Examining variations in prescribing safety in UK general practice: cross sectional study using the Clinical Practice Research Datalink

Study question What is the prevalence of different types of potentially hazardous prescribing in general practice in the United Kingdom, and what is the variation between practices? Methods A cross sectional study included all adult patients potentially at risk of a prescribing or monitoring error defined by a combination of diagnoses and prescriptions in 526 general practices contributing to the Clinical Practice Research Datalink (CPRD) up to 1 April 2013. Primary outcomes were the prevalence of potentially hazardous prescriptions of anticoagulants, anti-platelets, NSAIDs, β blockers, glitazones, metformin, digoxin, antipsychotics, combined hormonal contraceptives, and oestrogens and monitoring by blood test less frequently than recommended for patients with repeated prescriptions of angiotensin converting enzyme inhibitors and loop diuretics, amiodarone, methotrexate, lithium, or warfarin. Study answer and limitations 49 927 of 949 552 patients at risk triggered at least one prescribing indicator (5.26%, 95% confidence interval 5.21% to 5.30%) and 21 501 of 182 721 (11.8%, 11.6% to 11.9%) triggered at least one monitoring indicator. The prevalence of different types of potentially hazardous prescribing ranged from almost zero to 10.2%, and for inadequate monitoring ranged from 10.4% to 41.9%. Older patients and those prescribed multiple repeat medications had significantly higher risks of triggering a prescribing indicator whereas younger patients with fewer repeat prescriptions had significantly higher risk of triggering a monitoring indicator. There was high variation between practices for some indicators. Though prescribing safety indicators describe prescribing patterns that can increase the risk of harm to the patient and should generally be avoided, there will always be exceptions where the indicator is clinically justified. Furthermore there is the possibility that some information is not captured by CPRD for some practices—for example, INR results in patients receiving warfarin. What this study adds The high prevalence for certain indicators emphasises existing prescribing risks and the need for their appropriate consideration within primary care, particularly for older patients and those taking multiple medications. The high variation between practices indicates potential for improvement through targeted practice level intervention. Funding, competing interests, data sharing National Institute for Health Research through the Greater Manchester Primary Care Patient Safety Translational Research Centre (grant No GMPSTRC-2012-1). Data from CPRD cannot be shared because of licensing restrictions.


MethodS
A cross sectional study included all adult patients potentially at risk of a prescribing or monitoring error defined by a combination of diagnoses and prescriptions in 526 general practices contributing to the Clinical Practice Research Datalink (CPRD) up to 1 April 2013. Primary outcomes were the prevalence of potentially hazardous prescriptions of anticoagulants, anti-platelets, NSAIDs, β blockers, glitazones, metformin, digoxin, antipsychotics, combined hormonal contraceptives, and oestrogens and monitoring by blood test less frequently than recommended for patients with repeated prescriptions of angiotensin converting enzyme inhibitors and loop diuretics, amiodarone, methotrexate, lithium, or warfarin.
Study anSwer and liMitationS 49 927 of 949 552 patients at risk triggered at least one prescribing indicator (5.26%, 95% confidence interval 5.21% to 5.30%) and 21 501 of 182 721 (11.8%, 11.6% to 11.9%) triggered at least one monitoring indicator. The prevalence of different types of potentially hazardous prescribing ranged from almost zero to 10.2%, and for inadequate monitoring ranged from 10.4% to 41.9%. Older patients and those prescribed multiple repeat medications had significantly higher risks of triggering a prescribing indicator whereas younger patients with fewer repeat prescriptions had significantly higher risk of triggering a monitoring indicator. There was high variation between practices for some indicators. Though prescribing safety indicators describe prescribing patterns that can increase the risk of harm to the patient and should generally be avoided, there will always be exceptions where the indicator is clinically justified. Furthermore there is the possibility that some information is not captured by CPRD for some practices-for example, INR results in patients receiving warfarin.

what thiS Study addS
The high prevalence for certain indicators emphasises existing prescribing risks and the need for their appropriate consideration within primary care, particularly for older patients and those taking multiple medications. The high variation between practices indicates potential for improvement through targeted practice level intervention.

Introduction
Prescribing errors in primary care can cause considerable harm, with adverse drug events accounting for around 7% of hospital admissions in the United Kingdom, and half of these are judged to be preventable. [1][2][3] In 2012 the Practice Study found that one in 20 prescription items was associated with a clinically important error and one in 550 was associated with a serious error. 4 Prescribing safety indicators (indicators) define prescribing patterns that can increase the risk of harm to the patient and should generally be avoided, though there will always be exceptions where the indicator is justified for clinical reasons.
Several sets of indicators have been developed for use in primary care in the UK. [5][6][7][8][9][10] The widespread use of electronic prescribing systems and primary care patient records facilitates the analysis of aggregate patient records to estimate the prevalence of indicators and their variation in prevalence between practices and to examine the patient and practice factors that influence their occurrence. In Scotland, a set of indicators was developed through consensus between GPs and pharmacists and their prevalence was measured in 315 general practices. 9 In England, a pharmacist led intervention was found to reduce the frequency of indicators in general practice (PINCER trial), 10

WhAT IS AlReAdy knoWn on ThIS TopIC
Prescribing safety indicators have been developed to identify patients at increased risk of hazardous prescribing in primary care Although these prescribing safety indicators have been investigated in experimental settings, they have not been assessed in a large UK-wide primary care database

WhAT ThIS STudy AddS
Variation in the prevalence of potentially high risk prescribing and lack of appropriate monitoring tests between practices was high, even after adjustment for patient and practice level variables, pointing towards important targets for improving patient safety in primary care In a broadly representative sample of 526 UK general practices, about 5% of patients at risk were found to have received a potentially inappropriate prescription and about 12% had no record of appropriate monitoring Older patients and those receiving multiple repeat prescriptions had the highest risk of triggering a prescribing safety indicator, whereas younger patients with fewer repeat prescriptions had higher risk of triggering a monitoring indicator Scotland, validated measures of high risk prescribing were the outcome measures in a trial of an intervention to reduce high risk prescribing of non-steroidal anti-inflammatory drugs (NSAIDs) and antiplatelet agents in general practice. 11 The set of indicators used for the PINCER trial was developed through consensus among GPs 8 and has been expanded and made available for GPs to use as an audit tool since 2013. 8 10 12 13 Currently GPs in the UK are encouraged to report clinical audit or quality improvement projects as part of their revalidation portfolio. 14 If these indicators are to be used by GPs as an audit tool there needs to be an agreement about what constitutes a clinically acceptable prevalence of high risk prescribing and whether there are potential exceptions.
Furthermore, the ability of indicators to discriminate between practices and the relation between their prevalence and patient and/or practice characteristics needs to be understood; identifying groups of patients at increased risk of indicators will enable a targeted approach to reducing their occurrence. As far as we know, indicators have not been investigated in a large sample of UK-wide general practices.
We used anonymised patient level data from the Clinical Practice Research Datalink (CPRD) 15 to investigate the prevalence and predictors of previously described indicators 8 10 12 13 in UK general practice. We measured the prevalence of indicators identifying patients at increased risk of a prescribing error during a time period leading up to 1 April 2013; examined the variation in the prevalence of the above indicators between practices using intraclass correlation coefficients and the meaning for the individual practice using reliability estimates; and examined the associations between the indicators and patient characteristics (age, sex, number of repeat medications) and practice characteristics (list size, practice level deprivation, geographical location).

Methods
We included in the analysis all patients aged 18 or over who were registered with one of the CPRD practices that had uploaded data of research quality on or after 1 April 2013. Each indicator consisted of a denominator and a numerator (  1 ). When the condition or diagnosis was reversible, such as requiring monitoring with repeat prescriptions, the denominator was the number of patients prescribed a repeat medication within a fixed time window before the audit date. For example, the denominator for repeat prescribing of angiotensin converting enzyme inhibitor (ACEI) or loop diuretic (M1, table 1 , fig 1) includes patients with at least one prescription of (ACEI) or loop diuretic between 6 and 15 months before the audit date and a repeat prescription between 0 and 6 months before the audit date. Polypharmacy was defined as the number of medications with at least two prescriptions within the 12 months leading up to 1 April 2013.

Statistical methods
We calculated the proportion of patients triggering each indicator (with 95% confidence intervals from a binomial distribution) for each indicator and for a composite prescribing (prescribing indicator, P1-P13) and monitoring (monitoring indicator, M1-M4) indicator. A further composite indicator also included the indicators relevant only to women in (P1-P13 plus P16-P19). Each patient counted just once in the composite indicator with a triggered indicator taking priority if a patient entered the denominator for more than one indicator. Therefore the composite indicator describes the number of patients triggering at least one indicator divided by all patients with the potential to trigger an indicator. The variation in prevalence between practices was quantified for each indicator by an empty mixed effects two level logistic regression model with patients nested within practices (using the xtmelogit command in Stata). The outcome was a binary variable (1 or 0) with patients triggering an indicator designated "1" and those with potential to trigger an indicator but not doing so designated "0." We used the post-estimation estat command in Stata to derive the intraclass correlation coefficients (ICCs) for each indicator. The intraclass correlation coefficient quantifies the proportion of the total variation in an indicator that is accounted for by the variation between practices (τ 2 )-that is, indicators with low values of intraclass correlation coefficient show that there is more variability within practices (σ 2 ) rather than between them (τ 2 ), in relation to the outcome ( fig 2). The within-variance in this binary outcome context is non-intuitive and fixed in the logistic regression models, and it does not reflect variability in the Formula for proportion of total variation in indicator that is accounted for by variation between practices performance of health professionals within a practice. Ideally, rather than being empty, the model would be adjusted by patient level variables reflecting the case mix before estimation of the intraclass correlation coefficient. As the adjusted model did not converge for some of the indicators, the empty model was the best option for consistency across all indicators. Adjustment for age, sex, and polypharmacy did not alter the intraclass correlation coefficient for most of the indicators where the model did converge. The intraclass correlation coefficient does not intuitively translate into an understanding of the implications for the individual practice. Therefore, to help with interpretation of the ICC, we calculated the reliability for different numbers of patients in the denominator using the Spearman-Brown prophecy formula as described previously ( fig 3) . 9 For many situations a reliability of 0.7 is considered acceptable, but higher values might be preferable in comparisons of prescribing safety between practices. 9 We calculated the reliability for a hypothetical practice with the median number of patients in the denominator, and appendix 1 shows the number of practices with reliabilities greater than 0.7, 0.8, and 0.9. However, the intraclass correlation coefficient, and therefore reliability, cannot provide insight on the underlying causes of the variation between practices that might, or might not, be clinical in nature. To further describe the heterogeneity between practices we also reported the prevalence predicted by the regression model for each indicator with 95% prediction intervals as described elsewhere. 18 The prediction intervals describe the expected range of prevalence for a new practice (with 95% confidence).
The composite indicators were further analysed with the same mixed effects two level logistic regression model. As described above each patient counted just once in this analysis with the outcome variable "1" designating patients triggering at least one indicator and "0" designating patients with the potential to trigger at least one indicator but not doing so. We analysed indicators related to prescribing (prescribing indicators) separately from indicators related to monitoring (monitoring indicators). The practice level predictors were list size, intervals of the index of multiple deprivation (IMD) based on 2010 estimates, 17 and location of practice by country or region of the UK. Patient level predictors were age, sex, and polypharmacy. patient involvement Given our specific aims, no patients were involved in setting the research question or the outcome measures, nor were they involved in the design and implementation of the study. We will be working with the Research User Group at the NIHR Greater Manchester Primary Care Patient Safety Translational Research Centre to advise on plans for dissemination of these findings.

Results
Data were available from 526 practices and almost five million patients. The mean list size between 1 April 2012 and 31 March 2013 was 9410 (standard deviation 4971). Most practices were in England (384, 73%), followed by Scotland (72, 14%), Wales (49, 9%), and Northern Ireland (21, 4%). More details on the make-up of the CPRD practices and their representativeness of the UK primary care have been provided elsewhere. 19 The choice of audit date did not greatly affect the prevalence of the indicator (comparing audit date 1 April 2013 with a rolling time window 1 May-1 September 2013) except for M4 (prescription of lithium without a lithium level test). The lower prevalence when we used the 1 April audit date was because of an increased frequency of lithium level tests during January 2013. Analysis that used earlier audit dates showed that the January increase in tests occurred every year as far back as 2005 (data available from authors). Tables 2 and 3 and figure 4 summarise the observed prevalence, predicted prevalence, and ICC for each indicator and the composite indicators. The distribution of the observed prevalence for each indicator by practices and by list size is summarised in appendix 1. The prevalence of the monitoring indicators was consistently higher than prescribing indicators (fig 4 ). The prescribing indicators with a higher prevalence were mainly those related to the prescribing of NSAIDs and aspirin without gastroprotection (P3-P4, P11, P13, fig 4 ), whereas the prescribing of combined hormone contraceptives (CHC) and oestrogens was associated with a lower prevalence (P16-P19, fig 4).
Overall, 49 927 of 949 552 (5.26%, 95% confidence interval 5.21% to 5.30%; table 2 ) patients triggered at least one prescribing indicator (P1-13). We excluded two indicators from the prescribing composite indicator (P1-P13, table 2 ) as they were nested within P7 and dealt with subgroup populations (P14-P15). Four were excluded because they were relevant only to women (P16-P19) and we wanted to evaluate the relation between sex and overall risk in a set of indicators that was relevant across both sexes. An alternative composite indicator included the indicators relevant to women (P1-P13 and P16-P19) and was triggered by 52 671 of 1 195 408 patients (4.41%, 4.37% to 4.44%; table 2 ).The corresponding composite monitoring indicator (M1-M4 excluding M5) was triggered by 21 501 of 182 721 patients (11.8%, 11.6% to 11.9%) (table 3 ). The decision to exclude M5 (warfarin without international normalised ratio (INR) testing) from the composite monitoring indicator was based on the large variation between practices (ICC=0.78) that was judged to reflect INR results not being recorded in CPRD (possibly because of different models of service delivery 20 ) rather than real differences in INR testing. Previous work found that INR tests might be recorded in secondary, rather than primary, care records. 21  Most patients triggered just one indicator (48 504 (97.1%), prescribing indicator; 21 399 (99.5%), monitoring indicator), but a few triggered two indicators (1253 (2.5%) prescribing indicator; 102 (0.5%) monitoring indicator) and 170 (0. 3%) patients triggered three or more prescribing indicator. Many practices had a prevalence of prescribing indicators over the 90th centile for just one prescribing indicator (158 (30%), P1-P13), but some practices were repeatedly over the 90th centile for different prescribing indicator; 100 (19%) practices were over the 90th centile twice, 79 (15%) three to four times, and 10 (2%) five to six times. For monitoring indicators, 166 (32%, M1-M4) practices had a prevalence over the 90th centile for just one indicator, 29 (6%) twice, and six (1%) three to four times. Monitoring indicators had larger ICCs than prescribing indicators; 0.17 (0.15 to 0.19) for the composite monitoring indicator and 0.04 (0.03 to 0.05) for the composite prescribing indicator (table 2 , fig 4 ). There was low between-practice variation in prescribing β blockers to patients with asthma (ICC P1: 0.03, 0.02 to 0.04), aspirin to patients with peptic ulcer without gastroprotection (ICC P3: 0.04, 0.02 to 0.06), or co-prescription of warfarin and aspirin without gastroprotection (ICC P11: 0.03, 0.02 to 0.04) (table 2 , fig 4 ). The reliability provides guidance about how confidently we can compare an individual practice with a known number of denominator patients with the prevalence over all practices. The implication is that for an indicator with lower ICC, higher numbers of denominator patients are required for a reliable comparison. For this reason, although the indicator related to prescribing β blockers to patients with asthma (P1) has a low ICC, the reliability is high because of the large denominator (number of patients diagnosed with asthma, table 2 ). Conversely prescribing amiodarone without a thyroid function test (M2) has a higher ICC but the small denominator (number of patients prescribed amiodarone) means that the reliability is low (table 3). Tables 4 and 5 show the prevalence of patients triggering at least one prescribing indicator or monitoring indicator (composite indicators) by characteristics of patients and practices, as well as unadjusted and adjusted odds ratios derived from the multilevel mixed effects logistic regression model. Polypharmacy was strongly associated with increased risk of triggering the composite prescribing indicator (1.3% of patients with zero or one repeat medication compared with    table 4 ). Practice level variables had much less influence than patient level variables; there was no effect of list size on composite prescribing or monitoring indicators. Practice level index of multiple deprivation showed a small significantly increased prevalence for the composite prescribing indicator for practices in more deprived areas in the unadjusted model (table  4 ). For the composite monitoring indicator there was no significant association with practice level index of multiple deprivation in the unadjusted model but the adjusted model showed small significant increases for the practices above the third quintile and below the fifth quintile but not the most deprived (above the fifth  (table 5). We excluded the indicators relevant only to women from the composite prescribing indicator (P1-P13) to allow meaningful comparisons between sexes; however, this is not to say that the indicators relevant to women are less important. Table A in appendix 2 shows the odds ratios derived from the multilevel mixed effects logistic regression model for the  . Furthermore, the decrease in prevalence with increasing repeat medications occurred in the two to four repeated prescriptions group and did not increase further with increasing number of prescriptions. Figure 5 shows the prevalence of composite prescribing and monitoring indicators for each practice. A high prevalence for the composite prescribing indicator does not predict a high prevalence for the composite monitoring indicator or vice versa. discussion principal findings There is a high variation in prevalence between general practices for indicators of potentially hazardous prescribing and missed monitoring tests. It is unlikely that variation in case mix could explain the large differences between practices, though prescribing indicators can occur for valid clinical reasons. There is no parallel argument that a monitoring indicator might occur for clinical reasons but patient behaviour might be an important factor. Furthermore, monitoring indicators can occur because of poor recording of test results, whereas most prescribing is systematically recorded through the electronic prescribing system and is known to be accurate (reflected in the larger intraclass correlation coefficients (ICC) for monitoring indicators). 22 Many practices, however, do have zero prevalence for some indicators, and these results might help to set reasonable expectations for individual practices.

Prevalence of prescribing or monitoring indicators (%) Practice ranked by prevalence of prescribing indicator
An important question is whether or not it is appropriate to use these indicators to compare prescribing safety between practices. The ICC suggests that the differences between practices are an important source of variation in the prevalence of the indicators, especially for monitoring indicators, but the ICC is a statistical concept and further work is needed to estimate how often indicators result from oversight or error rather than a clinical decision. The reliability estimates (derived from the ICC) show that some practices have insufficient numbers of patients in the denominator for reliable prevalence estimates for some indicators and these practices should not be compared with other practices. For the composite indicators, however, the reliability was good across all practices (>0.8, appendix 1), suggesting that comparisons between all practices should focus on the composite indicators.
High levels of variation between practices might also reflect differences in service organisation-for example, M5 (prescribed warfarin and no INR) has an ICC of 0.78, possibly because in some practices that prescribe warfarin the INR results might be recorded outside the practice electronic record system and are not uploaded to CPRD. 20 21 One way to identify practices with high risk prescribing might be to look for those with a consistently higher prevalence across several different indicators; 17% of practices had prevalence over the 90th centile for three or more of the 13 prescribing indicators (P1-P13) and 7% of practices for two or more of the four monitoring indicators (M1-M4).
If a lower prevalence of these indicators is a measure of good practice then we would expect consistency between prescribing and monitoring indicators within practices. An alternative hypothesis is that since prescribing indicators are the culmination of a series of actions, whereas monitoring indicators result from inaction on the part of the practice or the patient, practices with a lower prevalence of monitoring indicators would have a higher prevalence of prescribing indicators. In fact there was no clear relation between the two types of indicator within practices ( fig 5). There were differences, however, in the type of patient triggering prescribing or monitoring indicators. Older patients and those receiving multiple repeat prescriptions had the highest risk of triggering the composite prescribing indicator, whereas younger patients with fewer repeat prescriptions had a higher risk of triggering the composite monitoring indicator. The implication is that different types of intervention might be required to deal with the two types of high risk prescribing.
Practice level variables had much smaller effects than patient level variables; patients attending practices in more deprived areas were more likely to trigger the composite prescribing indicator, possibly reflecting more prescribing within the practice leading to higher potential for a prescribing indicator. For the composite monitoring indicator, however, modest significant effects of practice level deprivation were seen only in the adjusted model, suggesting that patients attending practices in deprived areas have no increased likelihood of missing monitoring tests. The definition of the composite prescribing indicator did not substantially alter the predictions made by the model but M1 (prescribing ACEI without urea and electrolyte monitoring) was the major driver for a monitoring indicator and, given the number of patients at risk, might be a better marker of prescribing safety.
In this group of 526 practices, about 5% of patients with the potential to trigger a prescribing indicator did so and about 12% triggered a monitoring indicator. Whether or not these prevalence estimates might be generalised to the UK depends on how representative the practices participating in CPRD are in terms of characteristics of patients and practices. Practices in CPRD are slightly larger than the UK average (this might not cause bias as list size was not a significant predictor of prevalence) but are representative in terms of patient level deprivation and ethnicity. 23 All CPRD practices in this study use Vision software to manage their patient data, and the choice of software has been associated with differences in recorded clinical data. 24 The regional differences in the choice of software impact on the geographical representativeness of CPRD; Yorkshire, Humberside, and North East England are under-represented. 23 Given this, it is reassuring that there are few regional differences in the prevalence of the indicator. The lower prevalence of monitoring indicator in the North East and higher prevalence in London might simply reflect selection bias given the small number of practices. While there might be an argument for generalising these prevalence estimates to the UK, we need to be more cautious in making the same case for the reliability estimates. The variability between practices, from which both the ICC and reliability are derived, might be specific to this group of practices and hence a generalisation to the whole of the UK primary care might be difficult to justify.

Comparison with other studies
It is useful to compare these findings with studies in other UK practices. Consideration, however, must be given to the date of analysis as the quality of electronic data has improved over recent years. The audit date chosen here is a compromise between being as recent as possible but also allowing time for most practices to have had their data uploaded. The prevalence of these indicators was not sensitive to the choice of audit date, suggesting that seasonality and the Quality Outcomes Framework (QOF) return date had little impact. (The QOF incentive scheme provides UK GPs with financial incentives related to performance indicators and practices can review their patient records before the annual return in April 19 ). The only exception was M4 (prescribed lithium without lithium level check), where the increase in lithium level tests during January might relate to preparation for the QOF return or the end of the financial year. The PINCER trial found similar prevalences for P1, P4, P16, M1, and M2 (using identical definitions) in 72 practices in Nottinghamshire, Staffordshire, and Cheshire during 2006-07. 10 There were, however, some differences; the prevalence for M3 (methotrexate without full blood count or liver function test) was 35-42% in the PINCER trial but 16% in the CPRD; and for M5 (warfarin and INR) 6-7% in PINCER and 32% in the CPRD. (CPRD data might not capture all INR test results as discussed above, whereas the PINCER trial excluded practices that were known to undertake their own INR monitoring from this indicator). A study of 315 practices in Scotland used different indicator definitions measured relative to 31 March 2007; there were similar findings for the comparable indicators. 9 The prescription of gastroprotectants, however, was treated differently. We excluded patients from the denominator if they had been prescribed gastroprotection, whereas the Scottish indicator excluded them from the numerator. 9 If the same definition as the Scottish indicator for prescribing NSAIDS to elderly patients was applied, the prevalence was similar between the two datasets (data available from authors).

limitations of the study
While it is relatively straight forward for GPs within the UK to compare their practice with these results, it is a weakness of this study that international comparisons with these data would be challenging. Such comparisons would require access to routinely collected patient data (even CPRD might be missing some data-for example, INR tests 20 21 ). In many countries healthcare data are collected primarily for payment or insurance purposes and the clinical utility of the indicators, and their representativeness of the general population, might be uncertain.

Conclusions and policy implications
These results emphasise the need to give due consideration to the risks of prescribing multiple medications and the importance of regular medication reviews, especially for patients with multiple morbidity. Here we provide a baseline prevalence from which to determine whether or not prescribing safety is improving.
These findings are also relevant to policy makers looking for ways to compare and potentially reward practices with respect to prescribing safety, although careful consideration needs to be given to any such initiative. Prescribing tends to be an individual rather than practice responsibility and prescriptions might be initiated in secondary care rather than by the GP. Further work would also be needed to estimate clearly how frequently indicators result from a clinical decision rather than an oversight or error and which indicators pose the most risk to the patient.
This study is based on data from the Clinical Practice Research Datalink obtained under licence from the UK Medicines and Healthcare Products Regulatory Agency. The interpretation and conclusions contained in this paper are those of the authors alone.
Contributors: EK, DMA, SJS, AA, SR, and AJA designed the study. SJS extracted the data, did the statistical analyses, wrote the manuscript, and is guarantor. SR, AJA, and DMA provided the indicator definitions. DMA also identified the medication product codes. EK and AA provided statistical advice. EK, DMA, AA, SR, and AJA edited the manuscript.
Funding: This study was funded by the National Institute for Health Research through the Greater Manchester Primary Care Patient Safety Translational Research Centre (NIHR GM PSTRC), grant No GMPSTRC-2012-1. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR, or the Department of Health.
Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare no relationships or activities that could appear to have influenced the submitted work.
Ethical approval: The study was approved by the independent scientific advisory committee (ISAC) for Clinical Practice Research Datalink research (reference No 14_049R). No further ethics approval was required for the analysis of the data.
Transparency declaration: SJS affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.
Data sharing: Clinical Practice Research Datalink data cannot be shared because of licensing restrictions. This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/.