Leader: Massimo Anelli (Università Bocconi); Other collaborator(s):
Ageing increases the political support for parties proposing welfare expenditure targeted to the old and no structural reforms to remove unproductive rents. An “ageing trap” of low growth and unsustainable welfare state may follow. To study this phenomenon, we use data on individual attitudes and voting behavior, and information on the parties’ policy platforms. On the demand side, older voters may shift towards more ‘conservative’ parties. On the supply side, parties may change their policy proposals to match the age structure of their electorate, for instance, radical-right forces may turn left on retirement issues in the aftermath of financial crises. We plan to exploit discontinuities in recent pension reforms for causal identification.
Brief description of the activities and of the intermediate results
In these months, we have focused on developing a strategy to quantify the degree of intergenerational distribution implied by policy proposals contained in party manifestos for several European countries. We have identified the Comparative Manifestos Project as the base upon which we want to build our method and measure. The Comparative Manifestos Project (CMP) provides the corpus of all party manifestos for the past decades and multiple countries already parsed in sentences. While many other policy dimensions have been already classified and quantified by researchers in the past, no measure of intergenerational redistribution is available. We have decided to follow the CMP methodology to classify each party manifesto sentence as redistributing towards older generations or redistributing towards the younger generations. With the help of computer science researchers we are studying and developing a new method based on Natural Language Processing (NLP) to classify sentences from these parsed party manifestos. We have first experimented the classification of a small random sample of party manifestos with the GPT model. To do so, we engineered and tested an effective prompt that can replicate pretty closely the existing human classification for existing non-age specific policy dimensions. Once we have validated this new classification method for existing policy dimensions, we will applied it to the classification of the intergenerational redistribution dimension. With the help of a researcher hired for this project we have manually classified the intergenerational target of 10'000 sentences from a random sample of manifestos representative of 16 european countries, 5 main party family and the time dimension. We have then compared the performance of human classifiers against GPT and several other open-source Large Language Models
Main policy, industrial and scientific implications
Coming soon
In these months, we have focused on developing a strategy to quantify the degree of intergenerational distribution implied by policy proposals contained in party manifestos for several European countries. We have identified the Comparative Manifestos Project as the base upon which we want to build our method and measure. The Comparative Manifestos Project (CMP) provides the corpus of all party manifestos for the past decades and multiple countries already parsed in sentences. While many other policy dimensions have already been classified and quantified by researchers, no measure of intergenerational redistribution is available.
We decided to follow the CMP methodology to classify each party manifesto sentence as redistributing towards older generations or redistributing towards younger generations. With the help of computer science researchers, we developed a new method based on Natural Language Processing (NLP) to classify sentences from these parsed party manifestos. We first experimented with classifying a small random sample of party manifestos using the GPT model. To do so, we engineered and tested an effective prompt that closely replicates the existing human classification for non-age-specific policy dimensions. Once we validated this new
classification method for existing policy dimensions, we applied it to the classification of the intergenerational redistribution dimension. With the help of a researcher hired for this project, we manually classified the intergenerational target of 10,000 sentences from a random sample of manifestos representative of 16 European countries, 5 main party families, and the time dimension. We then compared the performance of human classifiers against GPT and several other open-source Large Language Models (LLMs).
Building on these efforts, we successfully validated the LLM-based classification method and refined the prompting strategy to maximize accuracy. After comparative testing, we identified the most effective open-source LLM and deployed it to classify the entire CMP corpus. This scaling allowed us to systematically analyze manifestos across decades, countries, and party families, generating comprehensive data on intergenerational policy preferences.
Using the classified sentences, we quantified the age-targeting of welfare-related spending proposals in the manifestos. This enabled us to calculate indexes of age-specific targeting for each party in each country and election in our dataset. To complement this, we calculated the median age of voters for every country-election. These voter demographics were then used to explore the relationship between the age composition of electorates and the policy proposals presented in party platforms.
Our analysis reveals a significant correlation between the median age of voters and the degree of age-targeting in party manifestos, with distinct patterns emerging across different political families. These findings highlight how demographic trends influence the policy proposals of political parties and provide valuable insights into intergenerational redistribution in political platforms.
Coming soon