Shared Medical Decision Making Reconsidered: Challenging an Overly Cognitivist Perspective with a Linguistic Approach

ABSTRACT This article critically examines the four patterns of shared medical decision making (physician-dominated; physician-defined, patient-made; patient-defined, physician-made; and patient-dominated) suggested by Lippa et al. (2017). The aim of the study is to challenge these patterns with a new data set of conversations between physicians and cancer patients in a hospital ward. We recorded 13 physician–patient-conversations during the medical round in an Austrian hospital, which in total lasted about 1.5 h (language: German). We then categorized the medical decisions found in the data following Lippa et al.’s instructions and further analyzed them with a fine-grained linguistic approach. The study revealed no patient-dominated decisions and one decision, which could not be categorized with one of the patterns. Results from the linguistic approach call into question the generalizability, distinctiveness and validity of the patterns. Finally, the relationship between shared decision making and clinical distributed cognition is discussed.

Medical decision making is a vital matter in medicine, which has been profoundly re-conceptualized in the past decades (Will, 2011b). With the flourishing of discourses on patient rights and patient autonomy, the paternalistic model gradually gave way (Will, 2011a). The paternalistic model -following the principle "doctor knows best" -placed the responsibility for medical decisions solely on the physicians (Fritzsche et al., 2020). Then, for a short period of time the informative model was endorsed. Within this model, physicians had to give all relevant information to their patients so that they could make an informed decision, but patients also had to take full responsibility for their decisions (Schweickhartdt & Fritzsche, 2007). After the paternalistic and informative model, which are located rather at each end on the continuum of patient autonomy, most researchers and practitioners arrived at the conclusion that it is best when physicians and patients share medical decisions (Elwyn et al., 2000(Elwyn et al., , 2014Légaré et al., 2008).
While various models of shared decision making (SDM) emerged over the last decades, Bomhof-Roordink et al. (2019) identified key components in their systematic review of SDM models. The basis for SDM is a sustainable partnership between physician and patient. Further, physicians shall describe treatment options and create choice awareness while patients shall inform their physician about their preferences. Together, physicians and patients deliberate and arrive at a decision or determine a next step.
Effective provider-patient communication is not only essential to deliver quality health care (Miller & Peck, 2019) but also enables SDM, i.e., the establishment of a sustainable physician-patient-relationship, reception/delivery of viable information, deliberation during a decision making process and so on. Provider-patient communication is a complex matter, which is not only influenced by the interpersonal context of patient and health care provider, but also by environmental aspects and social contexts as Street (2003) showed in his ecological perspective on communication in medical encounters. With regard to medical decisions, the inevitable power structure between physicians and patients is also relevant here (Giese, 2002).
The need for shared decision making has been emphasized as it is deemed to enhance the satisfaction with and investment in the outcome of both patient and physician (Drolet & White, 2012;Gattellari et al., 2001), and enables patients to increase their knowledge and undergo the decision making process with more confidence (Elwyn et al., 2012). SDM also plays an essential role in the treatment of cancer patients (Elkin et al., 2007;Leydon et al., 2000;Stacey et al., 2012) as they are often confronted with different options that include trade-off decisions regarding quality of life and lifespan (Coulter, 2003). Lippa et al. (2017) identified four patterns of medical decision making analyzing two data sets of medical decision making collected in a clinic specialized in treating multiple sclerosis and emergency departments: physician-dominated (I); physician-defined, patient-made (II); patient-defined, physicianmade (III); patient-dominated (IV). Lippa et al.'s patterns received considerable attention (Adams et al., 2018;Amon & Favela, 2019;Brooke et al., 2019;Feufel & Flach, 2019;Feufel, 2018;Papautsky, 2019;Willekens et al., 2017) and -following a cognitivist approach -attempt to categorize (shared) medical decisions along the continuum "physician vs. patient responsibility".
The model by Lippa et al. (2017), i.e., the expanded decision space model, conceptualizes SDM within the paradigmatic frame of distributed cognition. This theory assumes that cognition is not limited to individual minds but "locates thought as an emergent property of people interacting with other actors and the environment" (Lippa et al., 2017(Lippa et al., , p. 1035. According to the expanded decision space model, physician and patient have mutually acceptable options for action and negotiate the decision within a certain decision space. As depicted in Figure 1, this space is determined by actions physicians and patients have available that are acceptable and feasible (i.e., their action space) and the states they consider possible and desirable/justifiable (i.e., their state space). The action space is limited through personal capacities and adopted norms, while the state space is determined by physical constraints and values.
In order to make a decision, physician and patient therefore have to be aware of their action and state spaces and inform each other about aspects of their state and action spaces that only one person is aware of (i.e., distribute cognition). Thenby eliminating all undesirable and impossible states as well as all unacceptable and impossible actions -physician and patient have only a few options left from which they negotiate their final decision. Lippa et al. (2017) postulate that their patterns of distributed cognition constitute a continuum of shared decision making and that distributed clinical cognition always results into physician and patient being involved in the decision making.
The cognitivist approach assumes that decision makers are aware of their pre-existing action and state spaces at all timesan assumption not endorsed in the linguistic approach -and that the possible fit among these respective spaces is detected in joint conversation. The linguistic approach -following the tradition of social constructionism -challenges these assumptions and, in contrast, holds that reality is not just communicated, but created in joint conversation. Rabenstein and Gerlach (2016, p. 209) consider decision making processes as "practical doing". Their approach emphasizes the processuality and potential incompleteness of decisions and assumes that decisions are embedded in opportunity structures with constellations of different actors. It considers medical decision making from a praxeological perspective, i.e., one that does not aim at categorizing such decision making in the first place (what kind of decision is it), but at reconstructing how exactly this decision is brought about in language (Przyborski & Slunecko, 2009).
We do acknowledge that there are cognitive aspects involved in shared decision making, but question whether the four patterns of distributed decision making are able to adequately reflect the complexity of medical decision making. The symmetry of these patterns -a quite typical feature of representational theorizing -and the relative parsimony of both the model's assumptions and the resulting categories of decision making that follow from it, probably add to its appeal for clinical research. We compare these opposing approaches by applying the cognitivist as well as the linguistic approach to a new set of data, and to address this issue, ask ourselves the following research questions: • Are the patterns of distributed decision making generalizable to a new set of data? • Are the patterns of distributed decision making valid when applied to decision making processes in cancer care? • What conclusions can be drawn from the results of this study in regard to the relationship between SDM and distributed cognition?

Data collection
With the approval of the hospital's ethics committee and the written consent of both patients and physicians, one of us (B. B.) audiotaped conversations during the medical round with a digital recorder at the department of internal medicine in a church-affiliated hospital in Austria in May 2018. The conversations were conducted in German, the Austrian national language. Included were all patients with a cancer diagnosis who were admitted to the hospital department in the week chosen for data collection and agreed to be part of the study (only two patients declined). Three different physicians conducted the medical round with eight different patients, which resulted in thirteen twoto twelve-minutes-long physician-patient-conversations. This sample is not representative but was regarded as sufficient for our analytical approach and the sample consisted of thirty-six examples of medical decision making.

Sample description
The physicians conducting the medical rounds were all 55-to 60-year-old men and had several years of experience. The patients -three men and five women -were between 67 and 77 years old (M = 73, SD = 2.87), all of them were suffering from different stages and types of cancer (bile duct, ovarian, peritoneal, bladder, colon, pancreas, prostate, myelodysplastic syndrome and carcinoma of unknown primary). All patients, except one, who was admitted for the first time, had been in the care of the hospital for at least 7 months (M = 19.86, SD = 8.53) and therefore were acquainted with the resident physicians. All patients had additional health insurance and, therefore, stayed in a superior department of the hospital.

Generalizability of the patterns of distributed decision making
First, we tested the patterns in terms of their generalizability by trying to apply them to our new set of oncology data. For this purpose, the physician-patient-conversations were transcribed using GAT-2's convention for the basic transcript (Selting et al., 2011). Whereas Lippa et al. (2017) used Grounded Theory (Strauss & Corbin, 2008) to analyze their data and "decision making" emerged as a core category through open coding, we used the classification taxonomy for medical decisions developed by Ofstad et al. (2016) to identify medically relevant decisions from the transcripts. Decisions were included in the data, if they fell into one of the following categories: Gathering additional information, evaluating test result, defining problem, drug related, therapeutic procedure related, legal and insurance related, contact related, advice and precaution, treatment goal, deferment. 1 Two coders were familiarized with the categories and trained with a few examples before analyzing the transcripts for medically relevant decisions independently: Coder 1 identified 34 medically relevant decisions, whereas coder 2 identified the same 34 decisions and five more. The five additional decisions were discussed and both coders finally agreed that only two of them actually were medically relevant, which resulted in 36 medically relevant decisions in total. Using the instructions provided by Lippa et al. (2017), these decisions were then categorized according to the patterns of distributed decision making by allocating the responsibility for each of the three stages ("identification of decision point", "definition of parameters", and "final decision") to either physician or patient.
According to Lippa et al. (2017), the decision is physiciandominated, if the physician is responsible for all three stages; vice versa, the same logic pertains to patient-dominated decisions. For physician-defined, patient-made decisions (II) and patientdefined, physician-made decisions (III), it does not matter if physician or patient identify the decision point. In pattern II, the physician defines the decision parameters while the patient makes the final decision whereas in pattern III, the patient defines the decision parameters while the physician makes the final decision. Again, the categorization was done by two raters, who achieved an interrater reliability of κ = .94, which is an almost perfect observer agreement according to Landis and Koch (1977).

Validity of the decision making patterns
After categorizing all medically relevant decisions, we employed a linguistic analysis of conversation widely used in Germanspeaking countries for all decision making segments in our data: Deppermann's (2008) "Gesprächsanalyse" aims to reconstruct ways of interpersonal communicating, to identify regularities in conversations, to reconstruct communicative tasks, and to define communication practices on an empirical basis. The decision making sequences in our data showed a lot of variation in terms of phase of the conversation, speaker, action, and recipient. Therefore, we needed a method with which we could look at the conversations in depths but also remain on a meta level of analysis where we could still compare all these different segments. The underlying method for "Gesprächsanalyse" is sequence analysis, i.e., following a conversation's sequential process and examine the following categories (Deppermann, 2008): order of turns considering regularities of turn taking, interactive construction of turns, thematical linkage of turns, determination of segments and phases of conversations.
After we analyzed the decision making sequences linguistically, we compared the beforehand assigned pattern of distributed decision making with the linguistic description and interpretation of the decision making process to test the validity of the patterns and assess if the "label" matches the linguistic analysis.

Generalizability of the decision making patterns
In total, 36 medically relevant decisions were identified in the thirteen recorded physician-patient-conversations. The decisions revolved around matters such as changing medication or chemotherapy, arranging tests or examinations, and deciding on the next steps, etc. We used these decisions to apply the patterns of distributed cognition to a new set of data and test their generalizability. The distribution of the decisions over the four patterns are shown in Table 1. Unfortunately, no comparison with Lippa et al. (2017) is possible in terms of the distribution of the decisions over the patterns, as Lippa et al. (2017) did not indicate how many decisions they had in total or in each category, respectively.
The data did not include any patient-dominated decisions. Lippa et al. (2017) state that most patient-dominated decisions were to seek or discontinue care. These decisions usually are made at home and do not occur in the presence of a physician. Even before using our fine-grained linguistic approach, we encountered some difficulties when we tried to categorize our decision making material into the patterns of distributed cognition using Lippa et al.'s (2017) categorization method. We elaborate on these difficulties in the discussion section. Lippa et al. (2017) suggested that physician-dominated decisions occur when the decision is highly technical and does not require patient action. Twenty-five out of the 36 decisions in the data set were categorized with this pattern. In the following example depicted in Figure 2, a physician needs to make a technical decision and does so by deliberating with another physician. The sequence is located in the beginning of the conversation and starts after physician and patient greeted each other.

Physician-dominated decision from a linguistic perspective
In the beginning of the sequence physician and patient talk about the patient's condition and the patient has a question regarding his blood. After the physician starts answering the patient's question, he is interrupted by another physician (DB), who wants to know if he should order a component of the chemotherapy at the same percentage as the last time. The interruption is recognizable by the fact that he does not link his request thematically to anything said before. The physician confirms his request and after a short break continues to answer the patient's question about the blood.
"Then Let's Do Eighty per cent Again". According to Lippa et al. (2017) and shown in Table 2, this decision is physician-dominated. Looking at this decision making sequence closely, it becomes apparent that this decision is "physician-made" rather than "physician-dominated". It is not a shared decision at all as the patient has no real part in it despite being present while it was made by the two physicians. It seems as if a different conversation was inserted into the ongoing one. The patient would have to do considerable interactional work to gain back the floor and, for example, ask the physicians about the reason for this decision.
"Excuse Me, Professor". Excluding the patient from a conversation that he is part of should -if necessary -be accompanied by a metacommunicative utterance directed also toward the patient (Lalouschek, 2013) to establish or maintain an appreciative physician-patient-relationship. For example: "Excuse me. It is necessary that I interrupt your conversation for a moment because DA* and I have to decide on the composition of your chemotherapy now so I can order it and you are able to receive it on time." Physician-defined, patient-made decision from a linguistic perspective As described by Lippa et al. (2017), physician-defined, patientmade decisions characteristically concern biomedical information but require patient action. In total, four decisions from the data set were categorized with this pattern. In the following example, a patient's blood test is required. The selected sequence depicted in Figure 3 is happening after the physician explains to the patient what she must do when she gets released from the hospital.
In this sequence, the physician tells the patient that she has to come back to the hospital in ten days to do a blood test. Even though the patient is basically required to do this and is also aware of it (line 042: "then I have to come here"), the physician formulates his request very carefully (line 046: "is that doable?") and subjunctively (line 048: "reasonable would be", line 054: "then we would register you"). He suggests a date for the examination but highlights that it is only a proposal (line 48: "if I may suggest something"; line 050: "for example, yes."). The patient thinks aloud what that would mean for her and is aware of the inconvenience (line 59: "and they will take my blood there?") this means for her but is also very cooperative (approving back-channelbehavior in lines 040, 042, 044, 052, 056). After they arrive at a date for the blood test, the physician explains the next steps and what the patient has to expect when they take her blood sample.
"If I May Suggest Something". According to Lippa et al. (2017) and shown in Table 3, this decision is physician-defined, patient-made. This sequence shows the ambivalence that is often present in medical decisions: On the one hand, the patient is required to come in for the examination; on the other hand, the physician is seemingly open for the patient's wishes and eager to please her. This discrepancy occurs because there are certain things patients are required to do when they want to receive hospital care, but the request for patient autonomy leads physicians to give patients the impression that a decision is theirs to make. Therefore, the responsibility for the decision is seemingly given to the patient, but it is only formally made by the patient.

Patient-defined, physician-made decision from a linguistic perspective
Patient-defined, physician-made decisions often occur when a decision depends on a patient's case history and/or phenomenological information but also requires the physician's action (Lippa et al., 2017). Six patient-defined, physician-made decisions occurred in the data set. In the following example depicted in Figure 4, the patient demands to receive sleeping pills, which the physician has to prescribe.
The physician (DD*) wants to initiate a closing sequence (line 219), but the patient has one more request. She wants to receive sleeping pills (line 222 + 224). The physician asks what medication the patient normally takes. The patient does not recall immediately (line 232) but the nurse suggests a medication's name (line 233), which turns out to be the one the patient meant (line 236). The patient's daughter confirms this (line 235). The nurse also informs the patient that this medication had been reactivated by another physician (line 240), who is not present, which is confirmed by the physician present (line 241). "I Want Sleeping Pills". This pronounced statement suggests that this is a very important matter for the patient. Interestingly, the nurse -perhaps anticipating the importance of this matter for the patient -had addressed this topic already at the beginning of the conversation (not shown in this transcript). At that time, the physician was talking to the patient about her current condition and did not seize the nurse's interjection. The topic was only revisited when the patient raised it at the end of the conversation in line 222.
"Dr. Waldherr Activated It Again for Today". According to Lippa et al. (2017) and shown in Table 4, this decision is patientdefined, physician-made. Upon closer look, thus, it becomes evident that the decision the patient is fighting for (i.e., to receive sleeping pills) has already been made by another physician earlier. Therefore, it seems that stage I of the decision matter -identification of decision point -started much further back than we considered when categorizing the decision with the pattern of distributed medical decision making initially. Quite obviously,  some matters are not decided within one conversation but are negotiated over a longer period -an aspect not accounted for in the expanded decision space model, in which decisions tend to be conceived of rather isolated events.

Patient-dominated decision from a linguistic perspective
Patient-dominated decisions are dependent upon the patient's evaluation of the state space and do not require physician action (Lippa et al., 2017). We could not find any decisions in  the material, which fell in that category. There was, however, one decision which could not be categorized with any of the patterns (see Table 5). With the help of our linguistic lens, we wanted to understand the reason for this categorization problem.
The sequence depicted in Figure 5 is located rather at the beginning of the conversation. The physician sums up the patient's current state and tells her about her latest test results. Then he wants to close his summary by informing the patient about what is going to happen in the next few days, but the patient objects.

Patient-defined, physician-made decision
The selected sequence starts with the physician announcing the next steps for the patient (line 19). After a short pause, the patient objects because she is "going home tomorrow" (line 21). The nurse backs the patient by saying that this has already been settled with the chief physician (line 024). The physician admits that the day before it was still undecided (line 027) and then agrees to let the patient go home but insists on doing a blood test in the morning before she leaves (line 030).
"But I'm Going Home Tomorrow". Despite the patient's open objection to the physician's announcement the ensuing decision could neither be categorized as patient-dominated nor as one of the other possibilities in the frame of Lippa et al.'s (2017) assigning method, because the parameters are defined by three parties. The linguistic analysis provides further insight into this unusual decision making sequence as, from a linguistic point of view, it is an other-initiated repair with the patient initiating the repair in line 021 (Kendrick, 2015). This phenomenon explains why the patient is able to disagree with the physician so directly. Furthermore, the patient's objection is based on an earlier agreement with the chief physician.
Hence, the decision that the patient will be released from hospital care was already made before this conversation by another physician; cognitivistically speaking, thus, it is "physician-dominated".
"This Is Settled With the Professor". Another important aspect here is the influence of the nurse who confirms the patient's objection and thus helps her make it credible. If the nurse did not support the patient's perspective, the physician could stick to his version and keep the patient in the hospital longer. Again, the linguistic analysis points to possible pitfalls of a historically flat cognitivistic account which cannot sufficiently encompass such contingencies.

Discussion
In this study, we aimed to challenge the expanded decision space model and the four patterns of distributed decision making derived from it in terms of generalizability and validity. Although it was, in principle, possible to categorize all but one of the medical decisions in our data into three of the patterns of distributed decision making, we encountered a recurrent "We will wait over the weekend and continue with chemotherapy next week." Patient: "But I go home tomorrow" Nurse: "This is settled with the professor." Physician: "You could either go home for the weekend . . . " "Okay, but we do a blood test tomorrow morning."

Allocation of Responsibility
Physician ( problem: In many cases, it was difficult to distinguish if a decision (e.g., Physician: "We would like to add a new medication so that the chemo is better tolerated." Patient: "Okay.") was physiciandominated (I) or physician-defined, patient-made (II). The physicians often have a decision ratified by the patients and virtually transfer responsibility for it to them, but do not inform them about alternative options. In effect, the patients have little room for maneuver. This observation corresponds with Tate's (2018) finding that physicians phrase suggestions for treatment actions more cautiously in the beginning, when the physician-patientrelationship is not yet established, but rather pronounce them more later on. In terms of generalizability, it is of course a limitation that we could not assign all decisions from our data and that one pattern did not occur at all. However, this could be due to the sample size or the specific characteristics of cancer patients which could differ from the ones of patients in other chronic care settings such as the Multiple Sclerosis clinic examined by Lippa et al. (2017).
In cancer care, it is definitely more common that relatives are present during physician-patient-conversations than in settings like the emergency department observed by Lippa et al. (2017), but also in other setting medical decisions are usually distributed over several different people (Rapley, 2008). The fact that other people -like nurses, other physicians, or relatives -are not accounted for in the expanded decision space model is considered a weakness in terms of its ecological validity.
The lack of patient-dominated decisions may be due to a limited or specialized data set, but very likely these kinds of decisions are at least not as common or frequent as the symmetry of the patterns of distributed decision making lead us to believe. Our results indicate that the expanded decision space model -and perhaps other cognitivistic models, too -led by a "rhetoric of equality", overestimate the agentive capabilities of patients and disregard the still active power imbalance between physicians and patients (Giese, 2002). The expanded decision space model conceptualizes physician and patient as equal partners during the decision making process and therefore obtains symmetric decision patterns.
From a linguistic point of view, there was a lot of variation in the decision making sequences in terms of when the decision occurred during the consultation (begin, midst, ending), in terms of the action of the turn (e.g., request, recommendation, suggestion), selected recipient etc. While the expanded decision space model does not take these differences into account, these variations have major implications for the patients regarding their options for responding and their capabilities to participate, resist or accept. Hence, the cognitivist approach tends to only consider the result but not the dynamics which lead to a certain decision and thereby tends to overestimate the patients' possibilities to participate in the decision making process. Lippa et al. (2017Lippa et al. ( , p. 1045 argue that their patterns of distributed decision making "form a continuum of shared decision making". We observed in our data set a variety of decision making processes that did not include the key elements of shared decision making (Bomhof-Roordink et al., 2019). We therefore argue that the patterns of distributed decision making also represent decisions outside this continuum. Furthermore, we found decisions within pattern I, which were purely unilateral decisions, i.e., without any patient involvement. Though Lippa et al. (2017) share this observation in their data sets they do not indicate how to handle these decisions in terms of categorizingperhaps in order to avoid having decisions without distributed cognition. We consider this a limitation both in terms of the validity as well as of the distinctiveness of Lippa et al.'s suggestions. Lippa et al. (2017) implicate that distributed clinical cognition automatically leads to SDM, as both physician and patient are involved in the decision making process. We hold that clinical cognitive distribution did not occur in all of the decisions, which we would not categorize as SDM, but it did in others, which still do not fulfill the quality requirements of SDM. We would therefore be very cautious with the implication that distributed clinical cognition automatically leads to SDM. Numerous researchers worked out models how to facilitate SDM and how to remove the barriers for SDM (Elwyn et al., 2000(Elwyn et al., , 2012(Elwyn et al., , 2014Légaré & Witteman, 2013;Légaré et al., 2008;Stacey et al., 2012) that we consider it a step backwards to implicate that SDM is ensured through cognitive distribution alone.

Limitations
Firstly, only male physicians were conducting the conversations; in future research it would be interesting to put more focus on gender relations and their impact on decision making as different dynamics could occur. Secondly, the sample consisted of only thirteen conversations, which limits the explanatory power of the results. Nevertheless, the data contained thirty-six decision making situations, which we thought to be sufficient for our analytical approach. Thirdly, only patients with additional insurance were part of the study, a fact that could also be influential for the degree of their being -at least rhetorically -involved in the decision making and should be addressed in further research. Fourthly, patients were older than their physicians. Age relations and physicians' experience should also be considered in further studies from a linguistic angle.

Conclusion
Despite the advantages models can offer in terms of understanding complex phenomena like (shared) decision making, we think that a deeper, more comprehensive understanding is needed to examine medical decision making, at least in cancer care. The expanded decision making model overestimates patients' power in decision making processes, does not consider multiparty dynamics, and fails to depict decision making processes, which are negotiated over a longer period of time. These are important aspects of SDM in cancer care, which considerably limit the ecological validity of the expanded decision space model. In our oncology data set, the generalizability, validity, and distinctiveness of the patterns of distributed decision making were not satisfactory.
More importantly, distributed cognition between physician and patient should not be equated with SDM. Our results showed that while distributed cognition may occur in SDM processes, distributed cognition alone does not ensure SDM and says little about the quality of the interaction. SDM needs physicians and patients, who are willing to share medical decisions, and have to be actively worked out between them (e.g., as proposed by Elwyn et al., 2014).