Gender digital divide in a patriarchal society: what can we learn from Blinder–Oaxaca decomposition?

Abstract

Based on the self-reported usage time, this paper aims at analyzing the gender differences in computer and internet use at home. Using data from the 2012 Survey of Digital Divide in Taiwan, we apply a regression-based decomposition method to identify the underlying causes of observed usage differential between males and females. Conditioned on adoption, it is found that compared with their high-income counterparts, low-income females in Taiwan do not spend more time on internet surfing as a result of the high opportunity cost of leisure time. A further decomposition analysis suggests while the gender-specific factors are not the main causes of gender differences in computer and internet use, differences in internet experience and opportunity cost of leisure time between the two gender groups are root causes of observed gender usage differential. The present study adds to the literature by providing a framework that can be easily extended to understanding the root causes, such as that of gender digital divide, for patriarchal societies where female self-stereotyping is a pronounced problem.

Appendix: Identifying the contributions of gender-specific factors

In Sect. 3.3, the contribution of gender-specific factors to CI usage differentials is captured by the differences in estimated constant terms between males and females. Due to an identification problem, we apply the method of “normalized” regression proposed by Yun (2005) to identify the contribution of gender-specific factors. This appendix provides a brief introduction of the method of “normalized” regression. To discuss with ease, we assume that the usage regression only includes constant term and one set of categorical variables, D, with K categories and the first category is treated as the left-out reference group. The OLS regression can be modeled as:

$$Y = \alpha + \sum\limits_{k = 1}^{K} {\alpha_{k} D_{k} + \varepsilon }$$

(15)

where α is the constant term to be estimated; and parameter α ₁is constrained to be zero since D ₁ is the omitted group in Eq. (15).

We transform Eq. (15) into the “normalized” regression as follows:

$$Y = \alpha + \bar{\alpha } = \sum\limits_{k = 1}^{K} {(\alpha_{k} - \bar{\alpha })D_{k} + \varepsilon }$$

(16)

where $\bar{\alpha } = \sum\nolimits_{k = 1}^{K} {\alpha_{k} /K}$ is the mean coefficient of the categorical variables D. In Eq. (16), $(\alpha_{k} - \bar{\alpha })$ represents the deviation of the OLS estimates of the dummy variable from the mean. Due to the nature of dummy variables, Eq. (16) is mathematically equivalent to Eq. (15). The changes in the OLS estimates of dummy variable will be offset by the changes in intercept.

By dividing the total sample into two groups (e.g., m and f) and using Blinder–Oaxaca decomposition technique, the different form of endowment and coefficient effects for constant term and dummy variables are expressed as:

$$\begin{aligned} \bar{Y}_{m} - \bar{Y}_{f} = \underbrace {{\hat{\alpha }_{m} - \hat{\alpha }_{f} + \bar{\alpha }_{m} - \bar{\alpha }_{f} }}_{{{\text{Coefficient\, effect}}\;{\text{for}}\;{\text{constant}}\;{\text{term}}}} + \underbrace {{\sum\limits_{k = 1}^{K} {\bar{D}_{mk} (\hat{\alpha }_{mk} - \hat{\alpha }_{fk} + \bar{\alpha }_{m} - \bar{\alpha }_{f} )} }}_{\text{Coefficient\, effect\, for\, dummy\, variable}} \hfill \\ + \underbrace {{\sum\limits_{k = 1}^{K} {\bar{D}_{mk} - \bar{D}_{fk} (\hat{\alpha }_{fk} - \bar{\alpha }_{f} )} }}_{\text{Endowment\, effect\, for\, dummy\, variable}} \hfill \\ \end{aligned}$$

(17)

Similarly, Eq. (17) can be extended to the case that consists of multiple categorical variables as in Eq. (14). Applying “normalized” regression, the detailed coefficient effect of the constant term is invariant to the choice of left-out reference groups, and thus resolves the identification problem of contribution of gender-specific factors to CI usage differentials.