A research blog based on: “Does online sustainability communication work? Insights from 6 years of social media discourse between tenants and their housing providers”
The short version: Dutch housing associations are legally required to renovate over a million homes for energy efficiency — and they need tenant support to do it. So what happens when they take their message to Facebook? We analysed nearly 3,200 tenant comments across 92 housing associations to find out. The answer is more nuanced — and more instructive — than you might expect.
The Netherlands has a serious sustainability challenge. Over 2.4 million social housing dwellings — about 30% of the entire national housing stock — need to become significantly more energy efficient. But here’s the catch: housing associations can’t just renovate. They’re legally required to secure agreement from at least 70% of affected tenants first.
That means communication isn’t just good practice. It’s a legal prerequisite.
So housing associations have turned to social media — particularly Facebook — to build that support, inform tenants, and gauge public sentiment. But does it actually work? Do tenants engage meaningfully? Do they even respond to what’s being posted?
These questions sit at the heart of this research.
We collected data from the public Facebook pages of 92 Dutch social housing associations, spanning January 2018 to May 2023. After filtering for sustainability-related posts (energy, insulation, heat pumps, solar panels, renovation), we were left with:
To make sense of all this text, we built a machine learning pipeline that classified every single comment along three dimensions:
| Dimension | What it captures | How we measured it |
|---|---|---|
| Sentiment | The emotional tone of the comment | Ensemble of 3 pre-trained multilingual models |
| Intent | What the commenter was trying to do | Fine-tuned BERT models trained on 13,000 synthetic comments |
| Relatedness | Whether the comment actually responds to the post | Cosine similarity using sentence embeddings |
We then used unsupervised clustering (k-means) to group comments into natural discourse types, and validated patterns with multinomial logistic regression.
The data revealed six distinct types of how tenants engage online. Think of these as six different conversations happening simultaneously in the same comment section.