Although often overlooked, alcohol is a human carcinogen [1]. It causes 2.5 million deaths each year worldwide [2], and the United Nations has identified its harmful use as one of the four most important risk factors for noncommunicable diseases [3]. In Hong Kong, although alcohol consumption is still low, it is not uncommon. Alcohol is easily accessible by the general population due to a lack of regulation of minimum age for off-premises sales (with “premises” defined as restaurants and bars granted with a liquor license) [4,5]. Local studies have shown that almost one-third of adults and one-fourth of secondary school students drink alcohol [6]. The adverse health effects are daunting. According to the data released by the Department of Health in 2006-10, there was an annual average of more than 2000 episodes of in-patient discharges and deaths due to an alcohol-related disease [6].

Nonetheless, the Hong Kong Special Administrative Region (HKSAR) Government halved the duty on beer and wine with an alcoholic strength of not more than 30% in 2007 [7], and abolished it altogether in 2008 [8]. The duty rate on spirits with an alcoholic strength of more than 30% remained at 100% [8]. These unprecedented and anti-public health policies were aimed to help Hong Kong develop into an international wine trading hub, at the cost of increased alcohol-related harms to public health [9]. The Working Group on Alcohol and Health of the Department of Health noticed a surge in the alcohol consumption per capita in Hong Kong in 2008, which was attributed to the lowered price of beer and wine in the same year [4]. This echoed a meta-analysis of 112 studies, which found an inverse relation of alcohol tax or price with consumption [10]. In Finland, the one-third reduction of excise duties on alcoholic beverages in 2004 resulted in a clear rise in alcohol consumption and related harms [11], including hospitalization rate [12] and number of sudden deaths [13]. The drastic change of beer and wine tax policy in Hong Kong presents a unique opportunity to study alcohol drinking using infodemiology.

Infodemiology is a portmanteau of information and epidemiology, which is, according to Eysenbach, “the science of distribution and determinants of information in an electronic medium, specifically the Internet, or in a population, with the ultimate aim to inform public health and public policy” [14,15]. It is based on a bidirectional relation between population health status/attitudes/behavior and information patterns on the Internet [15,16]. Originally used to identify inaccurate health information on the Internet [17], it was later found that search engine query data could predict influenza epidemics [18]. Given its implication in public health and policy, infodemiology has since been used as a complement to traditional epidemiological studies [15,16], using analytical methods and metrics such as keyword prevalence and prevalence ratio [14]. The longitudinal tracking of infodemiology metrics for surveillance and trend analysis is called infoveillance [14,15].

Over the past decade, efforts have been made to overcome the difficulties in aggregating and analyzing the vast, unstructured information from the online database [15,16]. Examples of infodemiology within the public health sector include detection of disease outbreak or incidence by tracking search queries [16,18-20], investigating online search behavior for suicide-related information [21], monitoring public reactions towards health-related policy and campaigns [22,23], and identifying public health concerns from user posts in social networking sites such as Twitter [24]. Several computer tools have also been developed that allow more effective infodemiological analysis (eg, Technosocial Predictive Analytics [25], Infovigil [15], Global Public Health Intelligence Network, and HealthMap [26]). All these are underpinned by the technology of Web 2.0, which features individuation, open property, sociality, and microcontent [27]. One established example of Web 2.0 is online social networking.

Online social networking is constantly evolving. With proper mining and analytic technique [28], it might be possible to extract useful information from these networking sites for research purpose. Blogs (also called weblogs) are a kind of social networking website that is regarded as a relatively new form of mainstream personal communication [29,30]. They are characterized as being personalized, Web-based, community-supported, and automated [31]. Blog contents are versatile. They could be about the blogger’s life, commentaries, ideas, and emotions. They are also used to form and maintain community forums [32]. While some blogs are employed for political, educational, and commercial purposes [33], most are likened to diaries and are referred to as personal blogs [34]. People using blog services are referred to as bloggers.

Like other social networking sites, blogs are featured by their time stamps, consumer-generated content, and expansile database, making them potentially useful for longitudinal data retrieval and analysis [35,36]. In fact, blog analysis has become increasingly popular in various domains. Examples of its application include assessing a company’s image strength or customer product, monitoring public opinion in presidential elections, evaluating public reaction to disasters, tracing online hate groups or people with suicidal intent, studying youth cultures, and analyzing linguistic patterns [34,37-40]. The prerequisite of blog analysis for public surveillance is that a large proportion of the population should use the social Web services regularly, so that online information can be kept up-to-date and truly reflect the contemporary interest or concern of the community [25].

Internet use is ubiquitous in Hong Kong. In a survey by the HKSAR Government, Internet penetration of local population continued to rise over the years, from 56.9% in 2005 to 72.8% in 2012. Almost 70% of the online population were adolescents and adults aged 10-44 [41]. They also constituted the largest population engaging in online social networking activities including blogs and forums [42]. This was consistent with an older study by Blog-You.com and IN-Media in 2005, which found over 90% of local bloggers were aged 16-35 [43]. Infodemiology using blog searching data is therefore useful for studying important health issues among Hong Kong adolescents and adults, such as alcohol use.

There are different methods of blog analysis. Some (eg, time series scanning, semantic analysis) require special software and significant investment of time and computing resources [35]. This has greatly reduced its practicability in public health research done by clinicians who might not otherwise know much about computer programming. On the other hand, blog searching using blog search engines, which are freely available online, provides a technically easy, straightforward, and user-friendly method to extract data from blogs. Unlike Web search engines, blog search engines mainly index blog posts and are dedicated to searching information from blog posts only, ignoring the rest of the database [44]. Since each blog post is time-stamped, a blog search engine could search for date-specific blog posts, allowing longitudinal blog tracking, even retrospectively.

Local public health research using blog search and Chinese keywords is rarely done at the moment. Should blog searching data be correlated with local epidemiological data, clinicians (and policy makers) could readily replace traditional surveillance methods with blog searching—which is real-time, cheap, and fast—for public health tracking. Even if there is no correlation, an explanatory study like this would still contribute to the development of health informatics by demonstrating the challenges of Chinese blog searches in an otherwise English-dominated research environment. The clear rise of alcohol consumption in Hong Kong due to zero beer and wine tax provides a sound framework under which blog searching data can be validated against local epidemiology.

Using currently available search tools and Web resources, this study aimed to: (1) assess and explain the online use of alcohol-related Chinese keywords, and (2) validate blog searching as an infoveillance method to survey changes in the drinking patterns (eg, alcohol type) of Hong Kong people (represented by bloggers in a Hong Kong–based Chinese blog service provider) following changes to the beer and wine tax in 2007-8.

Following are the hypotheses of this study:

(H1) The online popularity of alcohol-related concepts, in particular “beer or wine”, increased among Chinese bloggers after 2007-8 when local tax policy on beer or wine changed.

(H2) Infodemiological data (represented by prevalence rate and ratio of alcohol-related concepts) correlated significantly with local epidemiological data (represented by per capita alcohol consumption).

To the best of our knowledge, this was the first research done in Hong Kong using blog searches to study a public health–related topic. The issue of alcohol drinking was chosen because of its public health interest and clear impacts by tax policy changes. Blogs were targeted for data extraction because regional interest could be maximized by choosing a blog service provider with the highest local rank and visitors, unlike other social networking sites such as Twitter or Facebook, which tend to cover a wide geographical area.

There were two main sets of data in this study: (1) infodemiology and (2) epidemiology. Infodemiological data stemmed from existing blogs indexed by a specific search engine, whereas epidemiological data were obtained from government documents covering the issues of public health. To reduce expertise and technology investment, this study used a free Web-based blog search engine, Google Blog Search, to extract data from the archives of a Hong Kong–based blog service provider, MySinaBlog, from 2005-2010. Three groups of Chinese keywords were used, each representing a specific alcohol-related concept. They were (1) “alcohol” (ie, the control concept), (2) “beer or wine”, and (3) “spirit”. The resulting blog posts were analyzed quantitatively using infodemiological metrics and correlation coefficients, and qualitatively by manual effort. The infodemiological metrics were (1) apparent prevalence, (2) actual prevalence, (3) prevalence rate, and (4) prevalence ratio. Pearson and Spearman correlations were calculated for prevalence rates and ratios with respect to per capita alcohol consumption in the same years. Manual analysis included (1) blog author characteristics (ie, authorship, sex, and age), and (2) blog content (ie, frequency of keywords, description of a discrete episode of alcohol drinking, drinking amount, and genres). The prevalence rate and ratio were used to assess the online popularity of alcohol-related concepts, whereas the correlation analysis and manual analysis were used to validate blog searching data as an infoveillance method for population survey.

Table 2 shows the per capita alcohol consumption extracted from a report released by the Department of Health of the HKSAR Government in 2011 [4]. It was adopted in our study because it was (1) freely accessible, (2) presented in a longitudinal form, and (3) subgrouped according to alcohol types. Data from 2011 were not available, and no updates of the data were seen hitherto.

Once blog searching was done, the blog posts were saved in .html files for subsequent analysis to avoid discrepancy due to time lag. The manual analysis focused on (1) blog author characteristics (ie, authorship, sex, and age), and (2) blog content (ie, frequency of keywords, description of a discrete episode of alcohol drinking, drinking amount, and genres). They would provide further information about the validity of utilizing blog searching data in an epidemiological survey. Their subcategories and criteria are listed in Table 4.

Blog posts with “alcohol” keywords were not included for manual analysis since a large portion of them were expected to overlap with those that contained “beer or wine” and “spirit” keywords. They might not be particularly helpful in analyzing the drinking pattern (eg, choice of alcohol) of the population.

It was noteworthy that most of the free text analytic tools did not support Chinese language and had no way to identify position of the keywords within a blog (eg, header, main body, sidebar, footer, and comment). Currently available concordancers for Chinese language were not too user-friendly as they lacked an external encoder/decoder, keyword-in-context (KWIC) format, or built-in dictionaries for semantic analysis or opinion mining [52]. This was why manual analysis was chosen in this study as a preliminary measure to explore the blog author characteristics and blog content.

The blog search was done on April 12, 2013, and the manual analysis was completed by researcher, KL Chan, in the subsequent week. The results are described below.

The online popularity of alcohol-related concepts was best represented by their prevalence rate and ratio, which normalized the effect of any changes in the total number of blogs [14]. In general, the concept “alcohol” was most popular in 2008. The concept “spirit” also experienced a short-lasting and somewhat erratic rise in its online popularity in 2008-2009. The concept “beer or wine”, in contrast, became increasingly unwelcome with an overall declining trend in its online popularity over the years. The hypothesis that alcohol-related concepts became more popular after 2007-2008 was true only for “alcohol” and “spirit” but not “beer or wine”.

One possible reason for the increase in the online popularity of “spirit” keywords in 2008-2009 was the presence of spam in the same years. This was masked in the apparent prevalence, which was used to calculate the prevalence rate and ratio. Indeed, after excluding the data in 2008-2009, the prevalence rate and ratio of the concept “spirit” remained relatively stable at a low level. The false elevation in the online popularity of “spirit” might also explain the peak prevalence rate of “alcohol” in 2008, although the latter appeared much larger in amplitude and there was still a chance for a genuine rise in keywords which were not included in “beer or wine” or “spirit” but “alcohol”. The downhill course in the online popularity of “beer or wine” was probably a true reflection of bloggers’ decreased interest over the topic. However, its relation to local drinking pattern remained doubtful, since many of the blog posts were in fact not describing a discrete episode of alcohol drinking. The same argument held true for the concepts “spirit” and “alcohol”.

Validation of blog searching data depended on correlation analysis and manual analysis of blog author characteristics and contents. The prevalence rate of “alcohol” was the only parameter that had a significant nonlinear and a marginally significant linear correlation with per capita alcohol consumption. The other correlations were all nonsignificant, although many of them demonstrated moderate to strong strengths. The hypothesis that infodemiological data correlated significantly with local epidemiological data was true only for “alcohol” prevalence rate. The statistical nonsignificance of other infodemiological metrics might be explained by the small number of blog posts relative to the population. This in turn could be attributed to the following:

From the manual analysis, most blogs were managed by single authors, meaning that the number of blogs could be used to represent the number of individual attendants in a population survey. The sex distribution of the single blog authors was close to the local population, but their age range was slightly inclined to 18 years old and above [54]. It should be noted, however, that many of the bloggers did not disclose their identity online, making validation difficult.

Many of the blog posts were not about a discrete episode of alcohol drinking but personal diary, opinion, or emotion outlet. This was not surprising as new genres of blogs continued to emerge [55]. Rather than just being informative, many blogs were affective in nature requiring semantic analysis for meaningful interpretation [25]. While it was unlikely that bloggers recorded their alcohol intake on each occasion, they might reveal their understanding on alcohol drinking when commenting on a particular event, answering a particular question, and describing a childhood incident, etc, hence, the presence of alcohol-related keywords. While there were inadequate clues to support that changes in the online popularity of the alcohol-related keywords were related to an altered drinking pattern of the local population, one should not ignore its social implications and disregard its role in evaluating public reactions towards health-related policy including the zero beer and wine tax.

Using blogs as the source of information had several inherent limitations. For example, demographic data of individual blogger such as gender, age, and race might be deficient or disguised; bloggers tended to share common interests and backgrounds that were probably different from those of the general population; and acquisition of precise data such as drinking patterns was often difficult. In order to construct a larger framework in a timely and efficient manner, informatics researchers often had to compromise the individuality of each blogger by using certain infodemiological metrics. Moreover, language usage by bloggers tended to be complex and not easily decoded by the frequency of some pre-determined keywords. As a solution, semantic analysis of individual blog post might be useful to explore bloggers’ opinions towards drinking, provided enough technical support. Face-to-face interviews and questionnaires might be conducted with individual bloggers to elaborate their viewpoints, preferably those who appeared to have the largest influence within a specific blog circle (using a social network analytic tool).

No single blog search engine indexed all blogs [56]. Despite its automaticity, a search engine might be subject to editorial choice and hence bias [57]. There were concerns that even in the same search engine, the search results may be different over time [14,44]. In our case, it might be explained by the (1) inherent limitation in the search algorithms of Google, which gave only an approximate estimate for query with large results, and (2) inconsistency of the search database due to a variable number of splogs (or spam blogs) and blogs that were previously not linked [35,44,56]. Of note, a large part of the Google Search algorithm was unknown to the public, aggravating sampling uncertainty in our study.

One challenge with the use of Chinese language in blog searching was that it tended to have a wide range of expressions owing to geographical difference and translation from English. Also, only a limited number of blog analytic tools supported the Chinese language. A self-designed research program with well-informed blog search algorithms and analytic functions especially for Chinese blogs would be most desirable, which would depend heavily on the availability of expertise and resources.

Using blog searching data from a Hong Kong–based Chinese blog service provider, we concluded the following: (1) the online popularity of alcohol-related Chinese keywords was attributed to many different factors including spam, and hence not a specific reflection of local drinking patterns, (2) correlation between infodemiological data (represented by prevalence rates and ratios of alcohol-related concepts) and epidemiological data (represented by per capita alcohol consumption) was poor, and (3) many blog posts were affective rather than informative in nature. While blog searches using pre-defined Chinese keywords might not be an ideal method to survey epidemiological data such as alcohol consumption, semantic analysis of blog content would provide invaluable information on public reactions towards health-related policy, given enough expertise and resources.