Data cleaning and processing were performed using the R software [33]. Noninformative responses (eg, N/A and no comment) were removed from the data set. Survey responses that consisted of only a single word were removed from the analysis data set, as the underlying mechanisms of LDA are based on the co-occurrence of terms. These are categorized separately, as attitudes are easily ascertained from single-word responses (shown in Tables 1 and 2).

Preprocessing procedures, standard in natural language processing, were undertaken to create a document term matrix (DTM) on which to perform LDA. This included converting all words to lower case and removing white spaces, punctuations, and common stop-words from texts to leave only meaningful words. Frequent misspellings were replaced with their correct form, and common equivalent words were standardized (Multimedia Appendix 2).

The SMART (System for the Mechanical Analysis and Retrieval of Text) stop-word data set was used to identify uninformative words. However, bigrams (two-word terms) that contain many common stop-words were not removed to uncover attitude positions within responses (Multimedia Appendix 2). Numbers were not removed as numeric bigrams are potentially meaningful (eg, 100 percent and 3rd party). Finally, lemmatization of words was undertaken to convert words into their root form (lemma) and reduce sparsity within the DTM (eg, cards was replaced with card; Multimedia Appendix 2). Lemmatization was preferred over stemming, as lemmas are more human-readable than stems, which are not always complete words (eg, storing would be replaced with the stem stor).

LDA was performed on a DTM of unigrams (one-word terms) and bigrams (two-word terms) within the texts. Bigrams are included in the DTM as this creates more human-interpretable topics, and many words are context-specific, for example, big brother/change mind, or formulate attitude positions in combination, for example, would change/wouldn’t change. In the example given, the term brother alone would provide little insight into data sharing attitudes, yet big brother signifies a potential invasion of privacy and distrust. Several papers have reported improved results for including bigrams and higher n-grams in different topic models [34-36]. Only terms that occurred in more than 2 documents were retained, leaving 993 unique terms for the store loyalty card question and 554 unique terms for the health and fitness app question.

Selecting an appropriate number of topics is a key challenge for LDA. According to Green et al [37], “too few topics will produce results that are overly broad, whereas choosing too many will result in the ‘over-clustering’ of a corpus into many small, highly-similar topics” that are difficult to interpret in a meaningful way. The number of topics (k) is conventionally chosen as the model with the lowest value of perplexity when applying different models of candidate k to held-out data [30,38]. This perplexity measure captures how well a probability distribution or probability model predicts a sample, indicating how surprised the model is by new data. For LDA, it is equivalent to the inverse of the geometric mean per-word likelihood calculated on the held-out data [30]. For this analysis, ten-fold crossvalidation was used to select an appropriate number of topics (k) from a candidate list of 15 k ranging from 2 to 100, optimizing for perplexity. Candidate k increases in smaller intervals between lower values, as greater change is expected between these values. The crossvalidation process randomly divides the data set into 10 approximately equally sized folds and uses 9 of these to train the model, using the held-out fold to test the model. This process was repeated 10 times such that each fold was used as the testing set once.

Some research has found that the models that produce the most semantically meaningful topics—in that topics are easily interpreted by humans and terms representing concepts are given high probabilities within the model—are not necessarily the models with the best perplexity scores [39]. Hence, a measurement of coherence, which research finds corresponds well with human-interpretable topics [40], was also considered. Average topic probabilistic coherence measures topic quality based on how commonly topic terms co-occur, controlling for statistical independence [41-43]. This was compared for models of candidate k topics and balanced with perplexity scores.

Once an appropriate number of topics were selected for each LDA model, the final models were created using responses from all individuals for the given question. This method was chosen, rather than using commonly used training and testing approaches, to support the study aim of summarizing survey responses rather than creating a predictive model to categorize new data, as no new LifeInfo Survey data will be collected in the future. Moreover, the size of the data set is small compared with many others that use LDA, and splitting the data set into fewer responses could reduce the model quality.

For both topic number selection and the final models, the LDA hyperparameters were set at α=.1, influencing document-topic density, and β=.05, influencing word-topic density. The α prior was set at this relatively low value because LifeInfo Survey responses were short (refer to Multimedia Appendix 3 for plots showing the distribution of response lengths in words, averaging 12 words per document for the store loyalty card question and 9 words for the health and fitness app question), so we would expect there to be only a few topics formulating each document. β was also set to a relatively low value, as we expected a small number of words to be highly influential per topic given the short responses. α is modeled as asymmetrical, as we expected some topics to be more common than others within the survey responses. Previous work has found that asymmetrical α values provide substantial advantages to LDA results, whereas asymmetrical β priors provide no benefit [44].

To make the topics more easily interpretable, those created through LDA modeling were further categorized into thematic groups with the aid of hierarchical clustering of topics (Multimedia Appendix 4). These topics were given summarizing names, shown in Tables 3 and 4 by subheadings in bold, based on their content, considering both topic top terms and the contextual use of these terms in texts (Multimedia Appendix 5). The Hellinger distance [45] between topics was calculated based on their term ϕ values, and the 2 closest topics were clustered together. ϕ values are the probability of a term being used within a text given; therefore, topics that more frequently use the same terms are considered closer. This was done iteratively until hierarchical clusters split topics into meaningful thematic categories. In some cases, topics thematically aligned with a category that was found to be semantically different to the topic according to hierarchical clustering; these are marked with a superscript in the tables and dendrogram and were reassigned to their appropriate thematic category.

The recommended actions that researchers can take to address the key barriers to data sharing and linkage identified through LDA modeling are presented. These recommendations are based on the synthesis of participant suggestions and expertise regarding wider research on data sharing.

For the store loyalty card question, perplexity scores for each model of candidate number of topics (k) indicated that the best model has a number of topics within the range of k=20 to k=60, as the perplexity scores plateau at their minimum value within this range (Figure 1). For the health and fitness app question, perplexity scores were minimized to within the topic number range of 20-30 topics (Figure 2). The average probabilistic coherence scores within these ranges of k varied by only very small amounts, indicating that the models had near-equivalent topic quality (Figures 3 and 4). To maximize human interpretation and for comparability between the 2 questions, parsimonious models of 20 topics were chosen for both the store loyalty card question and health and fitness app question.

Table 3 shows the 20 topics created by the final LDA model of free-text responses to the store loyalty card question. The table includes the 15 top terms most representative of each topic, selected according to their ϕ values. As ϕ values are the probability of a term being used within a text given the topic, those terms with the highest ϕ values are representative of the topic content. The table includes 2 further topic measures: first, topic prevalence, which indicates how common each topic was within the survey responses (eg, prevalence 5.0 indicates 5% of survey responses fell into this topic), and second, probabilistic coherence, as an indicator of the semantic meaningfulness of each topic (this can range from 0 to 1, where higher numbers are more meaningful). Lower values of probabilistic coherence are owing to lower frequencies of topic term co-occurrences within texts and/or topic terms being highly frequent within the data set at large. Low coherence scores are associated with conceptually less defined topics [40]; however, they could also be caused by fuzzy topics with many different terms or semantic crossover within other topics. Table 4 shows the same information for the 20 topics created by LDA modeling of the health and fitness app question responses.

LDA modeling also assigns survey responses with probabilities for each topic, known as θ values A higher θ value for any given topic indicates that the response should more likely be categorized into that topic (Multimedia Appendix 5 shows example responses most associated with each topic). Some responses, especially those that are very short (consisting of only 1 substantive term), are given equal θ values across all topics, which can be regarded as uncategorizable (Multimedia Appendix 6). There are some differences in LDA topic frequency depending on demographic groups (eg, age and gender); further analysis demonstrating demographic patterns can be found in Multimedia Appendix 7. Topic stability is validated by comparing results across multiple LDA runs (Multimedia Appendix 8), indicating that the thematic categories reported are consistently produced in topics created through LDA.

Among those who responded negatively to sharing their lifestyle data for health research, a large proportion would be unwilling to change their mind. This is indicated first by the number of texts that LDA modeling categorized as the topic nothing would change mind, which refers to SC14 and HA17, constituting approximately 6.94% (134/1930) and 6.30% (76/1206) of the analyzed responses, respectively, and second, by the single-word answers excluded from the analysis, which were mainly the words nothing, no, and none for both the store loyalty card question and the health and fitness app question (Tables 1 and 2). These 3 single-word answers combined account for 90.3% (361/396) of the single-word responses about store loyalty cards and 85.8% (265/309) of single-word responses about health and fitness apps.

Others mentioned that they did not use these services and thus would not have data to share. This was most clearly identified within the topics related to health and fitness apps (HA16 and HA11) and was evident for store loyalty cards (SC16). For example:

For this group, only greater participation in big lifestyle data production would allow them to share their data.

Some respondents reported feeling that the proposed data linkage was a big brother and an invasion of privacy (SC8, SC1, HA2, and HA3). However, for the fitness app question, these topics were less clearly defined by LDA modeling, reflected in their low probabilistic coherence scores for these topics (0.05 and 0.06). Answers such as “not interested to a ‘big brother is watching’ on all aspects of my life” (SC8) and “[I] wouldn’t want to feel every area of my life is out of my control and being watched by an institution that already makes me feel like I have no autonomy” (HA3) indicate a dislike and feeling of surveillance through the data. Actions by researchers to address these feelings are limited as they are related to a broader distrust of big data; however, greater transparency as to how data are being used and assurance of anonymity may convince some users. Within the responses about store loyalty cards, SC1 specifically identified privacy concerns related to unwanted emails, phone calls, and text that can be addressed through data protection and security actions.

Many responses focused on concerns about sharing these data, which respondents perceived as highly personal information (HA20, HA14, SC19, and SC3). References to data being personal and confidential appeared across topic categories, particularly in the context of not wanting to store loyalty card data and health records linked, data protection and security, and concerns about who is able to access these data. For example:

The degree to which an individual believes that their data are personal and sensitive influences their willingness to share data and with whom. Medical information is personal, particularly for individuals with complex health conditions. Health and fitness app data and store loyalty card data are personal in different ways. Fitness apps are primarily used for personal monitoring, meaning data are not created with the idea that they might be shared with other actors, whereas for store loyalty card data, individuals exchange information about purchase history with shops in exchange for discounts and points. However, in the case of transaction data, Skatova et al [13] found that people regarded the graduality of transactional data as personal.

Health data are regarded as particularly sensitive and confidential [26], related to the theme of private information, and many responses mentioned worries that data would be accessed by other actors without their permission (SC2, HA15, and HA7). For example, “if the information was available only to the research team, and not to others, e.g. insurance companies, mortgage companies, even the medical team” (HA15).

Private companies, third parties, and health insurance companies were frequently mentioned by respondents as actors they feared would gain access to their data. This is expected, as research has found that these institutions are trusted least by the public to use data appropriately [24,27]. Many respondents believed these companies would use their data for profit, to increase the cost of premiums, or to deny treatment altogether. However, respondents also mentioned concerns that health care professionals would be able to access their lifestyle data. For example:

This finding is less expected, as research finds that health care providers are one of the most trusted actors for data sharing [24,27]. This indicates that there is no straightforward relationship between an individual’s willingness to share their data and their trust in the actors involved. A further barrier for data linkage is illuminated in that data are often created for specific purposes, and alternative uses of data outside of this domain can create suspicion.

There was also a common misunderstanding that sharing lifestyle data for research would allow all involved actors to access these data. This was reflected in the previous example and others. For example, one respondent asked, “Why would I want Tesco knowing my health records?” (SC17).

Similar findings were reported by Skatova et al [13] in their research on transaction data sharing. This indicates that one easily achievable action that could influence people to share their data would be to explicitly state that data would not be available to anyone but researchers.

Closely related to data access by others was a reported general belief that health records should not be linked with other data and should be accessed only by health professionals (SC17 and SC11). This again reflects the belief that data should only be used for its designated purpose. For example, one respondent felt that “Health records should be kept in health care” (SC17).

Data accuracy was another identified barrier. Many respondents indicated that their purchase history or fitness tracking was only partial, creating misleading data about lifestyle behaviors (SC7 and HA5). Indeed, missingness and data integrity have been identified as a challenge for research using big lifestyle data and a concern for participants [10,26]. Nevertheless, good results have been found by comparing or combining these data with more traditional collection forms, for example, modeling individual consumption from household-level data [10,46].

When looking at responses within this topic, most participants reported concerns that their data would make them appear less healthy than their real activity. For example:

Implicitly, respondents worried that they would be judged unfavorably on their lifestyle behaviors. Researchers could address these concerns by making explicit in the study protocol that (1) data are not expected to be complete, (2) detailed actions they will take to accommodate for this with modeling techniques or additional surveying, and (3) all data would be made unidentifiable foreclosing the possibility of judgment.

Topics that focused on data protection and security formulated a large proportion of the responses (approximately 17.31% (334/1930) of responses for the store loyalty card question and 19.90% (240/1206) for the health and fitness app question). Changes frequently mentioned were assurances that these data would be stored completely safely and securely and that they would be protected from hacks and data breaches. These are common concerns for data sharing, especially with data that are regarded particularly private or sensitive, such as health data [26]. Although these assurances were given as a condition for data sharing within the question wording (Multimedia Appendix 1), many responses highlighted their importance or were not convinced that this would happen. Action would, therefore, need to be taken to give participants greater confidence, which could be achieved through transparency in the research process and providing details of how these data will be protected.

Across both store loyalty cards and health and fitness app data sharing, respondents mentioned that understanding the research better and being given more control over their participation would influence them to change their mind. This includes being provided with more information (SC18 and HA8), giving permission only for specific research projects (SC4 and HA4), and a greater understanding of the reason or benefits of research (SC20, SC4, SC6, HA9, and HA1). These findings are supported by the research of Skatova et al [13], who found that support for data sharing is contingent on its context and purpose, highlighting the importance of well-informed participants. Given that the LifeInfo Survey aimed to assess attitudes toward data sharing in the future, rather than requesting participants to consent to data sharing at this time, the participant information sheet (Multimedia Appendix 9) needed to be suitably broad, which would not be the case when recruiting participants to an actual data linkage study.

Researchers have found that a strong social motive, such as improvements to health or treatment, in addition to clearly defining the purpose of research, are key motivating factors for personal data donation [47], whereas using health data for insurance, marketing purposes, or commercial exploitation is unacceptable to the public [23]. This was reflected in the unstructured answers from the LifeInfo Survey. A respondent answered that they would be supportive “if the data fed into important public health or similar research and was not used to further commercial gains by these giants of commerce” (HA4).

Again, although LDA separates topics, they are connected, and the involvement of private companies creates concerns about whether these entities will be able to access health data for profit once linked.

Control and consent for health data being used for research have been found to be key for public acceptance [25]. A respondent answered that the data “has to be linked to a condition or specific research project with additional approval provided in advance” (HA2).

This is something that can be adapted into the research process, allowing participants to consent to or deny the use of their data, given the specifications of the study.

In addition to a desire for more understanding, another thematic category was identified of individuals who did not understand the reason or relevance of data linkage (SC13 and SC10). For example, one respondent said, “Do not see any reason why they should be linked” (SC10).

These topics were only produced when modeling the store loyalty card question, although topic HA5 also encompassed responses that stated data are not relevant. This perhaps indicates that respondents found the link between purchasing and health to be less relevant than fitness tracking and health.