Chapter 5: Causal Reasoning

Definition

( Case 32 , Case 34, Case 35)

The process of arriving at a working diagnosis requires interpretation of new data in light of existing information and assessment of the relations among all of the clinical findings. In carrying out these tasks, the physician makes repeated attempts to “tie all the findings together.” The probabilistic approach, such as that described earlier, represents one of the methods used, and the use of standard, compiled knowledge rules or categorical reasoning (e.g., if a patient has rheumatoid arthritis, splenomegaly, and leukopenia, consider Felty syndrome) is another. Indeed, much of the reasoning in medicine depends on probabilistic and categorical associations between clinical variables. Still another powerful approach is causal reasoning.

Causal reasoning is an aspect of the diagnostic process based on the cause-and-effect relations between clinical variables or chains of variables. It is a function of the anatomic, physiologic, and biochemical mechanisms that operate in the normal workings of the human body and the pathophysiologic behavior of these mechanisms in disease. In addition to its basis in the mechanisms of normal and abnormal pathophysiology, causal reasoning relies on commonsense notions of causality, such as the beliefs that an effect is usually generated by a known cause, that cause and effect are related in time and space, and that a cause and its effect are generally similar in intensity and magnitude.[ 44], [ 45]–[ 46]

Certain kinds of clinical reasoning are better described in terms of a causal model than in terms of the probabilistic associations between variables. Take, for example, the interpretation of a normal creatinine level 1 hour after a patient develops acute anuria. A probabilistic model that explains this apparent contradiction (a normal creatinine value in the face of zero kidney function) would strain credibility.

Yet the physiological explanation is simple, complete, and revealing: Because creatinine is produced in the muscle at a constant rate even when kidney function is nil, an insufficient amount would accumulate over such a short time following renal shutdown to produce a perceptible rise in the serum level. This example is based on a causal or physiologic model of reality. The capacity to make inferences from the observed clinical findings also depends on the principles embedded in this model. Such models abound in all domains in medicine but are especially common in nephrology, cardiology, pulmonology, and endocrinology because of their strong underlying physiologic knowledge base.

When applying causal reasoning, the physician examines clinical variables and includes them if they help to explain the model. A model is created for each patient, although a single model can be applicable for many patients in many clinical settings. Such a model is a coherent system that is capable of explaining its components, the range of possible variations, and the nature of findings in a particular patient. If the same model is not applicable for the next patient with a problem similar to that for which it was created, additional features of the model might be required and some existing features might have to be deleted.

To explain how a causal model can support a clinician’s performance by simulating possible courses of the disease and its modification by treatment and how such a model can serve as a coherency criterion for hypotheses about the patient, an example is presented here from the domain of fluid and electrolyte equilibrium. Assume that a patient with clinical and laboratory findings suggestive of the syndrome of inappropriate secretion of antidiuretic hormone (SIADH) has a high urinary sodium excretion: Does this finding influence the suspected diagnosis?

We could assess this finding in a probabilistic framework (e.g., we might say that 85% to 95% of patients with SIADH have a high sodium excretion), or, alternatively, by understanding the pathophysiology of SIADH, we could examine how the finding “fits” with the diagnosis. If our model of SIADH contains (as it should) the concepts that such patients are volume expanded, that volume expansion promotes sodium excretion, and that sodium excretion in SIADH typically matches sodium intake, we would understand readily that a high urine sodium excretion not only is consistent with the diagnosis of SIADH, but also that it adds to the credibility of that diagnosis. We also would be in a position to explain a low urine sodium excretion if that were the finding instead. In this instance, despite the presence of SIADH, urine sodium is low, presumably because the patient is ingesting little salt.

Using a Causal Model

( Case 11, Case 32 , Case 33, Case 35, Case 36)

We often are alerted to the possibility that we should be using a causal model when abnormal findings or events violate normal expectations. This deviation produces the context within which further data gathering and interpreting takes place. To carry out this interpretive process, we generate a causal model, typically a chain of related features consisting of stimuli and their responses.[ 44],[ 46], [ 47]–[ 48] When invoking a causal hypothesis involving two or more variables, we assess the links between stimuli and responses for their strength. The strength of this link can be assessed by the satisfaction of several criteria.

Is the entire causal chain credible? Does a given change in a response correlate closely with the change in the stimulus? Is there substantial congruity of duration and magnitude between response and stimulus? Is there close contiguity in time and space between a response and a stimulus (did one event follow another sufficiently closely to allow us to accept that the first event caused the second)? When these tests are satisfied, one gains confidence that a given stimulus and a suspected response are related.[ 48], [ 49]–[ 50]

Note that in describing the outcome of causal reasoning we deliberately use the notion of “confidence” in the relationship between cause and effect. We do so because causality virtually never can be proved; the stronger that the elements of causality are, the more likely it is that the effect can be attributed to the cause. Just because a given effect commonly quickly follows, a stimulus is not sufficient justification for attributing the effect to the cause.[ 51] Common sense does not always lead us to correct conclusions about causality, as evidenced by the mistaken notions that getting chilled causes upper respiratory infections and that stormy weather causes arthritic pain to worsen. Similar considerations lead to caution when attributing a rare complication to a drug that is new to the market.

In medicine, we are always attempting both to validate and to debunk causal relations. Indeed, the final step in assessing a causal hypothesis is testing it for alternate possible explanations. Simply because a given causal hypothesis appears to explain a set of findings does not necessarily prove that this causal chain is the correct one. Alternative constructions of the causal chain must be sought and their strengths assessed before accepting one model and not another.

Where in the Diagnostic Process does Causal Reasoning Fit?

( Case 11, Case 33, Case 34)

Causal reasoning can be applied in several steps of the diagnostic process. Early in the process, probabilistic reasoning is more likely to be helpful than causal reasoning in generating hypotheses. Because causal models are dependent exclusively on fundamental knowledge about physiologic function and dysfunction and the cause-and-effect relations between clinical events, they are specific to disease entities and independent of the patient population. By contrast, probabilistic models are dependent on the specific population from which the patient is drawn. Because diagnostic hypotheses are so critically dependent on disease prevalence, causal reasoning is a rather weak approach when the required task is triggering such hypotheses, whereas probabilistic reasoning is quite strong.

The assertion, for example, that a 60-year-old heavy smoker with hemoptysis is far more likely to have lung cancer than a 20-year-old nonsmoker with the same symptom is based predominantly on disease prevalence rather than on the mechanisms of bleeding. Nonetheless, causal reasoning can be useful early in the diagnostic process when formulating a context: If the possibility of a pathophysiologic state has been triggered by some findings, the state may provide the context for further data gathering. When physical examination in a 37-year-old man admitted for cough, hiccups, and extensive lesions in the lung disclosed that the patient had bilateral gynecomastia, attention immediately shifted to the possibility that metastatic germ cell was the etiology of his pulmonary nodules.

Further studies proved the diagnosis to be correct. In addition, once a possible cause has been proposed, causal reasoning allows us to assess whether the cause can explain the observation. The SIADH example given before illustrates the interplay of these reasoning strategies. Once the diagnosis of SIADH was triggered, the causal model made it possible to check the appropriateness of either a high or a low urinary sodium excretion. Causal models also help us to understand when certain findings do not fit within the framework of a given hypothesis. Such a signal then becomes a trigger for generating new hypotheses.

When a complete or nearly complete causal model can be constructed, it can be useful in the process of hypothesis refinement. In one aspect of hypothesis refinement, namely the interpretation of diagnostic tests, causal models can be used to check the validity of probabilistic models constructed to assess the data from diagnostic tests. As described in the preceding chapter, probabilistic models require that each disease under consideration be mutually exclusive of all others and that conditional probabilities under consideration be independent of each other. Causal models, because they encode dependence among the parameters they encompass and because they provide an understanding of the relations between variables, can identify circumstances in which the independence assumptions of a probabilistic model are invalid and can provide valuable guidance for correcting a poorly constructed model.

Causal reasoning may be most valuable when diagnostic hypotheses are undergoing final checking and a “working diagnosis” is being formulated (see later discussion of diagnostic verification). In that phase of the diagnostic process, a diagnosis is assessed for its coherency, namely, whether the physiologic or causal associations are reasonable, appropriate, and complete. This step involves determining whether a patient’s findings are consistent with recognized pathophysiologic manifestations of a suspected disease.[ 19]

A causal model is essential to this process: In a patient suspected of having hyperthyroidism based on the combination of clinical findings and a plasma thyroxine concentration that is only slightly elevated, the finding of a suppressed level of thyroid-stimulating hormone is a critical finding to verify the diagnosis. In this instance, suppression of pituitary function by excessive circulating thyroid hormone is the causal link that helps to confirm the hypothesis of hyperthyroidism.

Proper application of the causal approach can yield a rigorous guide to therapy because the treatment can be based on efforts to reverse the string of events that produced the disordered state. If, for example, one understands that chloride depletion is a regular consequence of the enhanced bicarbonate reabsorption that accompanies sustained hypercapnia, it follows clearly that replacement of depleted chloride stores will be necessary during any process in which hypercapnia is rapidly reversed. Any probabilistic approach to this therapeutic problem would be, at best, necessarily complex or opaque and, at worst, grossly inadequate.

Explaining Relations Between Variables

( Case 11, Case 33, Case 34, Case 36)

An important strength of causal reasoning is its capacity to provide an explanation for a given finding, especially when the relation is not immediately obvious from either probabilistic associations or from already compiled knowledge or concepts. A causal model also makes it possible to tie various clinical findings together in a common framework: The effect of dietary sodium intake and sodium excretion in patients with SIADH, mentioned earlier, is such an example. A causal approach provides a consistency check among related findings: Two common findings may have a strong probabilistic (or statistical) relationship, yet they may be causally inconsistent. Causal reasoning can help to identify such discrepancies.