Intra-cluster correlation coefficients in primary care patients with type 2 diabetes and hypertension

Cluster randomized controlled trials (CRCTs) are increasingly being used to study the impact of practice intervention in primary care settings [1, 2]. It is known that the application of cluster randomization is less demanding to clinical practice and more administratively convenient [3]. In CRCTs, subjects are not allocated to intervention independently, but as a group within a cluster. Examples of clusters include communities, schools, clinics and hospitals [4]. Compared to individually randomized trials, CRCTs are not only more complex in design and analysis, but also require more participants to reach equivalent statistical power, as observations in the same cluster tend to be more “alike” [5, 6]. The condition whereby the effective sample size is less than the total number of participants is referred to as the clustering effect in CRCTs.

The clustering effect can be determined by calculating the intra-cluster correlation coefficient (ICC), ρ = \( \frac{b}{b+w} \) , based on measured outcomes, where b is the between-cluster variance and w is the within-cluster variance [5]. Therefore, the ICC can range from 0 to 1, depending on how varied the measured outcomes would be within a cluster. The sample size of CRCTs often needs to be inflated by the design effect, DE = 1 + (n – 1) ρ, where n is the average number of individuals per cluster. For instance, when there is no clustering effect on a measured outcome, the calculated ICC would be near to zero and the estimated DE may be very close to 1. A small DE indicates that the clustering effect does not inflate the sample size [7, 8].

There have been repeated calls to publish ICCs for future cluster-based interventional studies [7, 9]. Several researchers have published ICCs from large primary care trials conducted in high-income countries [2, 7, 10, 11]. In this paper, we report information on ICCs from a large study conducted at the public primary care clinics in Malaysia, which is an upper middle-income country. Based on the measured variables, we report the ICCs of processes of care, intermediate clinical outcomes and patients’ experience of self-management support amongst patients with type 2 diabetes (T2D) with or without hypertension and patients with hypertension only clustered within the selected primary care clinics. The objective of reporting ICCs is to aid sample size estimation for future CRCTs conducted at the primary healthcare level in low- and middle-income countries, as countries in this income category are more similar in the structure and resources of their healthcare systems compared to high-income countries.

This study used baseline data from a larger study called the Evaluation of Enhanced Primary Health Care (EnPHC) interventions in public health clinics (EnPHC-EVA: Facility). We reported ICC estimates for a range of process and outcome measures for T2D and hypertension based on an ICC reporting framework recommended by a group of researchers and statisticians with experience in CRCTs [12].

A total of 6722 subjects with T2D (with or without hypertension) and 5014 subjects with hypertension only were included in this study for the reporting of results on processes of care and intermediate clinical outcomes. For the patient exit survey, 956 patients with T2D and/or hypertension were interviewed. The baseline characteristics of the study population are described in Table 1. The mean age of the study population was 60 years old. The patients were predominantly female (62%) and of Malay (69%) ethnicity. The majority of the patients (82%) were overweight or obese, and about 42% of them had hyperlipidaemia. The average cluster size ranged from 141 to 168 for T2D patients with or without hypertension and from 94 to 126 for hypertension-only groups.

ICCs for processes of care, intermediate clinical outcomes and patient experience are presented in Tables 2, 3 and 4 respectively. Between the three categories of measures, the highest ICC was observed for patient-reported experience with a median ICC of 0.09 (interquartile range [IQR] 0.07, 0.13) followed by processes of care with a median of 0.04 (IQR 0.02, 0.16) and intermediate clinical outcomes with a median of 0.03 (IQR 0.01, 0.08) (see Fig. 1).

For processes of care, the ICC values ranged from 0.01 to 0.48. The highest ICC was for measurement of body mass index (BMI, 0.48) and liver function test within 12 months (0.37) in T2D patients with or without hypertension. Overall, the ICCs for common processes of care for patients with hypertension only were higher than those for T2D patients with or without hypertension. However, both groups of patients had similar ICCs (0.01 to 0.04) for antihypertensive medications use. The results also showed low ICC estimates for medication prescribing across all medication groups, for both T2D patients with or without hypertension and hypertension-only patients.

As for clinical outcomes, two variables had the highest ICCs of 0.09. They were the proportion of T2D patients with or without hypertension who achieved the target blood pressure of ≤130/80 mmHg and the mean diastolic blood pressure of hypertension-only patients. These ICCs for clinical outcomes were mostly similar between both disease groups, ranging from 0.01 to 0.09. Between binary and continuous data types, such as the proportion of T2D patients with or without hypertension who achieved HbA1c ≤ 7% versus the mean value of HbA1c, the ICC values were also similar at around 0.01. This showed that the variations between clusters were largely similar for intermediate clinical outcomes, whether between different patient groups or data types.

For patient experience, the ICC values were between 0.03 (Proportion of patients who are willing to recommend the clinic to their friends and family) and 0.25 (PACIC item 9, Given a copy of my treatment plan). The mean scores for each item in PACIC were at the lower end of the scale, ranging between 1.5 to 2.8, except for item 5 (Satisfied that my care was well organized), which had a mean score of 4.2. Both the “Patient activation” and “Delivery system design/Practice design” subscales had the same ICC values of 0.11. The ”Goal setting/Tailoring” subscale had a slightly higher ICC value (0.13) than the other two subscales. This showed that there were slightly more differences between clinics in engaging patients to set their treatment goal as part of their self-management support initiative. The average cluster size for this group of patients was 23.9 (range 23–29).

To the best of our knowledge, this study is amongst the first to report ICCs for T2D and hypertension processes of care and intermediate clinical outcomes in a middle-income country. An analysis of 31 primary care studies found that the context or setting in which the ICC is derived influences the size of the ICC [26]. Since the majority of research works on ICCs for T2D and hypertension care were from high-income countries [2, 11, 27], it was necessary to report those from our setting to determine if factors such as differences in organizational structure and resource limitations within a healthcare setting are determinants of the magnitude of ICCs. In this study, we found that estimates for process and intermediate outcome indicators were broadly similar to those reported for high-income countries, by Singh et al. and Littenberg et al. for diabetes and hypertension, except for HbA1c and blood pressure readings [7, 11].

We have reported confidence intervals (CIs) together with the ICC estimates, which are useful for investigators who want to perform sensitivity analyses during sample size estimation [26]. Measures of precision are recommended to be included when reporting ICCs, and CIs are the preferred mechanism for providing this information [12]. Bell and McKenzie recommended to examine a range of plausible ICC values and then select conservative options when calculating the sample size for cluster trials [24]. We present these results to expand the existing literature on ICCs in primary care, particularly for low- and middle-income settings.

Our findings were consistent with those reported in the previous literature, where ICCs for processes of care were in general higher than for clinical outcomes, albeit to a lesser extent compared to most studies [7, 11, 28]. A high ICC indicates that the implementation of the processes of care is highly associated with practices. This happens because process measures are largely influenced by health providers’ behaviours, which are likely to be collectively more similar within the same clinic. We found that ICCs for the processes of care for hypertension-only patients are generally higher than those for the same processes of care in T2D patients with or without hypertension. This suggests that the variation in management of hypertension-only patients differs more between clinics compared to that of patients with T2D. On the other hand, clinical outcomes are based on an interaction of differences in biological characteristics, diet, lifestyle practices, attitudes and behaviour towards health; thereby they will exhibit greater variability within clusters, causing the ICC to move towards zero. Amongst all process of care measures, BMI measurement, annual electrocardiography and liver function tests had the highest ICC values. This high variability between clinics reflects differences in provider practice styles, documentation and reminders, availability of resources and time constraints between clinics. Similarly, high ICC values were also reported by Gulliford et al. for the recording of annual weight and urine protein measurement for government primary care services in Trinidad and Tobago [29]. The observed lower median ICC for process measures (0.04) compared to that in the study by Singh and colleagues (0.09) [11] is attributable to the more detailed breakdown of medication categories within this study and the fact that medication prescription has low variation between clinics. This low variation in medication prescription is largely due to the use of a common national drug formulary for public clinics and guideline-adherent practices.

Prescription of glucose-lowering, antihypertensive and lipid-lowering medications had lower ICCs compared to the laboratory tests and assessments, indicating consistency in pharmacological management of patients with diabetes and hypertension across clinics. We have also found that subjective measures from the patient experience survey had higher ICCs compared to the objective process of care measures. This finding is in agreement with those reported by Campbell et al. [28]. We observed that patient-reported experience had the highest ICC values compared to processes of clinical care and patients’ clinical outcomes, suggesting that there was a higher variation between clinic in this aspect, reflecting potential differences in the approach and level of self-management support provided by each clinic’s healthcare providers. For example, the high ICC value for PACIC item 9 (Given a copy of my treatment plan) suggests that some of the clinics provide their patients with a document or booklet to increase individual awareness on disease control and treatment, whilst the rest of the clinics are less likely to provide such documents.

There are several limitations in this study. The Malaysian healthcare system consists of two sectors: a government-led and funded public sector and a private sector which is largely financed by out-of-pocket payments and private insurance [30]. Results from this study are reflective only of public primary care clinics in Malaysia. Data for processes of care and intermediate outcomes were extracted from medical records. Therefore, the results are more likely affected by quality of documentation and the methods of measurement, which may vary between clinics. Lastly, the estimates for low-prevalence processes of care or patient experience should be interpreted with caution due to greater imprecision when assessing measures that occur infrequently [28].

We have reported ICC estimates and their respective 95% CIs for a range of measures for T2D and hypertension, which can aid in sample size calculations for cluster randomized controlled trials. This enables more accurate sample size estimation, which allows researchers to calculate realistic recruitment targets whilst reducing the risk of conducting underpowered cluster trials. This is important, particularly in settings where resources are scarce.