Methodology

Several models are estimated to fully assess the deterrent effect of disability status on the likelihood of employment, controlling for other variables hypothesized to affect employment. The broad definition of disability is adopted to preserve the sample size of persons with disabilities. Again, employment is defined according to International Labour Organization (ILO) criteria (ILO, 2013).

First, a logistic model is estimated to tackle the impact of having any disability on the likelihood of labour force participation (model 1). The dependent variable in this case is a binary variable equals one if the individual is in the labour market (employed or unemployed) and zero otherwise. Two other versions of model 1 are estimated: one model controlling for the interactions between disability and gender, and disability and status of the individual in the household (model 2), and another model assessing the effect of disability types on the probability of labour force participation (model 3). Thus, the probability of participating in the labour market is given by:

$Pr(L{F}_i^{*}>0)=Pr(u_i>-X_i\beta )=Ʌ\left(X_i\beta \right);$

where Ʌ (.) is the standardized logistic distribution, X_i is the vector of regressors and β is the associated parameter vector. Following the literature, other control variables are included in the models. These variables include factors that determine worker productivity (such as age and education) and factors that determine the relative value of personal time (such as marital status and wealth) (Baldwin, 1999). Moreover, it is believed that women and men often have different incentives and face different barriers to participating in the labour force; hence, a gender binary variable is included in the models. A variable that reflects the level of urbanization is also included since urban and rural areas have different structures of labour markets. Furthermore, the status in the household (head or non-head), household size and unemployment rate at the governorate level[1] are included as well (Fadayomi and Olurinola, 2014).[2]

Once in the labour market, would persons with disabilities be employed or unemployed, and if employed, would they be more likely to work in the public sector? To answer these questions, other models are estimated with different dependent variables. One model is estimated where the dependent binary variable equals one if the individual is employed and zero if unemployed (model 4). Another version of model 4 is estimated to assess the effect of disability types on the probability of employment. Finally, model 5 tackles the impact of being a person with a disability on the likelihood of being employed in public sector. Yet, the dependent variable in these cases is only observed for those already in the labour force. Hence, to control for selection bias, probit models with selection[3] are applied, following Hotchkiss (2004), and Jones and Latreille (2011).

At the first stage of this model, a labour force participation equation is estimated where the propensity to be in the labour force (LF_i^*) is given by:

$Pr(L{F}_i^{*}>0)=Pr(u_i>-X_i\beta )=\Phi \left(X_i\beta \right);$

where Φ (.) is the standardized normal distribution.

Then, at the second stage, the outcome equation can be estimated conditional on labour force participation being observed (LF_i = 1). This outcome equation is given by:

$E_i^{*}=T_i\gamma +v_i;$

where the variable E_i, which is only observed if LF_i = 1, is related to the latent variable E_i^* as follows: E_i= 1 if E_i^* > 0 and E_i = 0 otherwise. E_i= 1 indicates the presence of an employed person or public sector worker, based on the specified model. It is assumed that u_i and v_i are normally distributed with zero means and unit variances, and the correlation between them is given by ρ. If there are unobservable factors affecting both equations, the correlation may be non-zero (ρ≠0). In this case, the results of a simple probit model of the outcome equation are biased. Yet, if the correlation turns out to be insignificant, we can depend on the results from the simple probit model.

Finally, to identify these models, at least one variable is included in the selection equation (labour force participation equation) that is not included in the outcome equation. Yet, in our case, the variables that determine the dependent variables are roughly the same as those determining labour force participation. Thus, it is difficult to find a suitable identifying variable. Nonetheless, a number of studies attempted to solve this identification problem by inserting variables as the unemployment rate at the governorate level, status in the household and household size in the selection equation (Hotchkiss, 2004; Jones, 2011). These variables may affect the probability of labour force participation, but not employment. On the other hand, an additional explanatory variable, economic activity, observed only for the employed, is included in the last model.

[1]    The unemployment rate at the governorate level is obtained from the annual bulletin of the labour force survey (CAPMAS, 2018).