Krishnakumar, Jaya (1); Chavez Juarez, Florian (2) (2017). 'An agent based modelling of multidimensional capabilities and an evaluation of their estimation by structural equation models' Paper presented at the annual conference of the HDCA, Cape Town 2017.
The empirical operationalisation remains a key challenge for the capability approach. Structural equation models (SEM) have been increasingly used to estimate the unobserved capabilities. However, the appropriateness of this approach still subject to discussion. In a previous study (Krishnakumar and Chávez Juárez, 2016, Social Indicators Research) we used an innovative approach to simulate a real life scenario using an agent-based model and then evaluate the performance of SEM methodology by comparing the estimates to the ‘true’ values we have previously simulated. Being the first one of its kind in this domain, the study was focussed on the suitability of the methodology for capturing capabilities, by adopting a relatively simplistic view of the real world: a single dimension of well-being namely education and some basic behavioural assumptions of the different agents.
In this study, we extend this previous work in two major directions that bring the model much closer to a real life situation. First, by adding the health and social relations dimensions and by introducing links among the different dimensions, thus obtaining a multidimensional model with three wellbeing dimensions. Second, we model this artificial world much closer to reality by calibrating many processes using actual data and by initiating the model with a true population. In this respect, we use the Mexican Family Life Survey to build our initial population and to estimate various phenomena we include in the model. These phenomena include the wage structure of the economy, the age-dependent health shocks, the education dependent fertility rates and so on. Finally, we also explore different ways of conceptualising capabilities in our modelling framework, and make comparisons among them. A first approach is based on the notion potential achievements, where we simulate both the actual outcome and simultaneously the potential outcome. We then consider the potential outcome as a measure of capabilities. The second approach is based on the notion of choices, where we capture for every period the set of choices individuals and families had. We also capture interaction effects among the different dimensions.
For the initialisation, we use information on the family structure, the type of locality where the family lives and individual level indicators such as age, education, cognitive abilities, gender, the role in the family and the type of school they are attending. Using this complete information enables us to start with a population that is very closed to the actual population in Mexico, including all the heterogeneities that might be present.
Once the model is initiated, education and health decisions are taken on an annual basis and the model is simulated for a long period of time. Our model contains an elaborate process of human capital accumulation by which previous human capital, health and cognitive abilities determine how much a child learns in a given school year.
The model reproduces several key distributions according to the observed data. For instance, our implementation of the human capital accumulation reproduces the distribution of years of education quite coherently. The age structure of the population is also reproduced over the simulated periods. Overall, the data generated by the model looks very similar to actual data from a country like Mexico, with the important difference that we can also observe generally unobservable variables such as capabilities and sets of choices.
Several outcome indicators are being considered for representing functionnings, such as years of education, type of schooling and the number of repeated grades in the education dimension, the number and intensity of health shocks and the amount of health care expenditures in the health dimension, and the size and quality of the social networks for the social relations dimension. All these variables are generally available in real world data, but do not directly relate to capabilities.
We then use these indicators to estimate the underlying capabilities by the means of two econometric approaches. First, we use conventional latent variable models to estimate capabilities defined by the notion of potential achievements (e.g. potential human capital). Second, we use latent class models to estimate capabilities according to the set of choices approach. In both cases, we will then compare the estimated capabilities to the true (simulated) capabilities in our model and assess the quality of the estimation.
Currently we are programming the last details of the models and expect to have the econometric results by the end of April 2017. Based on some preliminary analyses and the insights from our previous study, we expect both approaches to provide good estimates of the underlying capabilities.