Disability definition
The most common definition and classification of disability within the Indian government was determined with the enactment of the 1995 Act, and states that a person is considered to have a disability if they suffer ‘from not less than 40 % of any disability as certified by a medical authority’ [5]. Disability is considered to be blindness, low vision, leprosy-cured, hearing impairment, locomotor, mental retardation, or mental illness. In 1999, the National Welfare of Persons with Autism, Cerebral Pulsy, Mental Retardation and Multiple Disability Act, added two classes: people with autism and people with multiple disabilities [18]. The 2001 Indian Census states that “defining and measuring disability is a complex issue and it is not easy to communicate these concepts during the census process, in which only a limited amount of questioning time is possible with a household for obtaining detailed information on every individual”. The Census therefore used its own version of disability types, classified into five categories: (i) sight (ii) speech (iii) hearing (iv) movement and (v) mental [1]. This definition has been accepted by the government, both administratively and legally, and is thus used in this paper.
Data
The data source for this study is the 2001 Population Census, which included multiple questions on disability. Each person was asked if he/she had a physical or mental disability according to five categories: speech, sight, hearing, mental (mental), or movement (physical) [19]. If a person has two or more types of disabilities only one was recorded, and it was left to the respondent to decide which one they wanted to be classified into as the most dominant. This was a choice made by the Government of India’s Census Office at the time. Persons with temporary mental or locomotor inability (due to acute medical conditions) on the date of enumeration were not considered as disabled.
The dependent variable in this research is employment and it is defined as those that participated in “work” according for the Indian Census. The Census defines work as “participation in any economically productive activity with or without compensation, wages or profit. Such participation may be physical and/or mental in nature. Work involves not only actual work but also includes effective supervision and direction of work. It even includes part time help or unpaid work on farm, family enterprise or in any other economic activity”. There are several categories of “work” used by the Census including main worker, marginal worker, cultivator, agricultural laborers, household industry workers and other workers [20]. According to the Census 2001 metadata, ‘the reference period for determining a person as worker and non-worker is one year preceding the date of enumeration”.
The disability data were detailed by district, which is the first-level administrative unit within an Indian state. There are 890 districts within the 28 states and seven union territories of India. Of these, 47 are island districts (such as the Andaman and Nicobar Islands, which have no neighboring districts and therefore are not suitable for spatial analytical methods). Of the remaining 843 districts, 250 have no inhabitants. Percentages were calculated considering only the 593 remaining districts with reported inhabitants in the Census. Each district can have both rural and urban areas, and a small number are considered as exclusively rural or urban; thus the denominator for urban and rural percentages varied. In the urban/rural analysis rural and urban percentages were calculated for each variable. We started with a dataset that was stratified by rural and urban PwD from the Indian Census, so the districts did not have to classify as urban and rural. Table 2 reports the total 843 because all districts are used regardless of their inhabitants in the spatial model, since it will exclude those districts automatically.
The disability data were spatially joined to the 2001 Census geographic dataset (retrieved from Harvard GeoSpatial Library, Cambridge, Massachusetts). The data were projected using Kalianpur 1975 India Zone IIb, which is a spatial adjustment to view a specific part of the globe in a flat way. Data joining and projection were done in ArcGIS 10 (Environmental System Research Institute, Redlands, California). The regression was completed in GeoDaTM. The research was ethically approved by the Harvard School of Public Health’s Department of Global Health and Population as a part of the fulfillment for a Master’s of Science and it was determined that International Review Board submission was not necessary due to the analysis of secondary data from the Indian Census which is publically available.
The main variable of interest is the proportion of employed PwD in a district. The variable was calculated as a rate: the total number of employed persons with disability in a district as the numerator and the total number of persons with disability in the denominator. Employment is defined as six types of “workers”: main workers, marginal workers, cultivators, agricultural labourers, household industry workers and other workers. Other relevant variables, all at the district-level included: (i) proportion of female PwD; (ii) proportion of PwD who are literate; (iii) proportion of PwD by disability type; and (iv) proportion of PwD living in urban areas. These variables were calculated with all ages of PwD as the denominator, except for the literacy variable. Age restrictions were not included in the data because this information was not available. Therefore, this analysis should only be interpreted as the proportion of employed persons with disability of total persons with disability in a district. There is potential confounding due to variation in age distributions between districts but this is likely small. Population density (total number of persons with and without disability) was also considered and not included because persons with disability are looked as a separate population in this analysis. Further, data was available for only those that lived in urban and those that lived in rural areas (as separate data sets). This was used in a stratified analysis of rural and urban populations to identify any patterns in characteristics that predicted employment.
In the 2001 Census, literacy was defined as the ability to read in the local language. In this research it is used as a crude estimate to determine whether someone has at least a few years of schooling. Although the ideal variable would be years of education, this information was not collected for PwD in the 2001 Census. We would expect this variable to predict employment because of the intimate link between education and employment that is experienced throughout India, especially among PwD [8, 21]. Therefore we may expect literacy to positively predict employment.
Analytical approach
Linear regression models were used to assess potential determinants of the proportion of PwD employed at the district-level. Linear regression has several assumptions, which were assessed. Potential covariates were chosen based on three criteria: (i) evidence from the literature regarding common drivers of employment and those specific to PwD, in India and in other countries; (ii) special attention to variables that can contribute to the formulation of state and local policy; and (iii) the availability of data in the 2001 Indian Census at the district-level.
Three model formulations were considered. The first (labeled as Model 1) included (i) proportion of female PwD in a district; (ii) proportion of illiterate PwD in a district; (iii) proportion of PwD by disability type in a district, considering four categories: mental, movement, sight, and speech/hearing (combining speech and hearing in one category is plausible because they can be considered communication disorders that generally (but not always) occur together [22]; and (iv) proportion of PwD living in urban areas in a district. The second model (labeled as Model 2) included all variables from Model 1 and added state fixed effects in order to account for potential correlation between the proportion of PwD employed and state characteristics.
Since previous studies have shown that employment for PwD is more difficult in rural areas [11, 23], compared to urban areas, we stratified the analysis by area of residence. We considered the model formulation with greatest explanatory power (as defined by the R2 observed in Models 1 and 2) and ran two additional models, one for urban and another for rural PwD. Further stratification could have been pursued based on variable distribution. The purpose of the model was not to account for the differences in distribution of disability characteristics. It was a cross sectional look at how the distribution of these characteristics influenced employment.
Lastly, we considered that the proportion of PwD employed could vary spatially [13, 24, 25], and in this case the presence of spatial autocorrelation would violate basic assumptions of linear models [26]. Thus, we tested the residuals of each model for the presence of spatial autocorrelation using the global Moran’s I indicator. If the test was significant, we used spatial autoregressive models, and included spatial lag terms based on diagnostics provided by Langrage Multiplier tests [26]. Model goodness-of-fit was assessed by comparing the likelihood ratio and the Bruesch-Pagan test of each model. The Breusch-Pagan test is used to compare the standardized square of the OLS residuals regressed against the square of the original coefficients to determine the presence of heteroskedasticity in the error terms. All regression models were run in GeoDA, an open-source spatial analysis software.