The Retroduction

TAM-INS.04 · The Insufficient · The Approximate Mind

In Hinds County, Mississippi, a woman named Tamara Williams is pregnant with her second child. She is thirty-one. She works as a home health aide, which means she spends her days caring for other people’s aging parents while her own mother, who lives twenty minutes away, watches her four-year-old son. She drives a 2014 Honda Civic with a slow leak in the rear left tire that she keeps meaning to get fixed. She likes crime novels and sweet tea and singing in the car when nobody is riding with her.

Her risk of dying from pregnancy-related causes is roughly three and a half times higher than it would be if she were white.

The empirical record offers explanations. Access to care: the nearest obstetric unit is forty minutes away, and it closes at night. Comorbidities: hypertension, diagnosed two years ago, managed intermittently because the medication costs more some months than she can absorb after rent and childcare. Provider bias: documented extensively in the literature, measured in studies of differential treatment by race in clinical settings.

These explanations are real. Each one contributes. Together they do not close the gap. The disparity persists after controlling for income, after controlling for insurance status, after controlling for documented comorbidities, after controlling for education. The documented mechanisms account for some of the excess mortality. They do not account for all of it.

The conventional research response to this residual is: we need more data. More variables. Finer-grained measurement. Larger sample sizes. Better controls. The assumption is that the answer is in the empirical record and we have not found it yet because we have not looked carefully enough.

There is another possibility.

What the Residual Means

#

Roy Bhaskar’s critical realism makes a claim about the structure of reality that is simple to state and difficult to absorb.

Reality is stratified. The empirical domain contains what has been observed and recorded. The actual domain contains what has occurred, whether or not anyone observed it. The real domain contains the generative mechanisms that produce events, whether or not those events occur, whether or not anyone observes them.

The entire apparatus of modern research, including every AI system trained on the outputs of that research, operates at the empirical stratum. It can only work with what has been observed, documented, published, and digitized. If a mechanism operates at the level of the real but has never produced an event that was observed and recorded by the research enterprise, the mechanism is invisible. Not absent. Invisible.

The residual in the maternal mortality disparity is not noise. It is data. It tells you that mechanisms exist at the level of the real that the empirical record has not captured.

Not because they are exotic. Not because they are rare. Because the research enterprise, the studies, the grants, the institutional infrastructure that produces medical knowledge, was never structured to find them.

The gap between the observed outcome and the explanatory power of documented mechanisms is itself the most important finding. It is evidence that the instruments are insufficient for the reality they are trying to describe.

Retroduction

#

Bhaskar called the method for working backward from outcomes to undocumented mechanisms “retroduction.” It is distinct from both deduction and induction.

Deduction moves from general principles to specific predictions. If all swans are white, and this bird is a swan, then this bird is white. Induction moves from specific observations to general patterns. I have observed a thousand white swans, so swans are probably white.

Retroduction asks a different question. Given that this outcome exists, what mechanism must be operating to produce it? The mechanism may never have been directly observed. Its existence is inferred from the outcome it produces, the way a physicist infers the existence of a particle from the trail it leaves in a cloud chamber.

Applied to Tamara: given that the mortality disparity persists after documented mechanisms are accounted for, what mechanisms must be operating at the level of the real to produce the excess? The mechanisms have not been directly observed. They can be inferred.

Arline Geronimus named one such mechanism: weathering. The cumulative physiological toll of chronic stress associated with living in a society structured by racial hierarchy. Weathering is not a disease. It does not appear in any diagnostic manual. It is a process by which the body’s allostatic load, the wear and tear of sustained stress response, produces premature aging of organ systems, particularly cardiovascular and metabolic systems. It operates beneath and across every specific diagnosis. It is a mechanism at the level of the real that produces outcomes at the level of the empirical, outcomes that appear as individual clinical events, as hypertension, as preeclampsia, as cardiac arrest, each of which is documented separately without the connecting mechanism being named.

Weathering was identified retroductively. Geronimus did not observe the mechanism directly. She observed the outcomes, the age-related health differentials between Black and white women that could not be explained by documented risk factors, and reasoned backward to the mechanism that must exist to produce them.

The mechanism was operating long before it was named. Women were dying from it for decades while the research enterprise treated each death as an individual clinical event explained by individual risk factors. The mechanism was at the level of the real. The deaths were at the level of the empirical. The research lived at the empirical and could not see the real.

The Other Direction

#

Here is where this essay departs from the frame that social justice provides. And the departure matters.

In Greenwich, Connecticut, a man named Richard Chen is fifty-three years old. He runs a private equity fund. His annual physical produces excellent results. Blood pressure managed. Cholesterol managed. BMI in the normal range. He exercises four times a week with a trainer. He eats well because someone is paid to prepare his meals.

He has not slept through the night in two years. He cannot identify what he is afraid of. His marriage is functional in the way that a well-maintained machine is functional: everything works and nothing is alive. His children are at boarding school. He speaks to them on Sundays, conversations that follow a script neither party chose. He has a persistent sense that the structure of his life, the fund, the house, the schedule, the trainer, the meals, is an elaborate mechanism for preventing him from encountering himself.

His risk profile, by every clinical metric, is excellent. No AI screening tool will flag him. No triage system will route him to further evaluation. His empirical record is thick with reassuring data.

If he has a stroke in three years, his cardiologist will look at the chart and be surprised. The risk factors were managed. The numbers were good. The outcome was not predicted by the model.

The gap between Richard’s observed risk profile and his actual physiological trajectory is the same kind of gap as the one in Tamara’s maternal mortality data. Something is operating at the level of the real that the empirical record does not contain. The mechanism is different. The stratum gap is the same.

For Tamara, the empirical record is thin because the instruments were never pointed at her life with adequate resolution. The research was never funded, never designed, never conducted. The mechanisms are undocumented because of institutional neglect.

For Richard, the empirical record is thick but miscalibrated. The instruments were pointed at him extensively. They measured everything they were designed to measure. They were not designed to measure the physiological consequences of a life that is materially complete and existentially hollow. Wealth codes as protection in every model. The models encode the assumption. The assumption prevents the mechanism from surfacing.

Different failure mode. Same stratum gap. The empirical undershoots the real in both cases. The consequences fall differently because the resources available to compensate for the system’s blindness are distributed unequally.

This is the axiology the series has been building toward. Not social justice, though justice follows from it. Epistemic completeness. Every person’s causal mechanisms deserve investigation adequate to the actual structure of their life. The method is universal. The urgency is differential. The consequences of the gap fall harder on Tamara because she has fewer resources to compensate for the system’s inability to see what is happening to her. But the gap itself exists for Richard too, and his stroke will be no less real for having been invisible to every instrument his wealth purchased.

The Compound as Mechanism

#

The Intersectional Systemic Harm Index, described in earlier essays, performs retroduction without calling it that.

When the compounding score for a person exceeds what the individual barrier scores would predict, the excess is retroductive evidence. Something is operating in the compound that the decomposed view cannot see. The conventional assessment framework treats the excess as noise, as measurement error, as the imprecision of individual barrier scores adding up. The index treats it as signal.

The retroductive inference: the interaction between barriers is not a complication to be controlled for. It is the causal structure. Transportation plus digital divide plus economic strain plus social isolation do not produce four times the difficulty. They produce a cascade whose dynamics are not predictable from the individual components because the interaction is the mechanism.

The index does not care about the direction of the barriers. A person with high income, social isolation, caregiver burden, chronic pain, and stigmatized mental health needs compounds too. Richard compounds. The interaction effects do not check your tax bracket. They operate at the level of the real regardless of what the empirical record says about your risk profile.

The index was built before the philosophical framework was found. It was built from the experience of watching healthcare systems process people whose outcomes could not be explained by the variables the system was measuring. The practice generated the insight. The philosophy explains why the insight works.

This sequence, from practice to theory rather than theory to practice, is itself retroductive. The builders observed the outcome: conventional assessment consistently underestimates compound effects. They reasoned backward to the mechanism: the assessment framework decomposes what should not be decomposed, because the research tradition that produced it treats interaction effects as noise rather than signal. The theory, Bhaskar’s stratified ontology, names the structure. But the structure was discovered operationally, by people standing close enough to the affected lives to see what the instruments were missing.

What Retroductive Systems Would Look Like

#

An AI system built on retroductive principles would treat outcome disparities as the starting point for investigation rather than the endpoint.

Instead of asking “what diagnosis best matches this patient’s presentation,” it would ask: “this patient’s outcomes diverge from what the model predicts, and the divergence is not random; what mechanisms must be operating that the model does not contain?”

Instead of asking “what intervention closes this disparity,” it would ask: “the disparity persists after documented mechanisms are accounted for, so documented mechanisms are insufficient; what undocumented mechanisms does the residual point to, and what would the research enterprise need to look like to find them?”

Instead of defaulting to the most probable explanation within the existing ontology, it would flag cases where the explanatory gap exceeds a threshold. Not “probable diagnosis: X.” But: “the available categories do not adequately explain this presentation. The gap suggests mechanisms operating outside the current ontological frame. Route to expanded investigation.”

This is not speculative. It is a design specification that could be built today. The technical components exist: outcome tracking, residual analysis, flagging mechanisms for cases that exceed model predictions. What does not exist is the institutional willingness to build a system whose output is, in certain cases, the admission that the institution’s knowledge is insufficient.

I wonder whether the resistance to retroductive systems is not technical but psychological: whether the institutions that fund research and deploy AI systems can tolerate a tool that tells them, regularly and with specificity, that they do not know enough, and that the gaps in their knowledge are not random but structured by the same institutional incentives that produced the knowledge they do have.

Tamara and Richard

#

Tamara is twenty-eight weeks pregnant. She has a prenatal appointment next Tuesday. The AI risk assessment will process her chart and produce a score. The score will incorporate the documented risk factors: her hypertension, her BMI, her age, her insurance status. It will not incorporate weathering, because weathering is not a variable in any clinical prediction model. It will not incorporate the physiological toll of driving forty minutes each way to the obstetric unit while managing a job, a child, and a tire that needs fixing. It will not incorporate the ambient stress of navigating a healthcare system that has documented, in its own literature, that it treats her differently based on the color of her skin.

The system will produce a risk score. The score will be accurate within the system’s ontology. The ontology does not contain the mechanisms that will determine whether she lives or dies.

Richard has a physical scheduled for May. His results will be excellent. The AI wellness assessment will confirm that his risk factors are well-managed. It will not incorporate the accumulated physiological cost of a life lived inside a structure that prevents encounter with the self. It will not incorporate the cardiovascular consequences of two years of disrupted sleep whose cause no clinical intake form has a field for. It will not incorporate the chronic low-grade inflammatory state associated with emotional suppression in men who have been trained to experience vulnerability as failure.

The system will produce a wellness score. The score will be reassuring within the system’s ontology. The ontology does not contain the mechanisms that will determine whether he survives the next three years.

Two lives. Two stratum gaps. Two systems performing perfectly at the empirical level while the real operates beneath them, unobserved, unmeasured, and consequential.

The Bridge and the Tire

#

Tamara’s tire has a slow leak. She fills it at the gas station every few days. The fix is fifteen dollars. The fifteen dollars is not available this week because the copay for her prenatal visit was higher than she expected. The tire is not a medical variable. It is a variable in her life.

The bridge in Dr. Rao’s photograph, the stone bridge her grandfather helped build in Tamil Nadu, has stood for eighty years. It was built by people who knew the river. Not people who had studied the river. People who had lived beside it long enough to understand what it does when the monsoon is heavy and what it does when the monsoon is late.

There is a kind of knowledge that comes only from proximity to the thing being known. Retroduction is the method that takes that proximity seriously. It says: the people closest to the outcome, the people living inside the conditions the system is trying to describe, carry information about the mechanisms that no instrument pointed from the outside can capture. Their testimony is not anecdotal. It is retroductive evidence. They are reporting outcomes that the documented mechanisms cannot explain, and the gap between their reports and the model’s predictions is the beginning of the investigation, not the end.

Tamara knows something about her own pregnancy that no risk model contains. She knows what it feels like to carry a child while carrying everything else. That feeling is not a variable. It is the trace of the real showing through the empirical, and it will be there whether or not anyone builds a system capable of seeing it.

The tire has a slow leak. The system has a slow leak too. Different kinds of air escaping from different kinds of containers. Both of them consequential. Both of them fixable, if anyone decides the fix is worth the cost.

This is the fourth and final essay in The Insufficient, a sub-series of The Approximate Mind. The series examined what lies beneath the empirical record that AI systems are built to search. “The Skeptic” introduced a system whose resting state is non-belief. “The Traditions” populated it with seven philosophical operations drawn from traditions the AI ecosystem was not built to see. “The Intent” moved upstream to the commissioning decisions that determine what gets studied and known. This essay provides the method: retroduction, working backward from outcomes to the mechanisms the insufficient record has not captured. The method is universal. The urgency is differential. The gap between the empirical and the real is where the harm lives, and closing it requires not better data from the same stratum but the willingness to look deeper.

References

#

Critical Realism and Retroduction

Bhaskar, Roy. A Realist Theory of Science. Verso, 1975.

Bhaskar, Roy. The Possibility of Naturalism: A Philosophical Critique of the Contemporary Human Sciences. Harvester Press, 1979.

Danermark, Berth, et al. Explaining Society: Critical Realism in the Social Sciences. Routledge, 2002.

Pawson, Ray, and Nick Tilley. Realistic Evaluation. SAGE Publications, 1997.

Weathering and Health Disparities

Geronimus, Arline T. Weathering: The Extraordinary Stress of Ordinary Life in an Unjust Society. Little, Brown Spark, 2023.

Geronimus, Arline T. “The Weathering Hypothesis and the Health of African-American Women and Infants: Evidence and Speculations.” Ethnicity and Disease, vol. 2, no. 3, 1992, pp. 207-221.

Maternal Mortality

Petersen, Emily E., et al. “Racial/Ethnic Disparities in Pregnancy-Related Deaths: United States, 2007-2016.” Morbidity and Mortality Weekly Report, vol. 68, no. 35, 2019, pp. 762-765.

Crear-Perry, Joia, et al. “Social and Structural Determinants of Health Inequities in Maternal Health.” Journal of Women’s Health, vol. 30, no. 2, 2021, pp. 230-235.

Social Determinants and Epidemiology

Krieger, Nancy. Epidemiology and the People’s Health: Theory and Context. Oxford University Press, 2011.

Marmot, Michael. The Health Gap: The Challenge of an Unequal World. Bloomsbury, 2015.

Development and Structural Violence

Farmer, Paul. Pathologies of Power: Health, Human Rights, and the New War on the Poor. University of California Press, 2003.

Sen, Amartya. Development as Freedom. Anchor Books, 1999.

The Production of Knowledge and Ignorance

Proctor, Robert N., and Londa Schiebinger, eds. Agnotology: The Making and Unmaking of Ignorance. Stanford University Press, 2008.

Harding, Sandra. Objectivity and Diversity: Another Logic of Scientific Research. University of Chicago Press, 2015.