A Structural Equation Model of Female Empowerment

Abstract

This paper proposes a structural equation model to measure and explain female empowerment in Cambodia. Empowerment is defined as the decision-making ability of a woman regarding her strategic and non-strategic life choices. Grounded in the Capability Approach and in the gender economics literature this conceptualisation accounts for three key elements: resources, values/traditions, and decision-outcomes. These elements interact into a system of structural equations where a latent variable is specified to measure empowerment; decision-outcomes enter as partial metrics of empowerment; and resources, and values/traditions are modelled as exogenous factors. Stochastic dominance analysis is used to compare the empowerment status of women across life choices.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. For a discussion of the limitations of proxy indicators as metrics of empowerment see Klein and Ballon (2016).

2. The threat options reflect the wellbeing status of a household member if cooperation fails.

3. SEM’s are regression equations that express relationships between observed and unobserved or latent variables. The general structural equation model consists of two types of models: a measurement model that specifies the relationship between observed and latent variables, and a latent or structural model that shows the influence of latent variables on each other. MIMIC models are a special case of SEMs where the measurement model contains observed variables that are multiple indicators of the latent variable, and the structural model contains multiple causes (or explanatory variables) of the latent variable. These models have been widely applied in social psychology, sociology (Bollen, 1989; Bollen & Curran, 2006) and also in development economics (Krishnakumar & Ballon, 2008; Kuklys, 2005). In the area of gender studies we find very few examples, among which we may cite Pitt, Khandker, and Cartwright (2006) who used a MIMIC model to analyse the impact of microcredit programme participation on empowerment.

4. As is standard in the SEM literature, latent variables are denoted by a circle while observed indicators and observed explanatory variables are denoted by a rectangle. The direction of the arrow indicates the sense of causality. Bold symbols denote vectors.

5. Note that the model that was estimated is the one described in Figure 2, and for clarity purposes in Tables 4–7 we only report variables that were statistically significant (at 5, or 10% levels).

6. The root mean square error (RMSEA) is an absolute fit statistic that determines how well a model fits the sample data (see Browne & Cudeck, 1993; Steiger & Lind, 1980). It is regarded as one of the most informative fit indices as it favours parsimony. Browne and Cudeck (1993) recommend a value of less than 0.5 for reasonably well-fitting models. Other conventional fit indices include the comparative fit index (CFI) and the non-normed fit index also known as the Tucker-Lewis index (TLI). Unlike the RMSEA fit measure the CFI and the TLI are relative fit statistics and their calculation relies on a comparison with a baseline or null model (Bentler, 1990). The TLI assesses a model by comparing the chi-square value of the model with that of a null, and the CFI performs a comparison with the null model using the sample correlation matrix. In both cases, the null model assumes independence and specifies that all variables are uncorrelated. Values of these statistics range from 0 to 1 with values closer to 1 indicating good fit.

7. Factor loadings show the expected change in the decision outcome indicator following a one unit change in the latent variable or empowerment domain in our case.

8. It is important to mention that the scores are purely ordinal (but continuous) in nature, hence their actual value or metric has no intrinsic meaning, but only in comparison with another value.

9. When the observed indicators (y) are categorical, s is computed by a set of p probit regressions of each y on all x and w variables, followed by a set of p (p-1)/2 bivariate probit regressions of each pair of y variables on all x and w variables.