covering 800 households. The IFPRI sample comprised rural areas of only four districts with no representation from Balochistan and urban areas of the country. In these five years the sampled households were almost visited biannually. Another two-round panel data available in the country is that of the Pakistan Socio-Economic Survey (PSES) carried out by the Pakistan Institute of Development Economics (PIDE) in 1998-99 and 2001 in the rural as well as urban areas of Pakistan. Both the IFPRI and the PSES panels could not be continued after the above-mentioned rounds.

In 2001, the PIDE took a major initiative, with the financial assistance of the World Bank, to revisit the IFPRI panel households after a gap of 10 years. The sample was expanded from four to 16 districts, adding districts from all four provinces. Continuing to be a rural survey, it was named the Pakistan Rural Household Survey (PRHS). The second round of the PRHS was carried out in 2004 while the third round was completed in 2010. The third round marked the addition of the urban sample to the existing survey design of the PRHS, as a result--the Survey was named as the Pakistan Panel Household Survey (PPHS).

Attrition bias can affect the findings of the subsequent rounds of a panel survey, so it is important to examine the extent of sample attrition and determine whether it is random or has affected the representativeness of the panel sample. After conducting three rounds of the PRHSPPHS there is a need to evaluate the panel dataset for attrition bias. The present paper looks into the socio-demographic profile of the sample over the three rounds and evaluates the presence, or otherwise, of an attrition bias. The paper, thus, has three major objectives, which are to:

(a) Describe the sample size of three rounds of the panel survey

(b) Analyse the extent of sample attrition and analyse whether it is random, and

(c) Examine the socio-demographic dynamics of household covered in three rounds.

2. SELECTION OF DISTRICTS AND PRIMARY SAMPLING UNITS (PSUs)

As noted earlier, the IFPRI panel (1986-1991) was limited to the rural areas of four districts, namely Dir in Khyber Pakhtunkhwa (KP), Attock and Faisalabad in Punjab and Badin in Sindh. A rural sample based on these districts cannot be considered representative of the rural areas spread across more than 100 districts of the country. To give more representation to the uncovered areas 12 new districts were added to the PRHS-1 round carried out in 2001. From KP two new districts, Mardan and Lakki Marwat, were added to give representation to the Peshawar-Mardan valley and the Kohat-Dera Ismail Khan belt, respectively. The Hazara belt of KP still needs to be added for an even better representation. Three districts from south Punjab (Bahawalpur, Vehari and Muzaffargarh) and one district from central Punjab (Hafizabad) were also included in the PRHS-I. By this addition, all the three broad regions of Punjab, north, central and south, have their representation in the panel survey (Table 1). The three added districts from Sindh were Mirpurkhas, Nawabshah and Larkana. Balochistan was not part of the IFPRI panel so the PRHS included three districts from Balochistan, namely Loralai, Khuzdar and Gawadar (Table 1).

For the rural sample a village or deh is considered as the PSU. Table 1 presents the number of rural PSUs by district. It is noteworthy that there were 43 PSUs (or village/deh) in four districts of the IFPRI panel (Attock, Dir, Badin and Faisalabad). From the 12 new districts, PRHS selected 98 more PSUs (villages/deh) randomly. The total rural PSUs, after all the additions and inclusions, now stand at 141 as can be seen in Table 1. For details regarding each selected PSU, their respective tehsils, districts and provinces see Table A1, A2, A3 and A4 in the Annexure.

It is worth mentioning here that the second round of the panel survey, PRHS-II, was carried out only in the rural areas of Punjab and Sindh. Because of security concerns the other two provinces, K.P and Balochistan, could not be covered in this round.

The urban sample was added in the third round (PPHS) carried out in 2010 in all 16 districts. A selected district was the stratum for the urban sample. All the urban localities in each district were divided into enumeration blocks, consisting of 200 to 250 households in each block. In total, 75 urban enumeration blocks (PSUs) were selected randomly for the third round (PPHS-2010).

The scatter of the selected districts, as can be seen from Figure 1, is a good indicator of the geographical coverage of the districts covered under the PPHS. The sample covers the whole of the country, strengthening its representativeness.

[FIGURE 1 OMITTED]

3. HANDLING THE SPLIT HOUSEHOLDS

Before discussing the sample size, it is important to understand how the split households have been dealt with in the panel survey. A split household is defined as a new household where at least one member of an original panel household has moved in and is living permanently. This movement of a member from a panel household to a new household could be due to his/her decision to live separately with his/her family or due to marriage of a female member. If split households are not handled properly, the demographic composition of the sampled households is likely to change over time.

In the rounds two and three of the PRHS-PPHS split households were also interviewed. They, however, were only those households that were residing in the same village as the original panel household. In other words, movement of panel households or their members residing out of the sampled villages were not followed because of the high costs involved in this type of follow-up.

4. SAMPLE SIZE OVER THE DIFFERENT ROUNDS

The size of the sample for each round of the panel survey is shown in Table 2. The total size varies from 2721 households in 2001 to 4142 households in 2010. These variations, as discussed earlier, are for three reasons. First, the PRHS-11 carried out in 2004 was limited to two provinces, Punjab and Sindh, while the other two rounds covered all four provinces. Second, in the PRHS-1I as well as the PPHS-2010, split households were also interviewed (Table 2). Third, urban sample was added in the third round, PPHS, 2010.

As can be seen from Table 2, in the PRHS-I, carried out in 2001, the total sample consisted of 2721 rural households. The sample size decreased to 1614 households in PRHS-II (2004) because of the non-coverage of two provinces. However, 293 split households were interviewed in PRHS-II to raise the total sample size to 1907 households. Table 2 shows that in the PPHS-2010 the total rural households interviewed in four provinces were 2800, out of which 2198 were panel households and the remaining 602 were split households. With the addition of 1342 urban households, the total sample size of the PPHS 2010 accounted for a total of 4142 households (Table 2).

Four features of the three rounds of the panel data are noteworthy, which are as follows:

(i) Urban households, which have been included for the first time in the sample in the third round (PPHS) held in 2010, are not panel households. Essentially, the urban sample can be analysed as a cross-sectional dataset at present and after their coverage in the next round of the survey they can be treated as panel households.

(ii) Split households are not strictly panel households, particularly those where a female has moved due to her marriage. Thus, the matching of split households with the original panel households is not a straightforward exercise. While doing any analysis the split households need to be handled carefully.

(iii) Only the rural sampled households in Punjab and Sindh are covered in all three rounds, so the analysis of the three-wave data is restricted to these two provinces.

(iv) For the analysis of all rural areas covering four provinces, panel data are available for the 2001 and 2010 rounds.

5. SCOPE OF THE PANEL SURVEY

The scope of the panel survey is examined in terms of the types of information (modules) gathered through the structured questionnaires. In all three rounds, two separate questionnaires for male and female respondents were prepared and different modules were included in these questionnaires (Table 3). A two-member team of enumerators, one male and one female, visited each sampled household to gather information. Female enumerators were responsible to fill the household roster and pass it immediately to her male counterpart. Education and employment modules were included in both male and female questionnaires but the relevant information regarding children (under 5 years old), both male and female, was recorded in the female questionnaire. One major objective of the PRHS-PPHS panel survey has been to examine the movement into or out of poverty therefore a detailed consumption expenditure module has been a part of the female questionnaire in all the three rounds. Expenditures on durable items, however, were recorded in the male questionnaire. Health and migration modules were included in PRHS-I and PPHS 2010 rounds. A module on household-run businesses and enterprises was part of the latter two rounds as well.

Each round of the survey has had certain specific areas of focus. Agriculture, for example, was the main focus of the PRHS-I when information even at the plot level was collected from the land operating households. In the other two rounds only a brief agriculture module was included. The main focus of the PRHS-II was mental health, dowry, inheritance and marriage-related transfers. The PPHS-2010 was conducted at a time when inflation was high and the nation had also faced some natural disasters including droughts and floods. In the latest round modules on shocks, food security, subjective wellbeing and overall security were specially included in the questionnaire.

In short, the scope of the three rounds of the panel survey is wide. A variety of social, demographic and economic issues can be explored from these rounds. While some core modules are common to all rounds, there are others that are specific to a certain round. Some of the information is, thus, cross-sectional in nature but can be linked to the household socio-demographic dynamics made available through the core modules.

6. AN ANALYSIS OF THE SAMPLE ATTRITION

As shown earlier, in the PRHS-PPHS data have been collected from the same households over three points of time- 2001, 2004 and 2010. It is common in such surveys that some participants (households) drop out from the original sample for a variety of reasons including geographical movement and refusal to continue being part of the panel. This attrition of the original sample represents a potential threat of bias if the attritors are systematically different from the non-attritors. It can lead to 'attrition bias' because the remaining sample becomes different from the original sample [Miller and Hollist (2007)]. If the participating units, however, are not dropped out systematically, meaning that there are no distinctive characteristics among the attriting units, then there is no attrition bias even though the sample has decreased between waves. It is, therefore, important to examine the attrition bias in our panel survey.

6.1. Theoretical Considerations (1)

Attrition in panel surveys is one type of non-response. At a conceptual level, many of the insights regarding the non-response in cross-sections carry over to panels. According to Fitzgerald, et al. (1998), attrition bias is associated with models of selection bias. Their statistical framework for the analysis of attrition bias, which has been used by several other studies [see for example, Alderman, et al. (20000; Thomas, et al. (2001); Aughinbaugh (2004)], makes a distinction between selection of variables observed in the data and variables that are unobserved. Alderman, et al. (2000) believe that, 'if there is sample attrition, then it has to be seen whether or not there is selection of observables. Selection of observables includes selection based on endogenous observables, which occurs prior to attrition (e.g. in the first round of the survey). Even if there is selection of observables, this does not necessarily bias the estimates of interest. Thus, one needs to test for possible attrition bias in the estimates of interest as well' [Alderman, et al. (2000)].

Assume that the object of interest is a conditional population density f(y|x) where y is scalar dependent variable and x is a scalar independent variable (for illustration, but in practice making x a vector is straightforward):

y = [[beta].sub.0] + [[beta].sub.1] + [epsilon], y observed if A=0 ... (1)

where A is an attrition indicator equal to 1 if an observation is missing its value y because of attrition, and equal to zero if an observation is not missing its value y. Since (1) can be estimated only if A=0 that is, one can only determine g(y|x, (A= 0)), one needs additional information or restrictions to infer f(x) from g(x), which can be derived from the probability of attrition, PR(A=0\y, x, z), where z is an auxiliary variable (or vector) that is assumed to be observable for all units but not included in x. This leads us to the estimation of the following form:

[A.sup.*] = [[delta].sub.0] + [[delta].sub.1] x + [[delta].sub.2] z + V ... (2)

A = I if [A.sup.*][greater than or equal to] 0 ... (3)

If there is selection of observables, the critical variable is z, a variable that affects attrition propensities and is also related to the density of y; conditional on x. In this sense, z is "endogenous to y". Indeed, a lagged value of y can play the role of z if it does not have structural relationship with attrition. Two sufficient conditions for the absence of attrition bias due to attrition of observables are either (1) z does not affect A or (2) z is independent of y conditional on x. Specification test can be carried out of either of these two conditions. One test is simply to determine whether candidates for z (for example, lagged value of y) significantly affect A. Another test is based on Beketti, el al. (1988), and is known as BGLW test. It has been applied by Fitzgerald, et al. (1998) and Alderman, et al. (2000). In the BGLW test, the value of y at the initial wave of the survey (yn) is regressed on x and on A. This test is closely related to the test based on regressing A and x and y., (which is z in this case); in fact, two equations are simply inverses of one another [Fitzgerald, et al. (1998)]. Clearly, if there is no evidence of attrition bias from these specification tests, then one has the desired information on f(y\x).

6.2. Extent of Attrition

Table 4 presents the attrition rate for different rounds. Between 2001 and 2010, the attrition rate was around 20 percent while the rate for the 2004 to 2010 period was 25 percent, suggesting some households had dropped in 2004 and re-entered the panel in 2010. For the 2004-10 period, the highest attrition rate is found in Balochistan hinting towards more movement of sampled households than in other provinces.

6.3. Attrition Bias

As stated earlier, the urban sample was included in the panel survey in 2010 for the first time and hence the attrition issue is related to the rural sample. It has also been noted that the PRFTS-II was limited to two large provinces, Punjab and Sindh. All the rural areas were covered in round I (2001) and round III (2010). The attrition bias is examined between the two waves 2001 and 2010. Five models have been estimated where the dependent variable is whether attrition occurred between these two rounds (1= yes; 0 = no), results for which are presented in Table 5. The sample used in these models consists of all 2001 households and all regressors are measured in 2001.

Following Thomas, et al. (2001) and Arif and Bilquees (2006), the first model of attrition includes the only one covariate, In(PCE), where per capita consumption (PCE) is used as a measure of households' economic status. Table 5 presents coefficient estimates from the logit regressions. The first model indicates that there is a statistically significant negative relationship between PCE and the probability of leaving the panel. On average, lower economic status households were more likely to attrite between the two waves, so without weighting, the PPHS-2010 would be lesser representative of lower economic status households than would be a random household survey.

In model 2, two variables, ln(PCE) and ln(househo!d size) have been included. Both PCE and family size (in 2001) are positively and significantly associated with a household staying part of the subsequent round of the panel survey. The third model in Table 5 adds one dummy, that of a household consisting of only one or two members. The association between attrition and PCE and household size still remains negatively significant. On the other hand, small size households (with 1 or 2 members) show a significant association with attrition.

Model 4 included measures related to three characteristics of the head of the household, which are age, sex and literacy. None of these variables turned out to be statistically significant. Two economic variables, ownership of livestock and land, and provincial dummies are added in model 5. Both the economic variables are significantly associated with keeping households part of the panel and maintaining them as non-attritors (see Table 5). Among the provinces, households in Balochistan are more likely to leave the sample than households located in other provinces. It is evident from the multivariate analyses that there is a positive association between leaving the panel and small household size. Improving economic status of the household is statistically significant to keep the household in the sample, so it is mainly the poorer households that are attriting.

As discussed in the beginning of this section, BGLW test, introduced and used initially by Becketti, et al. (1988), is the other method of testing the attrition bias. This test examines whether those who subsequently leave the sample are systematically different from those who stay in terms of their initial behavioural relationships. We estimate the consumption (InPCE) equations as well as poverty equations, dividing the survey participants into two subsets--all 2001 households, and those still in the sample in 2010, labelled as 'Always in' or non-attritors.

Tables 6 and 7 present estimates of OLS regression for consumption equations and logit estimates for poverty equations respectively. A standard set of household and the head of the household characteristics, including age, and literacy of the head of the household, family size, and ownership of dwelling unit and livestock have been entered as independent variables into these equations. All the estimates are significant, as can be seen from Table 6 and Table 7. These estimates indicate a number of associations that are consistent with widely-held perceptions about consumption behaviour and poverty. For example, age and literacy of the head of the households have a positive impact on consumption while they are negatively associated with poverty. A similar pattern of association was also found for family size as it has a positive association with poverty but a negative relation with the per capita consumption expenditure. The ownership of both livestock and land has a positive association with per capita expenditure, but a negative relation with the incidence of poverty.

Our interest here, however, is more in the difference that the attritors might have made to the sample. To ascertain this we apply the t-difference test with the following hypotheses and assumption:

[H.sub.0]: No significant difference between attritor and non-attritor.

[H.sub.1]: Significant difference exists between attritor and non-attritor.

Assumption: unequal sample size, unequal variance.

The t-difference test results (see last columns of Table 6 and 7) show that there are no significant differences between the set of coefficients for the sub-sample of those missing in the follow-up versus the sub-sample of those re-interviewed for indicators of either consumption or poverty. These estimates, therefore, suggest that the coefficient estimates of standard background variables are not affected by sample attrition.

7. CONCLUSION

The PRHS-PPHS panel is a rich source of information regarding a range of socioeconomic and demographic processes, and a means to understand their dynamics over time. Along with having a few core modules the panel questionnaire is flexible enough to accommodate any particular area of interest in a specific round without affecting the overall efficiency of the survey design. Addition of the urban sample in 2010 to the previously all rural sample has made the panel design even more comprehensive. With three rounds having been carried out so far, in 2001, 2004 and 2010, the panel sample retains its qualities despite all the attritions and the phenomenon of split households.

ANNEXURES

Table A1
Sample list for Pakistan Panel Household Survey 2010: Punjab

Province    Code   District       Code   Telisil          Code

Punjab       1     Faisalabad      1     Faisalabad        1
                                         Jaranawala        2
                                         Gojra             3
                                         Summandri         4

                   Attack          2     Feth Jang         5
                                         Pindi Ghaip       6

                   Hafizabad       5     Pindi Bhatian     11

                   Veliari         6     Mailsi            12

Punjab       1     Muzafar Garh    7     Ali Pur           13

                   Bahawalpur      8     Ahmed Pur East    14

Province    Village             Code

Punjab      Saddon 206RB         1
            Sing Pura            2
            Jarwanwala Chak      3
            Subdarawala 363JB    4
            Khalishabad 356JB    5
            Summandri            6
            Khirala Kalan        7
            Thathi Gogra         8
            Kareema              9
            Hattar               10
            Makyal               11
            Gulyal               13
            Dhock Qazi           14
            Khatteshah           53
            Nasowal              54
            Khidde               55
            Bahoman              56
            Daulu Kalan          57
            Bagh Khona           58
            Shah Behlol          59
            Purniki              60
            Thata Karam Dad      61
            Mona                 62
            Chak No 118-WB       63
            Chak No 190 WB       64
            Kot Soro             65
            Chak No 195 WB       66
            Mandan               67
            Kot Muzzfar          68
            Muradabad            69
            Chak No 109 WB       70
            Chak N0I66-WB        71
            Maqsooda             72
Punjab      Mail Manjeeth        73
            Makhan Bela          74
            Tibbah Barrah        75
            Malik Arain          76
            Kohar Faqiran        77
            NauAbad              78
            Kundi                79
            Nabi Pur             81
            Kotla Afghan         82
            Ghunia               83
            Chak No 157-N.P.     84
            Haji Jhabali         85
            Mad Rashid           87
            Mukhawara            88
            Pipli Rajan          89
            Qadir Pur            90
            Ladpan Wali          91
            Chak Dawancha        92

Table A2
Sample list for Pakistan Panel Household Survey 2010: Sindh

Province   Code  District  Code     Tehsil     Code  Village       Code

Sindh       2     Badin     3        Badin      7    Kerandi        21
                                   Golarchi     8    Kalhorki       22
                                                     Shaikhpur      23
                                                     Khoro          24
                                                     Khirdi         25
                                                     Bhameri        26
                                                     Walhar         27
                                                     Parharki       28
                                                     Golarchi       29
                                                     Lucky          30
                                                     Nurlut         31
                                                     Mitho Debo     32
                                                     Sorahdi        33
                                                     Chakri         34
                                                     Fatehpur       35
                                                     Mari Wasayo    36
                                                     Bajhshan       37
                                                     Khirion        39
                                                     Kandiari       40
                  Navvab    9     Daulat Pur    15   Jagpal         93
                  Shalt                              Kandhari       94
                                                     Khar           95
                                                     Sindal Kamal   96
                                                     Kaka           97
                                                     Bogri          98
                                                     Manhro         99
                                                     Uttar Sawri   100
                 Mir Pur    10       KotG.      16   Deh 277       101
                  Kltas            Mohammad          Deh 320       102
                                                     Deh 346       103
                                                     Deh 339A      104
                                                     Deh 306       105
                                                     Deh 302       106
                                                     Deh 285       107
                                                     Deh 257       108
                 Larkana    11    Qantber Ali   17   Chacha        109
                                   Rato Dero    18   Dera          112
                                                     Laktia        113
                                                     Do-Abo        114
                                                     Nather        115
                                                     Haslla        116
                                                     Sanjar Abro   117
                                                     Khan Walt     118
                                                     Khuda Bux     120
                                                     Naudero       121
                                                     Saidu Dero    122

Table A3
Sample list for Pakistan Panel Household Survey 2010: Khyber Pakhtunkhwa

Province   Code  District   Code  Tehsil       Code  Village      Code

KP          3    Dir         4    Blambut       9    Katigram      41
                                  Adenzal
                                                     Batam         42
                                                     Shalt Alam
                                                       Baba        43
                                                     Bakandi       44
                                                     Khanpur       45
                                                     Kamangara     46
                                                     Malakand      47
                                                     Khema         48
                                                     Khazana       49
                                                     Shehzadi      50
                                                     Munjal        51
                 Mardan      12   Taklit Bhai   19   Khan Killi   125
                                                     Dagal        126
                                                     Jangirabad   127
                                                     Saidabad     129
                                                     Mian Killi   130
                                                     Fethabad     131
                                                     Seri Behial  133
                 L. Marwat   13   L. Marwat     20   Nar Akbar    135
                                                     Nar Langar   136
                                                     Alwal Khel   138
                                                     Gorka        141
                                                     Ghazi Khel   142

Table A4
Sample list for Pakistan Panel Household Survey 2010: Balochistan

Province      Code  District  Code  Tehsil   Code  Village         Code

Balochistan    4    Loralai    14   Loralai   21   Sanghri         145
                                                   Urd Shahboza    146
                                                   Sor Ghand       147
                                                   Nigang          148
                                                   Marah Khurd     149
                                                   Mekhtar         150
                                                   Tor             151
                    Khuzdar    15   Khuzdar   22   Bajori Kalan    153
                                                   Ghorawah        154
                                                   Bhat            155
                                                   Kliat Kapper    156
                                                   Sabzal Khan     157
                                                   Khorri          159
                                                   Par Pakdari     160
                    Gawadar    16   Gawadar   23   Ankra           161
                                                   Chibab Rekhani  162
                                                   Dhorgati        163
                                                   Grandani        164
                                                   Nigar Sharif    165
                                                   Shinkani Dar    167
                                                   Sur Bandar      168

REFERENCES

Afzal, M. and T. Ahmed (1974) Limitations of Vital Registration System in Pakistan against Sample Population Estimation Project: A Case Study of Rawalpindi. The Pakistan Development Review 13:3.

Alderman, H., J. Behrman, H. Kholer, J. Mauccio and S. Watkins (2000) Attrition in Longitudinal Household Survey Data: Some Tests for Three Developing Country Samples. The World Bank, Development Research Group Rural Development. (Policy Research Working Paper 2447).

Arif, G. M. and F. Bilquees (2006) An Analysis of Sample Attrition in PSES Panel Data. Pakistan Institute of Development Economics, Islamabad. (MIMAP Technical Papers Series No. 20).

Aughinbaugh, A. (2004) The Impact of Attrition on the Children of the NLSY97. The Journal of Human Resources 39:2.

Becketti, S., W. Gould, L. Lillard, and F. Welch (1988) The Panel Study of Income Dynamics after Fourteen Years: An Evaluation. Journal of Labour Economics 6.

Fitzgerald, J., P. Gottschalk, and R. Moffit (1998) An Analysis of Sample Attrition in Panel Data. The Journal of Human Resources 33:2.

Miller, R. and C. Hollist (2007) Attrition Bias. Department of Child, Youth and Family Studies, University of Nebraska-Lincoln.

Thomas, D., E. Frankenberg, and J. Smith (2001) Lost but not Forgotten: Attrition in the Indonesian Family Cycle Survey. The Journal of Human Resources 36:3, 556-592.

(1) This sub-section depends heavily on Arif and Biquees (2006) who have examined the attrition bias between two rounds of the Pakistan Socio-Economic Survey (PSES) carried out in 1998-99 and 2001 by the Pakistan Institute of Development Economics.

Durr-e-Nayab <dnayab@gmail.com> is Chief of Research at the Pakistan Institute of Development Economics, Islamabad. G. M. Arif <gmarifW;pide.org.pk> is Joint Director at the Pakistan Institute of Development Economics, Islamabad.

Authors' Note: The authors are thankful to Shujaat Farooq for his help in the analysis regarding attrition of the sample. Thanks are due to Syed Majid Ali and Saman Nazir as well for their help in the tabulation for this paper. Usual disclaimer applies.

Table 1
Primary Sampling Units (PSUs) by Province and District

                                  Number of PSUs

Province       Districts          Rural   Urban (c)

Punjab         Faisalabad (a)       6        16
               Attock (a)           7         4
               Hafizabad1 (b)      10         4
               Vehari1 (b)         10         4
               Muzaffargarh (b)     9         4
               Bahawalpur (b)       9         7
Sindh          Badin (a)           19         3
               Nawab Shah (b)       8         4
               Mirpur Khas (b)      8         4
               Larkana (b)         11         7
KP             Dir (a)             11         2
               Mardan (b)           7         6
               Lakki Marwat (b)     5         2
Balochistan    Loralai (b)          7         2
               Khuzdar (b)          7         3
               Gwadar (b)           7         3
               Total               141       75

Note: PR.HS-I (2001) and PPHS (2010) covered all districts.
PRHS-II (2004) was limited to 10 districts of Punjab and Sindh,

(a). Districts included in the IFPRI panel.

(b). New districts added since 2001.

(c). Included only in PPHS-2010.

Table 2
Households Covered during the Three Waves of the Panel Survey

                        PRHS-II 2004

                        Panel    Split
               PRHS-I   House-   House-
                2001    holds    holds    Total

Pakistan        2721     1614     293      1907
Punjab          1071     933      146      1079
Sindh           808      681      147      828
KP              447       --       --       --
Balochistan     395       --       --       --

                                 PPHS-2010

                                 Total
               Panel    Split    Rural    Urban
               House-   House-   house-   House-   Total
               holds    holds    holds    holds    Sample

Pakistan        2198     602      2800     1342     4142
Punjab          893      328      1221     657      1878
Sindh           663      189      852      359      1211
KP              377       58      435      166      601
Balochistan     265       27      292      160      452

Source: PRHS 2001, 2004 and PPHS 2010 micro-datasets.

Table 3
Scope of the Panel Survey: Modules included in Household
Questionnaires

                                   PRHS-(2001)         PRHS-II (2004)

Modules                          Male     Female     Male     Female

Household Roster                [check]   [check]   [check]   [check]
Education                       [check]   [check]   [check]   [check]
Agriculture                     [check]      x      [check]      x
Non-Farm Enterprises            [check]      x         x         x
Employment                      [check]   [check]   [check]   [check]
Migration                       [check]      x      [check]      x
Consumption                     [check]   [check]   [check]   [check]
Credit                          [check]      x      [check]      x
Livestock Ownership                x      [check]      x      [check]
Housing                            x      [check]      x         x
Health                             x      [check]      x      [check]
Dowry and Inheritance              x      [check]      x      [check]
Mental Health                      x         x         x      [check]
Marital History and Marriage
  Related Transfers                x         x         x      [check]
Shocks and Coping Strategies       x         x         x         x
Household Assets                   x         x         x         x
Household Food Security            x         x         x         x
Security                           x         x         x         x
Subjective Welfare                 x         x         x         x
Business and Enterprises           x         x         x         x
Transfer/Assistance from
  Programme and Individuals        x         x         x         x

                                       PPHS (2010)

Modules                          Male     Female

Household Roster                [check]   [check]
Education                       [check]   [check]
Agriculture                     [check]      x
Non-Farm Enterprises            [check]      x
Employment                      [check]   [check]
Migration                       [check]      x
Consumption                     [check]   [check]
Credit                          [check]      x
Livestock Ownership                x      [check]
Housing                            x      [check]
Health                             x      [check]
Dowry and Inheritance              x         x
Mental Health                      x         x
Marital History and Marriage
  Related Transfers                x         x
Shocks and Coping Strategies       x      [check]
Household Assets                   x      [check]
Household Food Security            x      [check]
Security                        [check]   [check]
Subjective Welfare              [check]   [check]
Business and Enterprises        [check]      x
Transfer/Assistance from
  Programme and Individuals     [check]      x

Table 4
Sample Attrition Rates of Panel Households--Rural

(%)

               2001-2004   2001-2010   2004-2010

Pakistan         14.1        19.6        24.9
Punjab           12.9        17.1        23.8
Sindh            15.7        18.3        26.2
KPK.              --         16.1         --
Balochistan       --         33.2         --

Source: Authors' computations based on PRHS 2001 and
PPHS 2010 micro-datasets.

Table 5
Determinants of Attrition through Logit Regression

Correlates (2001/02)         Model 1       Model 2       Model 3

Log per capita
consumption                 -0.286 *      -0.342 *      -0.353 *
Log household size                        -0.257 *     -0.177 ***
Households with 1 or 2
family members only
(yes=l)                                                 0.416 ***
Age of head of
household (years)
Age-square of head of
household
Female headed
households (yes=l)
Literacy of the head
(literate=l)
Livestock owned (yes=l)
land owned (yes=l)

                                         Provinces (Punjab as ref.)

Sindh
KPK
Balochistan
Constant                      0.580       1.458 **       1.36 **
LR chi-square               11.93(1)      19.35(2)      21.63(3)
Log likelihood              -1353.789     -1350.079     -1348.941
Observations                  2,714         2,714         2,714

Correlates (2001/02)         Model 4       Model 5

Log per capita
consumption                 -0.214 **    -0.152 ***
Log household size           -0.014         0.056
Households with 1 or 2
family members only
(yes=l)                     0.426 ***       0.353
Age of head of
household (years)             0.001         0.003
Age-square of head of
household                     0.000         0.000
Female headed
households (yes=l)            0.378       0.493 ***
Literacy of the head
(literate=l)                 -0.138         0.010
Livestock owned (yes=l)     -0.443 *      -0.451 *
land owned (yes=l)          -0.280 *      -0.377 *

Sindh                                      -0.009
KPK                                        -0.021
Balochistan                                0.910 *
Constant                      0.926         0.222
LR chi-square               53.71 (9)    102.63 (12)
Log likelihood              -1332.229     -1307.268
Observations                  2,711         2,711

Source: Authors' computations based on PRHS 2001 and
PPHS 2010 micro-datasets.

Note: *** P<0.01; ** P<0.05, * P<0.10.

Table 6
Household Expenditure: OLS Regression Model 2001-2010

                                      Full Sample

Variables                 Coefficients   St. Error

Age (years)                  -0.001        0.004
[Age.sup.2]                  0.000         0.000
Literacy (literate=l)       0.196 *        0.023
Family Size                 -0.032 *       0.003
Land Ownership (yes=l)      0.255 *        0.023
Livestock                   0.142 *        0.025
Own House (yes=l)          -0.104 **       0.047
Constant                    6.838 *        0.105
F-stat                                56.46
R-square                              0.1305
Observations                          2.642

                                      Always in' (Non-attrition)
                                                     t-difference
Variables                 Coefficients   St. Error       test

Age (years)                  0.001         0.004        -0.500
[Age.sup.2]                  0.000         0.000        0.000
Literacy (literate=l)       0.190 *        0.025        0.251
Family Size                 -0.036 *       0.003        1.333
Land Ownership (yes=l)      0.252 *        0.025        0.125
Livestock                   0.133 *        0.028        0.341
Own House (yes=l)          -0.134 **       0.055        0.592
Constant                    6.870 *        0.117        -0.290
F-stat                                47.66               --
R-square                              0.1367              --
Observations                          2.115               --

Source: Authors' computations based on PRHS 2001 and
PPHS 2010 micro-datasets.

*** P<0.01; ** P<0.05, * PO.10.

Table 7
Correlates of Poverty: Logistic Regression Model 2001-2010

                                      Full Sample

Correlates                Coefficients   St. Error

Age (years)                  0.025         0.019
[Age.sup.2]                0.000 ***       0.000
Literacy (literate=l)       -0.545 *       0.102
Family Size                 0.093 *        0.011
Land Ownership (yes=l)      -0.827 *       0.102
Livestock (yes=l)           -0.592 *       0.105
Own House (yes=l)           0.538 **       0.210
Constant                    -1.817 *       0.483
LR chi-square                         206.39
Log likelihood                        -1374.198
Observations                          2,642

                                      Always in'(Non-
                                      attritors)     t-difference

Correlates                Coefficients   St. Error       test

Age (years)                  0.022         0.022        0.147
[Age.sup.2]                  0.000         0.000        0.000
Literacy (literate=l)       -0.504 *       0.117        -0.376
Family Size                 0.108 *        0.013        -1.257
Land Ownership (yes=l)      -0.840 *       0.116        0.120
Livestock (yes=l)           -0.504 *       0.122        -0.780
Own House (yes=l)           0.639 **       0.263        -0.430
Constant                    -1.994 *       0.568        0.339
LR chi-square                         160.22              --
Log likelihood                        -1058.706           --
Observations                          2,115               --

Source: Authors' computations based on PRHS 2001
and PPHS 2010 micro-datasets.

*** P<0.01; ** P<0.05; * P<0.1.