Theoretical Sampling¶
Core Idea¶
Theoretical sampling is the structural pattern in which the next case to study is selected by what it would teach about the emerging theory, not by what it would say about the population. The selection criterion is informativeness for concept development — pick the cases that would maximally challenge, refine, or extend the current model — and the procedure is interleaved with analysis: each new case is chosen in light of what previous cases revealed, what gaps remain, and what boundary conditions are still untested. The catalogue of cases is not fixed in advance; it grows by what the developing theory needs. The essential commitment is closing the analysis-to-sampling loop: a researcher or system running theoretical sampling treats sample selection as a control variable, steered by the current state of the model — where is uncertainty highest, which categories are still under-saturated, which boundary cases would discriminate competing explanations.
The pattern requires three roles. There is the emerging model whose gaps and boundary conditions drive selection. There is the informativeness criterion that ranks candidate cases by their expected concept-development yield. And there is the saturation condition that signals when to stop — when the next case is expected to add no new structural insight, only confirmation. The dual pattern is representative sampling, which fixes the sampling frame in advance based on population properties and ignores what individual cases reveal mid-collection. The contrast is sharp and load-bearing: representative sampling optimizes for inference about a population and treats each draw as interchangeable evidence about that population, whereas theoretical sampling optimizes for the development of a model and treats each case as a deliberately chosen probe of the model's current weak points. The two have different optimal procedures and different stopping conditions, and they are often actively in tension, because the cases that best estimate a population — the typical ones at the center of the distribution — are frequently the least informative for a model that is already confident there. The structure reads as substrate-neutral: it names a control-loop shape, state-driven selection with a saturation stop, without imposing any value direction, which is why it recurs across qualitative and quantitative substrates alike under different names.
How would you explain it like I'm…
Push The Curious Button
Study What Teaches Most
Sample What The Theory Needs
Structural Signature¶
an emerging model whose gaps and boundary conditions drive selection — a pool of candidate cases — an informativeness criterion ranking cases by expected concept-development yield — a selection step interleaved with analysis — an analysis-to-sampling feedback loop — a saturation condition that stops when new cases add only confirmation
The pattern is present when each of the following holds:
- An emerging model. A developing theory, classifier, or hypothesis whose current uncertainties and under-saturated categories define what is worth learning next.
- A candidate case pool. A set of available observations — informants, unlabeled examples, feasible experiments, reachable sources — from which the next is drawn.
- An informativeness criterion. A ranking of candidates by their expected yield for concept development: category under-saturation, predictive uncertainty, expected information gain, discriminating power between competing hypotheses. Selection is steered by the model's weak points, not by representativeness.
- An interleaved selection step. Each case is chosen in light of what prior cases revealed; the catalogue grows by what the model needs rather than being fixed in advance.
- A closed analysis-sampling loop. Analysis informs the next selection and each selection feeds back into analysis. Batch-sampling-then-analyzing breaks the loop and is not theoretical sampling.
- A saturation stop. A structural-yield condition — not a sample-size quota — that halts when the next case is expected to add no new structural insight, only confirmation.
These compose into state-driven case selection with a saturation stop, the dual of representative sampling (which fixes the frame in advance for population inference). The two are often in tension, since the cases best for estimating a population are frequently least informative for a model already confident there.
What It Is Not¶
- Not representative sampling. See
sampling_representativeness. Representative sampling fixes the frame in advance and draws interchangeable cases to estimate a population. Theoretical sampling selects each next case for its informativeness to an emerging model, deliberately over-weighting the boundary. They are duals with opposite optimal procedures and opposite stopping rules. - Not selection bias. See
selection_bias. Selection bias is a flaw — non-representative cases distorting a population inference one intends to make. Theoretical sampling's non-representativeness is deliberate and principled, because the goal is concept development, not population estimation; the same skewed sample is a defect for one goal and the method for the other. - Not statistical inference. See
statistical_inference. Statistical inference draws conclusions about a population from a sample under a probability model. Theoretical sampling is a case-selection control loop for building a theory; it produces no population estimate and uses a yield-based saturation stop, not a confidence interval. - Not experimental design. See
experimental_design. Experimental design pre-specifies conditions, randomization, and sample size to license inference. Theoretical sampling grows the sample adaptively by what the model needs and stops at saturation rather than at a pre-set power calculation. - Not batch field research. Running all data collection first and analyzing second breaks the loop. Theoretical sampling requires analysis interleaved between selections, so each case is chosen in light of what prior cases revealed; a pre-fixed, never-revised sample is not theoretical sampling however the cases were chosen.
- Common misclassification. Treating a theoretically-sampled set as evidence about frequencies. The catch: the sample was boundary-weighted to develop a model, so it badly misrepresents population proportions. Using it for population claims is exactly the objective-mixing error; estimation needs a representative sample, not this one.
Broad Use¶
- Grounded-theory qualitative research. The canonical use: having coded one round of interviews, the researcher picks the next informant by what category needs saturating — a deviant case to test a mechanism, a contrasting institution to probe a boundary, a similar case to confirm a thin pattern.
- Active learning in machine learning. A model selects which unlabeled examples to query by expected information gain — uncertainty sampling, query-by-committee, expected model change — with the model's current state determining which next examples are most informative. Structurally identical, and the active-learning literature explicitly cites theoretical sampling as a conceptual ancestor.
- Adaptive experimental design and Bayesian optimization. The next experiment is chosen to maximize expected information about the parameters of interest given the posterior so far, used in drug discovery, materials science, and policy evaluation.
- Investigative journalism. Having uncovered a partial pattern, the reporter seeks the next source by what would confirm, extend, or break the emerging story — a whistleblower in a different region, a counter-example to a hypothesized motive, a document that would discriminate competing accounts.
- Security testing and red-teaming. The tester explores the attack surface guided by what the current threat model misses; fuzzing campaigns prioritize inputs by coverage gain over already-observed program paths.
- Clinical case selection. A physician investigating an emerging syndrome seeks the next patient case by what variant would discriminate between competing etiological hypotheses, not by what would be a representative patient.
- Software debugging. Developers seek failing cases that discriminate between competing fault hypotheses — a minimum reproducible example, an edge-case configuration — by current diagnostic state rather than by frequency in production.
Clarity¶
The pattern makes visible a distinction that statistical thinking tends to collapse: the goal of sampling is not always inference about a population. When the goal is concept development, the sampling criterion is informativeness for the developing theory, and representative samples are often actively bad, because they over-weight the center of the distribution where the model is already confident and starve the boundary where it is uncertain. Naming theoretical sampling exposes this as a deliberate, principled choice rather than as a methodological lapse: the non-random selection is the point, not a flaw to be apologized for. The frame also names the steering loop: analysis informs sampling, and sampling informs analysis. This naming has a sharp diagnostic consequence — a researcher who runs all interviews first and codes them second has broken the loop; they have done field research but not theoretical sampling, because the case selection was never steered by the developing model. The same is true of any system that pre-specifies its sample and never revises it. The frame further makes visible the saturation criterion, which answers the otherwise slippery question of when to stop. Saturation is not "I have enough data" in the sense of a sample-size target; it is "the next case is now expected to add no new structural insight." This is a structural-yield condition rather than a sample-size condition, and distinguishing the two prevents the common error of stopping when a quota is met rather than when the model has converged, or continuing past convergence in pursuit of a number.
Manages Complexity¶
The pattern collapses a family of cross-substrate problems — qualitative research, active learning, adaptive trials, investigative journalism, fuzzing, case-study design, diagnostic medicine, debugging — into one structural problem: use the current state of the model to steer the choice of the next observation, and stop when new observations no longer change the model. This is a substantial compression, because it lets a practitioner who has internalized the loop in one substrate recognize and run it in another by relabeling the model, the cases, and the yield criterion, rather than learning each domain's methodology as an unrelated craft. The compression also collapses a family of failure modes into one structural failure. Confirmation-biased case selection, premature saturation, sampling so broadly that no case is deeply analyzed, and sampling so narrowly that no boundary case is tested are, structurally, all the same failure: the analysis-sampling loop is open, mis-calibrated, or running on the wrong objective. Recognizing them as one failure with a few distinguishable causes makes them easier to detect and repair, because the practitioner checks the loop — is analysis actually happening between samples, is the informativeness criterion genuinely testing the model rather than confirming it, is the stopping rule a yield condition or a quota — rather than treating each pathology as a separate problem. The complexity of designing a data-collection strategy is managed by reducing it to the design of a control loop with three components and a small set of characteristic failure modes, all stated in substrate-neutral terms.
Abstract Reasoning¶
The pattern licenses a stable set of reasoning moves. Treat sampling as control, not given: which case to study next is a decision variable, not a fixed schedule, and recognizing this is the precondition for everything else. Drive selection by gap structure: ask what the current model predicts uncertainly or differently across cases, because those are the high-yield ones — the cases that would move the model rather than merely thicken it. Distinguish development from estimation goals: theoretical sampling is for concept development and representative sampling is for population estimation, and mixing them silently underwrites bad research, because a sample optimized for one is typically wrong for the other. Watch for saturation: when new cases yield no new structural insight — not merely no new variance — the loop has converged and should stop. Audit the steering policy: the loop can be biased if the informativeness criterion is mis-specified, so the reasoner must ask whether the cases being selected genuinely test the current model or only confirm it, since a criterion that rewards confirmation produces premature, false saturation. And interleave with analysis: the loop only works if analysis happens between samples, so batch sampling breaks the mechanism by severing the feedback that makes each selection responsive to the model's current state. These moves are stated in terms of the model, the criterion, and the saturation condition rather than in terms of any substrate, which is why they apply equally to a grounded-theory interview schedule and an active-learning query budget. The frame also clarifies an internal distinction: selection may be uncertainty-driven, as in active learning that queries the highest-uncertainty examples, or gap-driven, as in grounded theory that samples the under-saturated category — different operationalizations of the same structural commitment to steer by the model's current weak points.
Knowledge Transfer¶
The pattern transfers across substrates as a reusable control-loop discipline, and several of the transfers are explicit in the literature rather than merely analogical. Grounded theory to machine learning: the active-learning literature directly cites theoretical sampling as a conceptual ancestor, and the shared structure — model-driven selection of the next observation — is recognized as the same pattern under different vocabulary. Adaptive trials to policy evaluation: the Bayesian-optimization apparatus developed for clinical trials has been ported to A/B testing, policy pilots, and educational interventions under names like bandit-based experimentation, carrying the same state-driven selection and yield-based stopping. Investigative journalism to security research: red-team campaigns model their next-target selection on what the current attack-surface map needs to learn, mirroring investigative sourcing. Qualitative research to software debugging: the deliberate search for deviant cases in grounded theory maps onto the search for minimum reproducible examples in software, both being attempts to find the case that most discriminates competing hypotheses. And active learning to diagnostic medicine: emerging-disease surveillance protocols use information-gain criteria to prioritize which patient subsets to phenotype deeply, applying the same structural logic.
The role mappings that make these transfers reliable are direct. The emerging model maps to the developing grounded theory, the partially-trained classifier, the posterior over parameters, the reporter's working account of the scandal, the current threat model, the candidate diagnosis. The informativeness criterion maps to category under-saturation, predictive uncertainty, expected information gain, discriminating power between competing hypotheses, coverage gain over observed paths. The case pool maps to the available informants, the unlabeled examples, the feasible experiments, the reachable sources, the inputs to fuzz. The saturation condition maps to the point at which a new interview adds no new code, a new label no longer shifts the model, a new experiment no longer narrows the posterior. Because the loop and its three roles are shared, a concrete procedure transfers without modification: a grounded-theory researcher who selects the consultant most often deferred to precisely because that case would test whether a code applies upward or only downward in a hierarchy is, structurally, doing the same thing as an active-learning system that queries the human labeler on the images of highest predictive uncertainty — steering the next observation by the current state of the model. What transfers is the recognition that case selection can be a control variable rather than a fixed schedule, the three-role decomposition that makes the loop designable, and the distinction between a yield-based saturation stop and a sample-size quota — together with the standing audit that keeps the loop closed, the criterion honest, and the objective matched to whether the goal is to develop a model or to estimate a population.
Examples¶
Formal/abstract¶
Uncertainty-sampling active learning is theoretical sampling realized as a machine-learning algorithm, and the active-learning literature explicitly cites theoretical sampling as its conceptual ancestor — the same loop under different vocabulary. The emerging model is a partially-trained classifier whose current decision boundary defines what it does and does not yet know. The candidate case pool is a large set of unlabeled examples, each of which could be sent to a human annotator at a cost. The informativeness criterion is the crux and is exactly the frame's "steer by the model's weak points": rather than labeling a representative random sample, the algorithm ranks unlabeled examples by predictive uncertainty — the points nearest the decision boundary, where the model is least confident — because those are the cases that would most move the model, not merely thicken it. The interleaved selection step is the active-learning loop proper: query the most uncertain example, get its label, retrain, recompute uncertainties, query again — each selection chosen in light of what prior labels revealed. The closed analysis-sampling loop is what distinguishes this from ordinary supervised learning, where the training set is fixed in advance; breaking the loop (labeling a batch once, never re-querying) forfeits the entire advantage. The saturation stop is a yield condition, not a quota: halt when new labels no longer shift the decision boundary, which is the convergence analogue of grounded theory's "no new codes." The frame's audit applies with teeth — if the uncertainty criterion is mis-specified so that it rewards already-confident regions, the loop produces premature, false saturation, exactly the open-or-miscalibrated-loop failure the frame names. And the dual contrast is sharp: a representative sample would over-label the easy center where the classifier is already right and starve the boundary, which is why representativeness is actively bad for model development here.
Mapped back: The classifier is the emerging model, predictive uncertainty is the informativeness criterion, querying-then-retraining is the closed analysis-sampling loop, and "labels no longer move the boundary" is the yield-based saturation stop — theoretical sampling's control loop running in ML, citing its grounded-theory ancestor.
Applied/industry¶
Grounded-theory qualitative research is the originating home, and an organizational study makes the three roles and the saturation stop concrete. A researcher studying how decisions get made in a company begins with a few interviews, codes them, and finds a tentative pattern: junior staff defer upward. The emerging model is this developing theory, and crucially its gaps now drive what comes next — the open question is whether the deference code applies upward only or in both directions across the hierarchy. The candidate case pool is the set of available informants. The informativeness criterion is gap-driven rather than representative: instead of interviewing a demographically representative cross-section, the researcher deliberately selects the consultant most often deferred to, precisely because that case would test the boundary of the emerging code — a deviant or boundary case chosen for its discriminating power, not its typicality. The interleaved selection step is the grounded-theory discipline of choosing each next informant in light of the prior round's coding, so the sample grows by what the theory needs. The closed loop is load-bearing and is exactly where the method is often violated: a researcher who runs all interviews first and codes them second has done field research but not theoretical sampling, because selection was never steered by the developing model. The saturation stop is the structural-yield condition — halt when a new interview adds no new code, only confirmation — and the frame distinguishes this sharply from stopping at a pre-set quota of "twenty interviews," which is the common error. The identical loop, relabeled, runs in investigative journalism (select the next source by what would discriminate competing accounts of a motive) and in security red-teaming (select the next attack target by what the current threat model misses). A journalist hunting the counter-example source and a grounded theorist selecting the deviant informant are structurally making the same move — steer the next observation by the model's current weak point.
Mapped back: The deference theory is the emerging model, the most-deferred-to consultant is the boundary case selected by the informativeness criterion, choosing each informant from prior coding is the closed loop, and "no new codes" is the saturation stop — the canonical control loop the active-learning case mirrors algorithmically.
Structural Tensions¶
T1 — Model Development versus Population Estimation (Conflicting Objectives). Theoretical sampling optimizes for developing a model; representative sampling optimizes for estimating a population. The two objectives prescribe opposite case selections — the typical cases best for estimation are often least informative for a model already confident at the center. The failure mode is silently mixing the goals: using a theoretically-sampled set to make population claims (the boundary-weighted sample badly misrepresents frequencies), or using a representative sample for concept development (over-labeling the easy center, starving the uncertain boundary). The diagnostic is to ask which goal is actually in force: a sample optimized for one objective is typically wrong for the other, so the choice of selection criterion must follow from whether the aim is to learn a model or to estimate a distribution.
T2 — Informative Case versus Confirming Case (Criterion Honesty). Selection should target the model's weak points — the cases that would move it — but a subtly mis-specified criterion rewards cases that confirm what the model already believes. Confirmation feels like progress while adding no structural insight. The failure mode is a steering policy that selects agreeable cases, producing premature, false saturation: the model stops being challenged and its convergence is an artifact of biased selection, not genuine coverage. The diagnostic is to ask, of each selected case, whether it tests the current model or merely thickens it: where the informativeness criterion rewards confirmation, the loop has gone confirmatory, and the repair is to re-target selection at uncertainty and under-saturated categories rather than at cases expected to agree.
T3 — Saturation versus Quota (When to Stop). The stopping rule is a structural-yield condition — halt when new cases add no new insight — not a sample-size target. These are different stops that get conflated. The failure mode runs both ways: stopping at a pre-set quota ("twenty interviews") while the model is still moving, or continuing past genuine convergence to hit a number. Both substitute a count for a yield judgment. The diagnostic is to ask whether the next case is expected to change the model or only confirm it: saturation is "no new structural insight," reached when a new interview adds no code or a new label no longer shifts the boundary, and the stopping decision must track that yield rather than an arithmetic threshold set in advance.
T4 — Closed Loop versus Batch (the Interleaving Requirement). The mechanism depends on interleaving analysis with selection, so each case is chosen in light of what prior cases revealed. Collect everything first and analyze second, and the loop is severed — the selection was never steered by the developing model. The failure mode is doing field research or bulk labeling and calling it theoretical sampling: real work is done, but the feedback that makes each selection responsive to the model's state is missing. The diagnostic is to ask whether analysis actually happens between selections: if the sample was fixed in advance and never revised mid-collection, the loop is open, and the procedure has forfeited the entire advantage of state-driven selection regardless of how the cases were chosen.
T5 — Sample Broadly versus Sample Deeply (Coverage versus Depth). The loop must both test boundary cases (breadth across the model's uncertain regions) and analyze each case deeply enough to extract its insight. These compete for finite effort. The failure mode runs both ways: sampling so broadly that no case is analyzed deeply (cases pile up unmined), or so narrowly that no boundary case is tested (the model thickens at one point and its limits go unprobed). The diagnostic is to ask whether the current bottleneck is untested boundaries or under-analyzed cases: breadth without depth leaves the informativeness of each case unrealized, while depth without breadth leaves the model's boundary conditions unexamined, and the loop must balance the two rather than maximizing either.
T6 — Uncertainty-Driven versus Gap-Driven Selection (Operationalizing Informativeness). "Steer by the model's weak points" has two distinct operationalizations: query the highest-uncertainty cases (active learning) or sample the under-saturated category (grounded theory). They can diverge — the most uncertain case is not always in the most under-developed category. The failure mode is assuming one operationalization is the criterion itself, optimizing predictive uncertainty when the real gap is an unexplored category, or saturating categories while ignoring high-uncertainty discriminating cases. The diagnostic is to ask what "informative" means for this model's current weakness: where uncertainty and gap-structure point at different cases, the selection policy must be matched to which kind of weak point actually limits the model, since the two proxies for informativeness are not interchangeable.
Structural–Framed Character¶
Theoretical sampling sits just on the structural side of the middle of the structural–framed spectrum — a mixed-structural prime with a 0.4 aggregate, which the rationale calls borderline framed. Its core is a clean control-loop shape: an emerging model whose gaps drive selection, an informativeness criterion ranking candidate cases by expected concept-development yield, and a saturation condition that stops the loop. That state-driven-selection-with-a-saturation-stop structure is genuinely substrate-neutral, but the prime's grounded-theory origin and lexicon keep it from the structural floor.
One diagnostic is fully clean: evaluative_weight is 0, since the entry is explicit that the structure names a control-loop shape with no imposed value direction — theoretical and representative sampling are duals suited to different goals, not a rigorous and a loose version of one thing. The other four sit at the partial 0.5 mark. Human_practice_bound is 0.5 because, although the originating cases are human qualitative research, the identical structure runs in fully mechanical substrates — active machine learning selects the next label by expected model information gain and stops at a confidence plateau, and Bayesian optimization steers sampling toward acquisition-function maxima — so the pattern is not bound to human practice, with a documented citation lineage from sociology to ML. Vocab_travels is 0.5 because the grounded-theory lexicon (theoretical sampling, saturation, constant comparison) is recognized under entirely different vocabulary (acquisition function, query strategy, exploration) in the quantitative substrates, i.e. translation is required. Institutional_origin is 0.5 because the prime was formalized as a grounded-theory method, a soft disciplinary origin. Import_vs_recognize is 0.5 because invoking the prime half-imports a methodological frame and half-recognizes an analysis-steers-selection control loop already present. This profile — value-neutral with genuine mechanical-substrate carry, but tinted by a grounded-theory lexicon and origin — is exactly what the mixed-structural label with its 0.4 aggregate records, sitting close to the framed boundary without crossing it.
Substrate Independence¶
Theoretical sampling is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its breadth is good: the model-driven case selection with a saturation condition recurs across grounded-theory qualitative research, active learning in machine learning, adaptive experimental design and Bayesian optimization, investigative journalism, security testing and red-teaming, clinical case selection, and software debugging — and the active-learning and Bayesian-optimization cases run in fully mechanical substrates where a partially-trained model selects the next observation by expected information gain with no human in the loop. The signature is highly relational — an emerging model whose gaps drive selection, a candidate pool, an informativeness criterion, an interleaved selection step, a closed analysis-sampling loop, a saturation stop — stated medium-neutrally, so the three-role decomposition and the yield-versus-quota distinction carry intact. Transfer is concrete and unusually well documented: there is a direct citation lineage from grounded-theory sociology to the active-learning literature, which names theoretical sampling as a conceptual ancestor, and the role mappings (emerging model, informativeness criterion, saturation condition) hold from a deviant-informant interview to an uncertainty-sampling query. What holds it a notch below 5 is the residue of its grounded-theory origin: the home lexicon (theoretical sampling, saturation, constant comparison) is recognized only under translation in the quantitative substrates (acquisition function, query strategy), and the prime takes its formalization from one methodological tradition. Recognized rather than translated across its range with genuine mechanical-substrate carry, it earns a composite 4.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 4 / 5
- Transfer evidence — 4 / 5
Neighborhood in Abstraction Space¶
Theoretical Sampling sits among the more crowded primes in the catalog (38th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Sampling, Inference & Statistical Bias (12 primes)
Nearest neighbors
- Sampling (Representativeness) — 0.76
- Selection Bias — 0.71
- Overfitting — 0.71
- Statistical Inference — 0.71
- Inductive Reasoning — 0.71
Computed from structural-signature embeddings · 2026-06-14
Not to Be Confused With¶
The defining confusion — and the embedding-nearest neighbor at similarity 0.94 — is with sampling_representativeness, which the prime treats as its explicit dual. The two are easy to conflate because both are about which cases to collect, but they optimize for opposite objectives and prescribe opposite procedures. Representative sampling aims at population inference: it fixes the sampling frame in advance so the sample's distribution mirrors the population's, treats each draw as interchangeable evidence about that population, and stops at a sample size set by the desired precision. Theoretical sampling aims at model development: it selects each next case for what it would teach the emerging theory, deliberately over-weighting the boundary and the under-saturated category where the model is weakest, treats each case as a chosen probe rather than an interchangeable draw, and stops at saturation — when new cases add no new structural insight. The tension is not incidental but structural: the cases that best estimate a population are the typical ones at the center of the distribution, which are frequently the least informative for a model already confident there. So a representative sample is often actively bad for concept development (it over-labels the easy center, starves the uncertain boundary), and a theoretical sample is actively bad for population estimation (its boundary-weighting misrepresents frequencies). Recognizing them as duals, rather than as a "rigorous" and a "loose" version of one thing, is what lets a practitioner pick the criterion matched to whether the goal is to learn a model or to estimate a distribution.
A second confusion, more insidious because it inverts a value judgment, is with selection_bias. From a statistical-inference standpoint, theoretical sampling looks exactly like selection bias: cases are chosen non-randomly, on purpose, in a way that does not represent the population. The distinction is the objective against which the selection is judged. Selection bias is a defect — non-representative selection that distorts a population inference the analyst intends to draw, smuggling in a skew that corrupts the conclusion. Theoretical sampling's non-representativeness is deliberate, principled, and correct for its goal, because the analyst is not drawing a population inference at all; they are developing a model, for which boundary-weighted selection is precisely right. The same physical sample — say, an over-representation of deviant cases — is a bias if you meant to estimate frequencies and the method if you meant to test a theory's boundary. The error the distinction guards against is bidirectional: apologizing for theoretical sampling as if it were a methodological lapse (when its non-randomness is the point), or, more dangerously, letting a theoretically-sampled set quietly become the basis for a population claim (where its deliberate skew is now a genuine selection bias). The objective determines which it is.
A third confusion is with statistical_inference and its apparatus (including experimental_design), because all three concern how to collect and reason from data. The divide is between producing an estimate about a population and steering the construction of a model. Statistical inference draws calibrated conclusions about a population from a sample under a probability model, with machinery — confidence intervals, power, significance — built around representativeness and pre-specified design. Theoretical sampling produces no population estimate and uses none of that machinery; its output is a developed theory or classifier, its stopping rule is yield-based saturation rather than a power calculation, and its sample grows adaptively rather than being fixed in advance. Experimental design in particular pre-commits to conditions, randomization, and sample size precisely to protect an inference, whereas theoretical sampling deliberately revises its selections mid-collection in response to what the data reveal — a move that would invalidate a classical inferential design. The adaptive cousins of statistics (Bayesian optimization, bandit experimentation) are where the two genuinely meet, and indeed those are recognized descendants of the same control-loop logic; but classical fixed-design inference and theoretical sampling are distinct, and importing one's standards into the other (demanding pre-registration and representativeness of a grounded-theory study, or reading saturation as if it licensed a population estimate) misapplies the wrong yardstick.
For a practitioner, the single decisive question routes all four distinctions: is the goal to estimate a population or to develop a model? If estimation, use representative sampling, guard against selection bias, and reason with statistical inference and pre-specified design. If model development, use theoretical sampling — steer by the model's weak points, accept and intend the non-representativeness, interleave analysis between selections, and stop at saturation, not at a quota or a confidence threshold. The gravest error is letting a sample built for one objective be used to answer the other.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.