Introduction and Overview

Most of what you have done in HSCI 841 so far has been variable-oriented in the loose sense: you have looked at codes, themes, and concepts across transcripts and asked how they are patterned. Lesson 11 takes a different turn. The three methods in this lesson — analytic induction, qualitative comparative analysis (QCA), and ethnographic decision modelling — are case-oriented in a strong and explicit sense. The unit of analysis is the case, the analytic move is comparison across cases, and the goal is not theme description but defensible causal or quasi-causal inference from a small number of cases.

This is the part of the qualitative toolkit that does the work HSCI 230, 341, and 410 do not do well. Your earlier epidemiology training was built around large-N statistical inference: you estimated effects of variables on outcomes net of other variables, you tested hypotheses with confidence intervals, you screened for confounders. That apparatus assumes you have enough cases for the central limit theorem to do its work and that the causal structure of the world is best approximated by linear, additive, and probabilistic relationships. For many public-health questions that assumption is fine. For others — especially questions about decision processes, configurations of conditions, and outcomes that depend on combinations of factors rather than net effects — it is wrong, and the methods in this lesson are the methodologically defensible alternatives.

Section 1 takes up the oldest of the three: analytic induction, formulated by the Polish-American sociologist Florian Znaniecki in 1934, demonstrated by Alfred Lindesmith on opiate addiction in 1947, and applied by Donald Cressey to financial trust violation in 1953. Analytic induction is the ancestor of all the case-oriented qualitative-causal methods that follow. Its central claim is methodologically aggressive: a defensible qualitative hypothesis must account for every case in the dataset, and a single counter-case is enough to force revision. We will reconstruct the logic carefully, work it through a worked example on the loneliness dataset, and confront W. S. Robinson's 1951 critique, which most qualitative methodologists have spent the last seventy years either accepting or working around.

1.1 The Procedure

Bernard, Wutich, and Ryan (2017, pp. 339–341) describe analytic induction as a six-step iterative procedure. The version below is closely faithful to theirs and lines up with how Lindesmith (1947) and Cressey (1953) actually worked.

1. Define the phenomenon to be explained. The definition has to be tight enough to allow you to decide whether any given case is or is not an instance. “Loneliness” is too loose to start with. “Sustained loneliness lasting at least six months that the participant explicitly names as such” is tight enough.
2. Formulate a hypothesis about the phenomenon. The hypothesis names the conditions under which the phenomenon occurs (or does not occur). Lindesmith's initial hypothesis was that opiate addiction develops when a person uses opiates regularly and experiences withdrawal.
3. Examine one case. Determine whether the hypothesis fits.
4. If the hypothesis fits, examine another case. Continue until a case is found that does not fit. The fit cases do not confirm the hypothesis; they merely fail to disconfirm it. The work of the method is in the disconfirmations.
5. When a non-fit case is found, do one of two things: either revise the hypothesis to accommodate the new case, or redefine the phenomenon to exclude it. Both are legitimate moves but they are not equivalent: revising the hypothesis broadens explanatory scope, while redefining the phenomenon narrows it. Cressey moves repeatedly between the two.
6. Continue until all cases in the dataset fit the hypothesis under the current definition. The terminal state is a definition-plus-hypothesis pair that explains 100% of the cases you have.

The procedure produces, in principle, a statement of necessary and sufficient conditions. If the hypothesis is true and all cases fit, then whenever the conditions hold the phenomenon occurs (sufficiency) and whenever the phenomenon occurs the conditions hold (necessity). This is what makes analytic induction methodologically aggressive: most quantitative methods aim only at probabilistic association, not necessary-and-sufficient logical relationships.

1.2 Znaniecki (1934): The Original Statement

Florian Znaniecki formulated analytic induction in The Method of Sociology (1934) as a sociological counterpart to what he called “enumerative induction” — the statistical inference that produced means and rates. Znaniecki's complaint was that enumerative induction tells you about populations but not about phenomena: a regression coefficient describes how cases distribute around a mean, but it does not tell you what a phenomenon is or what it requires to occur. He argued that sociology's task was to produce universal generalizations — statements that hold for every instance of the phenomenon — and that universal generalizations require a case-by-case examination logic incompatible with statistical sampling.

The methodological move Znaniecki made is now uncontroversial in qualitative methodology but was provocative at the time: he claimed that a single well-analyzed case could refute a universal generalization, the way a single black swan refutes the claim that all swans are white. He took this from Karl Popper's contemporaneous work on falsification (which he engages indirectly) and from the older Baconian tradition of seeking instantiae crucis, crucial instances. The link to Popper is important because it positions analytic induction not as a soft alternative to quantitative work but as a logically more demanding cousin: where quantitative analysis tolerates noise, analytic induction tolerates none.

1.3 Lindesmith (1947): Opiate Addiction

The most famous application of analytic induction is Alfred Lindesmith's Opiate Addiction (1947, expanded as Addiction and Opiates in 1968). Lindesmith began with the prevailing hypothesis of the 1940s: that opiate addiction was a function of the pleasurable euphoria opiates produce, and that addicts continued using because they sought the high. He interviewed addicts — a fully qualitative dataset by today's standards — and found cases that did not fit. Some people who experienced significant euphoria from opiates did not become addicted. Some hospitalized patients who received heavy opiate doses for medical reasons did not become addicted even when they experienced withdrawal symptoms.

Lindesmith revised the hypothesis. His new claim was that addiction develops when a person uses opiates, experiences withdrawal, recognizes the withdrawal as caused by the absence of the drug, and uses the drug specifically to relieve withdrawal. The hospitalized patients who did not become addicted had not recognized the withdrawal-drug connection, because they did not know what they were receiving. The cognitive recognition step — the conscious linking of withdrawal to the absence of the substance — was, Lindesmith argued, the necessary and sufficient condition.

What is methodologically significant is not whether Lindesmith was right (subsequent neuroscience suggests the picture is more complicated), but that his procedure was disciplined: every case had to fit the hypothesis, and every case that did not fit forced revision. He famously claimed to have examined hundreds of cases and that the hypothesis withstood all of them, after several rounds of revision. The work took the form of an actual sociological monograph in which the case-by-case progression was traceable in the text.

1.4 Cressey (1953): Other People's Money

Donald Cressey's Other People's Money (1953) applied analytic induction to financial trust violation — what we would now call white-collar embezzlement. Cressey interviewed 133 incarcerated trust violators in three federal prisons. His initial hypothesis was that trust violators were characterized by unshareable financial problems. Some cases fit and some did not. Through repeated revisions, Cressey arrived at a three-part necessary-and-sufficient condition: trust violation occurs when (a) the person has a non-shareable financial problem, (b) the person has knowledge of how trust violation could solve the problem, and (c) the person can rationalize the violation as something other than trust violation.

The rationalization step is what made the work famous in criminology. Cressey distinguished embezzlers who told themselves “I am borrowing this money and will pay it back” from embezzlers who told themselves “the employer owes me this anyway.” The rationalizations were not post-hoc; they were causally implicated in the violation, because cases that lacked an available rationalization did not proceed to trust violation even when problems and knowledge were present.

Cressey's method is a model of the form. The three conditions are jointly necessary and sufficient; if any one is absent, no violation occurs; if all three are present, violation does occur. The methodological lesson for your capstone is that analytic induction is most powerful when applied to a focused outcome with a small number of candidate conditions, and when the analyst is willing to revise both the hypothesis and the case-set as the analysis proceeds.

1.5 Necessary, Sufficient, Necessary-and-Sufficient

Because the rest of the lesson depends on this vocabulary, take a moment with it. A condition X is necessary for outcome Y if Y never occurs without X. Smoking is necessary for nothing in the strict sense — people get lung cancer without smoking. Female sex is necessary for ovarian cancer. The classical epidemiology causal criterion of “temporality” (HSCI 341 L12) is a necessary condition for causation.

A condition X is sufficient for outcome Y if X's presence guarantees Y. Decapitation is sufficient for death (the example is unsubtle for a reason). In public health, very few real-world conditions are individually sufficient; this is one reason regression-based causal inference is so dominant — it estimates effects of single conditions without claiming sufficiency. Most of what regression analyzes is INUS conditions in Mackie's (1965) terminology: insufficient but non-redundant parts of unnecessary but sufficient configurations.

A condition X (or set of conditions) is necessary and sufficient for Y if Y occurs whenever X is present and never when X is absent. Analytic induction aims here. Cressey claims his three conditions are jointly necessary and sufficient for trust violation; Lindesmith claims the cognitive-recognition condition is necessary and sufficient for opiate addiction. This is a much stronger logical claim than “the odds ratio is 4.2 with a 95% confidence interval of 2.1 to 8.6,” which is what HSCI 410 trained you to produce.

1.6 Worked Example: Loneliness and Professional Help-Seeking

Let us walk a small analytic induction through the loneliness dataset. The phenomenon to be explained is professional help-seeking for loneliness, defined operationally as “the participant describes consulting a clinician, therapist, counsellor, or other paid professional explicitly because of loneliness or its symptoms.”

Initial hypothesis (H1): Loneliness leads to professional help-seeking. Read across the 20 transcripts. Maya (P01) does not seek professional help despite reporting significant loneliness; she relies on her roommates and an online community. Counter-case found at case 1.

Revision to H2: Loneliness leads to professional help-seeking when the loneliness reaches a clinical threshold (sleep disruption, functional impairment, suicidal ideation). Diana (P04), an early-career professional with documented insomnia and intrusive thoughts, does not seek professional help; she explicitly describes not wanting to be “the kind of person who pays someone to listen to them.” Counter-case at case 4.

Revision to H3: Loneliness leads to professional help-seeking when (a) it reaches a clinical threshold AND (b) the participant does not hold strong cultural prohibitions against professional help. Marcus (P08), an older man who has lost his wife, describes severe loneliness and seeks help through his church first, his GP second, and is referred to grief counselling. He fits H3. Margaret (P12), a working-class single mother, describes severe loneliness and has no cultural prohibitions but explicitly says she cannot afford counselling and the wait-list for publicly funded service is 9 months. Counter-case at case 12.

Revision to H4: Loneliness leads to professional help-seeking when (a) it reaches a clinical threshold AND (b) the participant does not hold strong cultural prohibitions against professional help AND (c) the structural gate-keeping (cost, wait-list, geography, language) is permeable. Continue across the remaining transcripts. The hypothesis withstands the next several but fails on Amira (P15), the Syrian refugee, who meets all three conditions but does not seek help because she does not trust that the system will treat her confidentially — she fears immigration consequences.

Revision to H5: Loneliness leads to professional help-seeking when (a) clinical threshold, (b) cultural permission, (c) permeable structural gates, AND (d) the participant trusts the system not to inflict secondary harms.

This is how analytic induction proceeds. By the time you have moved through 20 transcripts you have a four- or five-part conjunctive hypothesis that fits every case. The conditions are jointly necessary and sufficient for the phenomenon as you have come to define it. You may, alternatively, decide at some point to redefine the phenomenon — for example, to restrict it to participants for whom professional help is even imaginable as an option, which would let you drop the trust condition. Either move is defensible; both should be documented in the audit trail.

1.7 Robinson's Critique (1951)

The most influential critique of analytic induction is W. S. Robinson's 1951 paper “The Logical Structure of Analytic Induction.” Robinson argued that analytic induction does not actually produce what it claims to produce, for two reasons.

First, analytic induction only examines cases in which the phenomenon occurs. Lindesmith interviewed addicts; he did not interview a comparison group of non-addicts. Cressey interviewed trust violators; he did not interview a matched sample of people in similar positions who did not violate trust. Without negative cases — cases where the conditions are present but the phenomenon does not occur — you cannot establish sufficiency. You can only establish necessity (because every case of Y has X). The method, as practised, gives you necessary conditions but not sufficient ones.

Second, the revision step is unconstrained. When you find a counter-case, you can always revise the hypothesis or redefine the phenomenon to accommodate it; there is no logical limit on how baroque the resulting formula can become. A sufficiently committed analyst can always tune the hypothesis to fit any finite case-set. The resulting statement is therefore not a universal generalization but a description of this particular case-set, which is exactly the criticism Znaniecki had levelled at enumerative induction.

The standard contemporary response, which Bernard, Wutich, and Ryan endorse, is twofold. (1) Treat analytic induction as a heuristic for hypothesis generation, not as a logically deductive proof procedure. The hypothesis you produce is a candidate for further testing — ideally including negative-case sampling that addresses Robinson's first objection. (2) Sample for non-instances as well as instances. (3) Pre-register the hypothesis revision rules to constrain the second concern.

Introduction and Overview

Qualitative comparative analysis (QCA) is the second case-oriented technique you will learn this week. It was developed by Charles Ragin in The Comparative Method (1987) and elaborated in Fuzzy-Set Social Science (2000) and Redesigning Social Inquiry (2008). For a hands-on introduction across csQCA, mvQCA, and fsQCA see Rihoux & Ragin (2009). QCA's central insight is that many real-world causal stories are not about net effects of single variables (the regression idiom) but about combinations of conditions that are jointly sufficient for an outcome. A risk factor that is irrelevant on its own can be essential in combination with another. A protective factor that works in one configuration can be neutralized in another. QCA gives you a disciplined, Boolean-algebra-based way to find and report those configurations.

QCA occupies an interesting middle ground between the two traditions HSCI 841 has been navigating all term. It is qualitative in that the unit of analysis is a case (a person, a clinic, a region, a policy regime) characterized by a vector of qualitative attributes. It is comparative in that the analytic move is to compare cases across configurations. It is formal in that the analytic engine is Boolean algebra and the output is a set of minimized sufficiency formulas. And it is computational in that contemporary practice uses software (the R package QCA, the standalone fsQCA software, or the R package SetMethods) to do the minimization. It is the closest qualitative method comes to producing the kind of formal output an epidemiology audience recognizes as “a result.”

This section walks through the QCA logic, the truth table, Boolean minimization, the distinction between necessity and sufficiency in Ragin's terms, the consistency-and-coverage metrics that QCA uses instead of p-values, and the crisp-set vs. fuzzy-set distinction. It connects QCA explicitly to your HSCI 341 L12 training on confounding and causal inference, and it sets up the QCA option for your Week 11 capstone milestone.

2.1 What QCA Does That Regression Cannot

To motivate QCA, contrast it with what you know from HSCI 410. A logistic regression of a binary outcome Y on three predictors X1, X2, X3 yields three coefficients, each representing the net effect of one variable holding the others constant. The model assumes the effects are additive on the log-odds scale (unless you include interaction terms, in which case it assumes the interactions are additive on top of the main effects). It treats every case as a draw from a population with that net-effect structure.

Many causal structures in public health do not look like that. Consider obesity policy: a school nutrition program might reduce childhood obesity only when (a) the local food environment supports it, AND (b) parental income is above a threshold, AND (c) the school has stable administrative leadership. The program alone has no effect; the food environment alone has no effect; income alone has no effect. The three together produce the outcome. A logistic regression with three main effects and three two-way interactions and one three-way interaction can in principle capture this, but only with enough cases to estimate eight coefficients (plus an intercept), and only if you remembered to put the interactions in.

QCA approaches the same problem differently. Each case is represented as a configuration: a vector of present/absent (1/0) values on the conditions. The analyst tabulates how cases distribute across configurations and asks: which configurations consistently produce the outcome? The answer is a Boolean expression. In our hypothetical example, the answer might be: FoodEnv · Income · Leadership → OutcomeReduction. There is no main effect of any single condition. There is one sufficient configuration. QCA finds it; regression cannot, without prior knowledge of the interaction structure.

Ragin (2008) names four features of causal structures that QCA handles well and regression handles poorly:

• Conjunctural causation: the outcome depends on combinations of conditions, not on single variables in isolation.
• Equifinality: there is more than one combination of conditions sufficient for the outcome (multiple paths).
• Causal asymmetry: the conditions producing presence of the outcome are different from the conditions producing its absence (the negation is not just the inverse).
• Limited diversity: there are not enough cases to populate every theoretically possible configuration, so the method must explicitly distinguish what the data show from what they cannot show.

The first three features are essentially invisible to a standard regression unless you specify them in advance. The fourth feature — limited diversity — is what makes QCA usable with small-N datasets where regression is not.

2.2 The Truth Table

The fundamental data object in QCA is the truth table: a row for every theoretically possible configuration of conditions, with a column indicating how many cases of each kind exist and whether they produce the outcome. With k binary conditions there are 2^k rows. With 4 conditions there are 16 rows; with 5 there are 32. Most empirical truth tables are sparse — many rows have zero cases, reflecting that the social world does not produce every combination.

Consider a toy QCA on the loneliness dataset. Let us code each of the 20 transcripts on four binary conditions and one binary outcome:

• B — bereaved (lost a primary attachment figure in the past 5 years): 1/0
• L — lives alone: 1/0
• I — immigrant (born outside Canada, arrived in adulthood): 1/0
• C — has a current caregiving role (children at home, elder care, partner care): 1/0
• Y (outcome) — describes loneliness as existential rather than situational (a feature of one's being-in-the-world rather than a fixable circumstance): 1/0

Each transcript is coded as a single row of zeros and ones. The 16 possible configurations are then summarized in a truth table that shows, for each configuration, how many cases of that kind exist and whether the outcome occurs.

The rows with zero cases are logical remainders — theoretically possible configurations that do not appear in the data. They are not the same as rows with cases that produced no outcome; they are rows we cannot speak to. Ragin's QCA distinguishes the conservative solution (ignoring remainders) from the parsimonious solution (treating remainders as freely available for simplification) and the intermediate solution (allowing simplification only with remainders consistent with theoretical expectations). The three solutions are reported alongside one another.

2.3 Boolean Minimization

The truth table summarizes the data, but the goal of QCA is a compact Boolean expression of the configurations sufficient for the outcome. The procedure that produces this is called Boolean minimization, and it works through pairwise comparison and the application of one logical rule:

If two configurations differ on exactly one condition and produce the same outcome, that condition is irrelevant to the outcome in the presence of the other shared conditions, and the two configurations can be combined into a single shorter expression.

Concretely: configurations ABC and ABc (capital = present, lowercase = absent) both produce Y. They differ only on C. Therefore the simpler expression AB is sufficient for Y: it does not matter whether C is present or absent, as long as A and B are. This is the Quine-McCluskey algorithm in disguise, the same algorithm electrical engineers use to minimize digital logic circuits.

In our loneliness truth table, the rows that produce Y = 1 (existential loneliness) are rows 1, 2, 4, 8 (and possibly others; this is a partial illustration). Applying the minimization rule: rows 1 and 2 differ only on I, both produce Y, so we get B · L · c regardless of I. Rows 4 and (a hypothetical row with the same pattern but B = 1) differ only on B, both produce Y, so we get L · I · c. Combining further: the minimized sufficiency formula might read (B · L) + (L · I), which translates to: existential loneliness occurs when (a) one is bereaved and lives alone, OR (b) one lives alone and is an immigrant. Two paths, both not involving a current caregiving role.

That is what QCA produces. Note three things. (1) The result is a Boolean expression, not a regression coefficient. (2) It identifies multiple sufficient paths — this is equifinality, and it is the feature that distinguishes QCA most sharply from regression. (3) It distinguishes “present” from “absent” for each condition; the lowercase c in the formula tells you that the absence of caregiving is itself a part of the sufficient configuration, not a missing variable.

2.4 Necessity and Sufficiency in QCA

QCA separates the analysis of necessary conditions from the analysis of sufficient conditions, and uses different metrics for each. Necessity asks: is condition X present in every case where Y occurs? Sufficiency asks: does every case where X is present produce Y? In set-theoretic terms, necessity is set-superset (the set of cases with X contains the set of cases with Y) and sufficiency is set-subset (the set of cases with X is contained in the set of cases with Y).

The metrics QCA uses are consistency and coverage. Consistency is the fraction of cases with the configuration that produce the outcome (analogous to positive predictive value). Coverage is the fraction of cases with the outcome that are explained by the configuration (analogous to sensitivity). The conventional thresholds in Ragin (2008) are consistency ≥ 0.80 for accepting a sufficiency claim and consistency ≥ 0.90 for accepting a necessity claim, with coverage interpreted descriptively rather than against a threshold.

2.5 Crisp-Set vs. Fuzzy-Set QCA

Crisp-set QCA (csQCA), described above, codes each case as 1 or 0 on each condition. This is fine when the conditions are dichotomous in the world (alive/dead; vaccinated/unvaccinated; bereaved within 5 years/not). It is less satisfying when the conditions are continuous and the dichotomy is forced (income above/below the median; loneliness severe/mild).

Fuzzy-set QCA (fsQCA) generalizes the 1/0 coding to a continuous degree-of-membership score between 0 and 1. A case with income of $45,000 might score 0.4 on the “high income” set, where a case at $200,000 scores 0.95. The Boolean operations — intersection (AND), union (OR), negation (NOT) — have set-theoretic analogues for fuzzy sets (minimum, maximum, 1-minus-X). The truth-table logic still applies but the cases now have partial membership in configurations, and the consistency and coverage metrics are computed accordingly (Ragin, 2008; Schneider & Wagemann, 2012).

For your capstone, csQCA is the appropriate starting point. Twenty transcripts is small enough that the dichotomies are defensible and fuzzy-set calibration would introduce more measurement uncertainty than it removes. If you wanted to apply fsQCA later in your research career to a 60-case dataset on, say, harm-reduction program implementation, you would need to read Schneider & Wagemann (2012) carefully and think hard about the calibration of fuzzy memberships.

2.6 Running QCA in R

The R package QCA (Duşa, 2019) is the most mature implementation of both csQCA and fsQCA. The package SetMethods (Oana & Schneider, 2018) extends it with methods for theory-evaluation and case-selection. Both are free, both work on Windows/Mac/Linux, and both produce outputs publishable in Sociological Methods & Research, Implementation Science, or Social Science & Medicine.

# Install QCA and SetMethods (one-time)
install.packages(c("QCA", "SetMethods", "venn"))

# Load
library(QCA)
library(SetMethods)
library(tidyverse)

# Toy dataset: 20 loneliness transcripts, 4 conditions, 1 outcome
# B = bereaved, L = lives alone, I = immigrant, C = caregiving role
# Y = existential loneliness
loneliness_qca <- tribble(
  ~case,  ~B, ~L, ~I, ~C, ~Y,
  "P01",  0,  0,  0,  0,  0,
  "P02",  0,  1,  0,  0,  1,
  "P03",  0,  0,  0,  1,  0,
  "P04",  0,  0,  0,  0,  0,
  "P05",  1,  1,  0,  0,  1,
  "P06",  0,  1,  0,  1,  0,
  "P07",  0,  1,  0,  0,  1,
  "P08",  1,  1,  0,  0,  1,
  "P09",  0,  0,  0,  1,  0,
  "P10",  0,  1,  0,  0,  1,
  "P11",  1,  1,  0,  0,  1,
  "P12",  0,  1,  0,  1,  1,
  "P13",  0,  0,  0,  1,  0,
  "P14",  0,  1,  1,  0,  1,
  "P15",  0,  1,  1,  0,  1,
  "P16",  1,  0,  1,  0,  1,
  "P17",  0,  0,  0,  0,  1,
  "P18",  1,  1,  1,  0,  1,
  "P19",  0,  0,  0,  0,  0,
  "P20",  0,  1,  0,  0,  0
)

# Move case ID to row names (required by QCA package)
df <- as.data.frame(loneliness_qca)
rownames(df) <- df$case
df$case <- NULL

# Truth table for sufficient conditions for Y=1
tt <- truthTable(
  df,
  outcome  = "Y",
  conditions = c("B", "L", "I", "C"),
  incl.cut = 0.8,           # consistency threshold
  n.cut    = 1,             # minimum cases per row
  show.cases = TRUE,
  sort.by  = "OUT, n"
)
print(tt)

# Read the truth table: each row is one configuration of B,L,I,C
# OUT = 1 means the configuration produces Y at consistency >= 0.8
# OUT = 0 means it does not
# OUT = ? means the row has fewer than n.cut cases (logical remainder)

# Conservative solution (no remainders)
sol_cons <- minimize(tt, details = TRUE)
print(sol_cons)

# Parsimonious solution (all remainders available)
sol_pars <- minimize(tt, details = TRUE, include = "?")
print(sol_pars)

# Intermediate solution (remainders consistent with theory)
sol_int <- minimize(tt, details = TRUE, include = "?",
                    dir.exp = c("B"=1, "L"=1, "I"=1, "C"=0))
print(sol_int)

# The output gives the minimized formula plus consistency and coverage
# for the overall solution and for each path

# Necessity for Y=1: which conditions are present in nearly every Y=1 case?
nec <- superSubset(df, outcome = "Y",
                   conditions = c("B", "L", "I", "C"),
                   incl.cut = 0.9,
                   cov.cut  = 0.5)
print(nec)

# Read: which single conditions, or simple combinations, are necessary?
# A condition is necessary if consistency >= 0.9

2.7 QCA and HSCI 341 L12 (Confounding and Causal Inference)

In HSCI 341 L12 you learned the classical epidemiology toolkit for causal inference: temporality, biological plausibility, dose-response, confounding control, mediation analysis, instrumental variables, the potential-outcomes framework. The framework is fundamentally variable-oriented: it asks what the effect of one variable is on an outcome, net of other variables.

QCA does causal inference, but in a different mode. Where HSCI 341 asks “what is the effect of X on Y holding Z constant?”, QCA asks “what configurations of X, Y, Z are sufficient for the outcome?” The first question presumes that the causal structure is additive in expectation and that the effects are estimable; the second presumes that the causal structure is conjunctural and that the effects of variables depend on their configurations. Neither is universally right. Both are tools that should be in the methodologically omnivorous public-health researcher's kit.

When QCA outperforms regression:

• Small N (10–50 cases): the regression coefficients have too little statistical power to be meaningful, but the QCA truth table is fully populated.
• Strong conjunctural causation: when the effect of X depends critically on Y, a regression with interactions can in principle catch it, but only with prior specification; QCA finds it without prior specification.
• Equifinality: when multiple distinct configurations produce the same outcome, QCA reports all of them; regression collapses them into an average effect.
• Asymmetric causation: when the conditions producing presence of the outcome differ from those producing its absence, QCA can be run separately on Y and ~Y; regression assumes the same coefficients apply to both.

When regression outperforms QCA:

• Large N with continuous outcomes and a known causal structure: regression's efficiency is unbeatable.
• Effect-size estimation: QCA gives you sufficiency claims; it does not give you effect sizes that can be aggregated across studies in a meta-analysis.
• Counterfactual reasoning over single variables: the potential-outcomes framework operates at the level of individual variables; QCA at the level of configurations.

Introduction and Overview

The third method in this lesson takes a different analytic stance from the first two. Where analytic induction tests propositional hypotheses and QCA identifies sufficient configurations, ethnographic decision modelling takes seriously the idea that people make decisions according to discoverable rules — rules they can sometimes articulate, sometimes act on without articulating, and sometimes systematically misrepresent. The output of the method is a decision tree: a branching sequence of yes/no questions that, applied to a new case, predicts what the decision-maker will do.

The method was developed by Christina Gladwin in Ethnographic Decision Tree Modeling (1989), drawing on earlier work by James Spradley (1979) and the broader cognitive-anthropology tradition that produced componential analysis and cultural domain analysis. Gladwin's foundational claim is that human decisions are not the inscrutable outputs of black-box psychology but the products of articulable choice rules that can be elicited, formalized, and tested for predictive accuracy on out-of-sample cases. The method has been applied to agricultural decisions (the original work), fertility decisions, medication adherence, vaccine acceptance, contraceptive choice, treatment-seeking under HIV, and health-services utilization in low-resource settings.

For public-health researchers the appeal is concrete. Decision-tree models predict behaviour at the individual level with the kind of mechanistic specificity that variable-oriented regression cannot match. The trees are interpretable in a way logistic-regression coefficients are not: they say, in essentially plain language, what people do under what conditions. When the goal is intervention design, the tree is more useful than the regression because it identifies the specific decision nodes where an intervention could plausibly change the outcome.

3.1 The Foundational Idea: Decisions Follow Discoverable Rules

Gladwin's (1989) starting point is that when people make decisions repeatedly under similar conditions, they develop and apply choice rules. The rules are not necessarily conscious in the moment of decision. They are, however, articulable on reflection — when asked “what would you do if…” with carefully constructed hypothetical cases, decision-makers can typically say what they would do and why. The job of the ethnographic decision modeller is to elicit those rules systematically and to formalize them as a decision tree.

The claim is not that everyone uses the same rules. Decision-tree modelling assumes that within a defined cultural or social group, there will be considerable consistency in the rules people apply, and where there is inconsistency, the rules will branch based on case features the analyst can identify. The output is not a single universal decision tree but a tree that classifies the cases in this group with high accuracy and that predicts new cases from the same group accurately.

Compare this to two adjacent methods. Logistic regression predicts a probability of decision conditional on covariates, but it does not represent the decision process; the coefficients are post-hoc summaries of population-level associations. Bayesian decision theory specifies a normative procedure for combining beliefs and utilities, but it does not describe what people actually do. Ethnographic decision modelling sits between the two: it is descriptive (what do people actually do?) rather than predictive in the regression sense, and it represents process rather than aggregating it.

3.2 The Procedure (Gladwin 1989)

Gladwin's procedure unfolds in three phases: elicitation, formalization, and testing.

Phase 1: Elicitation

Begin with semi-structured group interviews with members of the target population. The interview centres on hypothetical cases: “What would you do if…” The cases are designed to probe specific decision dimensions. For example, in a study of treatment-seeking for chronic illness in rural BC, the cases might be:

• “Suppose you have had a cough for three weeks and it is interfering with your sleep. Would you see a doctor?”
• “Suppose you have had the cough for three weeks but the nearest clinic is 90 minutes away and you do not have a car. Would you go?”
• “Suppose you have had the cough for three weeks, transportation is fine, but you do not have a family doctor. Would you go to walk-in?”

The interview moves through the dimensions one at a time, holding others constant, varying the dimension of interest, and asking the participant to articulate the threshold at which their answer changes. The output of Phase 1 is a list of decision dimensions (cough duration, transportation, doctor relationship, severity, insurance) and the participant-articulated rules for each.

Phase 2: Formalization

Cross-case patterning identifies the recurring rules. If most participants say they will seek care when (cough duration ≥ 2 weeks) AND (transportation is available) AND (a clinical relationship exists), this is a candidate rule. The analyst formalizes the rule as a sequence of yes/no questions, with branches leading to terminal classifications (seek care / do not seek care / seek alternative).

The formalization typically takes the form of a tree, with the most discriminating question at the root. If “is the symptom severe?” cleanly separates seek-care from do-not-seek-care cases, it goes at the top. Within each branch, the next-most-discriminating question follows. Within those branches, the next. The tree continues until every case in the building set is classified.

Phase 3: Testing

The model is then tested on a held-out set of cases — either new participants interviewed about new hypotheticals, or real cases observed in the field. The performance metric is the proportion of cases the tree classifies correctly. Gladwin's convention is that a defensible tree should classify at least 80–85% of out-of-sample cases correctly. Below that, the model is revised.

Revision is similar to analytic induction: you find the misclassified cases, identify what feature of those cases the tree is missing, add a branch or refine an existing one, and re-test. The iteration continues until the out-of-sample accuracy threshold is met.

3.3 Worked Example: Help-Seeking in the Loneliness Dataset

Let us apply Gladwin's procedure to a focused question in the capstone dataset: do participants describe reaching out to their existing personal network when their loneliness becomes severe, or do they describe withdrawing further? The outcome has two terminal classifications: reach out and withdraw.

Reading through the transcripts, several candidate decision dimensions emerge:

• Does the participant have an existing relationship they describe as “close” or “safe”?
• Has the participant tried reaching out before and felt rebuffed or burdensome?
• Does the participant frame loneliness as something other people would understand, or as something they would judge?
• Is the participant currently in a life-stage where reaching out is socially normal (recent loss, illness, new parenthood)?
• Does the participant have language to name the loneliness?

A first-pass tree on the 20 transcripts, using 12 of them for building and 8 for testing:

1. Does the participant have at least one relationship they describe as close/safe?
   • NO → withdraw
   • YES → go to question 2
2. Has the participant tried reaching out before and felt rebuffed?
   • YES → withdraw
   • NO → go to question 3
3. Does the participant have a stigma-permissive frame for loneliness (loss, illness, new parenthood)?
   • YES → reach out
   • NO → go to question 4
4. Does the participant have language to name the loneliness?
   • YES → reach out (with hedging)
   • NO → withdraw

Apply the tree to the held-out 8 transcripts and count correct classifications. If the tree gets 7 out of 8, the model is performing at 88% and meets Gladwin's threshold. If it gets 5 out of 8, the model is at 63% and needs revision. The revision examines the 3 misclassifications, identifies what feature of those cases the tree is missing, and adds or modifies a branch.

For example, if the misclassified cases are all participants who have language and a stigma-permissive frame but still withdraw, the tree is missing a condition. Inspection suggests the missing condition is “structural availability”: the participant has a close relationship in principle, but the person is geographically distant, in a different time zone, in a different language, or busy with their own caregiving demands. Adding a branch on availability raises the out-of-sample accuracy to 88%.

install.packages("DiagrammeR")
library(DiagrammeR)

tree <- "
digraph loneliness_helpseek {
  graph [rankdir = TB, fontname = 'Open Sans']
  node  [shape = box, style = filled, fillcolor = '#E6F3F0', fontname = 'Open Sans']

  Q1 [label = 'Close/safe relationship?']
  Q2 [label = 'Tried before, felt rebuffed?']
  Q3 [label = 'Stigma-permissive frame?']
  Q4 [label = 'Language for loneliness?']

  W [label = 'Withdraw', fillcolor = '#FDEAEF', shape = oval]
  R [label = 'Reach out', fillcolor = '#D1F0EA', shape = oval]
  H [label = 'Reach out (hedged)', fillcolor = '#FFF8E1', shape = oval]

  Q1 -> W [label = 'No']
  Q1 -> Q2 [label = 'Yes']
  Q2 -> W [label = 'Yes']
  Q2 -> Q3 [label = 'No']
  Q3 -> R [label = 'Yes']
  Q3 -> Q4 [label = 'No']
  Q4 -> H [label = 'Yes']
  Q4 -> W [label = 'No']
}
"

grViz(tree)

# Export to PNG for inclusion in the capstone memo
library(DiagrammeRsvg)
library(rsvg)
svg <- export_svg(grViz(tree))
rsvg_png(charToRaw(svg), file = "loneliness_decision_tree.png", width = 1200)

3.4 Health-Relevant Applications

Decision-tree modelling has been applied to a wide range of health-decision domains. A non-exhaustive list, with the kind of question each can answer:

Introduction and Overview

The previous three sections introduced the methods. This section turns them on the capstone dataset. The exercises here are not optional — they are the Week 11 capstone milestone. You will produce either a QCA truth table with Boolean minimization OR a Gladwin-style decision tree, applied to a focused outcome in the 20 loneliness transcripts, together with a 700–900 word interpretive memo. The capstone callout below specifies the deliverable in detail.

Whichever option you choose, the work in this section is about converting an interpretive reading of 20 transcripts into a formal analytic object — a truth table or a tree — that an epidemiology audience would recognize as a result. The discipline of producing the formal object is what gives Lesson 11 its distinctive value in your capstone trajectory. Most of what you have done in Lessons 5–10 has been theme-driven and descriptive. This week's work is causal-claim-driven, and the formal output makes the causal claim auditable.

4.1 Choosing Between Option A (QCA) and Option B (Decision Tree)

Both options are defensible. The principled grounds for choosing one over the other turn on what kind of causal claim you want to make and what the dataset can support.

If you cannot decide between the two, write down (a) the outcome you have in mind and (b) one paragraph naming whether it is an attribute or a decision. The exercise will usually resolve which option fits. If you remain uncertain after that, choose QCA: it produces a more compact deliverable, and the Boolean-minimization output is more uniformly receivable by an epidemiology audience.

4.2 Option A in Detail: QCA Truth Table

The QCA option requires you to define an outcome and a small set of conditions, code all 20 transcripts on each, and run the analysis in R. The methodological discipline is in the operationalization of the conditions: each condition has to have a defensible 0/1 coding rule, and the rule has to be applicable across all 20 transcripts.

Step 1: Pick an outcome

Candidate outcomes from the dataset:

• Existential loneliness (vs. situational): does the participant describe loneliness as a feature of their life-stage or being, or as a feature of fixable circumstances?
• Mentions seeking professional help: does the participant describe consulting a clinician, therapist, counsellor, or other paid professional explicitly because of loneliness?
• Mentions an online or digital community as a coping resource.
• Frames loneliness as a public-health or societal issue (vs. an entirely personal matter).

Pick one. The choice will shape the conditions. Once you have an outcome, write a paragraph defining it operationally — what counts as a 1, what counts as a 0, and how to handle ambiguous cases.

Step 2: Pick 3–4 conditions

Choose conditions that are (a) theoretically motivated by the loneliness literature, (b) codable reliably from the transcripts, and (c) likely to vary across the 20 cases. Candidate conditions:

• Bereaved (lost a primary attachment figure in the past 5 years)
• Lives alone
• Immigrant (born outside Canada, arrived in adulthood)
• Caregiving role (children at home, elder care, partner care)
• Recent life transition (job loss, retirement, relationship dissolution, geographic move within the past 2 years)
• Has stable employment / financial security
• Identifies as part of a marginalized identity group (immigration, LGBTQ+, disability, racialization)

Step 3: Code all 20 transcripts

For each transcript, code the outcome and each condition as 0 or 1. Document the coding rule for each condition in a brief codebook (a paragraph each). Where the coding is ambiguous, document the ambiguity. The codebook goes in an appendix of your eventual capstone paper.

Step 4: Build the truth table and minimize

Use the R code from Section 2.6. Run the conservative, parsimonious, and intermediate solutions. Note the consistency and coverage of each path. Decide which solution to report. The convention in published QCA is to report all three; for a 700–900 word memo, reporting the intermediate solution and noting the existence of the others is acceptable.

Step 5: Interpret

The Boolean formula is the analytic result. Interpretation is the work of saying what it means substantively. For each path in the solution formula, describe in plain language what kind of case the path represents (e.g., “bereaved widows who live alone”), what fraction of the outcome-positive cases the path covers, and what the path implies for loneliness theory or intervention design. The interpretation should occupy roughly half of the memo.

4.3 Option B in Detail: Decision Tree

The decision-tree option requires you to identify a focused decision in the transcripts, build a tree from 4–6 transcripts that contain explicit decision-process talk, and test the tree on the remaining transcripts.

Step 1: Pick a decision

Candidate decisions:

• Whether to reach out to existing personal contacts when loneliness becomes severe
• Whether to engage with an online community or platform
• Whether to seek professional help (clinician, counsellor, therapist)
• Whether to disclose loneliness to a family member
• Whether to attend a community/religious gathering specifically as a loneliness intervention

Step 2: Build from 4–6 transcripts

Identify the 4–6 transcripts where the participant most explicitly describes the decision process. Read carefully; in each transcript, identify the considerations the participant names. Extract a list of candidate decision dimensions (severity, prior experience, perceived burden, structural availability, stigma frame, language).

Step 3: Formalize the tree

Identify the most discriminating question — the one that, asked first, separates the largest number of cases into different paths. Put it at the root. Then within each branch, identify the next-most-discriminating question. Continue until each of your 4–6 cases is classified. Draw the tree in DiagrammeR or by hand.

Step 4: Test out-of-sample

Apply the finished tree to the remaining transcripts (those you did not use to build it). For each transcript, walk through the tree and see what classification it produces. Compare to the actual classification (what the participant actually described doing). Count the matches and mismatches.

Step 5: Revise if needed

If the tree classifies fewer than 80% of out-of-sample cases correctly, identify what feature of the mismatched cases the tree is missing. Add or refine a branch. Re-test. Iterate until you reach the 80% threshold or you reach the conclusion that the tree cannot accurately predict in this sample and the negative result is itself the finding.

Step 6: Interpret

The tree is the analytic result. The interpretive memo discusses what the decision nodes are, which one is most discriminating (and why), what the misclassified cases reveal, and what the tree implies for designing an intervention that would shift the decision. As with Option A, the interpretation occupies roughly half of the memo.

4.4 The Week 11 Capstone Milestone

The full deliverable is below. The brief and rubric for grading are linked at the bottom of the callout.

4.5 Anticipating Limitations

Whichever option you choose, the limitations of the analysis are real and should be reported. The most important ones:

• Sample size. Twenty transcripts is at the small end of QCA defensibility and is just enough for a decision tree. The conclusions are illustrative rather than definitive. The memo should state this.
• Synthetic dataset. The transcripts are instructional composites, not data from real participants. The patterns you find are patterns in a designed dataset, not necessarily patterns in the world. The memo should disclose this.
• Operationalization choices. The 0/1 codings of conditions involve interpretive judgement. The codebook makes the judgements transparent but does not eliminate them. The memo should acknowledge that other coders applying the same rules might produce slightly different truth tables, and that this is exactly the situation Lesson 5's intercoder reliability check is designed to assess.
• Causal-claim status. Neither QCA nor decision-tree modelling on 20 cases establishes causation. They identify sufficient configurations or decision rules in this sample; the inferential extension to a broader population requires further sampling. The memo should be epistemically modest.

Bringing It All Together

Lesson 11 has brought three case-oriented, logically formal qualitative methods into the course: analytic induction (Section 1), qualitative comparative analysis (Section 2), and ethnographic decision modelling (Section 3). All three sit in a methodological space the rest of HSCI 841 has approached but not fully occupied: the space where qualitative inquiry produces formal causal claims defensible to an epidemiology audience. Section 4 turned the methods on the loneliness dataset and set up the Week 11 capstone milestone.

The arc of this lesson connects backward to HSCI 341 L12 (where you learned the classical epidemiology causal-inference toolkit) and forward to Lesson 12 (where computational and LLM-assisted methods will let you scale these moves to larger datasets). Where HSCI 341 trained you in variable-oriented effect estimation, this week trained you in case-oriented configurational analysis — the methodological complement that handles conjunctural causation, equifinality, and asymmetric causation in ways regression cannot.

The final reflection below asks you to look across the three methods and name what you carry forward into your own future research. There is no single right answer.

This glossary collects the key concepts, methods, and people introduced in Lesson 11. Use it as a reference while you work through the material or as a review before the final assessment. Type in the search box to filter entries.