The way people search for health information is constantly changing. Not long ago, people relied almost solely on physicians’ advice to address patients’ medical needs because medicine was a highly specialized field of knowledge, inaccessible and incomprehensible to the general population. However, recent studies have indicated that the picture is gradually changing [1].

According to an annual report published by the Taiwan Network Information Center in 2012, 15.94 million Taiwanese used the Internet, accounting for 77.25% of the Taiwanese population aged 12 years and older [2]. Among those Internet users, many had used the Internet for searching health-related information. Research reported that, in the United States in 2009, 74% of adults used the Internet, and 61% had searched for health-related information online. In particular, 49% had accessed a website for understanding a specific medical condition or problem [3]. As reported by the Pew Research Center in 2013, approximately 59% of Americans had searched online health information in the preceding year [4]. The Internet offers a wide variety of health-related information, allowing the general population to reaffirm what they have heard from their doctors and to make a thoroughly informed decision [5-7].

Compared with the traditional ways of gaining health-related information, such as visiting an outpatient clinic, reading newspapers or magazines, listening to radio or watching television shows, getting advice from neighbors or community or family members, the Internet is inexpensive, accessible, and conveys diverse opinions. On the other hand, more and more physicians use the convenience of the Internet to search drug information [8] and Web-based evidence sources such as UpToDate [9] for providing better medical care to patients.

Relevant studies on how increasing Internet use affects outpatient clinic visits have been inconclusive. Azocar et al [10] in 2003 concluded that patients’ use of a behavioral health website significantly motivated them to increase their use of health care services. In 2008, Lee [11] reported that increasing use of the Internet was correlated with increasing outpatient clinic visits, although only age and sex were controlled in the data analysis. Another study analyzed the association between Internet use and the Internet users’ health-seeking behaviors for those who searched the Web for different purposes. The authors reported that, for every 1-hour increment of searching the Internet, the likelihood for patients to visit physicians was increased by about 10% [12].

In contrast, some studies found a negative or no effect of using the Internet on the number of outpatient clinic visits. A survey conducted in the United States showed that 94% of the participants said their Internet use did not change the number of outpatient clinic visits they made [13]. Another survey conducted in Japan reported that 88.9% of the respondents thought there was no association between their Internet use and their frequency of visiting outpatient clinics or making phone calls to their physicians to inquire about health-related issues [14].

Given that the association between Internet use and health-seeking behaviors is still controversial, and that prior studies were conducted without sufficiently adjusting for potential confounding variables, we conducted this survey in which we derived the data from a sample randomly selected from the general population in Taiwan, with adjustment for most of the potential confounding variables. The aim of this study was to examine the role of Internet use in health-seeking behaviors as indicated by the number of outpatient clinic visits.

We derived the data for this study from the sixth cycle’s second year (2011) survey of the research project Taiwan Social Change Survey (TSCS) [15]. The project was conducted by the Institute of Sociology, Academia Sinica, and sponsored by the Ministry of Science and Technology (formerly known as the National Science Council) in Taiwan. Each year, 2 modules were used for the questionnaires. For example, in 2009, the 2 modules were Social Inequality and Religion. A total of 2026 Social Inequality questionnaires and 1924 Religion questionnaires were finally answered by the participants randomly selected from Taiwan’s general population.

The 2011 TSCS survey contained 2 modules: (1) the Family questionnaire, and (2) the Health questionnaire. Participants in the 2011 TSCS survey were randomly selected from the general population in Taiwan. Face-to-face structured interviews were conducted for all selected participants. In addition to shared questions, some of the survey questions in the questionnaire had been separated into 2 sets: set-A questions and set-B questions. Each set-A question has a similar set-B question. For example, question D1a1 in set A and question D1b1 in set B both inquire about participants’ smoking status.

This was a cross-sectional study. The Health questionnaires collected for this study were answered by the participants who were randomly selected from Taiwan’s general population. Participants in this study were numbered: the odd-numbered participants were assigned to answer set-A questions and the shared questions; and the even-numbered participants were assigned to answer set-B questions and the shared questions in the Health questionnaires. Our study used a secondary dataset collected in 2011 by the Institute of Sociology in Academia Sinica.

We selected the following variables: (1) reported confounding variables such as age [16,17], sex [13,16], annual income [16], educational level [14], self-reported health status, and chronic disease [17,18]; (2) background information of a participant such as residence and marital status; (3) social support variables such as the total number of family members and self-perceived neighbor support; and (4) attitudes toward the health care system.

For balancing the variables between the group of questionnaires without missing data and the group of questionnaires with missing data, we built a model for propensity score matching using multivariate logistic regression. Age and sex were the independent variables, and whether a questionnaire answered by a participant had missing data (the uncompleted group) or not (the completed group) was the dependent variable in the multivariate logistic regression model for propensity score matching. We obtained each participant’s propensity score of being assigned to the completed group based on the multivariate logistic regression model. A participant in the completed group was matched to a participant in the uncompleted group using 1-to-1 nearest-neighbor matching without replacement and a caliper of 0.18 of the pooled standard deviation of the logit of the propensity scores. We examined whether the propensity score model had good discrimination by using the area under the receiver operating characteristic curve. We expected age and sex in the propensity score model to be balanced between the completed and uncompleted groups. The questionnaires in the completed group that were matched to the questionnaires in the uncompleted group were retained for further analysis.

For examining the linear association between an independent variable and the outcome variable, we calculated the Pearson correlation coefficient or Spearman rank correlation coefficient depending on the scale of an independent variable. The outpatient clinic visit was the outcome variable. We coded the outpatient clinic visit using a Likert scale ranging from 1 to 5, indicating the frequency of outpatient clinic visits from “never” to “several times a month,” respectively. An independent variable with a P value of the correlation coefficient <.30 was eligible to enter the multivariate linear regression model. In addition, we examined the collinearity between 2 independent variables using the Spearman rank correlation coefficient: “How is your health status?” and “Do you have chronic diseases?” Only one of them stayed in the model if a significant collinearity was identified.

We conducted multivariate linear regression analysis for examining the association between Internet use and outpatient clinic visits, including the confounding variables selected by linearity and collinearity checks. We regarded P≤.05 as statistically significant. We conducted all statistical analyses in this study using STATA/MP 11.0 (StataCorp LP) for Windows PC. This study was approved by the Research Ethics Committee of National Taiwan University Hospital (201510102W).

A total of 2199 participants joined this study to answer the Health questionnaires. Among the 2199 Health questionnaires, 1064 (48.39%) were even numbered and assigned to answer the set-B questions and the shared questions. Given that set-A and set-B questions were established using different scales—that is, a scale from 1 to 5 was established for set-A questions but a scale from 1 to 6 was established for set-B questions—we thus excluded the participants who answered set-B questions from further analysis. Among the 1135 participants who were odd numbered and assigned to answer the set-A questions and the shared questions on the Health questionnaires, 23 (2.03%) did not provide their age and were thereby excluded. Therefore, we finally included 1112 Health questionnaires for propensity score matching using multivariate logistic regression (Figure 1).

A total of 1112 Health questionnaires were eligible for this study: 555 questionnaires did not have missing data (the completed group); 557 questionnaires (50.09%) had missing data (the uncompleted group). The participants in the completed group were significantly younger than those in the uncompleted group (P<.001). More male participants were in the completed group than in the uncompleted group with a borderline significance (P=.06). To avoid the potential risk of selection bias, we conducted propensity score matching to balance age and sex between the 555 questionnaires in the completed group and the 557 questionnaires in the uncompleted group (Table 1).

We established a propensity score model for matching a questionnaire answered by a participant in the completed group (n=555) to a questionnaire answered by a participant in the uncompleted group (n=557). The propensity score model for the completed group and uncompleted group included 2 independent variables: age and sex. It showed good discrimination (area under the receiver operating characteristic curve=0.83). We identified 293 matched pairs: 293 respondents from the completed group and 293 respondents from the uncompleted group. The mean (SD) propensity scores before matching were 0.67 (0.22) for the completed group and 0.33 (0.25) for the uncompleted group (P<.001). The mean (SD) propensity scores after matching were 0.55 (0.22) for the 293 participants from the completed and 0.51 (0.22) for the 293 participants from the uncompleted groups (P=.07). Age and sex were not significantly different between the 293 participants from the completed group and the 293 participants from the uncompleted group. The 293 participants from the completed group were eligible for bivariate analysis and multivariate linear regression analysis.

In examining the linear relationship between each independent variable and the number of outpatient clinic visits (Table 2), we found that participants who rated themselves healthier were negatively associated with making more outpatient clinic visits (P<.003). In comparison, people with chronic diseases, such as diabetes, hypertension, or hyperlipidemia, tended to have made more outpatient clinic visits (P<.001). The collinearity check using the Spearman rank correlation coefficient for “What is your health status?” and “Do you have chronic diseases?” showed a correlation coefficient of –.18 (P=.002). We therefore excluded “What is your health status?” from further analysis.

After controlling for other confounding variables using multivariate linear regression for the 293 participants from the completed group, we found that Internet use was significantly associated with more outpatient clinic visits (P=.04) (Table 3). The adjusted R² was .1195, indicating that 11.95% of the variance could be accounted for by this multivariate linear regression model.