Figure 1 depicts a conceptual framework that explicates the role of the social context in health behavior change [9]. We chose a social ecological framework to illustrate social contextual factors across multiple levels of influence [14-17]. Among these were individual factors, which include material circumstances such as owning one’s own car or having adequate resources for child care. Interpersonal factors, such as the presence of social ties, family roles and responsibilities, and social norms, are likely to be powerful correlates of health behaviors and may vary by factors reflecting cultural differences (race/ethnicity, acculturation). Organizational factors may reflect the work setting, for example, job stress, control, and exposure to a hazardous work environment. Neighborhood and community factors measured on an individual level include access to a safe neighborhood. Finally, larger societal forces, such as racial discrimination, may also shape health behaviors and outcomes.

Social contextual factors, in turn, may influence health behaviors directly or indirectly through individual psychosocial factors. Social cognitive theory [18,19], the theory of reasoned action [20,21], and the transtheoretical model of behavior change [22,23] are guiding models that highlight specific individual psychosocial factors that predict a change in behavior. Psychosocial mediating variables in large part influence intentions to change behavior, which are highly associated with the likelihood of change [24,25].

Figure 1

Conceptual model

Recruitment for Open Doors to Health began in 2004. Participants provided informed consent and completed an interviewer-administered survey in either English or Spanish. Participants received US $25 compensation. Eligibility criteria for the study survey included (1) living in the housing community, (2) being at least 18 years old, (3) being fluent in English or Spanish, and (4) not having cancer. An initial sample of 3688 subjects was drawn. Of them, 747 (20%) were deemed ineligible, leaving 2941 eligible individuals. Of these, 828 (28%) refused participation, and 559 (19%) could not be reached, leaving 1554 residents who completed the baseline survey. This yielded an overall 53% response rate, with a range of 34% to 92% across the housing sites. The study protocol was approved by the Human Subjects Committee at the Harvard School of Public Health.

Sociodemographic variables collected included gender, date of birth, race/ethnicity (categorized as black, white, Hispanic, and other), and highest level of education completed. We also assessed poverty status and financial situation with two measures. Yearly household income (six response options ranging from less than US $10000 to at least US $50000) and the number of people supported by this income were used to measure poverty status (dichotomized as being above or below the poverty level based on the 2005 federal poverty guidelines on income and household size) [28]. Participants were also asked about their perception of the financial status of their household (comfortable with some extras, enough but no extras, have to cut back, or cannot make ends meet).

We assessed employment status in several ways. Participants were asked if they were working, and, if so, (1) whether they worked full-time or part-time and (2) the number of hours worked in a week, including overtime or extra hours. Hours worked were categorized as 0, less than 20, 20 to < 37, and 37+ hours per week. Participants were also asked about the number of jobs (beyond their main job) they worked (0, 1, or more than 1) and whether they supervised employees.

Lastly, we assessed immigrant status by asking participants their birthplace, the number of years they lived in the United States, and their first or native language.

Each participant was asked about several social contextual factors. Neighborhood safety was assessed by asking whether participants felt safe walking alone in their neighborhood during the day and at night [29]. For both daytime and nighttime, participants were asked “How safe do you feel walking alone in your neighborhood?” Response options included “safe,” “a little unsafe,” and “unsafe.” For analysis purposes, we combined the response categories of “a little unsafe” and “unsafe” for daytime safety due to the small number of responses in the latter category.

To assess social cohesion in the housing community, we asked respondents to report their agreement with five statements: (1) people around here are willing to help their neighbors; (2) this is a close-knit neighborhood; (3) people in this neighborhood can be trusted; (4) people in this neighborhood generally do not get along with each other; and (5) people in this neighborhood do not share the same values. Item responses were reversed for the first three statements and then responses to the five items were averaged. The summary score ranged from 1 to 4, with a higher score indicating higher social cohesion [30].

Marital status, number of close friends, number of close family members, and active membership in organizations (religious, professional, community, civic, etc) were combined to form a continuous measure of the number of social network ties ranging from 0 to 4, with a higher score indicating a greater social network [31]. Social support was assessed by asking participants about emotional support from family and friends, support when sick, help with household tasks, financial support, and help getting to the doctor. Responses to questions of social support were measured on a 5-point Likert scale ranging from “yes, always have someone to help” to “no, no one like that.” A single social support variable was created by adding the number of responses to the five questions that indicated at least some support, with a range of 0 to 5. Higher scores indicated greater social support [32].

Participants were asked about their various family roles, which included “earning money to support the family,” “taking care of children,” and “taking care of another household.” The measure of multiple roles was computed as the number of family roles for which the participant was mostly or fully responsible (0 to 3). To determine role conflicts, participants were asked whether their daily activities made conflicting demands on them (ie, role conflict) [9].

Health status was captured by asking participants whether health problems make it difficult for them to exercise (yes/no).

Participants were also asked to report the number of hours per day (during the week and weekend) that they watched television [33].

We assessed computer ownership and frequency of use (daily, weekly, monthly, less than monthly, and never). Use was recoded as regular (daily and weekly), intermittent (monthly and less than monthly), and never. For multivariable modeling purposes, this variable was further dichotomized as regular versus intermittent or no use. Participants were also asked where they most often use a computer: home, work, housing site, library, friend’s house, community center, or other. The latter five response options were coded as “other” for the purpose of these analyses.

On the basis of the cluster design, data for all analyses were weighted up to the population size within each housing site (with a total weighted size of 2270). Frequency distributions and estimates of means and standard deviations were assessed for distributional assumptions and outliers. Bivariate associations between the dependent variables, computer ownership and use of a computer, and independent variables were assessed, and variables found to be significant at the P = .15 level in bivariate analyses were retained for use in multivariable modeling. Based on the bivariate associations and consideration of effect modifiers (ie, interaction effects) and confounders, multivariable logistic models of the dependent variables were developed. Bivariate associations with computer use were assessed using three category polytomous logistic regression models with a generalized logit assumption. Based on the assessment of these models, the sample size of the intermittent users and consideration of cluster model constraints, we dichotomized the computer use variable to regular versus intermittent and nonusers. Thus, all multivariable models are dichotomous logistic models. All analyses were conducted to adjust for age as a potential confounder. All analyses were conducted using SUDAAN version 9.0.1 (Research Triangle Institute, Research Triangle Park, NC, USA) and SAS statistical software version 9.1 (SAS Institute, Cary, NC, USA) for clustered data.

The majority of the study participants were female (74%), not working or disabled (63%), and earned less than US $20000 per year (74%). The mean age of the participants was 51.0 ± 21.4 years. Almost half of the participants were black (43%), and an equal number were Hispanic (43%); 52% of participants were born in the United States. A slight majority of participants lived above the poverty level (51%); however, 43% considered themselves to be under financial hardship.

More than half (51%) of participants owned a computer, and 42% reported regular computer use (Table 1); 50% of regular users used the computer most often at home. Computer ownership was highest (greater than 70%) among participants who were less than 49 years old, those with at least some college education, and those who were employed. A large number of adults 65+ (86%) and those below the poverty level (58%) had never used a computer. Unemployed participants were more likely to use a computer in places such as a library or friend’s house than at home.

Our study focused on access to computers among low-income minority groups. We did not specifically ask about Internet access and use. Information regarding Internet use, type of Internet connection, and reasons for computer use would have further contextualized the communication experience of low-income minority adults. However, government reports show that about two thirds of households with computers also have Internet access [41]; rates of Internet access among our participants are therefore likely similar. Also, ownership does not imply use. It is possible that other people in the household (eg, children) actually use the computer. As our study was a cross-sectional design, we were unable to determine whether sociodemographic and social contextual factors causally influence computer ownership, frequency of use, and location. Because this study was conducted in urban, low-income minority public housing communities, its findings are only generalizable to similar settings. Nevertheless, this large study illustrates the level of access to computers among low-income, urban minority adults.

The racial/ethnic and socioeconomic disparities in access to communication technologies are narrowing, even among very low-income households, making communication technologies for health communication more feasible. However, as the number of technology-based prevention interventions that provide important health information increases, it will be imperative to continue to identify factors that contribute to disparities in access and to connect low-income racial/ethnic minorities to these technologies, particularly computers. In this study, computer ownership among low-income minorities was over 50%, showing noteworthy strides. This suggests that computer-based studies might be reasonable for this population provided that options for nontechnology modalities are also provided. While sociodemographic factors are commonly associated with computer access, a unique finding of our study is that it may be equally as important to consider specific social contextual factors when trying to increase access and use among low-income minorities, such as social network ties, household responsibilities, and neighborhood safety.