The empirical strategy follows a three-step procedure: (1) identification and measure-ment10 of the variables selected for each resilience block; (2) estimation of the latent variable representing each block, and of the resilience index, using multivariate methods (factor analysis, principal components analysis, optimal scaling, etc.); and (3) application of the CART methodology to estimate precise splitting rules based on a regression tree, to improve understanding of the whole process.11

The selection of relevant variables for estimating the latent indicator for each block is particularly complex. The multidimensionality of the concept gives rise to concerns about having variables that are relevant for more than one block, when a variable cannot be included more than once in the model. The conceptual framework described in the previous chapter simplifies this issue.

This subsection describes the separate estimation process for each component of the resilience framework.

Income and food access (IFA)

This indicator is directly related to a household's access to food. Economic access to food is considered the main food insecurity concern in Palestine (Abuelhaj, 2008; Mane et al., 2007). The traditional indicator for measuring food access capacity is income, but this study includes two additional indicators: the dietary diversity and food frequency score (DD) as a nutritional indicator; and the household food insecurity access scale (HFIAS) as an indicator of the household's perception of food security. Estimation of the IFA indicator involves use of the following variables:

• Average income per person per day (in Israeli new shekels (NIS)). This is an aggregated value of the different sources of income measured by the PPPS.

10 Not all the observed variables are taken directly from the raw data. Some are measured using complex procedures (e.g. the household food insecurity access scale and the dietary diversity and food frequency score).

11 The use of CART also made it possible to validate the decision process and identify the original variables (indicators) that play major roles in the different blocks defining resilience.

• DD. This is computed using specific methodologies for food security assessments applied to the weekly consumption of 20 food items. It can also be used as a proxy indicator for food access (Hoddinott and Yohannes, 2002).

• HFIAS . This indicator of a household's perception of food insecurity is measured through a set of nine questions developed by the Food and Nutrition Technical Assistance (FANTA) project12 (Coates et al., 2006).

Missing values in these variables are imputed using regression techniques.

All these indicators aim to measure food access, so the high correlation among them should produce a latent variable that fits the common pattern in the data. Because of this, to estimate the IFA latent indicator, a factor analysis is run using the principal factor method and the scoring method suggested by Bartlett (1937):

fs = r-1 A'V-lx, where r = A'^-1A, A is the unrotated loading matrix, ^ is the diagonal matrix of uniquenesses, and x is the vector of observed variables. The estimates produced by this method are unbiased, but may be less accurate than those produced by the regression method suggested by Thomson (1951).13 The regression-scored factors have the smallest mean square error, but may be biased. The first factor produced is quite meaningful and can be considered the underlying latent variable for food access. Table 21.1 shows the eigenvalue for each factor, and Table 21.2 shows the factor loadings for the original variables. The three indicators play almost the same role in estimating the IFA indicator, because their correlation coefficients are similar. As expected, HFIAS has a negative correlation because its score increases as food security decreases.

Factor |
Eigenvalue |

Factor 1 |
1.54613 |

Factor 2 |
0.75363 |

Factor 3 |
0.70024 |

Variable |
Factor 1 |
IFA |

Income |
0.4466 |
0.6779 |

DD |
0.4786 |
0.7308 |

HFIAS |
-0.4860 |
-0.7431 |

12 The FANTA project supports integrated food security and nutrition programming to improve the health and well-being of women and children. It is managed by the Academy for Educational Development and funded by the United States Agency for International Development (USAID).

13 The formula for Thompson's regression method is fT = A'£-1 x, where £ is the correlation matrix of x.

12 The FANTA project supports integrated food security and nutrition programming to improve the health and well-being of women and children. It is managed by the Academy for Educational Development and funded by the United States Agency for International Development (USAID).

13 The formula for Thompson's regression method is fT = A'£-1 x, where £ is the correlation matrix of x.

Access to public services (APS)

Public service provision is beyond households' control, but is a key factor for enhancing a household's resilience, such as by improving the effectiveness of that household's access to assets. As a result, better access to public services affects a household's capacity to manage risks and respond to crises. The following public services are considered in the analysis:

• Health. Measurement of health involves two indicators: physical access to health (need unmet, and need met within time limit, and need met after time limit); and the health care quality score (based on the quality of services provided in different health areas).

• Quality of education system (ordinal scale from 1 to 6).

• Perception of security (ordinal scale from 1 to 4). For this, a proxy index based on the general perception of security is constructed.

• Mobility and transport limitations (ordinal scale from 1 to 3). Different sources are used to generate an indicator from the mobility restriction questions in the questionnaire.

• Water, electricity and telecommunications networks. An indicator is developed for the number of services available.

A characteristic of the variables related to APS is that households in the same neighbourhood are expected to be highly correlated in terms of having the same availability of services. Missing values in these variables are therefore treated using the governorate-level mean.

In this case, use of the traditional multivariate methods (factor or principal components analysis) is impossible because the observed variables are not continuous. It was therefore decided to use the optimal scaling technique. The first factor (the APS indicator) obtained through PRINCALS is very satisfactory. Table 21.3 shows that all the original variables (transformed using optimal scaling) are, as expected, positively correlated with the estimated APS .

Social safety nets (SSN)

Social safety nets are a crucial aspect of the mitigation of crises in Palestine. More and more households are becoming dependent on assistance from international agencies,

Table 21.3 Correlation of APS with transformed variables.

Transformed variable APS

Physical access to health 0.6040

Health care quality 0.5984

Education system quality 0.4329

Perception of security 0.5317

Mobility constraints 0.5451

Water, electricity and telecoms 0.2838

charities and non-governmental organizations. Help received from friends and relatives is also substantial. Safety nets can therefore be taken to represent the system's capacity to mitigate shocks, and a general indicator for them has to be included in the estimation of resilience. The variables used to generate the SSN indicator are:

• Amount of cash and in-kind assistance (NIS/person/day).

• Quality of assistance (ordinal scale from 1 to 4).

• Job assistance (binary yes/no response).

• Monetary values of first and second types of assistance (NIS/person/day).

• Evaluation of the main type of assistance (ordinal scale from 1 to 4).

• Frequency of assistance (times assistance received in the last six months).

• Overall opinion of targeting (assistance targeted to the needy; to some not needy; or without distinction).

Missing values are treated using the governorate-level means and the level of income. It was decided to show the measurement scale of the variables as this is a crucial aspect of choosing the method for estimating the latent variable. The list above shows that types of variables were mixed, so for SSN, use of the PRINCALS algorithm is required. The first factor obtained is acceptable for estimating the latent variable SSN. Table 21.4 illustrates the correlation between estimated SSN and the transformed variables. All the variables are positively correlated with SSN and play important roles in the estimation. Nevertheless, the degree of satisfaction with the assistance received is the core of the generated variable.

Moreover, the correlation matrix for the various resilience blocks shows that SSN is negatively related to the others. This is because social cohesion increases as poverty increases. This aspect will be explained in more detail in the discussion of the results at the end of this subsection.

Assets (A)

Assets are part of a household's capital, and their availability is an important coping mechanism during periods of hardship. They therefore have to be considered as a key

Transformed variable |
SSN |

Amount of cash and in-kind assistance |
0.1669 |

Quality of assistance |
0.7347 |

Job assistance |
0.3794 |

First and second types of assistance |
0.7223 |

Evaluation of main assistance |
0.7304 |

Frequency of assistance |
0.6775 |

Opinion of targeting |
0.4462 |

factor in estimating resilience. Information on assets was not available from the PPPS data set; it was decided not to use proxies, so as not to contaminate the estimates.

Adaptive capacity (AC)

The adaptive capacity indicates a household's capacity to cope with and adapt to a certain shock, enabling that household to carry on performing its key functions. In other words, AC represents households' capacity to absorb shocks. For example, having more coping strategies means having a greater probability of mitigating food insecurity after, say, losing a job. The characteristic of adaptability is the buffer effect on household key functions. AC is measured by the following indicators:

• Diversity of income sources (count from 0 to 6). This indicates the number of income sources from different sectors (public, private, etc.); during a crisis, the more sources of income, the less the risk of losing the essential basis of the household's livelihood (i.e. income).

• Coping strategy index (count from 0 to 18). This represents the number of available coping strategies that have not yet been used. It does not consider whether or not a specific coping strategy has been adopted by the household (expost), but instead how many coping mechanisms are available to the household (ex ante).

• Capacity to keep up in the future (ordinal scale from 1 to 5). This is based on a household's perception of its own capacity to keep up in the future, considering the current socio-economic shocks in Palestine. It is a forward-looking variable, which allows households' expectations to be taken into account.

The variables are not continuous, even in this case, so the PRINCALS methodology is needed. Table 21.5 shows the correlation of the estimated AC with the transformed variables. The correlation coefficients correspond to the loadings of component 1; the coping strategy index and the capacity to keep up in the future are twice as important as the diversity of income sources.

Stability (S)

Stability is a widely used concept in the food security literature, although it is usually used to describe the stability of food supply. This chapter considers stability to be a cross-sectoral dimension of resilience. For example, an index of income stability may be its variability (increased, decreased or the same) over the last six months. The following variables are used to measure stability:

• Professional skills (count). The number of household members with at least a diploma14 is used as a proxy.

• Educational level (continuous). This is measured by the average number of years of schooling of household members.

• Employment ratio (from 0 to 1). This is the ratio of the number of employed household members to the household size.

14 A weighting system is used to give major relevance to household members with a master's or PhD degree.

Table 21.5 Correlation of AC with transformed variables.

Transformed variable AC

Diversity of income sources 0.3659

Coping strategy index 0.7551

Capacity to keep up in the future 0.7800

• Number of household members to have lost jobs (count). This is the number to have lost their employment in the last six months.

• Income stability (increased; the same; decreased). This is measured by income variation over the last six months.

• Assistance dependency (ratio from 0 to 1). This is the ratio of the monetary value of assistance to total income.

• Assistance stability (increased; the same; decreased). This is the variation in the quality of assistance over the last six months.

• Health stability (count from 1 to 8). This is measured by the number of institutions providing medical care.

• Education system stability (increased; the same; decreased). This is the variation in the quality of education over the last six months.

The variables have mixed measurement scales, so the PRINCALS algorithm is used to generate the S index. The correlations in Table 21.6 show that professional skills, educational level and employment ratio play the major roles in estimating S. Obviously, the correlations with the number of household members to lose their jobs and assistance dependency are negative: stability decreases as dependency on assistance increases, and as more household members lose their jobs. More surprising is the negative correlation with stability of the education system. This is because households with higher educational levels perceive a greater worsening of the education system than those with lower educational levels.

Table 21.6 Correlation of S with transformed variables.

Transformed variable S

Professional skills 0.7234

Educational level 0.7930

Employment ratio 0.6786 Household members to lose jobs -0.0609

Income stability 0.2112

Assistance dependency -0.3723

Assistance stability 0.3116

Health stability 0.1198

Education system stability -0.1315

Estimation of resilience (R)

The variables estimated above become covariates in estimation of the resilience index. Considering that all the estimated components are normally distributed with mean 0 and variance 1, it is easy to apply traditional factor or principal components analysis.15 A factor analysis is run using the iterated principal factor method, which re-estimates communalities iteratively.16

The results obtained are very satisfactory. Table 21.7 shows that factor 1 alone explains more then 75% of the variance, with factors 2 and 3 accounting for 16% and 8%, respectively. Table 21.8 presents some interesting results in terms of interpretation, especially of the first two factors. The first factor seems to represent fairly well the household's level of well-being. Of the resilience building blocks, only SSN is not positively related to the first factor, because it is negatively correlated with the other variables. This is obvious given that social cohesion usually increases as households become poorer. Social safety nets are a positive feature of resilience, however, so they are captured in the second factor, where SSN becomes positive. Even adaptive capacity (AC) assumes a positive value in the second factor. It can be imagined that when a household becomes poorer it acquires adaptive capacities that it did not have when it was richer (e.g., household members consider doing lower-level jobs than they would have done before). This second factor probably captures mechanisms - mainly non-economic and informal - that are triggered when a household is under stress (i.e. when it is poorer and in need, income

Factors |
Eigenvalue |
% of a2 |

Factor 1 |
1.51766 |
0.7529 |

Factor 2 |
0.32620 |
0.1618 |

Factor 3 |
0.15795 |
0.0784 |

Factor 4 |
0.01399 |
0.0069 |

Factor 5 |
-0.00018 |
-0.0001 |

Var. |
Factor 1 |
Factor 2 |
Factor 3 |

IFA |
0.8077 |
-0.0089 |
0.1396 |

AC |
0.6039 |
0.3130 |
-0.1380 |

S |
0.5753 |
-0.1407 |
-0.2049 |

APS |
0.2808 |
0.1146 |
0.2759 |

SSN |
-0.3011 |
0.4419 |
-0.0363 |

15 It is assumed that all variability in an item should be used in the principal components analysis, while only the variability in an item that is common to other items is used in factor analysis. In most cases, these two methods yield very similar results, but principal components analysis is often preferred as a method for data reduction, and principal factors analysis for cases when the goal of the analysis is to detect the structure in the data.

16 Communality is the proportion of the variance of a particular item that is due to common factors (i.e. shared among several items).

15 It is assumed that all variability in an item should be used in the principal components analysis, while only the variability in an item that is common to other items is used in factor analysis. In most cases, these two methods yield very similar results, but principal components analysis is often preferred as a method for data reduction, and principal factors analysis for cases when the goal of the analysis is to detect the structure in the data.

16 Communality is the proportion of the variance of a particular item that is due to common factors (i.e. shared among several items).

and stability become negatively correlated). The third factor is rather more difficult to interpret. It seems to have something in common with the concept of utilization in food security because it is positively related to public services and level of income.

The factor analysis shows that resilience cannot be a one-dimensional concept. As explained, the positiveness of SSN is not captured by the first factor but by the second. Even if the first factor explains more then 75% of the variance, the other two factors have to be included in the measurement of resilience. It is possible to have a weighted sum of three scored factors because they are orthogonal to each other, so the risk of multicollinearity is avoided. To estimate resilience, the three factors must therefore first be generated using Thompson's regression method and then each must be multiplied by its own proportion of variance explained:

Resilience = 0.753 x Factor 1 + 0.162 x Factor2 + 0.078 x Factor3.

Factor analysis produces a resilience index that is normally distributed with mean 0 and variance depending on the variance and covariance of the scored factors. This is a common feature of factor and principal components analysis, and is a limitation when resilience indices are to be compared among different countries or, more generally, different samples. Nevertheless, the role of the resilience index becomes crucial for policy-makers seeking to analyse subsamples of the original population, such as regions or social groups.

This section concludes by presenting some of the estimates of the resilience index and its components in the five subregions of Palestine. Figure 21.3 shows the Epanechnikov

Kernel Density Estimation of Resilience

Kernel Density Estimation of Resilience

Resilience Index

North West Bank

--East Jerusalem

Gaza Strip

Resilience Index kernel = epanechnikov, bandwidth = .13

North West Bank

--East Jerusalem

Gaza Strip

Figure 21.3 Distribution of resilience in the five Palestinian subregions.

kernel density estimates of the resilience distribution. The initial presentation of the results is based on a non-parametric method, because such methods are normally more informative.

Figure 21.3 shows a clear difference in resilience distribution between East Jerusalem17 and the other subregions. However, South West Bank (WB) and Gaza Strip seem to have more or less the same resilience level, so the parametric approach is used to improve understanding of the differences among subregions and to obtain the relevant significance level. Table 21.9 shows the means and standard deviations (SDs) for resilience and its standardized components. The matrix in Table 21.10 shows the t statistics for pairwise comparison among the means of the different subregions.

The differences among subregional resilience levels are all significant except for that between Gaza Strip and South West Bank. The level of resilience in Gaza Strip would have been much lower if social safety nets had been excluded from the analysis. Figure 21.4 shows the differences among subregions, not only for resilience but also for its different components.

Jerusalem has the highest level of resilience, which depends mainly on income capacity, access to services and stability. Adaptive capacity seems to remain relatively similar to that of other subregions, however. At the opposite extreme, Gaza Strip shows very limited access to food and income, and a high level of dependency on safety nets (from both family ties and external assistance) and on services provided by the Palestinian Authority. South West Bank, with a low level of resilience that is equally distributed among the five pillars, shows a substantially lower level of adaptive capacities. North and Mid-West Bank seem to have relatively balanced distributions across the five pillars and more stable structures of their resilience levels.

The analysis also increases understanding of the current situation in different areas of West Bank and Gaza Strip, particularly Gaza Strip's dependence on external aid. The availability of and access to food depend substantially on the capacity of social safety net mechanisms, both traditional household-based ones and those provided by the international community. Physical access to markets also has a major influence on the overall capacity of Gaza Strip households to bounce back after acute crises.

The study of resilience according to its different components is crucial, especially for policy decisions. A more detailed analysis will therefore be conducted in the following section to explore the policy implications.

Was this article helpful?

## Post a comment