2. Pick the right model

Before K, choose the model. This is a substantive decision. It encodes what you think generates the text and what you want to learn, and reviewers will ask you to justify it. Pick the simplest model that matches your question and your data structure.

First: is a topic model even the right tool?

Good fit

• Exploratory: what themes exist in this corpus?
• Large collections where manual reading is infeasible.
• You want to discover structure, not impose categories.
• You need to track theme prevalence over time or across groups.

Poor fit

• Purely confirmatory counting against a fixed code list, where you do not need topics at all: a dictionary or supervised classifier is simpler. If you want to measure pre-theorized concepts as topics, a guided model fits, and the framing below explains where.
• Very short documents (< ~50 words) → use a short-text model, not standard LDA.
• Highly technical or formulaic text, or a need for crisp category boundaries.

Start here: do you already know your topics?

One question settles most of the choice. Do you know, from theory, the topics you are looking for, or are you trying to find out what is in the corpus?

• No. You want to discover what is there. Impose as little structure as possible and read what emerges. Use LDA, or HDP if you would rather not fix the number of topics in advance.
• Yes. You have concepts and want to measure them. Name each topic with a few seed words and let the model anchor a topic to them. This is the guided path, KeyATM or SeededLDA, and it buys measurement validity and reproducibility a post-hoc reading of LDA topics cannot. From there you can regress prevalence on covariates or model it over time within the same family.
• Yes, and you also care how the same concept is worded differently across groups. Use STM with a content covariate. Modeling the topic-word distribution as a function of a group is the one thing the guided models cannot do.

Discovery and measurement are not a ranking. They answer different questions, and strong papers often run a guided model for the concepts they theorized and an unsupervised model as a robustness check on what they might have missed.

Match the model to your question

Your question / data	Model	Why
"What themes are here?" (baseline)	LDA	Standard, fast, well understood
"Does theme prevalence depend on metadata?"	STM	Prevalence covariates with valid effect estimation; the social-science default
"Is the same theme worded differently by group?"	STM (content) / SAGE	Content covariates on the topic-word distribution
"Do themes co-occur?"	CTM / STM	Logistic-normal allows topic correlation
"How many themes are there?"	HDP	Infers K nonparametrically (a check on your K)
"How does theme vocabulary drift over time?"	DTM	Topics evolve across ordered time slices
"Do themes predict an outcome?"	SupervisedLDA / DMR	Response- or covariate-conditioned topics
"I already know the themes I expect"	KeyATM, SeededLDA	Seed words steer named topics for better validity and reproducibility (guided topics)
Tweets / survey answers / headlines	GSDMM, PT	Built for short, sparse documents

How far keyATM reaches, and where it stops

Within measurement, KeyATM is close to a complete toolkit on its own: keyword-anchored topics, prevalence regression on covariates, and a change-point model for time trends, all in one family (validated against the R package in the replication). Its boundary is its premise. Because it needs keywords, it does not discover unknown topics, so reach for LDA or HDP there. And it models how much each group discusses a topic, not how they word it, so reach for STM or SAGE when content is the question.

Why STM is the social-science default

For most published social-science work the answer is the Structural Topic Model. It is LDA's correlated cousin plus a regression layer, so a single fit gives you topics, their correlations, and how covariates move them, with proper uncertainty. The R stm package made it the field standard. topica gives you the same model (spectral init, prevalence and content covariates, estimateEffect, searchK, FREX) in Python, faster, with clustered standard errors and GLM links on top.

Justify it in the paper

State the model and why it fits your design in one or two sentences:

Because our research question concerns how immigration framing differs by party and shifts across the campaign, we use the Structural Topic Model (STM), which lets topic prevalence depend on covariates (party, week) and provides valid uncertainty for those effects via the method of composition.

A nonparametric (HDP) or alternative-model robustness check pre-empts the "why this model?" objection: "the substantive topics were stable when we instead used LDA / a different K."

→ Next: Choose and justify K.