ADDITIONAL INFORMATION ON COMPUTERS AND CONNECTIVITY

Constantijn W. A. Panis, Tora K. Bikson

Introduction

The Computation of Net Disparities Across Socioeconomic Groups

Both outcome variables--presence of a computer in the individual's household and use of network services anywhere--are binary (yes/no) variables. The use of linear regression techniques, such as OLS or analysis of variance (ANOVA), would be inappropriate, since these do not guarantee that predicted fractions are between 0 and 1. The most commonly used statistical models to estimate binary outcomes are the logistical regression (logit) and probit models. The choice between them is largely arbitrary; we opted for the probit model (Maddala, 1983).

The procedure is as follows. First, we estimated multivariate probit models to explain, say, presence of a computer in the household using the six categorical predictor variables listed above. Second, to determine net disparities by, say, sex, for each individual in the sample, we predict the probability that he or she has a computer in the household under the counterfactual assumption that everyone is female. That is, if everyone were female, but otherwise with the same characteristics that he or she actually has, what would be the probability that each person would have a computer? Third, we average these predicted probabilities over all individuals in the sample to obtain the predicted fraction of the population that would have a computer if everyone were female. This prediction is repeated under the counterfactual assumption that everyone is male, and averaged over all individuals in the sample to obtain an estimate of the fraction having a computer among males. The resulting fractions are termed "net" fractions, since they represent differences that are due only to sex, controlling for all other socioeconomic characteristics of interest. This procedure is repeated for the other five socioeconomic predictor variables. The same procedure is used to compute net percentages of network users.

The multivariate probit estimates are of interest in their own right. They allow us to test whether net differences across socioeconomic groups are statistically significant. Given the fact that our samples are very large (146,850 individuals in 1989 and 143,129 in 1993), virtually all estimated differences are indeed statistically significant.

We also needed to test whether the disparities have grown or narrowed between 1989 and 1993. To achieve this, we pool all individuals and estimate two sets of coefficients. The first set applies to all individuals; the second applies only to individuals in 1989. This second set is thus a full interaction of the year 1989 with all characteristics of interest. (The result is that the first set provides estimates for the 1993 sample, whereas estimates for the 1989 sample are given by the sum of the first and second sets.) The sign and significance of the interaction terms determine whether differences across groups have significantly narrowed or grown.

Table A.1 presents the probit estimates to explain the presence of a computer in the household. The omitted category is a non- Hispanic white male who lives in a rural area, is between the ages of 20 and 39, is a high school graduate, and lives in a household with a total income that is in the bottom quartile. Most coefficients are significant at the 5 percent confidence level. The exceptions are, first, the main and interaction effects for female, i.e., there is no significant net difference in access to a computer in the household between males and females in 1993 or 1989. Second, the 1989 interaction effects for race and ethnicity are neither individually nor jointly significantly different from zero. This implies that there has been no significant narrowing or widening of the disparities across racial and ethnic groups.

Similarly, Table A.2 presents probit estimates to explain the use of network services. Again, most coefficients are significantly different from zero. The chief exceptions are, first, the main and interaction effects for female, i.e., there is no significant net difference in use of network services between males and females in 1993 or 1989.

Second, in 1993, Native Americans are not significantly more or less likely to use network services than non-Hispanic whites (the omitted category). Third, the 1989 interaction effects for race and ethnicity are neither individually nor jointly significantly different from zero. This again implies that there has been no significant narrowing or widening of the disparities across racial and ethnic groups.

Table A.3 presents gross and net percentages of individuals with access to a computer in the household and of individuals using network services. The gross percentages were also graphically shown in the bar charts of Chapter Two. The net percentages control for all other socioeconomic characteristics, as explained above.

Regional and Other Variation in Computer Access and Network Use