For researchers
The book’s arguments, frameworks, and materials, for readers engaging it as scholarship
Most of this site supports readers working through the book. This page is for readers working with it: methodologists weighing its claims, applied researchers considering its frameworks for their own studies, and doctoral students deciding what kind of computational social scientist to become. (Teaching a course instead? See For instructors.)
The book is a textbook in form and a methods argument in substance. This page states the argument compactly enough to be examined, lists the instruments that can be lifted out of it, and points to the materials that back it.
What the book argues
Eight claims carry the book. Each is developed and given its evidence in the chapters cited; here they are compressed to the point of bluntness, so you can decide quickly whether they deserve your engagement, or your disagreement.
Computational social science is a methodology, not a toolkit. The book defines computational methods as approaches to inference that require writing and executing code, distinguished by the capacities that specification buys rather than by the presence of a computer. What organizes them is research design, in four stages: data management, concept representation, data analysis, and scientific communication, each constraining the options at the next. (Chs. 1, 12)
AI-assisted research is an environment problem with an epistemological root. Because large language models are opaque, they cannot be trusted the way human collaborators are trusted; what remains available is structuring the environment so AI contributions are constrained, traceable, and grounded in project materials. That is the Rules / Memory / Knowledge framework, and the book’s diagnosis follows from it: most disappointing AI performance is environmental disorganization, not model weakness. (Ch. 2)
Cheap data, costly validation. Digitization made data cheap to acquire without making it cheap to understand. Born-digital and repurposed data carry the biases of the platforms and institutions that generated them, and the ease of collection systematically obscures the effort validation requires. (Chs. 3–4)
Measurement is where epistemology meets code. Embeddings, networks, and topic models can each serve postpositivist hypothesis testing, constructivist exploration, transformative action research, or pragmatist problem-solving. The epistemological commitment is made in the measure, not the tool: deductively, inductively, or integratively. And computational measures fail first at the margins, which is where the book sends its validity questions. (Chs. 5–6)
Capacity is not purpose. A method’s capacities (iterating, modeling non-linear relationships, processing high-dimensional data) are separate from what a study asks it to do: describe, explain, predict, or integrate. Collapsing that distinction is how fit to data gets mistaken for fit to concepts. The same distinction grounds the book’s treatment of LLMs as annotators or synthetic respondents: benchmark-style evaluation shows a model reproduces patterns, not that it captures the mechanisms a social-scientific claim needs. (Chs. 7–8)
Scientific objectivity is a social achievement. It is maintained by communication structures that keep work open to scrutiny, not by individual rigor alone. Three properties operationalize this for computational work: pipelines should be inspectable, results reproducible, findings interpretable. An open pipeline is not one that is right; it is one whose wrongness is discoverable. (Chs. 9–10)
Quantitative is a verb, not a property. Treating methodological distinctions as properties of data rather than choices about observation is a category error identified in the 1960s, and both sides of the current AI debate commit it: critics who hold that models categorically cannot reach qualitative meaning, and enthusiasts who hold that they make qualitative coding obsolete. The productive question is whether a particular procedure captures a particular concept, for whom, and where it fails. (Ch. 12)
Integration, not replacement. Read against the case-study/statistics controversy and the mixed-methods movement, the computational transition is following a familiar arc: new capacities, enthusiasm, resistance, overselling, backlash, then resolution through integration. Methods survive when they answer questions researchers care about, and the “computational versus traditional” framing is heading for the same dissolution as its predecessors. (Ch. 12)
Running through all eight is a commitment the book treats as structural rather than supplementary: ethics as a dimension of every design decision, not a checklist stage (Ch. 11).
Frameworks you can take
Each of these is developed in the book as a named, usable instrument. They travel: none depends on a particular language, library, or model generation.
| Framework | What it gives you | Where |
|---|---|---|
| Four-stage research design | The organizing spine: data management, concept representation, data analysis, and scientific communication, with each stage constraining the next | Ch. 1 |
| Three researcher roles | Algorithmic methodologist, applied methodologist, computational social scientist: a typology for locating a contribution (and a career) | Ch. 1 |
| Rules / Memory / Knowledge | An architecture for AI-assisted projects: rules configure what an agent may do, memory preserves what happened and why, knowledge grounds reasoning in project commitments | Ch. 2 · templates |
| 3S framework | Source, structure, standardization: three questions that surface what a dataset actually commits a study to | Ch. 3 |
| Deductive / inductive / integrative measurement | Locates any computational measure’s epistemological commitment; pairs with the marginal-cases check | Ch. 5 |
| Three-instrument coding sequence | Dictionary, LLM, and classifier coding of the same corpus, closed by a validity audit; instrument disagreement is the evidence | Ch. 6 |
| Capacities and epistemic purposes | Separates what a method can do from what a study asks of it; four purposes imply four validation standards | Ch. 7 |
| Inspectable / reproducible / interpretable | The three properties that let a computational pipeline count as evidence | Ch. 9 |
For doctoral students, the roles typology is the practical entry point. The three roles ask for different training, publish in different venues, and are evaluated by different standards. The book is explicit about which perspective it writes from: a computational social scientist who often works as an applied methodologist. Deciding early which role a dissertation is actually making a contribution to saves a great deal of misdirected effort.
The materials
Everything the book produced is on this site, and the honest limits of each item are stated where it lives.
- CSS Empirical Studies Database: 122 published computational studies, each annotated for the research-design stages it illustrates, with the annotations searchable and the whole table downloadable as CSV or BibTeX. It is a teaching resource, not a systematic review: studies were selected to span methods, disciplines, and epistemic purposes, coverage is uneven across stages, and it thins after 2023. Treat it as a well-documented convenience sample, not a sampling frame.
- Worked notebooks and data: three pre-executed notebooks (text representation, topic modeling, network analysis) with every output committed from a real run, plus the bipartite policy-to-scholarship network they build on.
- Project skeleton: the Rules / Memory / Knowledge directory structure as adaptable files, which is also the structure this book was itself written in.
Site text and code examples may be reused for teaching and research with attribution; linked tools and cited works carry their own licenses.
How to cite
Cite the book. The database, notebooks, and templates are its supporting materials rather than separately identified artifacts:
Ma, J. (forthcoming). Computational & AI-Assisted Methods for Social Sciences: A Research Design Primer. SAGE.
Final publication details will replace this entry when the book is in print. For citing a specific page of this site, see About.
Engaging
Corrections go through the errata page: an error, a misattributed claim, a figure that misstates its source. Disagreement is a different and welcome category: the claims above are stated plainly so that disagreement has something to grip.