This study identified five patient clusters related to the evaluation of hospital outcomes, as well as a group of heterogeneous extreme negative outliers. The first cluster groups all patients who attributed the highest scores for general satisfaction, absence of malpractice, and treatment benefit. The second cluster gave high scores to treatment benefit and absence of malpractice, but those patients were less satisfied overall than those in the first cluster. By contrast, the third cluster grouped those who were satisfied and did not describe malpractice, even if they gave a lower score to treatment benefit. The fourth cluster gave relatively balanced scores to each of the three indicators. The fifth cluster included patients not highly satisfied overall or not satisfied with treatment benefit, but who declared less malpractice. Finally the last group represents outliers who were not satisfied overall, declared malpractice, and little treatment benefit. The two groups with systematically poorer evaluation across outcomes comprised almost one-quarter of all patients, indicating the clear potential to improve hospital services from the patient perspective.
This study identified more clusters than previous studies. However, previous research in this field is scarce and heterogeneous, with differences in patient groups, statistical approaches and numbers of response clusters [18–20]. One previous study identified five clusters of patients , while other have identified two  and three . The two-cluster study is not comparable to the current study, since it was restricted to one hospital and only included 47 patients with type 1 diabetes receiving kidney or pancreas-kidney transplant . The two other studies were more comparable, but a main difference was that these studies restricted cluster variables to experiences and satisfaction. Therefore, at the conceptual level these studies differ from the current study. However, a common cluster pattern was identified across these studies: a top-score cluster, a medium cluster, and a low-score cluster. More research is needed on response clusters in patient evaluations, but it is important to standardize the methodology, especially with regard to handling of outliers and choice of cluster variables. We recommend excluding outliers from cluster formation, but including them in the interpretation of patient clusters. Furthermore, we recommend using relevance and importance for patients as the main criteria when selecting cluster variables. The cluster solution should include most patients, meaning that several of the patient-reported experience variables examined in this study were inappropriate, such as cooperation with other health services, where item non-response was high. From a research perspective, the middle clusters in our study should be further explored in future studies. However, from a quality-improvement perspective, these groups are not the most interesting since the overall outcomes are rated highly.
The outlier group scored poorly on all outcome items. The most striking feature of this group was the extent of perceived malpractice by the hospital: on average, these patients perceived themselves to have been subject to a large extent of hospital malpractice. The only reason the outlier group is not a cluster in a statistical sense, is the amount of internal variance on general satisfaction and benefit of treatment. However, from a quality improvement perspective this group is highly relevant. Efforts to identify, monitor and reduce the outlier group and cluster 5 should be a goal of the quality improvement work performed in hospitals. Further qualitative research should be conducted to explore quality problems within these groups, which will give valuable information when tailoring and implementing quality initiatives in hospitals. At the policy level, large differences in patient-perceived outcomes challenge both the goal of high quality and equal distribution of health-care quality . In Norway, the cooperation reform began on 1 January 2012, with the aim of improving cooperation between primary and secondary health care. This reform clearly relates to the largest improvement areas for cluster 5, and so potential improvements following the reform should be evaluated in future research.
Cluster analysis can be criticized for being exploratory and atheoretical . The existing research provided our study with little basis for building theoretical and analytical models. Several analytical strategies and approaches were tested before the final solution was reached; this solution was rigorously tested for validity, which resulted in further minor adjustments. All in all, we believe that the resulting response clusters have validity, but that the middle clusters should be further explored in future research. Another potential limitation is the use of single outcome items as cluster variables. Single items are normally less reliable than multi-item scales. Furthermore, patient-reported outcomes normally include both multiple scales and pre–post measurement . However, psychometric evaluation is of less concern since we are clustering patients, not items. In addition, the use of factor scores in cluster analysis is debated, since research has shown that the most discriminatory variables are not well represented in most factor solutions . The outcome items in this study were developed and tested in Norway and found to be very important for Norwegian patients . Consequently, using these items as cluster variables appears to provide adequate validity. A third limitation is related to the response rate. More than half of the patients failed to respond to the survey. Findings from previous non-response research in Norwegian national surveys shows that low response rates have not caused serious bias in population estimates [28, 29]. The findings from a Norwegian follow-up study involving the same hospital population as the current study showed that postal respondents and non-respondents had almost the same scores . However, it is not unlikely that patients experiencing the poorest health care are underrepresented in patient experience surveys, for instance patients seriously harmed by adverse events. These patients constitute a minority of hospital patients, which implies only small effects on the overall survey estimates. However, there is a danger that the size of the outlier group and cluster 5 are underestimated. We stress the importance of allowing proxies to answer on behalf of patients not able to answer themselves, so that these patients can be adequately represented. The national survey included the opportunity for proxies to respond, but the current study did not have a particular focus on the association between patient clusters and non-response. This is a limitation of the study and an important area for future research. A final limitation is the fact that the study was restricted to identifying and statistically validating patient clusters. Consequently, substantial issues related to explaining and profiling clusters are warranted in future research. For instance, the importance of poor quality of care and disease severity for belonging to the outlier group and cluster 5 should be explored.
The purpose of national patient experience surveys is systematic measurement of patient experiences, as part of quality improvement, business control, free hospital choice and public accountability. The hospital level is the main level for the national surveys, and both case-mix adjusted comparisons of hospitals and unadjusted frequency based results for each hospital are presented in national reports and at several internet-sites. Patient clusters are highly relevant for both reporting types, and present a novel way of reporting and understanding patient evaluation of hospitals. The size of the improvement clusters (cluster 5 and the outlier group) can be computed, compared and presented at the hospital level. Research has shown that consumers have difficulties in understanding quality information , and that “less is more” in this respect . Therefore, a percentage at the hospital level showing the size of improvement clusters seems appropriate in the context of presenting information to consumers. On the other hand, more specific results are called for when reporting information to health providers aiming to evaluate and improve the quality of care . Consequently, both the size of cluster 5 and the outlier group should be included and constitute a fruitful supplement when reporting results to the responsible hospitals. The latter should also be supplemented with qualitative data to better understand the types of problems these clusters are facing, and profiling data to be able to target improvement initiatives within hospitals.
More research is needed to secure the usefulness of cluster analysis and reporting in this setting. The cluster approach should complement existing reports, and not only reproduce the same ratings of hospitals as the existing approach . The statistical construction of the measure at the hospital level should also be further explored, especially how to handle case-mix adjustment. Furthermore, the appropriate method for analyzing and presenting qualitative data and profiling data for hospitals should be closely examined in future research. To assess generalizability, the cluster analysis should be reproduced in future national surveys in Norway and national hospital surveys in other countries. The latter can be easily done by including the three outcome items in this study and by applying the same statistical procedures as in the current article.