Digital traces as an epidemiological tool

An early example showing how users’ online digital traces can be used to monitor and predict public health is the work of Polgreen and colleagues [17] who investigated the potential of influenza related queries submitted to the Yahoo! search engine to predict influenza cases and deaths in the US. Later work following a similar approach, but using Google search queries, formed the basis for the well-known Google flu tracking service [10]. More recent research has used the linguistic properties of social media content to track smoking [18], the use of alcohol [19] and other drugs [20], as well as life satisfaction [21, 22] and mental illnesses, such as depression [23]. Two comprehensive overview publications for these kinds of studies are [24] and [25].

While viruses such as H1N1 (flu) and mental illness cases have received considerable attention, relatively little research has focused on using digital traces to understand nutritional habits or health related issues such as obesity. From what we know about how people make food choices e.g. [26] we know this is a complex, multi-faceted process, influenced by biological, personal and socio-economical factors. Yet for the majority of people, aspects of taste and sensory appeal seem to be the drivers for decisions, followed by health concerns, nutritional value, and price [27], meaning that the decisions can be modeled using relatively simple heuristics [28]. All of this suggests that data regarding food preferences established via interactions with online recipes may shed light on food related, socio-economic problems, such as obesity.

Work in this direction includes West et al.’s [29] analysis of a large corpus of Microsoft Bing search-logs to determine the extent to which people search and access information in the online food community website Allrecipes.com. They discovered that people’s searches for food on the Web follow weekly and yearly trends, that they significantly differ between states and that there is a correlation between searches for food online and heart attacks recorded in the US. Similar trends were also observed by Wagner and colleagues [30] and Trattner et al. [14] using data collected via the German online food portal Kochbar.de. Said & Bellogin [31] analyzed a crawl of Allrecipes.com and using simple statistical techniques showed that there are statistically significant differences between states in the US on an ingredient level. Twitter has also been used as a basis to understand food behavior. Fried and others [32] and Abbar and others [33] both found correlations between food mentions on Twitter and health related issues such as diabetes and obesity in the US. Moreover correlations have been found between FourSquare check-in patterns at fast food restaurants and obesity patterns in the US [34]. De Choudhury et al. [35] showed that digital traces left behind by users on the social platform Instagram and poor online food consumption patterns are correlated with so-called “food deserts” (unhealthy regions) in the US. All of these analyses focused on the relationship between the occurrence of certain food types or ingredients perceived as being unhealthy and health-related metrics. Recent research has investigated how digital traces, predominately search queries, can be used to detect life-threatening diseases, such as cancer [36].

Relevant nutritional science research

Nutritional science and epidemiology have contributed considerable knowledge with respect to how food consumption patterns relate to public health. Based on this knowledge, a technical report by the WHO expert group on diet, nutrition and prevention of chronic diseases [37] reports on the extent to which macro nutrients impact on public health, recommending that intake of sugar- and energy-dense foods should be limited to reduce the likelihood of suffering from obesity. For type-2 diabetes the recommendation is to reduce the saturated and total fats consumed. Similar findings and recommendations are reported by Gross and colleagues [38] based a longitudinal ecologic correlation study on nutrient consumption in the United States between 1909 and 1997 obtained from the US Department of Agriculture. In their work, with the help of a multivariate nutrient-density model, they discovered that corn syrup (sugar) was positively associated with the prevalence of type 2 diabetes. Fiber however, was negatively associated with the prevalence of type 2 diabetes. For protein and fat, no associations with type 2 diabetes were found, a result similar to that obtained by Meyer et al. [39] in 2000.

Van Dam and Seidell [40] provide an overview of research studying the link between obesity and the intake of certain micro nutrients. Somewhat mirroring the recommendations of the WHO, they found that sugar consumption is associated with weight gain and that fiber is associated with a lesser degree of weight gain. For further studies in this context, we refer the reader to the already cited WHO technical report [37], as well as to the database of the American journal of clinical nutrition [41].

Differences to previous research

Our work extends previous computational research approaches in three distinct ways: (i) First, in contrast to previous work, we do not look at specific food types or specific ingredients liked by users, but rather rely on macro-nutritional properties of the kind of food people prefer (i.e. bookmark). This provides the potential for a different, richer picture of the factors affecting people’s health, perhaps building bridges to research from nutritional science and epidemiology. (ii) Second, we do not limit our study to one single snapshot of the data. All previous studies rely on data that reflects online food interactions within a single, short time frame of 1–3 months. We, instead, provide insights over a 9-year period. This allows us not only to show that the relationships are stable over time, but also to improve the modeling process by incorporating temporal aspects. This is important because there are several temporal trends in both bookmarking behavior and the incidence of obesity—we show these trends globally and across states and counties. (iii) Finally, there are methodological differences between our study and previous investigations in this area: We go beyond previous work, which offered simple correlation analyses between digital traces (e.g. the caloric values of foods mentioned in tweets per each state and obesity rate in the state), by using multi-level regression modeling approaches to understand the complex relationships between the micro nutritional properties of recipes and obesity levels at a user, county and state level over time.

Data pre-processing

Performing these analyses required a number of important pre-processing steps. First, we restricted the bookmarks to the same time periods as provided by the US census data. Next, we restricted the dataset further to include only the recipes for which full nutritional information was available. Although recipes in Allrecipes.com are peer-reviewed and macro nutritional facts are generated using the ESHA [46] research database, not all nutrition properties can be calculated as some of the ingredients do not match with the ESHA database. For our analysis, we relied only on those recipes for which all ingredients were present in the ESHA food database (out of 242,113 originally crawled recipes 58,263 had nutrition information available). We chose to focus on “fat”, “saturated fat”, “sugar” and “sodium” (measured in 100g per recipe) because they allow us to measure the healthiness of a recipe according to international standards introduced in 2007 by The Food Standard Agancy (FSA) [47].

Following the procedure described in Howard et al.’s paper [48], for each meal we calculated the nutritional content per portion by dividing the total content by the number of portions in the meal. This was done to obtain a so-called FSA health score that measures on a discrete scale the extent to which a recipe is healthy or probably unhealthy [47]. The FSA front of package labeling system introduced by the Food Standard Agancy (FSA) [47] relates to 4 macro-nutrients (sugar, sodium, fat and saturated fat). The scale is green (healthy), amber and red (unhealthy). In order to derive a single metric we follow the procedure of Sacks et al. [49] who first assign an integer value to each color (green = 1, amber = 2 and red = 3) then sum the scores for each macro-nutrient resulting in a final range from 4 (very healthy recipe) to 12 (very unhealthy recipe). We use this metric throughout our analysis to determine the healthiness of recipes and we refer to it as “FSA score”.

It was necessary to map users of Allrecipes.com to states and counties in the US. This was possible because users in Allrecipes.com typically provide information about their current home location. For the purpose of our study, we considered only users who could be objectively mapped to a state and a county in the US. This was done with the help of the Google Places Web API [50].

We continued to transform the data by calculating, for each US state and county, the mean value for each nutritional property (fat, saturated fat, sugar and sodium) for recipes bookmarked by users residing in that state. From these values we derived the FSA score representing the healthiness the recipes on average. This allowed us to create yearly datasets that contain obesity incidence level at state and county level as well as the corresponding interaction, macro-nutritional properties and FSA scores.

Finally, to remove noise from the dataset we only considered those observations where for a specific year the county had at least 30 Allrecipes.com users. Although this produced a dataset with imbalanced observations (some counties are only present in later years where the number of users increased beyond 30), the regression modeling approach used to analyze the data is suited to this condition.

30 users was chosen as a cut-off point to balance the need of having a representative number of users for each county without overly restricting the data to highly populated geographical locations (e.g. cities) and counties with few problems with obesity. The dataset contains a general bias whereby fewer users reside in locations with the worst obesity situations. Again our statistical analyses are robust to this problem. The overall descriptive statistics of our dataset for the years 2004 to 2012 can be viewed in Table 1. The dataset can be obtained via the supplementary information, see S1 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Basic statistics of the Allrecipes.com dataset with at least 30 users per county.

https://doi.org/10.1371/journal.pone.0179144.t001

The statistical analysis

It is common to precede a regression modeling procedure with an exploratory data analysis to understand the data distribution, identify potential outliers and detect issues which violate the assumptions such models make. A good example is the presence of strong multicolinearity between predictors. We report EDA results by presenting a correlation analysis (Pearson correlation) considering the variables obtained from the traces extracted via Allrecipes.com (bookmarks, nutritional information from bookmarked recipes, the FSA health score) and also from the census data (obesity prevalence levels).

Moving beyond the approaches applied in previous research efforts (e.g. [29, 33, 51], we perform a multi-level regression analysis to investigate the relationships between the predictors in more detail. A standard multiple regression modeling approach would be inappropriate because the hierarchical structure of our dataset violates the assumption of independence of the observations. Our data on obesity incidence, as well as the nutritional data of bookmarked recipes stems from different counties all over the US. The dataset, moreover, has repeated observations over time (from 2004 to 2012) and we know the nested geographical structure of counties within states. Using this geographical, temporal and nutritional information we built several multilevel regression models in order to discover how these factors relate to obesity.

Figs 1 and 2 depict the linear trends in obesity prevalence over time at state and county level, respectively. The figures are derived from US Census data and to make the trends legible only counties where at least 30 users had bookmarked at least one online recipe were included. In Fig 1 each data point reflects the mean obesity prevalence within a county, while the linear line is an extrapolation of these data points to show the trends over time at state level.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Linear trends of fat as a function of time (2004–2012) for selected states in the US.

The plots show a variety of intercepts and trends (slopes) over time. The general aggregated trend is shown with a dashed line.

https://doi.org/10.1371/journal.pone.0179144.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Trends of obesity prevalence levels as a function of time (2004–2012) for states and corresponding counties (presented as lines) in the US.

We only report states and counties with at least 30 users bookmarking recipes in each of the counties for each year.

https://doi.org/10.1371/journal.pone.0179144.g002

Additionally, Fig 3 shows the trends for fat (top-left plot), saturated fat (top-right plot), sugar (bottom-left plot) and sodium (bottom-right plot) as a function of time, for the same selected states as presented in Fig 1. What these plots show is that overall we see a gradual increase in fat and sugar content per 100g of food from 2004 up until some point between 2008 and 2010. After this, the aggregated quantity of these macro nutritional properties per 100g decreases. Similar trends are also observed in the correlations presented in the first columns of Figs 4 and 5. It is important to note that during this time period we also observe important differences in the trends for individual states, similar to what is shown in Fig 1 for obesity prevalence. The patterns for sodium and fat show different trends, with sodium showing a constant increase over time while saturated fat declines between the years 2008 and 2010 when the values are aggregated. Again, the trends are different for individual states, which is important for our modelling process.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Trends for fat (top-left plot), saturated fat (top-right plot), sugar (bottom-left plot) and sodium (bottom-right plot) measured in 100g per recipe as a function of time (2004–2012) for the same selected US states as depicted in Fig 1.

The general aggregated trend is shown with a dashed line.

https://doi.org/10.1371/journal.pone.0179144.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. County-level correlations (spearman).

Note: ***p < 0.001, *p < 0.05.

https://doi.org/10.1371/journal.pone.0179144.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. State-level correlations (spearman).

Note: ***p < 0.001, **p < 0.01, *p < 0.05.

https://doi.org/10.1371/journal.pone.0179144.g005

Fig 1 provides evidence to justify our approach of fitting multilevel regression models with random intercept and random slope for each state if we observe their evolution over time. For instance, while the state of Michigan had an average obesity prevalence level of around 28% in 2004 and a steep growing trend until 2012, the state of Colorado had an initial average obesity level of only around 16% in 2004 and a less steep rate of growth. In addition, we have states without observations for every year, such as Mississippi and Vermont, which had enough data (at least 30 users per county each year in Allrecipes.com) only starting from 2008 and 2010, respectively.

These special characteristics of the dataset (differing temporal patterns at state and county levels, missing observations for counties for some and sometimes all years, and prevalence of datapoints for particular geographic regions (e.g. California and Texas)) are shown clearly in Fig 2. Multi-level regression models are robust and well suited to modeling datasets with these properties.

Following the methodology in [52, 53], we build models of increasing complexity: unconditional-means models, unconditional growth models and finally a model with random intercept and slope including nutritional properties as fixed factors.

In total we report four longitudinal multilevel regression models, which predict level of obesity. Each of these models can be expressed in a three-level equation such as:

(1a)

(1b)

(1c)

In the set of Eqs 1a–1c, the variable Obesity_tij identifies the obesity prevalence level—variable collected from census information—at county level and it is observed at time t for county i and state j. The model has random intercept π_0ij and also a random slope π_1ij for the coefficient year. Level 1 equation also contains the fixed effects of nutritional variables (nutritionalVars_k) such as proteins and carbohydrates. We model them at level 1 since these measures vary every year. ϵ_tij corresponds to the error at level one, i.e., the variation in Obesity_tij data which is not explained by the model. Next, the county and state information is encoded in levels 2 and 3 within the random intercept π_0ij and the slope for year. More specifically, the intercept at level 1 is modeled at level 2 as a function of county π_0ij, and similarly the coefficient π_1ij accompanying year. Finally, the intercepts and slopes of the aforementioned models are modeled as function of state (counties are located within states), where β_00j and β_01j are related to the random intercept at level 1, whereas β_10j and β_11j are related to the random slope of year at level 1. We use a diagonal variance structure, which is the default in the lmer package [54].

Correlation analysis

We performed the correlation analysis at two geographical levels, county-wise and state-wise, as shown in the correlation matrices in Figs 4 and 5.

The final row of each matrix depict how obesity incidence correlates with the other variables. Although there are significant correlations with different variables, we highlight three important observations. First, the two strongest correlations with obesity prevalence are time (i.e. year) and the FSA healthiness score of the recipes bookmarked by the users. At county level, the correlation between obesity prevalence and time (year) is ρ = 0.21, p < .001, whereas the FSA score shows a correlation of ρ = 0.18, p < .001 with obesity. At state level, a similar but stronger correlation between the FSA score and obesity is observed ρ = 0.26, p < .001. Second, we discover a correlation between the nutritional variables (Fat, Saturated Fat, Sugar and Sodium) of the recipes bookmarked with Obesity prevalence. These are nevertheless not as pronounced as the correlation between obesity and the FSA score, which combines all four variables in to a single health measure. The rho values range from ρ = 0.093 (Fat) to ρ = 0.15 (Salt) at county level and from ρ = 0.093 (Sugar) to ρ = 0.20 (Saturated Fat) at state level.

Finally, we observe a significant correlation between all variables studied, although on state level two variables show non-significant correlations. Multicollinearity can become a problem for the interpretation of regression models [55]. Correspondingly, in our analyses we employ a stepwise procedure to discard variables too correlated and keep only those with larger effect and without large correlation to other predictors.

Multilevel regression analysis

We present the results of the multilevel regression analysis by presenting four of the models created. The modeling procedure followed the suggestion of several authors to build multilevel regression models by incrementally increasing the number of random and fixed effects and testing the model fitting using Akaike Information Criteria (AIC), Bayesian information Criteria (BIC), and likelihood-ratio test [52, 53]. We first checked the need for using multilevel models by comparing a null baseline model –fitted using a single intercept– against a multilevel model with random intercept which allowed the intercepts to vary across different counties/states.

A likelihood-ratio test between the baseline null model (AIC = 9363.07, BIC = 9373.92) and the multilevel model (AIC = 6796.61, BIC = 6818.30) shows that the second model fits the data significantly better (χ²(2) = 1285.25, p = 0), so we proceed with the multilevel regression analysis.

Table 2 summarizes four models for obesity which help to explain the effect of each factor. Model 1, an unconditional means model, accounts for the diversity of obesity only with the random effects based on geographical variables county and state. The random effect county:state has a between-county/state variance of 8.90, while the between-state variance is smaller, 4.87. Both variances are comparatively large with respect to a residual variance of only 1.79, what results in an intra-class correlation for county:state of ICC_{county: state} = 0.88 and for state of ICC_state = 0.31. These large intra-class correlations tell us that a multilevel model is necessary to explain the data (due to clustering effect within geographical regions) and also that a large amount of the variance of obesity is already explained by this simple model [56].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Multilevel models for obesity.

Models 3 and 4 incorporate a random intercept per county/state, a random slope for year, and fixed effects for the FSA score and Fat and Sugar. A likelihood ratio test shows significant differences between the models: Model 1 vs Model 2: χ²(5) = 585.64, p < 0.001; Model 2 vs Model 3: χ²(1) = 23.91, p < 0.001; Model 3 vs Model 4: χ²(1) = 14.67, p < 0.001. For the fixed effects, the number in parenthesis shows the standard error.

https://doi.org/10.1371/journal.pone.0179144.t002

Model 2 adds a random intercept for the temporal variable year and no other fixed effects. This type of model is known as unconditional growth model [52]. Compared to Model 1, it has a smaller residual variance of 0.97. This drop in residual variance tells us that 45.7% of the within-county and within-state variation in obesity is systematically associated with linear time, considering [57]. Since the variable of time, year, was centered in 2004 = 0, this means that in 2004 the average obesity was 24.89% as observed in the significant intercept, p < 0.001. The coefficient associated to year is = 0.3, which means that every consecutive year the obesity level is increased by 0.3%. Model 2 has a significantly better fit than Model 1 tested with likelihood ratio test, Model 1 vs Model 2: χ²(5) = 585.64, p < 0.001.

Our third model, Model 3, incorporates fixed effects to Model 2. In particular it adds the FSA health score to the model. As shown in Table 2 the effect is significant, implying that for every additional unit of the FSA score, the the obesity level increases by 1.26. What is worth noting here is that this effect is stronger than the effect of time (Year) though lower than the impact of state/county. This model has a significantly better fit than Model 2, assessed by the decreased in AIC, BIC and by the likelihood ratio test, Model 2 vs Model 3: χ²(1) = 23.91, p < 0.001.

Our fourth and final model, Model 4, includes nutritional variables. We tested all the nutritional variables, but we found that adding “Fat” and “Sugar” produced the best fit (again employing a stepwise procedure). The effect is rather small compared to the variance already explained by the geographical and temporal dimensions, but it is still significant and positive, implying that for every additional unit of fat/100g and sugar/100g of the recipe, the obesity level increases by 0.19% (for fat) and 0.08% (for sugar). This last model has a significantly better fit than Model 3, assessed by the decreased AIC and by the likelihood ratio test, Model 3 vs Model 4: χ²(1) = 14.67, p < 0.001.

Findings

Our findings can be summarised as follows:

• We demonstrate significant and meaningful (i.e. sensibly interpretable) relationships between the nutritional properties of bookmarked recipes (sugar content, fat content and a combined FSA-score for recipes) and obesity incidence.
• As our dataset spans a 9 year time period we show the stability of such data as predictors.
• We show that temporal patterns are also important with obesity rate increasing overall over time. We do acknowledge, however, that there are counties and states for which this is not the case.
• The geographical differences in obesity patterns mean that it is also important to account for these in the modeling process.
• The good fit achieved by our models suggests that combining interaction data, geographical data and temporal data can be a useful in monitoring obesity incidence.

Limitations

There are several limitations to the analyses we have presented and we wish to acknowledge these directly.

• The analyses relate only to the United States and in particular to one food portal—Allrecipes.com. There is no evidence that the findings would be repeated with other websites or in other countries.
• The user demographics of Allrecipes.com are skewed in general towards wealthy and densely populated areas. The socio-economic factors relating to obesity are well-documented—higher rates are found among groups with lower educational and income levels, among racial and ethnic minorities, and in high-poverty areas [64]. Although the modeling approaches applied are typically robust to these kinds of issues, the data available may have restricted the quality of the fit achieved.
• We restricted our focus to four micro/macro-nutritional components (fat, saturated-fat, sugar and sodium) as well as the FSA metric combining these. Other interesting variables we may have investigated include the WHO metric [48], which accounts for additional properties such as Fiber and Carbohydrates. The USDA Healthy Eating Index (HEI) [69] also encompasses different aspects in the model such as vegetables, fruits, etc.

Future work

We are currently extending and planning to extend our work in numerous ways:

• We are extending our models to include other interactive features (ratings, comment sentiment etc.) to establish if these offer better or additional insight into obesity incidence. In addition, we will investigate the use of nonlinear hierarchical models [70] to better account for the not strictly linear nature of the relations between nutritional elements and outcome variables such as obesity or diabetes.
• We are also examining the WHO [66] and HEI [69] metrics to see if these offer additional utility.
• We are investigating the possibility of repeating our analyses with other data sources. In particular we hope to repeat the analyses for data in a different country.
• We hope to investigate whether interaction data from Allrecipes.com can be combined with other interaction sources, such as longitudinal social media content. The intuition here is that these sources may offer different and complementary insight.
• We plan to develop more complex models to investigate the relation of climate. The intuition being that people living in locations with colder climates may be prone to eating more caloric food than people in warmer regions, without this implying larger obesity or diabetes levels. Analyzing this relation more closely might help understanding the effect of geographical factors.