METHODS

Subjects

From a random sample of 101 general practitioners working in the northern part of the Netherlands, 61 participated in an in depth study focusing on prescribing.¹⁹ There were 54 men and seven women and they had been in general practice for a mean (SD) of 14 (8.5) years (range 2–37). The participating doctors were comparable with the whole population of general practitioners in the Netherlands at the time of data collection with regard to mean age, practice size, and practice organisation.²⁰

Material

The original data were collected in 1992 and 1993. Part of the material collected then was used for the present study. All doctors had made treatment decisions for five written patient cases. For the diagnoses described in these cases there have been no significant changes in the recommended first choice treatment in recent years.^17,18,21,22

The written cases were based on actual patient cases and were tested in a pilot study with 30 general practitioners practising outside the study area. In this pilot the five written cases were selected from a set of 12 cases using the following criteria: (1) the written case included sufficient (diagnostic) information for making a treatment decision; and (2) a panel of two medical doctors and one pharmacist, all involved in academic teaching of pharmacotherapy, had agreed on what the recommended treatment would be according to the guidelines.^17,18

The cases described a patient's complaints, the circumstances, and diagnostic information (table 1). Two were dealing with uncomplicated urinary tract infections in non-pregnant women (cases 1 and 2) and were relatively short (100 words). Three patient cases dealt with stomach complaints (cases 3, 4, 5); these were longer (170– 315 words) and included contextual aspects such as patient demand. Co-medication was mentioned in cases 1 and 5. For the urinary tract infection cases it was stated that these women were not pregnant, but no information was given on other contraindications such as drug allergies.

Data collection

The “think aloud” method was used to elicit the doctors' decision processes. “Think aloud” is a process tracing method which reveals the cognitive processes engaged in while making decisions, without affecting the outcome of the decisions.^12,23,24 Subjects are asked to verbalise all thoughts that come into their mind while performing a decision task. Most people are able to do so quite easily after a little practice; the only side effect is that it may slow down the decision process. “Think aloud” gives more accurate information about decision processes than when doctors are asked to describe or explain their decisions, since most people are not fully aware of all the thoughts that crossed their mind before the final decision was made and tend to make hindsight rationalisations.^23,25,26 The most important limitation of the “think aloud” method is that not all thoughts can be verbalised. For example, recognition of and reaction to non-verbal stimuli, such as smells or patterns, are difficult to verbalise. Such stimuli were not present in our patient cases. Furthermore, automatic thought processes are often one step procedures of which only the outcome will be verbalised. These will be identified as such in our analysis.

The doctors were presented with the written patient cases and asked to verbalise aloud all thoughts and considerations that came into their minds while reading the case and while deciding upon the treatment. To get used to verbalising all thoughts each doctor started with an extra patient case which was not included in the analysis. The verbalisations were tape recorded and subsequently typed out verbatim. This gave us written transcripts of the decision processes.

Data analysis

Scoring the transcripts

For all 61 doctors there were separate transcripts for each of the five patient cases resulting in 305 transcripts. These 305 transcripts were cut up into segments of independent sentences or groups of closely associated sentences. Two of the authors (CLMW and HWS) independently coded these segments to indicate the type of thought it contained. Cohen's kappa was calculated to assess their agreement regarding the codes assigned to each segment. The following coding categories were used: 0 for verbalisation of the text of the written patient case; 1 for remarks not related to the treatment choice; 2 when a treatment option was mentioned; 3 when an aspect relevant for a treatment choice was considered; and 4 for the final treatment choice.

We distinguished between drug treatments, non-drug treatments, and no treatment. To establish the number of treatments mentioned, a class of drugs was counted as a separate treatment option if no further generic or brand names belonging to that class were mentioned by the doctor. When a brand name as well as its generic name was mentioned, these were counted as one treatment option. Aspects relevant to the treatment choice were divided into: (1) core aspects referring to effects, side effects, co-medication, co-morbidity, or other characteristics of the treatment itself; (2) contextual aspects referring to specific circumstances influencing the decision such as previous experiences of the patient or patient demand; and (3) habitual aspects—that is, references to habits or standard treatment (table 2).

Categorising the decision process

On the basis of the coded transcripts, the doctors' decision processes were categorised as one of seven types depending on the number of treatments and aspects considered (table 3). These categories refer to differences in scope and nature of the decision process. When only one treatment is mentioned, this indicates a decision process in which no comparison is made between different alternatives. This could be a type A process in which also no aspects are mentioned, or a type B or C process in which one or more aspects are mentioned. Type A is typically an automatic or habitual process. Type B and, especially, type C can be interpreted as satisficing strategies—that is, settling for the first treatment that comes into mind which satisfies some basic criteria such as being effective and not contraindicated for that patient. This is contrasted with optimising strategies where one tries to evaluate and compare several options in search of the optimal choice.²⁷ Types D, E, and F are processes in which more than one treatment was mentioned and again either no, one, or more than one aspect. Type F processes are closest to the optimising strategy. This thorough deliberation of different treatments on different aspects is seen as the normative model in decision theory.²⁷ Transcripts that mentioned no treatment at all were classified as type N.

The type F processes were described in more detail by classifying the steps made when going from one treatment or aspect to another as either alternative wise, aspect wise, or diagonal steps. The occurrence of many alternative wise steps indicates that the doctor first considers one alternative on a number of aspects before moving to another alternative. Such a strategy may inhibit a thorough comparison of treatments. Aspect wise steps indicate that treatments are explicitly compared with each other on an aspect. In diagonal steps such comparisons could be made implicitly—for example, when a doctor says “X is quite effective, but Y is a lot cheaper”.

Statistical analysis

The influence of the patient case on the number of treatments and type of aspects considered during the decision process was tested with non-parametric tests for more than two related samples (Friedman test for ordinal data and Cochran's Q test for dichotomous data).²⁸ The relation between the doctors' years of practice experience and number of treatments or aspects considered was tested for each case separately with non-parametric correlations (Spearman's rho). To test whether differences in the decision process were related to prescribing different treatments, χ² tests or Fisher's exact tests were used, making comparisons between first and second choice treatments for each case. The treatments prescribed were classified as either first choice or second choice treatments according to the Dutch national guidelines for general practice.^17,18,21,22

RESULTS

There was a high level of agreement in the coding of the transcripts: 92.6% of the text segments were classified in the same coding category by both judges resulting in a kappa value of 0.84. In total, 2.9% text segments were coded as not relevant (code 0 or 1) by one of the judges and relevant (code 2, 3 or 4) by the other. In addition, there were a few discrepancies (1.4%) between assignments of code 2 (mentioning a treatment option) and code 4 (the final treatment choice), and between subcategories of one code (h1.4%). These differences were resolved in discussion.

In 175 of the 305 transcripts (57%) only one treatment was considered, whereas two or more were considered 125 times (table 4). Core aspects were mentioned in just over half of the transcripts, contextual aspects in more than a third, whereas “habit” was explicitly mentioned in relation to 14% of the treatment decisions (table 5). In a quarter of the cases no explicit considerations were made, indicating an automatic or habitual process. The core considerations in case 1 mostly referred to the co-medication mentioned in this case (table 6), but many said they would rely on the pharmacist for checking possible drug-drug interactions. The possibility of a drug allergy was almost never explicitly contemplated, although this was relevant before prescribing an antibiotic (cases 1 and 2). Efficacy was the most often mentioned consideration in cases 2, 3, 4 and 5, followed by side effects and by mechanism of action for the stomach complaint cases (table 6). Costs were considered only in some of the stomach complaint cases. Contextual aspects were considered in almost half of the cases with stomach complaints (table 5). Patients 3 and 4 explicitly asked for drug treatment, whereas patient 5 implicitly emphasised the need for an effective drug. These patient demands were considered by 60.7%, 24.6%, and 19.7% of the doctors, respectively. Other frequently mentioned contextual aspects were “going away on holiday” (case 3), “previous use of an over the counter drug” (case 4), and “previous experience with an H₂ antagonist” (case 5). Finally, several general practitioners mentioned their own good experience or affinity with a specific drug for stomach complaints.

Common decision processes were type A, mentioning only one drug without any considerations; type B, considering one drug on only one aspect (either a core or contextual or habitual aspect); and type F, considering two or more drugs on two or more aspects (table 7). Aspect wise steps in the decision process, indicating that two treatments were explicitly compared on the same aspect, were made 109 times in the type F transcripts. Alternative wise steps were made 83 times, whereas diagonal steps going from mentioning one aspect in relation to one treatment to another aspect related to another treatment occurred 208 times. Again these aspects could be either core, contextual, or habitual in nature (table 2).

Background variables

The cases themselves had a clear effect on the number of treatments, type of aspects, and type of decision process (tables 4, 5, 6, 7). The urinary tract infection cases predominantly triggered more simplistic decision processes (types A and B). Alternatives were considered by only 16.4% of the doctors for case 1 and by 26.2% for case 2. In case 2 this occurred mostly in view of the recurrence of the infection. For the stomach complaint cases, two or more drugs were mentioned by 67.2%, 37.7%, and 59.0% of the doctors, the majority considering more than one aspect (type F). In particular, more contextual aspects were mentioned in relation to the stomach complaint cases.

The experience of the doctors was not consistently related to the number of treatments or aspects considered. Only for case 1 did doctors with more practice experience mention fewer core aspects (Spearman's rho correlation coefficient –0.276, p=0.031), and for case 4 doctors with more practice experience mentioned fewer treatment options (Spearman's rho correlation coefficient –0.256, p=0.046).

Relation to prescribing second choice treatments

For the urinary tract infection cases most of the general practitioners (91.8% and 90.2%) prescribed drugs recommended as first choice in the national guideline for general practice (table 8). Second choice drugs such as norfloxacin or co-trimoxazole were prescribed five times in case 1 and six times in case 2, four times by the same doctor. In eight cases this decision process could be classified as habitual.

For the cases with minor stomach complaints (nos 3 and 4), second choice drugs (e.g. H₂ antagonists) were prescribed by 65.6% and 54.1% of the doctors, while the others prescribed antacids or no drug treatment. For the patient with reflux complaints (case 5), most doctors prescribed an H₂ antagonist which would be the drug of first choice. Second choice drugs were more often prescribed after consideration of two or more alternatives in a type D, E, or F process (table 8). There was no relationship between the number of core or habitual aspects mentioned and the outcome of the decisions. Doctors more often mentioned contextual aspects before prescribing a second choice drug for case 3; 31 of the 40 doctors who prescribed a second choice drug considered one or more contextual aspects, whereas eight of the 21 doctors prescribing a first choice drug considered contextual aspects (p=0.004, Fisher's exact test). Furthermore, costs were only mentioned by doctors prescribing a first choice drug in cases 4 and 5.

DISCUSSION

The doctors in this study were not asked to explain or defend treatment decisions, but just to verbalise all thoughts when making a decision for an individual case. This gives insight into what is actively considered during this process and helps to uncover where and how prescribing decisions could benefit from support.

The decision behaviour of the general practitioners was not very extensive and, in almost 40% of the transcripts, it could be classified as habitual. More than one drug was considered in less than half of the transcripts. Usually only a limited number of considerations was made. Some considerations referred to core aspects of the drug or the patient case, others referred to more contextual aspects. When doctors evaluated different treatments on more than one aspect, they often made diagonal steps going from one aspect of one treatment to a different aspect of another. This indicates that they made shortcuts and did not necessarily compare treatments explicitly with each other, which is in line with previous findings among medical specialists.^29–32 There was no consistent influence of the doctors' practice experience on the decision process.

The behaviour observed clearly deviates from the decision theoretic norm of making an inventory of the treatment options, filtering out those options that are risky, and weighting the remainder on relevant aspects such as expected effectiveness, side effects, and suitability for the patient. One could argue that it is not pragmatic to follow this norm for repetitive decisions, and our study showed that simple decision rules may result in choosing a first choice treatment. On the other hand, it is obvious that prescriptions that ignore possible side effects, interactions, or contraindications should be avoided.

The type of aspects considered is partly influenced by the case descriptions since contextual aspects in particular will only be considered when included in a case. However, the doctors did not actively consider all information presented and made their own selection for each case. The doctors almost never asked for information on drug allergies relevant to the patients with a urinary tract infection, and possible drug-drug interactions were only considered by half of the doctors and seldom checked. Decision support can be of help by bringing such pertinent information to the prescriber's attention.^9,33,34

With regard to characteristics of the medication, most considerations referred to efficacy and mechanism of action, while possible side effects and costs were only mentioned occasionally. Although many doctors say that these latter aspects are important for their prescribing decisions,³⁵ it seems that only a minority consider these aspects when choosing a treatment for an individual patient. When looking at the available treatment guidelines, it becomes clear that side effects and costs are relevant criteria for differentiating between first choice and second choice treatments.^17,18,21,22 Many drugs are considered effective for the type of cases included in our study, and it is advised that drugs should be prescribed that are cheap and cause few (serious) side effects.

For the stomach complaint cases, two thirds of the second choice drugs were chosen after deliberation of other alternatives. Offering correct information on the expected benefits and costs may then be useful. When such support is given routinely, however, doctors may feel overloaded with information during the consultation.¹¹ Since general practitioners seldom focus on more than one core aspect, they may perceive it as a disruption when a support system forces them to pay attention to other aspects. Such “intrusions” on the normal procedure should be made either on request or sparingly with a specific purpose.

Second choice drugs were sometimes prescribed in view of specific circumstances. One could argue that this may be justified. For instance, in one case the patient had used a first choice drug himself without experiencing much effect. Computer systems do not include such contextual information and this may be perceived as a drawback by the doctors. On the other hand, it is clear that some doctors are more sensitive to contextual aspects than others. A decision support system may help these doctors to resist, for instance, patient demands for second choice treatments.

Finally, most general practitioners had their favourite drug for treating urinary tract infections and this was sometimes a second choice drug. Either they chose it habitually or evaluated it on one or two aspects—that is, following a kind of satisficing strategy. Habits are usually difficult to change and it is unlikely that doctors will consult a decision support system for habitual decisions. A support system could help to remind doctors of the first choice treatments. For those who are willing to change their habits, such reinforcements may be useful.³⁶

Conclusion

The decision process of general practitioners deviates from the decision theoretic norm of thoroughly evaluating all possible options, but these deviations do not always result in suboptimal prescribing behaviour. Divergence from optimal care can be due to unawareness of situations where certain drugs may or may not be beneficial, to focusing primarily on efficacy or contextual aspects and not on costs, or to habitual preferences for “second choice” treatments. Decision support can lead to improvements in the quality of prescribing when it interrupts the decision making process when needed—for example, by bringing relevant side effects, costs, interactions, or contraindications to the attention of the prescriber and by reminding them what is considered the treatment of first choice for a specific patient. It should be made clear to the users that support systems can never incorporate all circumstances of a patient, and that the doctor remains responsible for the final treatment decision.