Using routine data to monitor inequalities in an acute trust: a retrospective study

Hospital Episode Statistics (HES)

This study used a retrospective design to explore the feasibility of monitoring inequalities within an acute trust. We obtained routinely collected administrative data, HES, from April 2007- March 2010, from the NHS Connecting for Health secondary uses service (SUS). We hold Section 251 National Information Governance Board for Health and Social Care permission to hold these data for research purposes. We also hold South East Local Research Ethics Committee approval to analyse the data. HES have been collected on all patients admitted to NHS hospitals since 1989 and include demographic, diagnostic and procedural data [23]. Three years of data were used to gain the largest numbers of patients whilst minimising additional confounding because of changes over time in coding, medical practice or policy, although coding may have improved over the course of these three years which we cannot control for [24].

HES records represent the finished consultant episode – “a period of admitted patient care under a consultant or allied healthcare professional within an NHS trust” [23]. A stay in a hospital can be made up of one or multiple finished consultant episodes. These were linked together into admissions which is the unit of analysis used here. Transfers in from other hospitals were not included. A proportion of the admissions considered in this study would have ended as transfers, however post-transfer length of stay and readmissions were not considered in this study.

Procedures and database inclusions

Inequalities were examined within procedure groups rather than specialities so that patients undergoing similar procedures were compared to each other. Although patients can have multiple procedures within an admission, patients were grouped by their main operative procedure, the most resource intensive procedure of the admission [23]. This study looked at elective admissions for seven procedures - bariatric surgery, cholecystectomy, coronary angioplasty, primary hip replacement, inguinal hernia repairs, primary knee replacement and mastectomy (for breast cancer in women). These elective surgical procedures were chosen in discussion with service leads to include the most common, and those that were rapidly increasing in volume and those where inequalities have been reported from national data. By including common procedures we would increase the statistical power to identify inequalities should any exist. Non-elective patients, day admissions and patients under 18 were excluded from the analysis to try to reduce the variability in the sample and to take into account some of the case-mix within the procedure groups.

Measures of inequalities

This study explored variations in access, process and outcome measures by age, sex, ethnicity and social deprivation. Data on age, sex, ethnicity and patient postcode are routinely included within HES. Data on ethnicity, however, is often recorded as not stated or not known, though this improved nationally from 24% missing in 2004 to 9% in 2010 [25]. Due to small numbers in some groups, age bands were combined into 3 to 5 groups, based on the distribution of age for each procedure area. Ethnicity categories within HES are based on the ethnic groups used in the 2001 census [23]. Small numbers made it necessary to combine these into four categories – white, Asian, black and other and mixed backgrounds. The Carstairs index of deprivation was used to determine the social deprivation of the postcode of the patient’s home address as a proxy for the patient’s socioeconomic status [23]. The Carstairs index was used in the dataset as it is available at a smaller area level, the lower super output area. The scores in the original dataset were split into five population-weighted quintiles based on the national distribution. These were combined into 2 groups, quintiles 1 to 3 and quintiles 4 to 5, to provide sufficient numbers and enable comparison between the more and less deprived.

In addition to the four socio-demographic variables which were used to explore inequalities, data on comorbidities were used to take case-mix into account. The dataset included a measure of comorbidity. Each patient had a primary diagnosis, any other secondary diagnoses or comorbidities were used to derive a comorbidity score using the Charlson comorbidity index, taking into account both the number and severity of the comorbidities that a patient might have [26]. The weights used were derived from English administrative hospital data [27]. The comorbidity score was dichotomized into a binary variable ‘no comorbidities’ or ‘one or more comorbidities’.

Dependent variables: process and outcome measures

The process and outcome measures of hospital care used were waiting times, length of stay and readmissions. Waiting time is the time between the date on which the patient was put on the waiting list and the date on which they were admitted, and therefore includes any time when an individual might be suspended from the waiting list, a patient does not attend or if a patient is unable to have surgery because of ill health [23]. It is only valid for elective patients with planned admissions; non-elective patients were therefore excluded from all analyses conducted for this study. Length of stay represents the number of days the patient spends in the hospital during their admission. The continuous variables length of stay and waiting times were tested for normality. Common normalising transformations of the data, such as the reciprocal, square root and natural log were unsuccessful and binary variables were therefore created from these continuous variables. Under the NHS constitution patients should not wait more than 18 weeks from referral for treatment, however this could not be used to define a prolonged waiting time as too few people waited longer than this time [23]. 75th percentiles were therefore used for each procedure group. The 75^th percentile was also used to define a prolonged length of stay. A similar study looking at waiting times used the median as a cut-off point but we used the 75^th percentile as the tail of the distribution was of more interest [21]. 75th percentiles have been used in other studies to define a prolonged length of stay [16]. Readmissions were measured using the derived field of unplanned readmissions within 28 days of discharge. Those patients who died were excluded from the analysis when readmissions were analysed. Readmissions were not explored for inguinal hernia repairs as there were too few readmissions.

Data analysis

The data analysis was conducted using SPSS v.18 (SPSS Inc, Chicago, Illinois, USA). Statistical significance was set at p ≤ 0.01, to take the multiple analyses into account. Descriptive statistics were used to examine the distribution of the variables for the whole population and for each procedure group.

Logistic regression was used for each procedure group to explore the relationship between each of the explanatory variables (ethnicity, social deprivation, age, sex) and the process or outcome measure. Multiple logistic regression was then used to explore the independent effects of the explanatory variables and to adjust for comorbidity.

Power and sample size calculations

Power is the probability of rejecting a false null hypothesis i.e. it is the ability of the test to find an effect that is there. Power calculations were conducted retrospectively for one of the analyses - the relationship between social deprivation and a prolonged wait for coronary angioplasty, as Pell and Pell et al.’s 2000 study found a similar relationship [11]. Sample size calculations were also made, using this study as a pilot, to determine how many years of data would be needed to detect an effect.

Power and sample size calculations were calculated retrospectively using G*Power 3.1.2. For logistic regression this program uses the methodology described in Hsieh, Block & Larsen’s 1997 paper [28]. Rather than using complex calculations for logistic regression, this method is based on comparing proportions and then adjusting for a multifactorial model by a variance inflation factor [28].

Characteristics of the population

The 75^th percentiles of waiting times and length of stay that were used to define a prolonged wait or prolonged stay for each procedure are indicated by the upper limit of the inter-quartile range (IQR) in Table 1.

Logistic regression

Age

There was little evidence of inequalities in waiting times in terms of age for any of the procedures. Only for cholecystectomy were those who were 45-54 years old significantly more likely (aOR = 1.79, 95% CI: 1.16-2.74) to have a prolonged wait of over 74 days, compared with those who were 18-44 years old.

There was a largely consistent trend towards a more prolonged stay as age increased with the exception of bariatric surgery. This was most extreme for those undergoing hip and knee replacements, with those over 85 being far more likely to stay over 7 days than younger patients (aOR = 9.02, 95% CI: 4.00-20.33 and aOR = 11.62, 95% CI: 5.76-23.47 respectively).

The only procedure for which there was a significant relationship between readmissions and age was hip replacements. Those who were 75-84 years (aOR = 2.86, 95% CI: 1.34-6.07) and 85+ years (aOR = 8.08, 95% CI: 2.92-22.38) were more likely to be readmitted within 28 days than those who were 18-64 years.

Sex

In general, women did not significantly differ from men with regard to waiting times, length of stay or readmission rates. For coronary angioplasty, however, women were significantly more likely to be readmitted within 28 days compared to men (aOR = 1.87, 95% CI: 1.24-2.81).

Ethnicity

Overall different ethnic groups did not seem to experience significantly different waiting times, lengths of stay or readmissions rates. The only significant difference was that black Caribbean or African bariatric surgery patients seemed to be more likely to experience a prolonged wait of longer than 64 days than white bariatric surgery patients (aOR = 2.47, 95% CI: 1.36-4.49).

Social deprivation

There was little evidence of systematic inequalities by social deprivation. For coronary angioplasty, those from the two most deprived quintiles were 1.66 (95% CI: 1.25-2.20) times more likely to have a prolonged length of stay of over a day.

Power calculations and sample size calculations

Post-hoc power calculations were conducted for the analysis of social deprivation and waiting time for coronary angioplasty. The analysis had a power of 0.47. Based on this study, a sample size of 4,132 would be needed to be able to detect, with 80% power and alpha of 0.01, an odds ratio of 1.3 for the relationship between social deprivation and having a prolonged wait for coronary angioplasty. In this study we used three years of data which gave a sample of 2,238 coronary angioplasty patients. Therefore to get 4,132 patients approximately 6 years of data would be needed.

Findings

This study explored whether routine data can be used to monitor inequalities in an acute trust based on a case study. We found little evidence of inequalities in waiting time, length of stay or readmission rates by sex, ethnicity or social deprivation for common surgical procedures in the trust. We did identify some differences which may warrant further investigation. Overall we conclude that there are challenges in using routine data to monitor inequalities at this level due to limitations in sample size that reduce the power to detect differences. We also identified problems of data validity and relevance in studying inequalities.

We identified most variation in relation to age. Older age groups differed from younger age groups for almost all procedures for length of stay, and for readmissions after hip replacements. This is likely to represent differences in clinical need, as older patients will take longer to recover as well as social support at home contributing to some delayed discharges. We found that people aged 45 to 54 years were more likely to have a prolonged wait for a cholecystectomy; without additional clinical information this is difficult to interpret. It is possible that this group has less of a pressing clinical need compared than those who are 18-44 and therefore waited longer for the procedure.

The only evidence of inequalities by sex was for readmission rates for coronary angioplasty, where women were almost twice as likely to be readmitted. Gender differences have been observed in many different aspects of coronary artery disease and coronary heart disease in terms of epidemiology, diagnosis, treatment and outcomes [29]. Mortality and readmission rates have been found to be higher in women than in men following coronary revascularisation [30, 31]. One study, however, found that once baseline clinical risk and body size, a proxy for blood vessel size, were taken into account this was no longer significant [31]. In this analysis we did not look at the diagnosis code for the readmission to see whether it was related to the procedure, this may have provided more clinical information as to the cause of the readmission.

The only significant difference by ethnicity was that black African or Caribbean patients were more likely to have a prolonged wait than white patients for bariatric surgery. This may be due to residual confounding that could not be taken into account. For example we did not have information on primary care trust of origin or clinical factors that could impact on the waiting time.

There was no evidence of systematic inequalities by social deprivation, except for coronary angioplasty. Those from the most deprived areas were 1.7 times more likely to have a prolonged length of stay of over 1 day. There are many explanatory factors that could take account of this. For example, these patients could have a different baseline clinical risk (unaccounted for in our casemix adjustment), although patients from more deprived areas may be less likely to be classified as urgent [11].

Strengths and limitations

The strength of this methodology is that by using easily accessible routine HES data, potential inequalities can be monitored to ensure that the healthcare system is not reinforcing existing inequalities in health. It is a simple way to establish whether certain groups are likely to wait longer, stay longer and be readmitted, adjusting for other factors. The data also allowed for inequalities to be examined by procedure.

The main limitation of using routine data to monitor inequalities on a trust level was that the study was underpowered. Using data from one trust did not provide a large enough sample to detect a medium effect (OR 1.3), though it would have been able to detect larger effects (OR 1.5) for some of the more common procedures. This problem is exacerbated for rarer events such as readmissions. The acute trust that we analysed is one of the largest hospital trusts in England – if it is difficult to use three years of data from such a large trust, this method for monitoring inequalities is likely to be problematic in smaller trusts. More years of data may have given more patients but changes in coding, medicine and population over the years would make the analysis more complex. Pooling data from trusts in a county or neighbouring trusts may make this analysis more feasible – though it is hard to estimate how many trusts you would need to pool together due to the variation in trust size. The small numbers also meant that the grouping of age bands, ethnic and social deprivation groups was necessary, which may oversimplify the patient population and hide inequalities within the groups. Data on social deprivation showed a skewed distribution in our population with a high proportion of people from more deprived areas which will have limited our power to detect a social gradient.

Using routine data also means that the data quality and validity of the data can be problematic. There is, for example, the possibility with an administrative database such as HES that variables have been miscoded [32]. Missing data may be particularly problematic, with 10% of admissions in this study having missing ethnicity codes, and it is possible that this may introduce bias if some groups are more likely to be missing data. However, studies such as this which demonstrate the potential uses of HES data for ethnic monitoring may help to improve the collection of these data [33].

As is the case with many studies using routine data, the data were not collected for this purpose and therefore many variables that would be of interest were not available. For example, though we attempted to take into case-mix by taking into account co-morbidities and by looking at inequalities by procedure, it is very likely that this did not take all the variation of case-mix into account. More data on clinical status may have helped to control for confounding when looking at length of stay and readmissions.

Other limitations included that there is no clear way to dichotomize the outcome variables, length of stay and waiting time. Using 75^th percentiles may not inform us of the clinically important differences. For example, the finding that people from more deprived areas were more likely to spend two or more days in the hospital following a coronary angioplasty may not be clinically important. Additionally, multiple tests were performed and therefore it may not be surprising that significant results were found, though a p value of 0.01 was used to try to take this into account.

Even where inequalities were found, using routine data can make it difficult to determine whether these inequalities are inequitable. The main aim of this work was to explore the feasibility of monitoring inequalities not inequities, however the inequalities may be providing equity. More in-depth clinical data, which would not be found in routine administrative dataset may be needed to work out whether these differences are delivering equity or not. For example, this analysis found that older age groups differed from younger age groups for almost all procedures for length of stay. This is most likely to represent a natural difference in clinical need as older patients will take longer to recover; therefore this inequality is not necessarily inequitable. To determine whether it was inequitable, further investigation would be needed into why they are being readmitted and if these readmissions could be prevented.