Wordfish
Wordfish is a statistical model for scaling documents on latent dimensions, developed by Slapin and Proksch (2008). Unlike reference-based methods like Wordscores, Wordfish uses a Poisson generative model to jointly estimate word frequencies and document positions without requiring reference texts or manual annotation. It is particularly useful for estimating time-series changes in policy positions and can scale documents from multiple languages simultaneously.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Apply Wordfish when you want to scale documents on a latent dimension without reference texts, compare document positions over time (speeches of a legislature across years), or scale texts from multiple languages. Ideal when reference texts are unavailable or when you want to avoid the subjectivity of selecting reference materials. Works well with large corpora of comparable documents.
Strengths & limitations
- Reference-free: no need for reference texts, avoiding subjective anchor selection
- Time-dynamic: naturally extends to tracking position changes across time periods
- Multilingual capable: can scale documents in different languages simultaneously if word dictionaries align
- Theoretically grounded: based on explicit Poisson model with clear assumptions
- Word-level diagnostics: provides word discrimination scores, revealing which words drive the latent dimension
- Single dimension: standard Wordfish estimates one dimension; multi-dimensional extensions require additional modeling
- Poisson assumption: word counts may not follow Poisson distributions, especially with long documents or rare words
- Convergence challenges: EM estimation can be slow for large corpora and sensitive to initialization
- Interpretation ambiguity: unlike reference-based methods, the latent dimension is identified only up to reflection and rotation
Frequently asked
How does Wordfish differ from Wordscores?
Wordscores requires reference texts with known positions; Wordfish estimates positions from word distributions alone. Wordfish is better for unsupervised analysis and time-series applications but offers less external validation. Wordscores is more transparent but depends heavily on reference quality.
What do the word discrimination scores tell me?
Word discrimination (psi) measures how much a word's frequency varies with document position. High discrimination means the word strongly differentiates between documents on the latent dimension. However, high discrimination does not guarantee the dimension is meaningful; always validate against external measures.
Can I rotate the Wordfish dimension to align with known anchors?
Yes. After estimating positions, you can identify anchor documents (e.g., manifestos of known left and right parties) and linearly rotate the Wordfish dimension so the anchors align with expected positions. This aids interpretation without changing the underlying model fit.
How does document length affect Wordfish estimates?
Long documents will have more word count variation, potentially affecting convergence. Wordfish includes document-level offsets (beta_d) to account for different document lengths, so estimates should be robust. However, extremely short documents may have unreliable position estimates.
Can Wordfish scale documents in different languages?
Yes, if the documents share sufficient vocabulary or if you align dictionaries across languages. However, pure cross-lingual Wordfish assumes word meanings transfer across languages, which is often unrealistic. Language-specific Wordfish models followed by alignment may be more reliable.
Sources
- Slapin, J. B., & Proksch, S. O. (2008). A scaling model for estimating time-series party positions from texts. Journal of Politics, 70(3), 554-569. DOI: 10.1111/j.1540-5907.2008.00338.x ↗
- Proksch, S. O., & Slapin, J. B. (2009). How to avoid pitfalls in statistical machine learning for social science. Political Analysis, 20(3), 343-357. link ↗
- Benoit, K., Muhr, D., & Spirling, A. (2016). Crowd-sourced text analysis: Reproducible and distributed production of political data. American Political Science Review, 110(2), 278-295. DOI: 10.1017/S0003055416000058 ↗
How to cite this page
ScholarGate. (2026, June 3). Wordfish. ScholarGate. https://scholargate.app/en/psychometrics/wordfish
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Exploratory Structural Equation ModelingPsychometrics↔ compare
- Fuzzy ANOVAPsychometrics↔ compare
- Latent Transition AnalysisPsychometrics↔ compare
- Partial Least Squares Structural Equation ModelingPsychometrics↔ compare
- WordscoresPsychometrics↔ compare