The LoCard data used in this study were obtained from S Group, which is the largest commercial operator of retail grocery stores in Finland. According to S Group, their full coverage is 2.4 million households, meaning that 88% of households in Finland have registered purchases in their databases. The members of S Group’s loyalty card program are provided with an electronic customer card to be used when making purchases, and customers are rewarded for their purchases by getting a maximum 5% financial bonus that is refunded to them on a monthly basis. Individuals of the same household may link their purchases to the same loyalty account. In this study, only purchases of the household’s main cardholder were used.

Members of S Group’s loyalty card program (primary cardholders) across Finland were contacted via email and were invited to take part in the study, which involved consenting to the release of their grocery purchase data to be used for research purposes and voluntarily responding to the study questionnaire. Members who did not have an email address declared or who had prohibited the retailer from contact them with any marketing or research-related material were excluded. Cardholders under 18 years of age were also excluded. All invitations were sent by S Group as they had customers’ contact information.

The grocery purchase data used in this study covered the period from January 1, 2017, to December 31, 2018. Each purchase was associated with item description, time stamp, quantity (ie, weight, volume, or number of packages), and expenditure on the item.

All consenting participants were asked to fill out a web-based background questionnaire that included the following sociodemographic variables: education, marital status, size of the household, number and age of children, occupational status, income, and perceived health. The background data were complemented with information on participant sex, age, and postal code obtained from the retailer’s electronic database.

As part of the baseline questionnaire, all participants were asked to estimate their degree of loyalty as a share of purchases made in the retailer’s shops and supermarkets on a five-item ordinal scale. The response categories were as follows: “0%-20%,” “21%-40%,” “41%-60%,” “61%-80%,” and “81%-100%.”

The LoCard grocery purchase data required preprocessing to be usable in further analyses. First, we identified food groups from all the grocery product groups. Second, we regrouped the identified food groups into new groups that were formed on the basis of the commonly used food groupings in nutritional studies [31] and earlier findings on the associations between dietary components and health [32,33]. For instance, skimmed liquid milk and buttermilk were aggregated into “skimmed milk & sour milk” and foods and mixed dishes with red or processed meat as the main ingredient were aggregated into “red meat & processed meat.”

Out of 4234 grocery product groups, 865 (20.4%) were assigned into one of the new food groups used in this study. In addition, 42 food groups were left out as they involved either (1) a mixed dish or food group with no definite primary ingredient or (2) a rarely purchased product. The food groups used in this study included “vegetables,” “skimmed milk & sour milk,” “sugar-sweetened beverages,” “rye bread,” “red meat & processed meat,” “fat spreads,” and “sweets & chocolate.” These groups were used as indicators for evaluating the nutritional quality of household food purchases. A detailed description about the grouping of the food purchase data is included in Multimedia Appendix 1.

Population statistics on the general adult population were obtained from Statistics Finland using StatFin databases that can be freely accessed [34]. The databases include tabulated data on Finnish citizens and Finland in general that are collected on a yearly basis. Data from 2017 were used because of the availability of the latest data tables for all sociodemographic variables used in the analyses. For this study, individuals aged at least 18 years were included.

The FinHealth survey is a national population health study on Finnish citizens. The study encompasses a series of cross-sectional population surveys carried out every 5 years in Finland. The latest FinHealth survey was carried out at 50 localities in 2017, with a participation rate of 71% among those invited for the study [35]. The purpose of the FinHealth study is to collect up-to-date information about the health and well-being of adults residing in Finland and on the factors influencing their health and well-being. Each survey invites 10,000 randomly selected individuals aged over 18 years. The study consists of physical examinations and study questionnaires. The latest report (values used in this study) is restricted to adults aged 30 years or older to make the results comparable with earlier FinHealth studies. A subgroup of the participants was also invited to undergo a nutrition review; the FinDiet survey is a substudy (n=1655) of the FinHealth survey, which monitors the nutrition and dietary habits of the Finnish population [36].

The study was approved by the University of Helsinki Review Board in the Humanities and Social and Behavioral Sciences (Statement 21/2018). Informed consent was electronically obtained from all participants included in the study when they were invited via email to release their loyalty card data and fill out the background questionnaire. The data were pseudonymized by S Group before the researchers could obtain the data.

S Group had approximately 2.4 million primary loyalty card owners, and all of them were assessed for eligibility (Figure 1). Approximately half (1,214,663, 51%) of the loyalty card owners were contacted, and of these, 47,066 (4%) consented to participate. We did not have information on the number of valid email addresses or what proportion of emails reached the card owners (eg, by passing through trash email filters). Among the participants, 36,621 (78%) responded to the background questionnaire. Nearly all participants (46,825, 99.5%) purchased at least one grocery item from 2017 to 2018.

Table 1 shows the participant characteristics compared with those of the Finnish adult population and the FinHealth study participants. Discrepancies were found in sex, age, education, and occupational status when compared with the general Finnish adult population. Namely, there were more women, more individuals with a higher education, and more employed individuals in the LoCard sample. On the contrary, individuals aged under 30 years and over 70 years (correspondingly, retired individuals) were underrepresented in the LoCard sample. Selectivity associated with education was strong in the LoCard sample. The proportion of individuals having a basic education level was clearly lower in the LoCard sample (6% of participants had basic education) than in the Finnish adult population (25% had basic education). There were no major differences in the distribution of marital status. However, there were fewer individuals living in a household with children aged under 18 years in the LoCard sample. The LoCard sample was widely distributed across Finland and comparable to the geographical distribution of Finnish citizens (Multimedia Appendix 2 and Multimedia Appendix 3).

On comparing the LoCard sample to the FinHealth study participants, there were differences in sex, education, and marital status distributions, with more women and individuals with higher education and fewer married individuals in the LoCard sample. The age distributions were not comparable owing to the fact that the FinHealth study included only individuals aged at least 30 years. Distortion in the distribution of occupational status was similar in the two studies compared with the Finnish adult population.

The reweighted distributions of the sociodemographic variables demonstrated that the constructed poststratification weights corrected the deviations successfully, and thereafter, the sociodemographic distributions of the LoCard sample matched well with the Finnish adult population.

Table 2 shows the purchases of selected food groups in the original LoCard sample and in the weighted LoCard sample. Over 95% of the participants had purchased at least one food product in all food groups, except skimmed milk & sour milk. Skimmed milk & sour milk had been purchased by 74% of the participants. Among them, the median expenditure and the median weight were €23.0 (€1=US $1.13 in 2017) and 23.5 kg, respectively, during the 2-year follow-up.

After applying the poststratification weights, there was an increase in the purchase of red meat & processed meat and small increases in sugar-sweetened beverages and fat spreads. The purchase of vegetables and sweets & chocolate decreased as a result of reweighting. The largest change was seen in red meat & processed meat; the weighted amount of purchase increased from €387 to €417 (cost) and from 48 kg to 54 kg (weight), corresponding to relative percentage increases of 7.8% and 12.6%, respectively.

Table 1 shows the self-assessed degree of loyalty. Almost 40% (14,194/36,526) of the participants reported that they made 80% or more of their food purchases at S Group shops and supermarkets, and 64% (23,418/36,526) reported making at least 60% of their purchases at the retailer’s shops and supermarkets.

The RFM scores were significantly different among the five degree of loyalty groups, with the lowest scores in the lowest degree of loyalty group and a steady increasing trend toward the highest degree of loyalty group (F₄=4625.5, P<.001). The poststratification weights also differed significantly across the five groups (F₄=24.1, P<.001), indicating that the degree of loyalty was associated with individuals’ sociodemographic characteristics. However, the observed differences were rather small, with a maximum difference of six percentage points between the groups (Table 3). In the highest degree of loyalty group, there were slightly more young and married participants, and the percentage of households with children was higher, whereas the percentage of divorced or separated participants and those with a master’s degree declined with the degree of loyalty.

Table 4 shows food purchases in the degree of loyalty groups, and all showed significant associations (P<.001 for all food groups, except sweets & chocolate [P=.007]). The result was expected owing to the large sample size. The shares of vegetable, red meat & processed meat, and fat spread purchases increased as the degree of loyalty increased. In the other food groups, there were no major differences across the degree of loyalty groups.

Additionally, Table 4 shows that the quantity and expenditure regarding food groups increased steadily with the self-assessed degree of loyalty, suggesting that the self-assessment can be relied upon.

The findings of this study showed that individuals who consented to the release of their loyalty card data for research purposes tended to diverge from the general Finnish adult population. Similar to many other health and nutrition studies, including those encompassing electronic data collection tools [7,13,35,40,41], the LoCard participants manifested volunteer bias, with employed individuals, middle-aged individuals, women, and individuals with higher education being overrepresented in the sample. The LoCard sample included fewer retired individuals, fewer individuals with basic education, and fewer individuals who had children aged under 18 years living in the household. Compared with the Finnish national FinHealth and FinDiet studies, the selection mechanism appeared to be somewhat different in the LoCard sample. While employed individuals were overrepresented in all these three studies, the gender and education biases were stronger in the LoCard sample. Moreover, the LoCard sample had a rather similar distribution of marital status as among Finnish adults, whereas in the FinHealth study, married individuals were overrepresented [35,36,42].

However, the size (n=47,066) and heterogeneity of the LoCard sample enabled a successful correction of the differences seen in the sociodemographic variables. We developed the poststratification weights using all sociodemographic background variables available with the two-way joint distributions to correct the background distributions of the LoCard participants to make them closer to the Finnish adult population. The large sample size provided a sufficient number of participants for hard-to-reach population subgroups, and thus, it was possible to construct the poststratification weights for them as well. The highest weights were seen for unmarried men aged under 30 years, who indeed are often underrepresented or not enrolled in health studies [41].

Of the 1.1 million loyalty card holders contacted, approximately 4% (n=47,066) took part in the LoCard study. Although low, the participation rate was similar to that for other massive data collection methods [7]. The advantage of the use of digital tools is that they reach a large number of potential study participants with relatively low effort in data collection. After all, we reached substantially more individuals than in the majority of dietary studies using traditional data collection methods with minimum human involvement in data collection. A likely reason for the low participation rate was that the participants were contacted via email, which may not have reached them (invalid email address or contact email classified as “junk email”) or may have limited their participation and induced selection bias. Although 88% of households in Finland have an internet connection and 83% use email [43], email use varies according to sociodemographic profiles and is relatively low at 62% among individuals aged over 65 years and among individuals with basic education [44]. This may partly explain the baseline characteristics of the LoCard sample. However, it has been shown that the use of digital tools in recruitment and data collection does not increase the selection bias, but the traits of participants in health studies are rather similar regardless of the recruitment method used [13,40]. Moreover, it is likely that many simply ignored an email coming from a commercial party.

Important aspects are whether and when informed consent from loyalty card owners is needed. Recently, Aiello et al [45] published an interesting ecological study on the associations between loyalty card food purchase data and prescription records that were used as a proxy for real disease profiles in London. Their dataset included 1.6 million loyalty card users, and they used the anonymized data without the consent of the individuals. In our study, consent and a positive reply were required for two reasons. First, transparent use of loyalty card data on customers for a common good builds trust among them, researchers, and the company, and reduces the likelihood of negative publicity. Second, contact was needed to obtain information about participants’ background characteristics for use in further analyses. A future scenario could involve a consent request when the customer becomes a member of the loyalty card program. This would create an ethically sound and transparent research protocol for the use of customer data.

Poststratification weights were further applied in evaluating the purchases of the main food groups. The corrections demonstrated small changes in some food groups; the purchase of vegetables and sweets & chocolate decreased after the correction, whereas the purchase of red meat & processed meat, sugar-sweetened beverages, and fat spreads increased. The sociodemographic profiles of the LoCard participants and bias related to them might, at least partly, explain these results. The FinDiet study showed that women, who were overrepresented in the LoCard sample and thus had smaller weights, tended to consume more fruits and vegetables than men [36]. In line with this, after applying the poststratification weights, the purchase of vegetables decreased. It has also been shown that socioeconomically advantaged individuals, who likewise were overrepresented in the LoCard sample, consumed healthy foods, such as fruits and vegetables and low-fat dairy products, more frequently [46]. Moreover, the increased amount of red meat & processed meat purchase is likely related to male participants who tend to consume more meat [36,47]. In line with the overall findings of this study, the NutriNet-Santé study showed that the consumption of fruits and vegetables was higher and the consumption of meat was lower in the cohort than in the general population in France [13].

The degree of loyalty was fairly high in the LoCard sample with 64% (23,418/36,526) of the participants reporting making over 60% of their grocery purchases at the retailer’s shops and supermarkets. The food purchases were rather similar in the higher loyalty group (60% or higher), whereas individuals making less purchases in the retailer’s grocery stores showed some differences. In particular, individuals reporting the lowest degree of loyalty tended to buy fewer vegetables and fat spreads and fewer red meat & processed meat products. Although some variation was seen, the differences across the loyalty groups were smaller than expected. One reason could be the food groups selected for the current analyses. There could be other products, such as alcohol and tobacco, that are differently purchased. These results, together with the differences seen in the sociodemographic variables between the loyalty groups, underline the importance of estimating and accounting for the degree of loyalty in future studies using loyalty card data. A direct way to address the problem of coincidental purchases is to focus on a subsample with at least 60% loyalty. It is also important to note that loyalty card data can enable research on longitudinal trends in food purchases, which can be performed regardless of the degree of loyalty.

Although we used a large set of matching variables for developing the poststratification weights, some limitations concerning these remain. First, we were not able to compare or account for possible differences in income, as there was no comparable reference data available in Statistics Finland. Therefore, it remains unclear whether the LoCard sample was representative in terms of income, which is an important contributor to food purchase. The higher education level of the LoCard participants and the lower prevalence of young and retired individuals clearly suggest that the income levels might be overestimated in our sample. Second, although we matched families with children, the number of children and their ages, which can clearly affect a household’s food purchases, were not used in weighting. Importantly, we were able to correct the differences only in the observed sociodemographic variables, and thus, unidentifiable selection bias cannot be ruled out. This may include factors that would be associated with willingness to participate, such as special dietary restrictions and socially excluded people. In particular, among those participants who did not have complete background information and whose poststratification weights were thus based on sex and age only, the risk for unidentifiable selection bias could be even higher.

It is important to note that grocery purchases reflect consumption on a household level, which may consist of more than one person, and not everybody might eat the same foods. Thus, accuracy of loyalty card data in investigating individual diet may not be as high as that obtained with traditional data collection methods. However, good compatibility between respondent-collected household-level food purchase data and individual-level dietary data has been demonstrated [28,48]. Moreover, foods purchased from stores do not necessarily indicate foods consumed owing to many different reasons. These include, for example, foods that are not included in loyalty card data, such as dinner foods at restaurants or lunch foods at work. Cardholders may also buy foods that are consumed by others, for example, grandchildren, other relatives or close friends invited for dinner, and pets. Some foods are not consumed at all, resulting in food wastage [49,50].

Finally, the degree of loyalty remains a challenge. In this study, the degree of loyalty was self-estimated, and it is difficult to estimate possible bias related to this self-report. However, we showed that the RFM scores increased steadily with the groups of loyalty, indicating that higher frequency, higher engagement, and more money spent on grocery purchases were associated with a higher degree of loyalty. Moreover, a positive correlation was found between the proportional increases in money spent (€) and weight (kg) regarding food purchases and the degree of loyalty. These results suggest that this self-assessment seemed to provide a feasible estimate of the true values. In another study, the researchers defined loyalty through the frequency of purchases made in the supermarket combined with the amount of money spent on purchases. However, in this study, the degree of loyalty was not specifically defined [51].

Despite its limitations, we see real potential in the use of these automatically collected longitudinal food purchase data in the population-based assessment of dietary patterns, which are important determinants of health and carbon footprint [52]. Loyalty card data provide a cost-effective tool to reach large groups of individuals with minimum data collection efforts and to investigate diet-related behaviors with less information bias. Linking these data with other health data (such as electronic health records and health registers) would provide new opportunities to understand diet and related outcomes. However, such research settings include privacy concerns that need to be carefully addressed to guarantee individual anonymity and consent. In addition, loyalty card data enable the monitoring of longitudinal trends in food purchases including timely monitoring and evaluation of the impact of population-level steering instruments such as taxation.

Individuals who consented to the use of their loyalty card data for research purposes tended to differ from individuals in the general Finnish adult population. The sociodemographic distributions were toward similar characteristics, as is frequently seen in health and nutrition studies. However, the high volume of data enabled the inclusion of sociodemographically heterogeneous subgroups, potentially including hard-to-reach subgroups, and further correction of the differences so that distributions matched well with those of the general Finnish adult population. A potential confounder in studies using loyalty card data is the degree of loyalty, which in this study, was associated with food-purchasing profiles and also the participants’ background characteristics. This underlines the importance of obtaining sufficient background information when using loyalty card data for health research.

Despite the limitations, loyalty card data provide a cost-effective approach for large groups of individuals with minimum data collection effort and for the investigation of diet-related behaviors on a large scale with less information bias. Importantly, loyalty card data enable the monitoring of longitudinal trends in grocery purchases.