Further Research Routes
This chapter is part of a book in active development and has not yet been through the author’s review. Content may change as the review advances.
Marisol studies how cities regulate housing. Her question is causal: did a new eviction-notice rule, adopted by one city on a known date, change how many tenants were evicted? She cannot assign the rule to anyone, so she declared the observational causal pathway at Studio 5. The observational causal lesson names the comparison her situation offers but does not teach it: a neighboring city that kept the old rule, observed month by month before and after the same date. That comparison has a warrant and a known way of failing, and Chapter 41 teaches both.
What a further route is
A further route is a data strategy and answer strategy that a question may call for once the compass has placed it and you have declared a pathway, taught here as its own chapter. Marisol’s question is still causal, it is still about the tenants of the city that changed its rule, and her pathway has not changed. The route adds the machinery that pathway needs when a date, a rule, or a cutoff did the assigning: difference-in-differences (comparing how much each city changed), a discontinuity (a jump in outcomes at a cutoff), or an interrupted time series (a break in one place’s long trend).
Every route is classified by the two questions you answered in Studio 2: which kind of answer your question wants (descriptive or causal), and how far it should reach (the data at hand, a population, or unseen cases). The question sets the kind, and choosing a route never changes it. What a route changes is what your design can license. Whenever an answer reaches past what you directly observed, or claims a cause, the route still names the crossing that licenses it. Sampling licenses a population claim. Prediction-time honesty plus a held-out check licenses a claim about unseen cases. Assignment or identification licenses a causal claim. Your declared model, inquiry, data strategy, and answer strategy still structure the design (Blair et al. 2023). A causal question whose route cannot identify the effect stays causal: you write “causal, currently unidentified” and say what evidence would change that.
Two of the routes stretch the book’s usual answer. A qualitative inquiry returns a bounded account of meanings, practices, or processes, where every claim traces back to what people said or wrote. An evidence synthesis returns a map or a pooled estimate built from other people’s studies. Its kind follows your review question, and it inherits each study’s warrant. Combining studies never creates a causal license that none of them had.
The six routes
Read the row that sounds like your project. The milestones stay the same, and the last column shows the shape their products take on each route. The kind column describes the questions each route usually serves. It is not a rule: a survey experiment, an audit, or a review answers whatever kind your question asks for, and its chapter shows which pathway that puts you on.
| Route | Kind and reach it usually serves | Pathway it extends | Data and answer strategy in your Research Contract | What Studios 5–11 produce on it |
|---|---|---|---|---|
| Chapter 41: Natural experiments | Causal: the effect on the units the rule reached, in the periods you observed, or only at the cutoff for a discontinuity; a population only when those units were sampled from it | Observational causal | Records on both sides of an assignment you did not make, before and after a date or either side of a cutoff; a comparison of changes, a jump at the cutoff, or a break in a trend | A timing and cutoff document quoting the rule from its primary source; in the stress test, pre-period checks for a comparison of changes, a check that units did not sort around a cutoff, and a falsification test whose failing result you wrote down in advance |
| Chapter 42: Survey experiments | Set by your question. Descriptive when the experiment measures what respondents hold: a prevalence from a list experiment, stated preferences from a conjoint. Causal when you ask what a message, a frame, or an attribute does to their answers. Reach: your respondents, a population only when your sampling licenses it | Experimental descriptive when your question asks what respondents hold; experimental causal when it asks what a message, frame, or attribute does. The question decides, not the instrument | A questionnaire in which you randomize a wording, a vignette, an attribute, or a list; a difference in mean counts, AMCEs or marginal means, or a difference in mean support, each with an interval | The instrument written word for word with its randomization plan, a simulation of the interval your sample will give, a plan for failed attention and manipulation checks, and a permission record |
| Chapter 43: Audit studies | Usually descriptive: the net response gap for these decision makers, or this AI system build, in this period; a defined list only when you sampled from it. A question about the effect of the signal itself reads the same contrast as causal | Experimental descriptive; experimental causal when your question asks for the signal’s effect | Cases identical except for one attribute, sent to decision makers or to a model; the response gap, with an interval that treats the decision maker, the posting, or the case template as the case | For an AI audit, a run record (model build, date, every setting, and a scoring rule written before the first run); for an audit of people, a permission record, and usually a reanalysis of published open data |
| Chapter 44: Text as data | Descriptive: the corpus you hold, with the random human-coded subset as its crossing when humans read only part of it, or a population only with a defended corpus frame; unseen documents, the prediction position, when you train a classifier for new ones. A causal question makes text the outcome of a causal design, which must carry the identification | Observational descriptive, and prediction for classifiers | A corpus with a written frame and a codebook; human coding, or model labels corrected with a random human-coded subset, reported with an interval | A written corpus frame with its likely gaps, the codebook, an agreement check between two coders, a validation table, and a corrected share with its interval |
| Chapter 45: Qualitative inquiry | Descriptive accounts of meanings, practices, or processes in the cases you studied, the data at hand; transferable only as far as your description lets a reader judge, never a prevalence claim; a causal question stays causal, currently unidentified, unless a within-case design such as process tracing carries it | Observational descriptive | Interviews or documents from purposively chosen cases; thematic analysis with an evidence trail | A case-selection record, an interview guide, a consent and anonymization record, and an evidence trail from excerpt to code to theme to claim |
| Chapter 46: Evidence synthesis | Set by your review question. The included studies’ designs set what the answer can license: pooled randomized experiments can support a causal claim, while a pile of observational comparisons leaves a causal question currently unidentified. Reach: for a map, every eligible study, warranted by your search log; a pooled effect describes settings like those studied and reported | The pathway your review question calls for, applied to the included studies and run as a multi-stage design (Hybrid and complex designs) | A protocol with eligibility fixed before you search, a search log, and a map of the studies or a small meta-analysis | A review protocol, the search log, a flow diagram, a screening record with its second check, an extraction form, and a forest plot if you pool |
In each route chapter, the It is your turn section walks you through writing the route’s version of your Research Contract lines.
What AI does on these routes
A language model can draft a codebook for you to revise, suggest search strings for a review, or write the code that randomizes a survey experiment. Each route also has its own way for the tool to fail. It can call a plain before-and-after comparison “difference-in-differences”, return a supporting quote that appears nowhere in the document, or treat fifty runs of one résumé prompt as fifty different résumés. Each chapter names its failure and a check that can catch it. Which comparison you defend, which documents count, and whose accounts you gather stay your decisions. AI is your arm and your research assistant, not your brain.
Beyond this book
Some designs ask for more than one project can supply without a mentor. Instrumental variables, synthetic control, difference-in-differences with staggered adoption, ethnography, qualitative comparative analysis, network analysis, and Bayesian analysis each rest on assumptions or interpretive skills that take supervised practice to defend. The appendix Further Designs: A Field Guide gives each one a short entry: what it answers, what licenses the answer, what it demands beyond this book, what AI can do for it, and where to learn it. Use it to recognize a design your question seems to need, then decide whether to learn it with a mentor or keep your project on a route taught here.