Do patients’ preferences prevail in hospital selection?: a comparison between discrete choice experiments and revealed hospital choice

We focused on insured individuals who had received surgery for breast cancer or cataract in 2010 in one specific region of the Netherlands. We chose these conditions for two main reasons. First, as publicly reported quality information was only available for a small number of conditions in 2009, we selected breast cancer or cataract since the quality information had been collected for a relatively large number of patients compared to, for example, diseases of the adenoids and tonsils. Second, breast cancer and cataract patients had to exercise choice in very different settings, given the differences in medical urgency for treatment. In 2010 and across all Dutch hospitals, the waiting time between referral and first hospital appointment was, on average, 1.5 weeks for breast cancer and 4.9 weeks for cataract [13].

Patients were identified using claims data of 2010 and were eligible if they had received surgery in one of the four general hospitals located within a 15-km radius of the metropolitan area of Eindhoven in the Netherlands. This area had 750,000 inhabitants in 2010 [14] and was considered to be highly suitable due to the absence of academic hospitals as this would avoid potentially undesirable concentrations of high-risk patients (e.g., patients with comorbidities) which generally occur in these hospitals.

To measure stated preferences, we conducted discrete choice experiments (DCEs). In this methodology, respondents are presented with a series of hypothetical scenarios (choice sets). Each choice set contains two or more alternatives that are described by a systematic combination of attributes and levels. Within each choice set, respondents indicate their most preferred alternative which, according to the Random Utility Maximization framework [15], is assumed to maximize their utility thereby reflecting their latent preferences. In our study, patients had first received their surgery and were then invited to complete the DCE questionnaire (i.e., post-surgery). We conducted a DCE for each condition separately.

As we used claims data to identify eligible patients, we could also determine in which of the four hospitals patients had received their surgery (revealed hospital choice). To describe the four general hospitals, we used four of the five attributes as information for recommendation was not recorded in claims data. For waiting time, patient experiences and clinical outcome indicator, we derived the hospital’s actual score on these measures from national databases and included these values as level [17, 18]. For travel distance, we first computed the patient-hospital distance using postal codes and then transformed this distance into an appropriate level also used in the DCE (3, 8 or 15 km).

The questionnaires were completed by 1740 respondents in total (response rate: 46.5% (breast cancer) and 66.6% (cataract)). As shown in Table 2, most respondents were female, aged between 45 and 64 (breast cancer) or between 65 and 79 (cataract), had attained a low educational level and perceived their general health and mental health (breast cancer only) as good to very good. In addition, most respondents had not been provided by their GPs with hospital quality information. While the majority of the cataract group indicated that no one in particular had recommended the hospital in which they received their surgery, proportionally more persons in the breast cancer group relied on the advice of their GP. Furthermore, on average and relative to representative patient populations, respondents (both patient groups) were similar in terms of age and gender but had attained a lower educational level.

Both MXL models (breast cancer and cataract) demonstrated theoretical validity (Table 3). All attributes had their expected sign and levels that were significant. For dummy coded attributes with ordered levels (i.e., patient experiences and travel distance), the ordering of the point estimates was consistent with the ordering in levels; for example, the marginal utility for the “above average” level of patient experience (breast cancer) was larger than that of the “average” level.

Regarding the relative importance of attributes (Table 4), clinical outcome indicator was the most important attribute for both cataract and breast cancer as it yielded the largest MRS. For example, cataract patients were, on average, willing to wait an additional 210.5 working days (95% CI: 140.8–280.2) to be able to select the hospital that had scored 0.4% (i.e., favorable score) on the per-operatively performed vitrectomy’ indicator over a hospital with a 1.7% score (i.e., unfavorable score). For both patient groups, the second most important attribute was patient experiences. The importance of both quality-of-care attributes implied that patients were willing to wait longer for better quality of care. In addition, the least important attribute was travel distance (breast cancer) and recommendation (cataract). The latter was in line with the lower self-reported importance of the advice of GPs among cataract patients relative to breast cancer patients as described above and shown in Table 2. Moreover, the observed MRS values for cataract were larger and less concentrated in terms of range (range: 32.2 to 210.5) relative to those for breast cancer (range: 3.0 to 51.4). This implied that, on average and ceteris paribus, cataract patients were more responsive to a change in level at a given attribute compared to breast cancer patients.

Supplementary Material 2 provides a detailed description of assigned levels for the four general hospitals in real-life settings. In short, for breast cancer and given their assigned levels, hospital 2 and 4 were labelled as “nearby located” hospitals and hospital 3 as the “overall best quality of care” hospital. For cataract, hospital 4 was labelled as a “nearby located” hospital, while hospital 2 was labelled both as a “nearby located” hospital and as the “overall best quality of care” hospital.

As shown in Fig. 1, the predicted distribution of patients across the four hospitals differed from that observed in real-life settings in terms of absolute value and relative order, and these differences were more apparent for breast cancer than for cataract. While most breast cancer patients (average (95%CI)): 58.5% (53.9–63.1)) were expected to select the “best quality of care” hospital (i.e., hospital 3), only 21.7% of the group did so in real-life settings. Instead, most patients selected the hospitals labelled as a “nearby located” hospital (i.e., hospital 2 (27.3%) and hospital 4 (33.3%)). Hence, the relative order of all four hospitals differed between the predicted and observed situation. For cataract, most patients (average (95%CI): 53.6% (50.7–57.0)) were predicted to select hospital 2 (i.e., the “overall best quality of care” hospital). Similar to the predicted situation, the largest share of patients was observed for hospital 2 (34.4%) followed by hospital 4 (27.0%), although proportionally fewer patients selected hospital 2 than expected. Unlike for breast cancer, the relative order of the four hospitals only differed for the two hospitals with smallest share of cataract patients.

In the interest of brevity, results of all additional analyses are provided in full detail in Supplementary Material 3. In 4-attributes analyses and in IPW analyses, similar results were observed and did not change the overall conclusion of the main analyses: for both analyses, MRS values resembled those of the main analyses and the observed difference in predicted and observed distribution of patients across the four hospitals persisted (Figs. 2 and 3). Moreover, the assessment of explanatory power revealed that travel distance was the most important attribute in the revealed choices for both conditions: as shown in Table 5, this attribute accounted for 85.5% (breast cancer) and 95.5% (cataract) of the log-likelihood.

We set out to assess to what extent a patient’s choice of hospital in real-life settings corresponds with their preferred choice. Our results indicated that (1) while both breast cancer and cataract patients valued quality of care in their stated hospital choice, hospital quality was of lesser importance in their revealed choice. (2) In contrast, a patient’s travel distance was the most important attribute in the choice of hospital in real-life settings for both patient groups. (3) The predicted distribution of patients across the four general hospitals differed from that observed in real-life settings in terms of absolute value and, for breast cancer also in relative order. (4) For both conditions and relative to the main analyses, similar results were observed in both the 4-attributes analyses and in  the population weighted analyses.

The discrepancies between the relative importance of attributes in stated and revealed hospital choices, and between the predicted and revealed distribution of patients are consistent with the literature: patients state certain characteristics as important in hypothetical settings, but act upon others in real-life settings [6]. As described in more detail below, this may be explained by the fact that most patients rely on their GP’s advice in patient choice [5, 6, 31], and that the preferences of patients and the physicians who represent them may often not align [32].

With respect to the DCE, we observed the largest MRS values for both conditions for clinical outcome indicators followed by patient experiences. For instance, breast cancer patients were, on average, willing to wait an additional 38.6 working days (95% CI: 32.9–44.2) to select the hospital with a favorable score of 5% on the tumor-positive resection margin’ indicator over a hospital with a score of 20%. However, while highly valued in the DCE, quality information only played a minor role in the revealed hospital choices; a finding that supports previous research. Faber et al. have concluded that, although valued as important by patients, quality information rarely affects decisions in real-life settings [7].

With respect to the revealed hospital choices, the most important attribute was travel distance: additional analyses showed that this attribute contributed for 85.5% (breast cancer) and 95.5% (cataract) of the model’s log-likelihood. The apparent difference in the relative importance of attributes between stated and revealed hospital choices may explain why the predicted distribution of patients differed from that observed in real-life settings. For breast cancer, while most patients were expected to select the hospital labelled as “best quality of care“, they selected the hospitals labelled as “nearby located”. For cataract, a similar shift was observed, although it did not affect the relative order of hospitals in terms of shares of patients. The smaller shift may be explained by the fact that the hospital labelled as the one with the “best overall quality of care" “was also labelled as a “nearby located” hospital. Our finding that patients generally go to the closest hospital, is also in line with literature: research has shown that (1) patients generally go to the nearest provider, (2) that they prefer the status-quo option (i.e., prefer to be treated in the hospital where they already have been treated), and (3) that they are only more likely to switch hospitals if they have had a bad experience with their current provider or are faced with long waiting lists [5, 6].

While consistent with previous studies, the limited importance of quality information in practice is somewhat surprising in the light of our study protocol [10]. Some of the GPs involved in our study were also involved in the development of the patient report cards [10], and were likely to be motivated to stimulate patient choice while using these cards. Moreover, as part of our study protocol, all GPs were instructed to present patients with these report cards to stimulate patient choice [10]. Hence, patient experiences and clinical outcome indicator were expected to be of greater importance in the revealed hospital choices. The literature provides a possible explanation. Most patients rely on the advice of their choice of hospital [5, 6, 31]. However, research has shown that preferences of patients and those who act on their behalf (e.g., physicians) may often not align [32]. For instance, in a study conducted by Empel et al. physicians have underestimated the importance of patient experiences to patients regarding fertility care [33]. Moreover, studies have shown that GPs generally do not take quality information into account in their referrals, but rather rely on other factors such as their own preferences, close connections and good previous experiences with the specialists working at a given hospital [9‐11]. On the one hand and given the shown importance of the GP’s advice, these factors may, ceteris paribus, explain the lower relative impact of quality information observed in our study. On the other hand, it is reasonable to assume that GPs have close connections with the specialists working at nearby hospitals. As shown by the large explanatory power in our study, travel distance may capture the close relationship between the GP and the hospital departments and act as a proxy thereof in practice. This implies that patients are more willing to travel beyond their nearest hospital for better hospital quality than research suggests [5, 6].

Our study shows that the assumption of many current policy reforms - patients and referring physicians who choose on their behalf, choose hospitals in accordance to the preferences of patients – is unlikely to hold in practice. At the point of referral, other factors (e.g., time constraints) come into play that prevent the patient’s stated preferences to prevail.

For patients to adopt the active role which is often assumed in health policy, additional efforts are required. If patients are in fact willing to take charge in the decision-making process for a hospital, GPs need to incorporate time during their consultations to discuss possible options, ideally, while using decision support tools. This approach would tackle the two main reasons why most patients do not actively choose their providers: (1) the perceived limited degree of choice (e.g. due to the health insurer’s constraints) and (2) the lack of adequate and suitable information to support their choice (e.g. patients are often overwhelmed by the large amount of publicly available quality information [6]). A tailored decision support tool allows the GP to present all possible options given the patient’s health insurance coverage. Similarly, these tools may tailor comparative information to the needs of patients by presenting only a subset of the available quality indicators (i.e., the most important ones) [34]. However, some patients are not willing or are unable to become actively involved in patient choice and simply prefer their GP to decide on their behalf. Therefore, knowledge of the preferences of the given patient group are needed. This calls for patient-group specific preferences studies to gain insights into what subgroups patients prefer, followed by the development of decision support tools that present GPs with patient-group specific information.

Furthermore, the discrepancy between the predicted and observed distribution of patients should not be seen as potential evidence of the (non)-predictive value of DCEs. For such evidence, the context of the given decision requires the decision-maker to have freedom of choice and to be willing to exercise their choice in accordance to their preferences (see, for example, in health care [35]). Our study has investigated whether patients do in fact have the ability to choose their hospitals in accordance to their preferences and unconstrained by external parties.

As research comparing the stated and revealed choices of the same sample is scarce [6], our study adds new and unique evidence to the literature. From a methodological perspective, the use of the same sample has allowed us to rule out potential differences in preferences between the stated and revealed choices that may result from sample variability. In addition, we have used DCEs to quantify preferences: a meta-analysis concludes that DCEs are able to predict health-related decisions with a moderate accuracy [36]. Furthermore, we have performed IPW analyses to improve the representativeness of our sample; these analyses have not affected our general conclusion, and thus indicate that our findings are representative for the total patient population in terms of gender, age and educational level.

Our main limitation is that we have not been able to model the attribute recommendation in the revealed choices as this information has not been recorded in claims data. Although we have asked respondents in the questionnaire who had recommended the hospital in which they have been treated, we believe that this information cannot be validly collected for the non-chosen hospitals due to, for example, recall bias. As we have relied on the Random Utility Maximization framework in our models, we expect that an omitted attribute may lead to different marginal utility coefficients and estimated MRS of the included attributes in absolute terms, but it should not impact the relative order of attributes. Findings of our 4-attributes analyses support this hypothesis.

Our second limitation relates to the timing of our study. Regarding the data, we have used data that originates from 2010, 4 years after the 2006’s reform that has introduced provider competition in the Dutch health care system. Patients and GPs may have required more time to fully adapt to their new role with regard to patient choice. However, as our findings are consistent with previous studies that have been conducted across various health systems and time periods [5, 6, 31], we expect that our conclusions still hold in current daily practice. Regarding the experimental design, our design (i.e., a fold-over of an orthogonal main effects plan, forced choice sets) has been constructed in accordance with the common practice of 2010 [37]. Given that both MXL models (breast cancer and cataract) have demonstrated theoretical validity and in accordance with Rose and Bliemer who have stated that “given large enough samples, the underlying experimental design should not matter in terms of statistical power” (p612) [38], we believe that the use of a fold-over design does not affect our findings.

Furthermore, although we have considered as many patient characteristics as possible in our IPW analyses, we do not have data on health status. We expect that patients in poor health may be less likely to participate and thus may be underrepresented in our samples. Similarly, we lack any information on variables such as patient activation [39]. As patients with higher activation levels are more likely to actively choose providers than those with lower levels [6], we hypothesize that, for example, recommendation may play a smaller role in patient choice among the former relative to the latter.