Data and sample
To test the postulated hypothesis, this study used the Health Information National Trends Survey (HINTS), which has been administered every few years since 2003 by the U.S. National Cancer Institutes. Its ultimate goal is to learn the patterns of how adults find, understand, and use health information. It plans to achieve this goal by collecting data on health-related communications, patterns of communication with doctors, and behaviors related to the Internet, health services, and health information technology. It is one of the most comprehensive national-level datasets for these topics in existence [
59].
The data collection procedures for this dataset encompassed a complex, multistage sampling designed to represent the civilian, non-institutionalized population of the United States. It encompassed samples from both a telephone random digit dialing sample of phone numbers and the mail through a sample of addresses [
60]. For this study, the researcher used HINTS data collected between January 2008 and May 2008 (
N = 7,674). A subset of the sample for this study contained those (≥18) who (1) went to the Internet first to look for information about health and medical topics; (2) used healthcare services during the past 12 months; and (3) gave valid data. The final unweighted sample consists of 2,297 respondents.
Measures
A measurement of patient empowerment has many constructs, but is not well constructed, leaving “[uncertainty] about the best way to define and measure it” ([
30], p. 1, also see [
8,
29]). Moreover, Zimmerman [
22] acknowledged that it is unlikely that an empowerment measurement “would [universally] fit all (or most) persons” and “would [globally] fit all (or most) contexts” (p. 587). The measurement of psychological empowerment, however, may measure the consequences of the empowerment process, including an examination of “the effects of interventions designed to empower participants [and] empowering processes and mechanisms” ([
22], p. 585). The latent concept of empowerment can be “potentially measurable” using the manifested perceptions and behaviors as to empowering process as well as empowered outcomes ([
33], p. 1).
Literature revealed that an outcome measurement of the quality of healthcare services encompasses the structural properties such as facilities and healthcare professionals, the process and outcome of the treatment, and relational properties such as communications, information, and coordination [
39‐
41,
61], which overlap with the contextual factors in patient empowerment [
22,
25‐
28,
34,
35]. The patient empowerment measurements, which asked the patients’ perceptions of the quality of healthcare services, included Small et al.’s [
62] and Bulsara et al.’s [
63] patient empowerment scales. The former used one question to measure patients’ trust in their doctors and seven items to measure the patients’ perceptions of their doctors’ interpersonal care skills (e.g., listening to the patients, involving the patients in decisions, treating the patients with care, and taking the patients’ problems seriously) while the latter used a single item to ask about patients’ perceptions of their healthcare professionals’ willingness to include them in the decision-making process for treatment. Patients’ assessments of the quality of healthcare services may well capture the patients’ global perceptions of how the healthcare services (i.e., contextual factors) were facilitated to empower the patients to talk with their healthcare providers about the health information they found online like a measurement of self-assessed health. A measurement which asks individuals about universal health is considered to be as reliable and valid as biological measures, such as physical and laboratory examinations [
64]. Hence, the outcome measurement of the contextual factors in psychological empowerment was operationalized as “Overall, how would you rate the quality of health care you received in the last 12 months?” This question was measured on a 5-point Likert scale with higher numbers representing greater empowerment outcomes.
The patients’ self-assessed general health (from 1 = excellent to 5 = poor) and psychological distress were used to represent the health latent variable. Psychological distress contains six constructs (i.e., sad, nervous, restless, hopeless, taxing, and worthless), which the patients could have experienced over the past 30 days. Each item was measured on the 5-point scale (from 1 = all of the time to 4 = none of the time) and the scores were reversed for this study. The summation of these constructs ranged from 1 to 24. Hence, higher scores indicated poor physical and mental health.
The patients’ predisposing factors to be controlled were cancer history (binary) and socio-demographics. Socio-demographics included age (18–34, 35–54, 55–74, or 75+), gender (male vs. female), race and ethnicity (i.e., non-Hispanic whites, Hispanics, blacks, or Other), marital status (not married vs. married), education (i.e., high school, some college, or college+), job (unemployed vs. employed), household income (<20 K, <35 K, <50 K, <75 K, 75 K+), and U.S.-born (yes vs. no). Residential area (<250,000 for urban vs. ≥250,000 for rural) was descriptively analyzed, but excluded in the structural equation modeling (SEM) due to its non-significant impact on any paths of the model.
The binary dependent variable (i.e., a consultation with healthcare providers about health information found from the Internet) was operationalized as “In the past 12 months, have you talked to a doctor, nurse, or other health professional about any kind of health information you have gotten from the Internet?” This variable is only available from HINTS 2, HINTS 3, and HINTS 4 Cycle 1, for which the data were collected in 2005, 2008, and 2011, respectively [
65]. This study used HINTS 3 (
N = 7,674) because the total sample size was almost double that in HINTS 4 Cycle 1 (
N = 3,959), which reduces the sampling error and produces better estimates of the U.S. population given, in particular, the present study subsample as described in the Data and Sample section.
Data analysis
There were two steps to the analyses. First, the study variables’ univariate analysis for the sample characteristics and bivariate relationships of the study variables with respect to whether taking the health information found online to healthcare professionals were conducted in the SAS statistical software version 9.2 (see Additional file
1). In order to account for HINTS’s survey design and complex multistage sampling design, all of the data were weighted in the descriptive analyses, using post-stratification weights with Jackknife repeated replication methods. These methods allowed for accurate estimates of the variance for the full sample, which, in turn, affected the standard errors, p-values, and confidence levels in the inferential statistical analysis with HINTS [
59]. The univariate distribution and bivariate relationships with respect to consultations with healthcare professionals were examined using Rao-Scott chi-square tests for the categorical data and t-values for the summated psychological distress, which were regressed upon the dependent variable. The chi-square values and t-values were calculated using the PROC SURVEY-procedures.
Next, to test the hypothesis for the proposed indirect paths, SEM in Mplus was used, not only because it allowed us to test complex paths and multiple regressions for the model, but also because it is a comprehensive means for assessing and modifying theoretical models, which led to further theory development [
66]. This study used a robust maximum likelihood estimator (i.e., MLR option in Mplus) using Monte Carlo integration with 500 integration points. This method is robust for categorical data that has violated the underlying normality assumption because it produced robust standard errors [
67].
To see whether an indirect effect exists in the proposed model, the following four steps guided by Baron and Kenny [
68] were used: Confirm (1) IV was significantly correlated with DV (= c); (2) IV was significantly correlated with M (= a); (3) M affected DV while controlling for IV (= b); and (4) the total effect (=
c) equaled the summation of the direct effect (=
c′) and indirect effect (
a*
b).