Model Specification, Data and Methodology

This section presents the models, data and methodology utilised to estimate Indonesia’s poverty status movement of the households using the household head’s characteristics and

household characteristics for the period 2000 and 2007. The poverty estimation of spell approach noted earlier, the variables that explain the poverty dynamics and the related econometric issues for this analysis are discussed in detail.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1 2 3 4 5 6 7 8 9 10 2007

Food Schooling Medic

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1 2 3 4 5 6 7 8 9 10 2000

3.4.1 Model Specification

The factors affecting the probability of Indonesian households between 2000 and 2007 that fall in or move out of poverty are examined using the spells approach. There are four categories that represent the poverty status as the dependent variable, those are, chronic poor, transient poor (-), transient poor (+) and non-poor. Concerning these multi-categories of dependent variables, the multinomial model for categorical data is used. The model is

different from the ordinary binary model. In this study’s model, the dependent variable must be multinomially distributed whereas in the binary model, the dependent variable is Bernoulli or binomially distributed (Cameron & Trivedi, 2010).

The multinomial logit model is “the simplest model…because computation is simple and

parameter estimates are easier to interpret than in some other multinomial models” (Cameron

& Trivedi, 2010, p. 498). The analysis utilised here follows the method used by Herrera and Rorbaud (2005) as well as Bhatta and Sharma (2006). The multinomial logit is a more

preferable method for analysis because “…although poverty status is based on an underlying welfare measure…defined on an interval scale, it is not always appropriate to assume that

chronic poverty represent a higher level of deprivation than transient poverty…” (Bhatta &

Sharma, 2006, p. 11). Moreover, the multinomial logit models used in this study have multi- categories of the dependent variables with the same independent variables that may have different impact on each poverty status transition (Herrera & Rorbaud, 2005). The independent variables included in the models are based on previous studies and the determinants of poverty dynamics in the case of Indonesia.29 The multinomial logit model equation based on Cameron and Trivedi (2010) can be expressed as follows:

= Pr( = ) = ( , ), = 1, … , , = 1, … , (3.1)

The model specification for the determinants of poverty dynamics include the multinomial logit model. Where = Pr( = ) is the outcome probability of individual i at alternative or category j. In this study, the model uses the poverty status transition of the household i (poverty dynamics) between 2000 and 2007 as independent variable ( ), that is, 1 = chronic poor, 2 = transient poor (-), 3 = transient poor (+), and 4 = non-poor. In general, equation (3.1) is considered to analyse the probability of a household being in one of the poverty status transitions within a set of demographic and socio-economic characteristics. The

29 _{See studies by Alisjahbana and Yususf (2003), Herrera and Rorbaud (2005), Bhatta and Sharma (2006), Xing}

outcome probability ( ) is a function of regressors , that is, a vector of a household’s

demographic, socio-economic, shocks and policy variables that explain the impact experienced by a household in 2000 and a vector of changes in variables between 2000 and 2007, which may determine the household transition status of poverty dynamics for this period.

Equation (3.2) below specifies the measurement of the probability of the household’s

poverty status transition ( ) by multinomial logit model. It depends on regressors and the coefficient of the regressors (

.).The probability ( ) takes the value 0 < < 1. The

estimated coefficients of this categories in the model will be set to zero, named the base category. Thus, the coefficients from the other categories are interpreted with respect to the base category coefficient.

= exp(

′ ₎

∑₌₁exp( ′ ) (3.2)

The base category is the chronic poor, thus equation (3.2) can be rewritten as equation (3.3) where the , the “chronic poor” coefficient, is equal to zero. The coefficient of other categories ( ) can be viewed as parameters of binary logit estimation between the transient poor category and non-poor category relative to chronic poor.

Pr( = | = 1) =_{Pr = +Pr( =1)}Pr( = )

=

′ ₎

1+exp( ′ ) (3.3)

However, the interpretation can be easier and helpful if the coefficients are transformed to odds ratio or relative risk ratio (RRR). The relative risk ratio of being in transient poor and non-poor category rather than chronic poor is specified by

( )

( )

=

′ _{) (3.4)}

3.4.2 Data and Methodology

To distinguish between the poor and non-poor household by monetary approach, the consumption expenditure or income data is commonly utilised (World Bank, 2005). However, the consumption expenditure is preferable because expenditure is a direct

more problematic due to its inability to capture the basic needs fulfilment (see Townsend, 1962; Meyer & Sullivan, 2003; Duclos & Araar, 2006; Wisor, 2012). This study follows the method to differentiate the poor and non-poor households by comparing the expenditure per month per capita in the household to cover the cost of basic needs.

Figure 3.3. Structure of Poverty Status Transition for Indonesia’s Western Provinces.

This analysis uses panel data from the Indonesian Family Life Survey (IFLS) data which consists of the same households sample between 2000 and 2007. Firstly, the households are classified into poor and non-poor groups in 2000 based on whether per capita expenditure per month in the household can meet the minimum cost of food and non-food basic needs. The household panel data then are disaggregated to four categories of poverty dynamics those are, chronic poor, transient poor (-), transient poor (+), and non-poor. The chronic poor shows the households who are always poor in 2000 and 2007 while non-poor is for the households who never fall into poverty during that time. The transient poor (-) expresses the households who are not poor in 2000 but temporarily poor in 2007, otherwise the success movement of the households to temporarily move out from poverty is described by the transient poor (+). The structure of household panel data and its poverty status is shown in Figure 3.3.

Since this study utilizes the household panel data, there is a necessary preliminary data checking for attrition bias (Bhatta & Sharma, 2006; Bayudan-Dacuycuy & Lim, 2013b). The

bias is related to the selectivity bias which “results from using non randomly selected samples to estimate behavioral relationships as an ordinary specification error or "omitted

variables" bias” (Heckman, 1979, p. 153). To address this, the study follows the procedures used by Heckman and also Bayudan-Dacuycuy and Lim (see Heckman, 1979; Bayudan- Dacuycuy & Lim, 2013b). The first step involves probit estimation to obtain the Inverse

Mill’s Ratio (IMR), followed by adding the IMR variable in second step model, which is

Households in 2000 Non Poor in 2000 Poor in 2000 Poor in 2007 Transient Poor (-) Non Poor in 2007 Non Poor in 2007 Transient Poor (+) Poor in 2007 Chronic Poor

multinomal logit model, to test for the bias correction.30 The variable definitions and descriptive statistics of the variables used in this study are presented in the Appendix, Tables A.3.1 and A.3.2, respectively.