What predicts strong public support for climate action, and does it split the world into types?

The People's Climate Vote 2024, run by the United Nations Development Programme, is the world's largest public survey on climate change. It covers 73 countries with sample sizes of roughly 900 to 1,500 respondents per country, more than 73,000 people in total. The published dataset already holds weighted means: the share of each demographic group (a Country, Age band and Education level) that picked a given answer, not individual survey responses.

Over 20,000 rows held null values, concentrated in the weighted mean columns, and age bands above 60 or below 18 were only reported for a handful of countries. We considered predicting the missing values with a regression model, but rejected it: filling in invented answers to a public opinion survey risked biasing the analysis. We removed all null rows instead and kept only complete cases.

We tried two models against the same features. The linear regression predicts the exact weighted-mean percentage that chose "strengthen", a continuous target. The random forest predicts only whether a group clears the 85% "high support" bar, a binary target. Both share the same cleaning and encoding pipeline described above.

The random forest also ranks which features matter most for telling high-support groups apart. The three most important were how much a person's worry about climate change changed compared to last year, whether climate change affected a family decision such as relocating or a large purchase, and how often they think about climate change day to day. All three are about a person's own lived experience, not their country or demographic group.

Alongside prediction, we grouped the 73 countries into typologies by their collective climate attitudes, using only the "All Ages, All Education" rows so each country is represented once. Each country became one row, with one column per survey question holding the weighted mean of its most positive or proactive answer. We excluded Q8, a question about who is responsible for climate action, because it does not reflect a respondent's own attitude.

We tried k-means clustering on three versions of this table: the raw features, normalised features, and features reduced with PCA (principal component analysis, a way to compress many correlated columns into a few that keep most of the spread in the data). The raw features gave poorly separated clusters, and the normalised features gave clusters too sparse to interpret because of the high number of columns. PCA gave the most compact, interpretable clusters, so we combined it with k-means. The elbow method, which looks for the point where adding more clusters stops meaningfully reducing error, pointed to 6 clusters.

Our first coursework asked a narrower question: is there a consistent generational gap in urgency about the fossil fuel transition? We defined an age gap as the share of 18 to 35 year olds who want a "quick" or "very quick" transition, minus the same share among 36 to 59 year olds. We restricted the comparison to these two bands because they were the only ones reported consistently across countries. A positive gap means younger people are more urgent; a negative gap means the older band is.

The data is opinion, aggregated twice. The public dataset gives weighted means per demographic group, not individual responses, and our models group by Country, Education and Age on top of that. A model that predicts group-level averages well does not prove it would predict any one person's answer.

Complete-case cleaning drops information. Removing over 20,000 null rows and every "Don't know" answer keeps the analysis honest but shrinks the usable sample, and countries with patchier reporting are underrepresented as a result.

Feature importance is not causation. Knowing that worry and family impact rank highest for the random forest tells us what the model leaned on, not that raising someone's worry would raise their support.

Six clusters is a choice, not a law. The elbow method suggested 6 as a reasonable balance of detail and readability. A different cut, or a different year of the survey, could plausibly draw the lines elsewhere.

Sample: 73,000+ respondents across 73 countries, People's Climate Vote 2024 (UNDP).

This was group MSc coursework at the University of Leeds, submitted as two reports by team Luxray: Harshit Verma, Ayush Patne, Abhishek Arun Raut and Sourav Jagati.