Bo Strangert (RD16)

Case-based reasoning and qualitative modeling

for representation of generality and individuality 

In some previous notes (M6, RD14, RD15)  I argued for case-based reasoning and qualitative modeling regarding R&D of complex tasks. Specifically, some of the arguments for qualitative modeling are based on the distinctive qualities of complex tasks, that is, the variety of interacting factors in dynamic and uncertain states. Other arguments emphasize that complex tasks and cases commonly have unique qualities and often prevail in different contexts.

Methodically, the referred characteristics would imply some kind of case modeling to ensure that unique system qualities will not be misrepresented. However, case modeling is often conceived as a process of collecting a variety of empirical indications which after analysis eventually get a conceptual representation. Such an inductive approach is open to bias and can fail to grasp the distinctive system characteristics of a phenomenon. For example, one possible bias is to analytically concatenate indications linearly, although there are latent nonlinear relations among them. Another example of a risky inductive strategy is to discretionally quantify indicators, correlate them, and then select an alluring but biased sample from a large array of correlations.

There are many different reasons behind biases. A common reason is a need for methodical simplification, both theoretically and as a consequence of data overload. Psychological reasons involve thinking influenced by interests, conventions, and information economy. Therefore, the risk of bias in data collection and treatment should justify the use of an initial control-theoretic approach to case studies.

But what about the risk of bias when selecting an initial conceptual model? This question deserves of course a serious consideration in any modeling approach. An answer can unfold through a careful analysis of the purpose of a development project: which are the necessary system qualities of the task or object? Then these necessary qualities should become the basis of the initial model and ensure that continued conceptual formation, experimentation and design take a valid course.

A master plan for an initial concept development (CD) is easy to state in general terms. It is less easy to apply in real practice. And there are certainly no clear-cut boundaries between CD or empirical work (E) in mental reasoning. However, it is an important methodological difference between starting with an open survey of empirical correlations of variables and with a primary construction of a formal CD model that could embody significant relational structures and steer a succeeding empirical investigation.

The distinction was brought up in the referred projects M6 and RD15 due to the rare circumstance that the research team was unfamiliar with the work content and practices of the practice concerned. It prompted a need to search for integrating mechanisms that could meaningfully interrelate different qualitative core aspects of the healthcare activities.

Possible qualifications for case-based modeling

Preferably,  the concept model should be applicable to a diverse population of individuals or events, though in a sense that it will still represent significant features of individual cases.

A possible resolution can begin by identifying some general mechanisms or functions that involve complex integrative characteristics of the object to be modeled. This functional description would be supplemented with parameters including content-specific variables.

An example of this CD approach is described in M6 and RD15 about a patient’s capability to adapt to long-term illness (Figure 1).  The top conceptual structure is a higher-order control mechanism or set of functions, R_p, for regulating bodily and psychic functions regarding internal and external disturbances. This mechanism can sometimes be superseded by conscious decisions (C), expressing the patient’s will.

Figure 1. Possible sources of data in the patient model

(adapted from Ashby, 1956).

The next level of concepts include: (T) a patient’s inherent fixed capabilities to cope with disturbances, and (E) a set of essential variables for a patient’s states of health and wellbeing; (D) is a set of disturbances for a patient’s health and wellbeing.

Each set T and E includes many common variables for a population of cases, besides specific individual variables. Hence, the general applicability of the control model, including the sets R_p, T, E, and their possible common variables and interrelations should be considered first, before continuing with the search for individual case characteristics of T, E, D.

What is the difference between this modeling approach and a supposed inductive procedure? The question should be answered for the assumed conditions of a complex system: Is it possible to reach testable conclusions on multiple nonlinear interactions between functions and variables by merely interrelating empirical data? It would probably be difficult. (The term ’variable’ is here used in its logical meaning, i.e. ’variable’ can refer to both qualitative or quantitative entities.)

Figure 2 gives a simple nontechnical illustration of some difficulties to infer concepts of regulation from the corresponding data sources about a patients’ disturbances, capabilities to cope with them, and effects of coping on essential variables. We will use the same set of denotations D, T, E in the data representation as in the concept model (RD15, Figure 1) to facilitate the comparison between them.

Assume that data about the sets D, T and E are related. A typical standard method is then to try to represent the relations as a chain composed of two functions g ◦ f. That is, for y_j ∊ T, x_i ∊ D,  and z_k ∊ E  y = f(x) and z = g(y), where y_j is a reaction to a disturbance x_i, and z_k is a positive or negative effect of y_j.

For some data, it is not possible to prove or assume connections between variables of D, T, E; these separate data are denoted by x?, y? or z? in Figure 2. In many other cases, the expected composite function g(f(x)) is either incompletely represented by data or nonlinearly complicated by cross-couplings. Relations between physical components and operations are expected to be better identified and estimated in data compared with other aspects of healthcare.

Figure 2. Is data analysis or concept development the best way to start treating this simplified example of an ill-structured empirical data pattern?

Figure 2 illustrates some principled difficulties of using a simplistic data representation. First, it is limited by the availability of data sources and techniques. Next, and more seriously, it is probably incongruent with actual latent processes underlying the states and transformations of D, T and E. It is true that some incongruences may be confirmed by a closer analysis of empirical relations. But the real problem to come to grip with is when the data model and the analysis are not congruent with the nonlinear interactions, uncertainty and dynamics of complex systems.

Therefore, it seems not easy to identify and formulate regulatory mechanisms on the basis of data patterns like those in Figure 2 , if you only have to rely on unconditioned data analysis. Compare that procedure with the supporting CD approach in RD15, Figure 1.

The difficulties are multiplied when the task analysis is extended to infer consequences for design of support systems. The requirement of an appropriate healthcare planning was commented upon in M6 and RD15. It dealt with the design of an adapted regulatory system, R_t, coordinated with the patients’ R_p regulation.

Tentative conclusions on data analysis in long-term healthcare. Lack of domain knowledge and treatments is common, though physical aspects of healthcare are significantly better managed than psychological and social ones. In addition, much remains about understanding interactions between physical and psychosocial conditions.  Furthermore, the methodological problems in diagnosis and treatment include difficulties to identify and estimate critical states and processes. The uncertainty of complex systems contributes to low statistical reliability of observation and measurement, which is a complication that must be mastered.

Choice of path?

The preceding paragraphs express doubts about inferring complex integrative mechanisms by an approach based on experiential knowledge, even if it uses standard methods of data collection and analysis. Nonetheless, complex phenomena require advanced concept development to maintain efficient data collection and treatment. Unfortunately, the readiness among healthcare planners to use a hypothetical-deductive methodology does not seem very probable, except when physical healthcare and its domain experts are involved.

Organizational and soft psychosocial issues are often regarded as impossible to handle with natural science methodology, presumably because of lack of knowledge among those concerned. These issues are instead often characterized as in need of other, ”qualitative” tools and solutions. For a long time, the prevailing culture has included many unarticulated and fuzzy holistic views on diagnostics and treatment. It hinders progress.

The suggested analysis based on case-based reasoning and qualitative modeling is just one promising component in work and organization development. Another lead is how to manage development work. Much of long-term healthcare concerns context-sensitive tasks, which should promote a participative R&D design approach, closely adapted to the specific circumstances in each case. The requirements on systematic planning and accurate follow-up should be adhered to. Some reflections on that practice may appear in coming papers.

References on this website

M6. Argument för modellteoretisk ansats vid utveckling av vård och omsorg (2015)

RD14. On general sources of uncertainty in project planning (2015)

RD15. Model-theoretical reasoning to give structure to complex realities:

A case of project planning in healthcare (2015)