As countries develop, the burden of disease moves from infections toward diabetes, heart disease and other chronic illnesses. This shift is called the epidemiological transition. One of its most consistent features is a reversal in who gets sick. In middle-income countries, richer people are more likely to be obese and to eat Westernized, processed diets. In high-income countries the pattern flips, and obesity concentrates among the poor. Surveys can measure this gradient, but they arrive years late. Social media is fast and cheap, and it has been used to predict obesity. Every such study, however, has worked inside the United States and scored itself on how accurately it predicts state rates. None has asked whether the text preserves the direction of the socioeconomic gradient, which is the quantity the transition literature actually cares about.
We argue that it probably does not, at least not naively, and we propose to test why. A post about fried chicken or a latte is not a record of eating. It may be an everyday meal. It may equally be a status display, a wish, a guilty joke, or an advertisement. The mix of these mention types plausibly depends on wealth and on a country's stage of development. In transitional economies, Western food is aspirational and gets photographed. In rich countries, cheap processed food is mundane and rarely worth posting. Simply counting food words should therefore produce a spurious "pro-rich" gradient everywhere. That artifact would hide the pro-poor gradient that surveys document in high-income settings. The proposal is to build a cross-country, randomly sampled corpus of food discourse and to separate first-person habitual consumption from other mention types. The test is then whether this separation, and only this separation, makes social text reproduce survey-measured gradients in the United States, South Korea, China and India.
Core Claim
The main claim is about validity rather than architecture. The sign and size of the SES–diet association estimated from social text depend on which kind of food mention is counted, and the display share of mentions varies with development stage. The claim is falsifiable in both directions. It fails if raw food-term frequencies already match the survey gradient signs across all four countries, because then the mention-type step is unnecessary. It also fails if even consumption-only counts get the signs wrong, because then social text cannot track the transition at this resolution.
The motivating survey evidence comes from three sources. Pro-rich obesity inequality among Chinese elderly is documented in Assessing the income-related inequality in obesity among the elderly in China: A decomposition analysis, which reports a normalized concentration index of 0.075 with socioeconomic factors explaining 68.73% of the inequality. A study of effect of physical activity quotient on central obesity (waist hip ratio) in adolescent girls and boys of central India found that 44% of centrally obese adolescents came from the upper SES group, against 24% from the middle group. Dietary Pattern Extraction Using Natural Language Processing Techniques tracked a decade of Korean dietary change using topic models over survey recall records, including the emergence of a Western/processed pattern covering 15.5% of participants by 2016–2018.
The social-media work this proposal contrasts with is cross-sectional and US-only. Analyzing the Language of Food on Social Media classified state overweight status from meal-hashtag tweets at 80.39% accuracy. Towards using social media to identify individuals at risk for preventable chronic illness reached 82.35% with random forests, but its individual-level accuracy fell to 16% for users with BMI above 28.7. A Novel Approach to Characterize State-level Food Environment and Predict Obesity Rate Using Social Media Data: Correlational Study correlated Yelp-derived food-environment scores with state obesity at r = 0.796. None of the three infers SES. None examines a transitional country. All treat any food mention as evidence of eating. We differ on all three counts.
Research Program
The first component is sampling. Hashtag filtering selects people who show off their meals, which is exactly the bias we want to measure rather than absorb. We therefore follow Mental Health Surveillance over Social Media with Digital Cohorts, which built a demographically characterized cohort of 48,000 randomly sampled users rather than keyword-matched ones. We build comparable cohorts in each country, using X/Twitter for the US, Korea and India and Weibo for China. We geolocate users to sub-national units: counties, Korean si-gun-gu districts, Chinese prefectures and Indian districts. SES is assigned from area-level income for those units. We are explicit that this makes the study ecological at a finer grain than states, not individual-level.
The second component is constructing the diet measure. Each country needs a "traditional" and a "Westernized/processed" food axis. We anchor both in national food lists. For Korea we use the KNHANES food vocabulary and topics from the Choi et al. pipeline, which relied on fastText subword embeddings. For the other countries we use analogous survey food lists. Food mentions are induced with a short-text topic model. We choose the length- and frequency-weighted Dirichlet multinomial mixture of Exploring Semantic Topics in Dementia Caregiver Tweets. That model assumes each short post has a single topic and outperformed LDA and BERTopic on sparse tweets, with peak C_V coherence of 0.6054. This supplies the method; we do not claim a topic-model contribution. Posts in India are frequently code-mixed Hindi–English. Shared-task work such as codecrackers@DravidianLangTech 2025 shows that code-mixed text is a persistent failure point, so India is treated as the stress case.
The third component, and the key ingredient, is mention-type decomposition. We annotate several thousand food-mentioning posts per language into five types. The first is first-person habitual consumption. The second is one-off or celebratory display. The third is aspiration or desire. The fourth is ironic, guilty or figurative mention, such as "healthy dinner lol 🍕🍕🍕". The fifth is third-party or commercial mention. The ironic category is where signal inversion bites. Identifying Depressive Symptoms from Tweets: Figurative Language Enabled Multitask Learning Framework showed that jointly detecting sarcasm and metaphor improved symptom identification to 75.03 F1. We borrow that joint-training design as one candidate classifier. We do not assume it is the best one. A classifier is chosen per country on held-out annotation and then applied to the full cohort.
The fourth component is the gradient estimate. We summarize each gradient with a concentration index,
$$CI = \frac{2\,\mathrm{cov}(y_a, r_a)}{\bar{y}},$$
where $y_a$ is the share of Westernized-diet mentions among a given mention type in area $a$, $r_a$ is that area's fractional rank in the income distribution (population-weighted), and $\bar{y}$ is the mean share. A positive CI means Westernized-diet talk concentrates in richer areas. The same index is computed on survey data for the same areas. This follows the CFPS convention in the Xu et al. decomposition, so the social-text and survey numbers are directly comparable.
Baselines
The naive reference point is raw food-term frequency: the share of Westernized-food terms among all food mentions, with no mention-type filtering, computed on the same random cohorts. This is essentially the Fried et al. feature set moved to sub-national units. We run it with identical geolocation, SES assignment and topic vocabulary, so the only difference from our method is the missing filter. It is the most important competitor because our central claim is that it gives the wrong sign in high-income settings.
The second baseline replicates the hashtag-sampled corpus. We rerun the same raw-frequency estimate on meal-hashtag posts, as in Analyzing the Language of Food on Social Media, using the same areas and time window. Comparing it with the first baseline isolates the effect of sampling from the effect of mention type.
The third baseline is the strongest existing structured method: the food-environment approach of A Novel Approach to Characterize State-level Food Environment and Predict Obesity Rate Using Social Media Data: Correlational Study. It computes fast-food availability and weighted caloric density from restaurant listings. The paper reports Pearson r = 0.796 against US state obesity for the full feature set and r = 0.676 for fast-food availability alone. We recompute it at our sub-national units, using Yelp or a local equivalent (Dianping in China, Zomato in India), and derive its concentration index. At the state level it is a very strong predictor. It measures supply, not consumption, but it may still get the gradient sign right.
The fourth reference is a prediction baseline. On the existing task of classifying US states above or below the median overweight rate, the best reported figure in the file is 82.35% accuracy, from Towards using social media to identify individuals at risk for preventable chronic illness. We report this task only as a sanity check that our cohort and features are competitive. Beating it is not our goal. For the new task itself, recovering survey-measured SES–diet gradients across countries, no reported result found on Lacuna.
The ablation that isolates the key ingredient swaps the learned mention-type filter for a random filter that retains the same fraction of posts. This checks that any gain comes from which posts are removed, not from removing posts in general.
Beating the baselines has a concrete definition. The consumption-only estimate must match the survey CI sign in all four countries. Its sub-national rank correlation with survey diet or obesity measures must exceed each baseline's, with non-overlapping area-bootstrap intervals. The naive baseline must fail to match the sign in at least one high-income country. All text-based methods share one tuning budget: annotation-set hyperparameter search only, with no tuning against survey targets.
Why not just…?
Why not just use the national surveys? Surveys are the ground truth here, not a competitor to replace. Their weakness is latency. The Korean cycles in the Choi et al. study are nine years apart, and CFPS waves are years apart. The question is whether text can track changes between waves. The Korean temporal test below answers this directly. If social text cannot reproduce the 2007–09 to 2016–18 KNHANES shift, surveys win and we say so.
Why not just use restaurant and point-of-interest data? It is clean and already reaches r = 0.796 at the US state level. It is a real threat. Supply may track consumption well enough, and it may even recover the gradient sign. The third baseline tests this head-on at the same areas. Our expectation, which is uncertain, is weaker. In transitional cities, Western outlets cluster in rich districts regardless of who eats there, while in rich countries cheap food is everywhere. Supply data would then flatten the within-country gradient. If it matches our method, the practical conclusion is to use supply data. The mention-type finding would still stand as a statement about what social text measures.
Why not just prompt a large language model to classify the posts? This could work as well as a trained multitask model, and we do not argue otherwise. The only related evidence in the file is Deloitte at #SMM4H 2024: Can GPT-4 Detect COVID-19 Tweets Annotated by Itself?, where GPT-4 scored 0.51 accuracy. That was a different task, telling human annotations from GPT-4 annotations, so it says little here. We have therefore made the classifier interchangeable. Zero-shot and few-shot LLM classifiers are evaluated alongside the joint figurative-language model on the same annotated test sets. Whichever wins per language feeds the gradient estimate. The contribution is the decomposition and its validation, not a particular network.
Why not just count keywords, as prior work did? That is the first baseline, and the hypothesis predicts it fails specifically in high-income countries. The weight-prediction study Quantified Self Meets Social Media: Sharing of Weight Updates on Twitter used dictionary counts and explained roughly 25% of weight variance. It also relied on a small, highly self-selected cohort, which is the other failure our random sampling addresses.
Proposed Experiments
The first decisive check is the sign test. For each country we compare survey CI signs with the CI signs from the three text estimates: raw-frequency random cohort, hashtag corpus, and consumption-only. Survey anchors are BRFSS/NHANES for the US, KNHANES for Korea, and CFPS for China. For India we use the national survey closest to district resolution, treating the Rewa study only as corroborating direction. The prediction is specific. The raw estimate is pro-rich everywhere. The consumption-only estimate is pro-poor in the US and pro-rich in China and India. Korea sits nearer zero.
The second check tests the mechanism. We estimate the display-plus-aspiration share of Westernized-food mentions as a function of area income within each country. The prediction is that this share rises with income, and does so more steeply in China and India than in the US. If the display share is flat across income, the mechanism is wrong even if the signs come out right.
The third check is temporal. Using Korean posts from 2009–2011 and 2016–2018, we test whether the change in consumption-only Westernized share, and in its CI, matches the KNHANES change across the two cycles. This is the one test of whether text can fill survey gaps.
The fourth check is a placebo. Traditional staples such as rice and kimchi in Korea are not predicted to reverse gradient direction across countries in the way Westernized foods are. A method that shows spurious reversals for staples is picking up area-level posting artifacts rather than diet.
Uncertainty for every estimate comes from bootstrapping over areas. Annotation reliability is reported per mention type and per language. The code-mixed Indian data is reported separately, so that a failure there reads as a text-analysis limitation rather than an epidemiological one.
References
- Assessing the income-related inequality in obesity among the elderly in China: A decomposition analysis
- A study of effect of physical activity quotient on central obesity (waist hip ratio) in adolescent girls and boys of central India
- Dietary Pattern Extraction Using Natural Language Processing Techniques
- Analyzing the language of food on social media
- Towards Using Social Media to Identify Individuals at Risk for Preventable Chronic Illness
- A Novel Approach to Characterize State-level Food Environment and Predict Obesity Rate Using Social Media Data: Correlational Study
- Mental Health Surveillance over Social Media with Digital Cohorts
- Exploring Semantic Topics in Dementia Caregiver Tweets
- codecrackers@DravidianLangTech 2025: Sentiment Classification in Tamil and Tulu Code-Mixed Social Media Text Using Machine Learning
- Identifying Depressive Symptoms from Tweets: Figurative Language Enabled Multitask Learning Framework
- Deloitte at #SMM4H 2024: Can GPT-4 Detect COVID-19 Tweets Annotated by Itself?
- Quantified Self Meets Social Media: Sharing of Weight Updates on Twitter