Impact of a national primary care pay-for-performance scheme on ambulatory care sensitive hospital admissions: a small-area analysis in England

Objective We aimed to spatially describe hospital admissions for ambulatory care sensitive conditions (ACSC) in England at small-area geographical level and assess whether recorded practice performance under one of the world’s largest primary care pay-for-performance schemes led to reductions in these potentially avoidable hospitalisations for chronic conditions incentivised in the scheme. Setting We obtained numbers of ACSC hospital admissions from the Hospital Episode Statistics database and information on recorded practice performance from the Quality and Outcomes Framework (QOF) administrative dataset for 2015/2016. We fitted three sets of negative binomial models to examine ecological associations between incentivised ACSC admissions, general practice performance, deprivation, urbanity and other sociodemographic characteristics. Results Hospital admissions for QOF incentivised ACSCs varied within and between regions, with clusters of high numbers of hospital admissions for incentivised ACSCs identified across England. Our models indicated a very small effect of the QOF on reducing admissions for incentivised ACSCs (0.993, 95% CI 0.990 to 0.995), however, other factors, such as deprivation (1.021, 95% CI 1.020 to 1.021) and urbanicity (0.875, 95% CI 0.862 to 0.887), were far more important in explaining variations in admissions for ACSCs. People in deprived areas had a higher risk of being admitted in hospital for an incentivised ACSC condition. Conclusion Spatial analysis based on routinely collected data can be used to identify areas with high rates of potentially avoidable hospital admissions, providing valuable information for targeting resources and evaluating public health interventions. Our findings suggest that the QOF had a very small effect on reducing avoidable hospitalisation for incentivised conditions. Material deprivation and urbanicity were the strongest predictors of the variation in ACSC rates for all QOF incentivised conditions across England.


Supplemental Tables and Figures
. ICD codes used to define QOF incentivised ACSC hospital admissions Table S2. QOF indicators included in the aggregate performance scores and their changes over time Table S3 -Effects of QOF overall population achievement on hospital admissions for QOF incentivised ACSCs Table S4

Classification of ACSC hospital admissions for QOF incentivised conditions.
In HES, all admission data are related to episodes rather than persons as some individuals may have been admitted in hospitals multiple times over one calendar year and these presentations will be recorded as separate admissions. Classification of ACSC hospital admissions for the study used the International Classification of Diseases, 10th edition (ICD-10) and included all hospital admissions with primary diagnosis related to one of the 9 QOF incentivised conditions examined in this study. The specific 9 conditions were chosen based on the list of ACSCs which are used for measuring system performance in the NHS (1) plus a small number of diagnoses adopted from a previous study. (1) We also included complications of diabetes associated with hypoglycaemia. (2) Table S1: ICD codes used to define QOF incentivised ACSC hospital admissions J45, J46 Asthma I20, I24.0, I24.8, I24.9, I25 Coronary Heart Disease J20, J41, J42, J43, J44, J47 Chronic Obstructive Pulmonary Disease E10.0-E10.8, E11.0-E11.8, E13.0-E13.8, E14.0-E14. 8 Diabetes E162 Diabetes (hypoglycaemic) G40, G41 Epilepsy I11.0, I13.0, I50, J81 Heart Failure I10, I11.9 Hypertension I61 I62 I63 I64 I66 I672 I698 R470 Stroke

Description of indicators included in the composite measure of quality of care
Several indicators were used over the study period (2006 -2015) and were aggregated into a single score. Although these indicators have been revised or rephrased many times over the years, their underlying aim has consistently remained a) to identify patients with symptoms of each respective condition b) to provide assessment and monitoring to those patients. Throughout the years of the scheme, some indicators were dropped while others were revised. When an indicator was dropped from the scheme, population achievement was calculated for each year from the remaining indicators. When an indicator was revised, the revision was concerned with the time span that measurement or treatment was taken or provided (e.g. for some indicators the time span was specified for 15 months but after the revision the time span was reduced to 12 months). In the composite measure we included all recording/measurement and treatment indicators related to the conditions outlines in table 1.  The percentage of patients with asthma who have had an asthma review in the preceding 15 months that includes an assessment of asthma control using the 3  The percentage of patients with coronary heart disease whose notes have a record of total cholesterol in the previous 15 months 40 90 7

CHD 8 3-8 Outcome
The percentage of patients with coronary heart disease whose last measured total cholesterol (measured in the last 15 months) is 5 mmol/l or less 40 70 17 CHD 8 9 Outcome The percentage of patients with coronary heart disease whose last measured total cholesterol (measured in the last 15 months) is 5 mmol/l or less 44 70 17 CHD 9 3-8 Treatment The percentage of patients with coronary heart disease with a record in the last 15 months that aspirin, an alternative anti-platelet therapy, or an anti-coagulant is being taken (unless a contraindication or side effects are recorded) 40 90 7 CHD 9 9 Treatment The percentage of patients with coronary heart disease with a record in the last 15 months that aspirin, an alternative anti-platelet therapy, or an anti-coagulant is being taken (unless a contraindication or side effects are recorded) 50 90 7 The percentage of patients with coronary heart disease whose last measured total cholesterol (measured in the last 12 months) is 5 mmol/l or less 45 85 17 CHD 005 10-12 Treatment The percentage of patients with coronary heart disease with a record in the last 12 months that aspirin, an alternative anti-platelet therapy, or an anti-coagulant is being taken (unless a contraindication or side effects are recorded) 56 96 7 CHD 006 10-12 Treatment The percentage of patients with coronary heart disease with a record in the last 12 months that aspirin, an alternative anti-platelet therapy, or an anti-coagulant is being taken (unless a contraindication or side effects are recorded) The percentage of patients with COPD who have had a review, undertaken by a healthcare professional, including an assessment of breathlessness using the MRC dyspnoea score in the preceding 15 months. The percentage of patients with COPD who have had a review, undertaken by a healthcare professional, including an assessment of breathlessness using the MRC dyspnoea score in the preceding 12 months. The percentage of patients with diabetes with a record of a foot examination and risk classification: 1) low risk (normal sensation, palpable pulses), 2) increased risk (neuropathy or absent pulses), 3) high risk (neuropathy or absent pulses plus deformity or skin changes or previous ulcer) or 4) ulcerated foot within the preceding 15 months 40 90 4

DM 29 9 Measurement
The percentage of patients with diabetes with a record of a foot examination and risk classification: 1) low risk (normal sensation, palpable pulses), 2) increased risk (neuropathy or absent pulses), 3) high risk (neuropathy or absent pulses plus deformity or skin changes or previous ulcer) or 4 The percentage of patients with diabetes whose last measured total cholesterol within previous 15 months is 5 or less 40 70 6 DM 17 8 Outcome The percentage of patients with diabetes whose last measured total cholesterol within previous 15 months is 5 or less 40 70 6 DM 17 9 Outcome The percentage of patients with diabetes whose last measured total cholesterol within previous 15 months is 5  In those patients with a current diagnosis of heart failure due to LVD who are currently treated with an ACE-i or ARB, the percentage of patients who are additionally currently treated with a betablocker licensed for heart failure. The percentage of patients with hypertension aged 16 or over and who have not attained the age of 75 in whom there is an assessment of physical activity, using GPPAQ, in the preceding 12 months 40 80 5 HYP 005 10 Treatment The percentage of patients with hypertension aged 16 or over and who have not attained the age of 75 who score 'less than active' on GPPAQ in the preceding 12 months, who have a record of a brief intervention in the preceding 12 months 40 80 6

Lower Super Output Areas
The low geographical units to which the indices of deprivation are assigned are called lower super output areas (LSOAs) and they are designed to contain around 1500 inhabitants, on average. Following the 2011 census, there were 32,844 English LSOAs. (4) Urbanity information was also updated following the census, and we used a rural vs urban dichotomy for simplicity, with settlements with 10,000 people or more defined as urban. (5) Census-adjusted population estimates over time, by age groups and sex and for each English LSOA, were obtained from the Office of National Statistics.

Ethnicity
Ethnicity was taken from the 2011 census and was available at the LSOA level from the Nomis website. (8) Ethnicity classifies people according to their own perceived ethnic group and cultural background. The variable uses a harmonised country specific ethnic group question and the question is recommended when a show card is used in a face to face interview or self-completion survey (both paper and electronic).

Sex and population estimates
From the Office for National Statistics, we obtained mid-year population estimates for 2015 at the LSOA level. (6) The mid-year population estimates are the official set of population estimates for the UK and its constituent countries, the regions of England and Wales and local authorities. They are used directly as a base for other secondary population statistics, such as population projections, population estimates for the very old and population estimates for small geographical areas. A combination of registration, survey and administrative data are used to estimate the different components of population change. The data are provided for the whole population as well as by sex. We used the number of females in an LSOA in 2015 as recorded by ONS to calculate the percentage of females who were resident in each LSOA in 2015-2016.

Spatial Weighted analyses and attribution methodology
We used general practice level information from NHS Digital on several variables of interest to attribute it to our units of analyses, namely the LSOAs. We inputted the LSOA centroid coordinates (longitude and latitude) in R to create a 32,844x32,844 inverse distance matrix (in miles). This matrix features a detailed distance mapping of each LSOA with all other LSOAs and was used to quantify geographical proximity where nearby LSOAs have larger weights and also to generate prevalence and a measure for had closed down or merged, while 87 new practices (1.1%) emerged. To calculate attribution rates for all years in order to subsequently quantify prevalence and quality of care we used regression modelling under various assumptions to obtain attribution estimates for previous years. For each LSOA, if two or more practices were linked to it, we fitted Poisson and negative binomial models with list size and distance to practice as predictors, and the model that was the best fit to the data was selected. If a practice was present both in the analyses and the attribution dataset, we adjusted the attributed population for practice' list size in the respective year, thus assuming a constant attribution rate over time. If a practice was present in our analyses but not in the attribution dataset, we generated estimates using the models selected in step 1 across all years. If a LSOA was served entirely by a single practice, we assumed that this was the case in previous years. Those practices that emerged after our baseline year were used only to model the 2015 attribution in the area. Redistribution of patients to the other active practices within each year was achieved according to the selected regression model. In the same manner, for those practices that closed down or merged, their patients were re-distributed to the years in which they were active, according to their characteristics. Finally, the attribution counts estimated in the previous steps across practices and within each year were used to generate the weighted mean estimates for prevalence and quality of care. The algorithm is available from the corresponding author. We assumed that the attribution rates remained constant over time, and we used this assumption to model the contribution of each practice; even the ones that had closed or merged by 2015. This method can possibly introduce uncertainty in the estimates which we could not include in the models because of methodological limitations. Even though the limitations and assumptions made for our approach could attenuate the relationship between quality of care and suicide, we would expect any strong relationship between our variables of interest to be detected as it was for example detected for prevalence of depression.

Sensitivity analyses results
We conducted a sensitivity analyses under different assumptions to test the presence of The results from our sensitivity analyses were very similar to the results from the principal analysis and the coefficients had the same direction. First, due to the presence of many zeros we tested how a zero-inflated negative binomial (ZINB) fitted our data by running a regression with exactly the same covariates as our main model (i.e. negative binomial model presented in the main paper). The ZINB model assumes that the excess zero counts come from a logit or probit model and the remaining counts come from a negative binomial model. We modelled our counts as follows: We modelled excessive zeros using a logit model with the LSOA population as the only predictor and the remaining counts were modelled from the negative binomial model that we specified in the manuscript using deprivation, age, sex, rurality and ethnicity as predictors. The results from the zero inflated negative binomial regression indicate a very small association (0.993; 95CI 0.990, 0.995) between QOF Population Achievement and ACSC admissions for incentivised conditions. All other results were identical to the negative binomial regression mode. Over time our results indicated slightly smaller effect (0.998; 95CI 0.997, 0.999).