Treatment of first-time traumatic anterior shoulder dislocation: the UK TASH-D cohort study

Aims

The commissioned aims were:

• to study the association between surgical treatment and recurrence rates following a first-time TASD in young adults who had surgery compared with those who had not had surgery

• to identify clinical predictors of recurrent dislocation in young adults with a TASD for surgical and non-surgical patients.

Objectives

To use routinely collected data from two NHS computerised databases [i.e. the Clinical Practice Research Datalink (CPRD) and Hospital Episode Statistics (HES)] to study the association between surgical treatment and rates of re-dislocation, compared with no surgery in young adults with a first-time TASD. Potential predictors of re-dislocation in patients from each treatment group were further investigated.

The research questions were addressed by implementing a two-stage approach using two work packages.

Work package 1

The first phase of this project involved conducting an internal and external validation study to test the suitability of using the CPRD data set for identifying patients with a first-time TASD, and then externally validating the findings against published results from other settings. Relevant risk factors that were identifiable in the CPRD were recorded and used to inform the formal analysis regarding future predictors.

Work package 2

Data source

Population-based primary care data from the CPRD were used to identify a cohort of patients diagnosed with a traumatic shoulder dislocation (aged 16–35 years) in the UK from 1 April 1997 to 31 March 2015. The CPRD covers 11.3 million people from 674 UK general practices and provides a representative coverage of around 6.9% of the UK population, which is broadly representative of the population in terms of age, sex and ethnicity. 18 Patient and practice data are anonymised, but patient-level data are available on age, sex, geographic region, body mass index (BMI), smoking status (i.e. current smoker, ex-smoker, non-smoker) and drinking status (i.e. current drinker, ex-drinker, non-drinker). The CPRD data were linked to data from the Index of Multiple Deprivation (IMD) 200419 for English patients and the Charlson Comorbidity Index (CCI) score was calculated using predefined Read codes.

Participants

The Read codes were used to identify patients from the CPRD with a shoulder dislocation. These codes had been established a priori through consensus by specialist shoulder surgeons with clinical experience and experts in epidemiology research (see Appendix 1). To ensure that ‘primary’ or ‘first-time’ shoulder dislocations were captured, patients were required to have no recorded shoulder dislocations in their CPRD clinical data for 2 years prior to first entry of a shoulder dislocation Read code. The 2-year washout period was defined using the date that the general practice was classified as ‘up to standard’ and the date that the patient was first registered at the general practice. This first entry of a shoulder dislocation Read code was defined as the primary dislocation.

The patients had to be registered at ‘active’ CPRD practices. An ‘active’ practice was defined as a practice that had contributed to the CPRD database in the previous 6 months. No general practices were classified as active in the East Midlands, and so it was not possible to include patients from this region. Following the identification of the shoulder dislocation cohort from within the CPRD data set, a predefined set of patient exclusion criteria was applied to facilitate the validation (Table 1).

General practitioner questionnaire design and implementation

The GP questionnaire was designed with the assistance of GPs and based on a developed validation algorithm (see Appendices 2 and 3) and a random sample of 172 patients was selected from a list of the 6046 eligible patients. CPRD personnel then sent the questionnaire to each patient’s general practice for a clinician to complete by comparing the records on their CPRD computer system with the patient’s clinical records. The GPs assessed the use of shoulder dislocation codes for traumatic dislocation, confirmation of first-time shoulder dislocation, subsequent codes used for further events, physiotherapy referral codes and confirmation that physiotherapy took place.

Four written reminders were sent every 2 weeks to the general practices. Data from the returned questionnaires were double-entered into a data set by a statistician and a project manager. Any queries were resolved by an academic orthopaedic shoulder surgeon. In the instance that two questionnaires were received for the same patient with differing answers (on three occasions two questionnaires were sent back for one patient, i.e. three patients and six questionnaires; differing responses were only received for questions 6 and 7), clarification was sought from the general practice via CPRD personnel (clarification was received for one patient).

The following validation criteria were defined a priori to reflect that the coding of shoulder dislocations in the CPRD was of a sufficiently high quality to proceed with the main analyses planned in work package 2:

• a positive predictive value of ≥ 75% accuracy for shoulder dislocation coding in the CPRD

• a positive predictive value of ≥ 75% accuracy in coding ‘primary’ or ‘first-time’ traumatic shoulder dislocation within the CPRD.

Cohort

An initial cohort of 63,324 patients with codes for shoulder dislocation was identified from the CPRD database. A database manager and statistician assessed the cohort against clear predefined and important exclusion criteria (see Table 1). Unacceptable patients were defined as those whose records had not met quality standards and had been flagged by the CPRD as ‘unacceptable’. Unacceptable dates were defined as when it was impossible to find a shoulder dislocation code between the CPRD minimum and maximum acceptable dates, as defined by data management standard operating procedures for clinical research. 18 During this process, the majority of patients were excluded, either because they had a shoulder dislocation diagnosis outside the study time period (55%) or because they were outside the study age limits (16–35 years) (20%). The final cohort included 6046 patients aged 16–35 years who were diagnosed with a shoulder dislocation between 1 April 1997 and 31 March 2015 in England.

Internal validation

Of the 172 patients whose GP received a copy of the validation questionnaire, a response for 95 (55%) patients was received. For two patients, their GPs confirmed that they had transferred out of the practice and that no further information was available for them on the CPRD system.

Table 2 presents demographic characteristics for the following patient groups:

• the cohort of 6046 patients from the CPRD aged 16–35 years and diagnosed with a shoulder dislocation between 1 April 1997 and 31 March 2015 in England

• the 172 patients randomly selected to have their GPs receive a questionnaire

• the 97 patients for whom completed questionnaires were returned by their general practice

• the 75 patients for whom questionnaires were not returned by their general practice.

All four groups were similar with respect to demographic characteristics, including age, BMI and CCI score. The highest response rate (100%) was received from general practices in the South West of England, but otherwise the proportion of responses received reflected the regional distribution of patients included in the cohort. Data on the IMD 200419 were obtained and linked after the selection of the 172 records to be validated. There were no missing data on deprivation, as all practices sampled were ‘active practices’. A higher proportion of patients had been sampled from category 1 (affluent) and category 4 (somewhat deprived) than in the initial cohort of 6046 patients, but otherwise response rates were similar from all deprivation groups.

The distribution of CPRD Read codes used by GPs to code shoulder dislocations is given in Table 3. Codes S41..00, S41z.00 and 14G5.00 for dislocation of shoulder accounted for 82% of all shoulder dislocation coding in the data. Recurrent shoulder dislocation codes only identified another 10% of patients, indicating that the 2-year washout period was a successful approach to identifying primary or first-time shoulder dislocations. Of the seven patients who had a recurrent shoulder dislocation code and for whom a GP questionnaire response was obtained, four were confirmed as having had a primary shoulder dislocation and one was confirmed as not having had a shoulder dislocation at all.

Shoulder dislocation was confirmed as having been coded correctly in 89% (95% CI 83% to 95%) of all patients (Table 4). The remaining 11% (10 patients) had been miscoded and the patient had suffered other shoulder trauma or injuries, such as strains or dislocations of the acromioclavicular joint, as confirmed by their GP. Of all patients, a first-time or primary shoulder dislocation was confirmed in 76% (95% CI 67% to 85%) of cases. Subsequent dislocations occurring up to 2 years after the primary dislocation were recorded in the CPRD for 32% of patients. From the GP responses, an additional 11% of patients experienced a re-dislocation during this time that was not recorded in the CPRD.

Twenty-eight per cent of patients had been coded as having received physiotherapy within the CPRD, and GPs confirmed that physiotherapy had been given to 63% of these. However, a further 17 patients had received physiotherapy for their shoulder dislocation that was not recorded within the CPRD. Thus, 41% of patients known to be receiving physiotherapy were not recorded within the CPRD.

Data source

The CPRD of population-based primary care data was used to identify a cohort of patients diagnosed with a traumatic shoulder dislocation aged 16–70 years during 1995–2015 in the UK. A description of the CPRD and the potential risk factors available within it was presented in Chapter 2.

Participants

The cohort of 16,763 CPRD patients aged 16–70 years with a TASD during 1995–2015 in the UK was used. The patients were identified using predefined Read codes as described in Chapter 2 (see also Appendix 1).

Statistical analysis

Descriptive statistics were used to summarise the epidemiology of primary shoulder dislocations by demographic factors. The incidence rates by age and sex per 100,000 person-years and incidence rate ratios with 95% CIs and p-values were calculated for all age and sex groups using Stata^® software version 14.1 (StataCorp LP, College Station, TX, USA).

Incidence rate denominators were constructed using the patient-level data from the CPRD. Observation time per patient was calculated between 1995 and 2015 as the sum of total year-time contributed by all subjects, wherein person-years start as the latest of first registration date, practice up-to-standard date and 1 January 1995, and end as the earliest from patient transfer out date, practice last collection date, death date and 31 December 2015.

UK incidence rates

The incidence rates and incidence rate ratios by age and sex for primary shoulder dislocation patients in the UK are presented in Table 8. The overall incidence rate in males was seen to be 40.39 per 100,000 person-years (95% CI 40.38 to 40.41 per 100,000 person-years) and in females was 15.52 per 100,000 person-years (95% CI 15.51 to 15.52 per 100,000 person-years). The highest incidence observed was in 16- to 20-year-old males (80.55 per 100,000 person-years, 95% CI 80.45 to 80.65 per 100,000 person-years). The incidence in men decreased with an increase in age. A U-shaped pattern of incidence was observed in women. The incidence was 16.36 per 100,000 person-years in those aged 16–20 years. This decreased in women aged 21–50 years and then increased to 28.64 per 100,000 person-years in women aged 61–70 years. Overall, the incidence was significantly higher in men than in women in almost all age groups, with an overall incidence rate ratio of 2.60 (95% CI 2.52 to 2.69). The exception was found in men and women aged 61–70 years, in whom no significant difference in incidence was observed (p = 0.334).

Comparison of UK incidence data with Canadian incidence data

Incidence rates for TASD in the UK were also compared by age and sex with those reported in Canada. 16 A comparative summary of the characteristics of the UK and Canadian cohorts is shown in Table 9. The UK cohort included in the analysis consisted of 15,666 patients aged 16–70 years with a primary shoulder dislocation in the UK as recorded in the CPRD between 1 April 1997 and 31 March 2015. Denominators for incidence analyses were obtained from the CPRD by individual year, age and sex. Data on urban or rural residence were not available in the CPRD, and thus this comparison could not be made.

The demographic data for the UK cohort (Table 10) were similar in age and sex distribution to those observed in the Canadian cohort. 16 The median age in both cohorts was 35 years, with a similar IQR. In the UK, 72% of primary shoulder dislocations occurred in men; in the Canadian cohort, this was slightly higher, at 74%.

The English IMD 200419 is a composite deprivation index at the small-area level, based on seven domains: income, employment, health and disability, education, barriers to housing and services, living environment, and crime. The Canadian measure of deprivation is solely based on income. These two measures are not directly comparable but a similar pattern of deprivation is observed, with increasing numbers of patients linked to increasing affluence. There was a large number of missing data for the UK cohort because the IMD results were available for English patients only, whereas complete data on deprivation were available for the Canadian patients.

Figures 2 and 3 present the percentage of primary shoulder dislocation patients by age and sex in the UK and Canada, respectively. In the UK, the peak in numbers for men is spread over those aged 17–22 years, whereas there is a distinct peak in men aged 17 years in Canada. The pattern in the number of women with shoulder dislocations is similar in both cohorts, with a high percentage in those aged 16 years, which decreases up to the early 30s and then increases until the age of 70 years.

FIGURE 2.

The percentage of primary shoulder dislocation patients by age (16–70 years) and sex recorded within CPRD (UK) between 1 April 1997 and 31 March 2015. Reproduced from Shah et al. 17 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/.

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/

FIGURE 3.

The percentage of primary shoulder dislocation patients by age and sex recorded in Canada. Leroux T, Wasserstein D, Veillette C, Khoshbin A, Henry P, Chahal J, et al. , American Journal of Sports Medicine, vol. 42, issue 2, pp. 442–50, copyright © 2014 by American Orthopaedic Society for Sports Medicine, reprinted by Permission of SAGE Publications, Inc. 16

Leroux T, Wasserstein D, Veillette C, Khoshbin A, Henry P, Chahal J, et al. , American Journal of Sports Medicine, vol. 42, issue 2, pp. 442–50, copyright © 2014 by American Orthopaedic Society for Sports Medicine, Reprinted by Permission of SAGE Publications, Inc.)

The incidence rates by age and sex for primary shoulder dislocation patients in the UK and an extract of similar Canadian incidence data are presented in Table 11. The patterns of incidence by age and sex were similar and are also represented in Figures 4 and 5. Incidence rates were higher in the UK for all combinations of age groups and sex, except for men aged 16–20 years [UK men (81.6 per 100,000 person-years) vs. Canadian men (98.3 per 100,000 person-years)].

FIGURE 4.

The incidence rate of primary shoulder dislocation by age and sex among patients aged 16–70 years with a primary shoulder dislocation recorded within CPRD (UK) between 1 April 1997 and 31 March 2015.

FIGURE 5.

The overall IDR of primary anterior shoulder dislocation in Canada requiring closed reduction in accordance with both age and sex. The y-axis depicts the IDR per 100,000 person-years and the x-axis depicts each age category. IDR, incidence density rate. Leroux T, Wasserstein D, Veillette C, Khoshbin A, Henry P, Chahal J, et al. , American Journal of Sports Medicine, vol. 42, issue 2, pp. 442–50, copyright © 2014 by American Orthopaedic Society for Sports Medicine, reprinted by permission of SAGE Publications, Inc. 16

Leroux T, Wasserstein D, Veillette C, Khoshbin A, Henry P, Chahal J, et al. , American Journal of Sports Medicine, vol. 42, issue 2, pp. 442–50, copyright © 2014 by American Orthopaedic Society for Sports Medicine, Reprinted by Permission of SAGE Publications, Inc.)

Comparison of UK incidence data with US incidence data

The age and sex incidence rates for primary shoulder dislocation in the UK were also compared with published data from the USA. 15 A comparison of the characteristics between the UK and US cohorts is shown in Table 12. The UK cohort included in the analysis comprised 20,784 patients of all ages with a primary shoulder dislocation between 1 April 1997 and 31 March 2015. Data on ethnicity were not available within the CPRD, thus this comparison could not be made. Denominators for incidence analyses were obtained from the CPRD by individual year, age and sex.

The overall incidence rate in the UK was 26.6 per 100,000 person-years (95% CI 26.2 to 26.9 per 100,000 person-years), which was similar to the incidence rate reported in the USA (23.9 per 100,000 person-years, 95% CI 20.8 to 27.0 per 100,000 person-years). The peak incidence occurred in patients aged 20–29 years in the UK (41.8 per 100,000 person-years, 95% CI 40.6 to 43.1 per 100,000 person-years) (Table 13), which is similar to the peak in 20- to 29-year-olds observed in the USA (47.8 per 100,000 person-years, 95% CI 41.0 to 54.5 per 100,000 person-years). 15 Significantly higher rates of incidence were observed in the UK among patients aged > 50 years in comparison with those in the USA (see Table 13).

In the UK, the incidence of primary shoulder dislocation was significantly higher in men (34.3 per 100,000 person-years, 95% CI 33.7 to 34.9 per 100,000 person-years) than in women (19.0 per 100,000 person-years, 95% CI 18.6 to 19.5 per 100,000 person-years) (p < 0.001), which is similar to the pattern observed in the USA (see Table 13). Incidence of shoulder dislocation in men was similar between the UK and US cohorts, but incidence in women in the UK was much higher than incidence in women in the USA. In the UK cohort, 36% of shoulder dislocations occurred in women, in contrast to 28% of shoulder dislocations in the US cohort.

Figure 6 shows the peak of incidence in men aged 20–29 years in the UK (71.5 per 100,000 person-years), which was similar to the peak in men in the same age group in the USA (79.2 per 100,000 person-years) (Figure 7). The peak in women in the UK was observed in those aged > 90 years (71.7 per 100,000 person-years), in contrast with the 38.8 per 100,000 person-years in women aged 80–89 years in the USA. A possible reason for the differences in incidence may be caused by the UK study being based on primary care records and the US study being based on emergency department records.

FIGURE 6.

Shoulder dislocation incidence rates and 95% CIs by age and sex in CPRD (UK) between 1 April 1997 and 31 March 2015.

FIGURE 7.

Total weighted NEISS estimates of all US shoulder dislocations between 2002 and 2006 by age and sex, demonstrating a bimodal distribution with peaks among men (aged 20–29 years) and women (aged 80–89 years). The vertical bars denote the 95% CIs. 15 Zacchilli MA, Owens BD. Epidemiology of shoulder dislocations presenting to emergency departments in the United States. J Bone Joint Surg 1992;3:542–9. https://insights.ovid.com/pubmed?pmid=20194311. Reprinted by permission of Wolters Kluwer.

Zacchilli MA, Owens BD. Epidemiology of shoulder dislocations presenting to emergency departments in the United States. J Bone Joint Surg 1992;3:542–9. https://insights.ovid.com/pubmed?pmid=20194311. Reprinted by permission of Wolters Kluwer.

Study design

A population-based propensity-score-matched cohort study to control for confounding at baseline has been conducted. Young adults (aged 16–35 years) who presented with a first-time TASD were selected from two computerised NHS databases (CPRD and HES). Figure 8 shows a detailed illustration of the study plan, which is described in more detail in the following paragraphs.

FIGURE 8.

The UK TASH-D phase 2 plan design for patients who received surgery or no surgical treatment within 6 months of a diagnosis of a first-time TASD during 1 April 1997–26 April 2014, in England.

Using methodology identical to the internal validation study described in Chapter 2, to ensure that only primary dislocations had been captured, all participants had to have 2 years of clinical data within CPRD before their first TASD and at least 2 years’ follow-up from the event (re-dislocation). This 2-year washout period is required to ensure that a first dislocation code actually represents a first-time dislocation, as a recurrent dislocation usually occurs within 2 years of the primary event. This period therefore minimises the risk of a code being a second dislocation code. The period was defined using the date that the general practice was classified as ‘up to standard’ and the date that the patient was first registered at the practice. Patients had to be registered at ‘active’ practices, as defined in Chapter 2, Data source. Taking into account this washout period, the first Read code entry in CPRD for shoulder dislocation was then defined as the first dislocation. All events were collected using a pre-agreed validated list of CPRD Read codes (see Appendix 1). Patients experienced their first-time TASD between 1 April 1997 and 26 April 2014, allowing at least 2 years of follow-up for each patient to the end of the study on 26 April 2016.

A 6-week washout period for re-dislocation codes within CPRD was applied to all patients following their TASD to avoid duplicate records. Patients were highly unlikely to re-dislocate their shoulder during this period as their arms would be in slings and they would be following rehabilitation protocols, but would probably return to visit their GP for prescriptions for painkillers or referrals to physiotherapy and secondary care.

Outcome

The outcome was time to a shoulder re-dislocation, as defined by the CPRD Read codes given in Appendix 1. For the surgical group, the shoulder re-dislocation can occur between 6 weeks and 2 years following the date of surgery. For the non-surgical group, the shoulder re-dislocation could occur between 6 weeks and 2 years following the date of first-time TASD.

Any patients who died during the study were censored on their date of death.

Data sources

The analysis utilised two computerised NHS databases, one from primary care (the CPRD) and the other from secondary care (the HES). The characteristics of the CPRD have been described in Chapter 2, Data source. For HES data, each time a patient sees a health professional in a hospital, a record or ‘episode’ is created and added to the HES database. The HES record contains patient details, some diagnosis codes, treatments and lengths of hospital stay.

Clinical Practice Research Datalink has been linked with HES data to provide a HES-linked patient identifier. About 60% of CPRD-contributing practices have been linked to HES data. HES data provide a general patient identifier to facilitate linkage of hospital records belonging to the same individual. Management of the CPRD and HES databases was carried out by a senior data manager with expertise in the use of these data sets. The senior data manager developed an ad hoc code using Python (version 3.6, developed by Python Software Foundation, Wilmington, DE, USA) and Structured Query Language (SQL) (version 5.6.12, developed by the International Electrotechnical Commission and the International Organization for Standardisation; Geneva, Switzerland) to produce a final working data set that was analysed using standard statistical software packages Stata^® version 14.1 and R (version 3; The R Foundation for Statistical Computing, Vienna, Austria).

Sample size

The original sample size considerations for this study were based on data from a Cochrane systematic review comparing surgical and non-surgical treatment for an acute TASD. 5 From the pooled results, 3 out of 58 patients in the surgical arm had subsequent further surgery (5.17%) compared with 17 out of 61 patients in the non-surgical arm (27.9%), at a minimum follow-up of 2 years (risk ratio 0.22, 95% CI 0.08 to 0.64). The large effect size is based on the pooled results of three randomised controlled trials with uncertainty around the true size of the effect outside a clinical trial setting in routine general practice. Therefore, values were set to detect a smaller difference in subsequent surgery within 2 years, with a 25% re-dislocation rate in the non-surgical group, compared with a 20% re-dislocation rate in the surgical group (an absolute difference of 5%). A two-sided, log-rank test for equality of survival curves was used, with 90% power at a 5% significance level (alpha) and for which the outcome is time to re-dislocation with an anticipated 25% re-dislocation rate in the non-surgical control group compared with a 20% re-dislocation rate in the surgical group [equivalent to a hazard ratio (HR) of 0.78]. Allowing for a 10% loss to follow-up and assuming equal group sizes meant that the study required a total sample size of 3065 participants, with 656 expected re-dislocations. 5 It was assumed that 35% of the patients would receive surgery within 6 months after one dislocation (n = 1073). 23

Participants

Exclusion criteria

The following patients were excluded:

• those aged 16–35 years with a first-time TASD who cannot be linked by CPRD-HES

• those with < 2 years of follow-up in the CPRD

• those with prior shoulder surgery for a shoulder dislocation

• those whose instability was treated with rotator cuff repair surgery or fracture surgery prior to or following a TASD.

Exclusions were made in accordance with the criteria above and are described in Table 14. Following linkage to HES, there was a linkage loss of 1234 patients and six duplicates were identified and removed. A substantial number of patients had < 2 years of follow-up within the CPRD (n = 854), which may relate to these young people moving away from home to attend university or to start new jobs in new locations, which, in turn, requires a change of GP. A very small proportion of patients (3%) were excluded for having shoulder surgery for rotator cuff tears, fractures or prior dislocations. In total, 3759 patients remained available for analysis, which was greater than the minimum number of patients (n = 3065) required for a sufficient sample size and power (Figure 9).

TABLE 14 - Shoulder dislocation exclusions for patients aged 16–35 years during 1 April 1997–26 April 2014, within CPRD in England with the linkage to HES records

Exclusions	Excluded (n)	Total (N)	%
Eligible patients in the CPRD cohort for the internal validation study, following exclusions made in Table 1		6046	100
HES linkage loss	1234		20
Duplicate removal	6		< 1
< 2 years of follow-up within the CPRD	854		14
Prior shoulder surgery for a shoulder dislocation	155		3
Surgery for rotator cuff tear prior to TASD	1		< 1
Surgery for rotator cuff tear in the 6 months following TASD	4		< 1
Surgery for a shoulder fracture prior to TASD	14		< 1
Surgery for a shoulder fracture in the 6 months following TASD	19		< 1
Total exclusions	2287
Patients included in subsequent analyses		3759	62

FIGURE 9.

Shoulder dislocation exclusions flow chart for patients aged 16–35 years during 1 April 1997–26 April 2014, within CPRD in England with linkage to HES records.

Covariates

Demographic data were available from the CPRD on age, sex, BMI, IMD 2004,19 smoking status (i.e. current smoker, ex-smoker, non-smoker), drinking status (i.e. current drinker, non-drinker), geographic region, epilepsy and prescriptions for painkillers in the 3 months preceding the first-time TASD and 1 month following the first-time TASD. The CCI score was calculated using a list of predefined CPRD Read codes.

A consensus survey was conducted of specialist shoulder surgeons and shoulder physiotherapists who were all members of the British Elbow and Shoulder Society. The saturation point and a list of predictors was reached rapidly and this list is tabulated in Appendix 4. The list highlights the risk factors (and covariates) deemed most important. However, data were only reliably available on the following factors from the list: age, sex, geographic region, deprivation scores, time between first dislocation and surgery.

Missing data

Multiple imputation using chained equations was used to address potential bias and increase precision as a result of missing data on BMI, smoking, drinking and IMD. 24 The imputation equations included all potential factors, including the outcome and length of follow-up time. Fifty imputed data sets were generated and the resulting estimates were combined using Rubin’s rules.

Confounding by indication

In randomised controlled trials, each person has an equal probability of being in the treatment or the control group. Observational study designs, such as the one used for this study, are limited by an inherent imbalance of both known and unknown confounders, making some patients more likely to receive surgery than others. A surgeon typically uses information and risk factors on the patient at baseline to make a decision about whether or not to operate. Whether or not a patient receives surgery is therefore not random in this population-based setting.

As the type of surgery received is not randomly allocated in this study, propensity score matching methods were used to minimise confounding by indication. These propensity score methods were used to achieve comparability of groups with and without the intervention with respect to their observed baseline covariates, thus, controlling for confounding in estimating treatment effects. The use of these methods for the assessment of causality in epidemiological studies has been previously described. 25

The propensity score represents the probability that a patient received the intervention (surgery) conditionally, based on their covariate values. One feature of the propensity score is that it provides balance, so that at each value of the propensity score, the distribution of the covariates (used to define the score) is expected to be similar in the intervention group and non-surgical-intervention group. Comparing patients with the intervention and those without the intervention with the same propensity score gives an unbiased estimate of the effect of treatment.

A logistic equation was fitted for which the outcome was actually the main study exposure (surgical compared with non-surgical intervention) and an agreed list of covariates were introduced as potential confounders of the study outcome. 26,27 All of the covariates described in Covariates were included in the model.

Propensity scores were then used to match each patient receiving surgery to comparable non-surgical controls using a 0.2 standard deviations calliper, as demonstrated in previous simulation studies. 28 Matching was performed without replacement using the MatchIt package in R. 29 Each patient receiving surgery for TASD was matched to three comparable non-surgical controls.

The balance of covariates before and after matching was assessed by calculating the absolute standardised mean difference (SMD) for each covariate. The commonly used boundary for the absolute SMD to indicate acceptable balance is 10%, meaning that a standardised difference of < 10% is considered a good balance. The distribution of propensity scores before and after matching were also judged subjectively for sufficient overlap following matching using density plots.

If the analysis includes participants outside the boundaries of the overlap, this can lead to biased estimates. Participants outside these boundaries will be patients with very high or very low propensity scores. Thus, the best approach is to trim the patients included in the analysis, removing those with extreme propensity scores. This was carried out by calculating 1% of the extremes of the propensity score tails and removing patients outside the limits.

This is a standard method for minimising confounding by indication that not only provides participants with balanced baseline characteristics in both surgical and non-surgical groups, but also eliminates surgical patients with no comparable controls. 30

This methodology is now widely used in pharmacoepidemiology and drug safety, and has both strengths and limitations. The main advantages of propensity score matching are:

• Exclusion of non-comparable subjects (e.g. non-surgical participants with a very low propensity score who probably have some contraindication or are not fit for surgery and, therefore, should not be compared with those who actually underwent surgical repair).

• This method produces clearly comparable cohorts in terms of observed confounders and is highly visual and intuitive.

The main disadvantage (when compared with randomised trials) is the lack of adjustment/matching for unobserved confounders. In an observational setting, there is the potential risk that the choice of patient treatment by skilled clinicians is driven by unmeasured patient characteristics and risk factors that are not recorded in the observational data sets. This can affect the precision of the estimate of treatment efficacy and external validity. 31 The performance of a propensity score can be examined for homogeneity at different points on the propensity score scale. If the analysis has worked as anticipated, a similar treatment effect should be observed across the range of propensity score values. Another potential limitation is the potential loss of power if patients cannot be matched as part of the propensity score analysis.

The association between surgery and time to re-dislocation within a 2-year time frame was described using a Cox regression survival model, including the surgical patients and their three matched non-surgical controls. The model was stratified on matched sets, to allow for the correlation between matched pairs of surgical patients and controls. An assumption in the use of the Cox regression model is that of proportional hazards, which was assessed using Schoenfeld Residuals Test. Probability of survival up to 2 years was estimated in the surgical and non-surgical groups using Kaplan–Meier plots.

Immortal time bias

A common issue in epidemiological studies is that of immortal time bias, which describes a form of bias introduced from a period of time when the outcome or event of interest cannot occur by design. It usually occurs in a cohort study with two index dates, when the passing of time from the first date (i.e. inclusion date) to the date when a patient receives the intervention (i.e. exposure/treatment date) is by definition immortal in the exposed group. In the present study, immortal time bias would be introduced in the surgical group, arising in the time following a first dislocation to when they receive surgery, because during this time they cannot have the outcome of interest (otherwise they would have been classified as ‘non-surgical’). Although the patient is not truly ‘immortal’, they had to remain free from re-dislocations prior to receiving surgery, which introduces a bias of offering guaranteed survival time to the surgical group. The surgical patients will be artificially ‘safeguarded’ from having a re-dislocation until their date of surgery. By not being correctly classified, this immortal time would produce an artificial increase in re-dislocations in the non-surgical group, suggesting that surgery has a better outcome. The effects of immortal time bias have been confirmed and quantified by Suissa. 32

To address the problem of immortal time bias, time-varying exposures have been used. In the survival analysis, the time prior to surgery for the surgical group has been reclassified as ‘non-surgical’, and new (‘twin’) patients have been created and added to the ‘non-surgical’ group. This is deemed the best available method for the minimisation of immortal time bias. 33

Descriptive characteristics

A cohort of 3759 patients diagnosed with a first-time TASD during 1 April 1997–26 April 2014 was identified in the CPRD. The CPRD Read codes used to identify these patients can be found in Appendix 5.

An unexpected finding at this stage was that only 4% (n = 156) of the remaining patients had received surgery for their dislocation within 6 months. This was only 15% of the expected number of surgical patients. In addition, an identical proportion of patients (20%) had suffered a re-dislocation in the surgical and non-surgical groups instead of the anticipated 5% difference. Although this finding is useful for this commissioned study, indicating that surgery after one traumatic dislocation is not common in the general NHS population, it also means that, overall, this study was underpowered for the primary question at 6 months. The analysis was conducted as per the approved protocol, but a protocol amendment was added to carry out an additional sensitivity analysis at 12 months.

For the 156 patients who underwent shoulder surgery within 6 months, the OPCS 4.7 surgical codes identified within HES are given in Appendix 6. The descriptive characteristics of patients categorised by receiving surgery within 6 months or no surgery are described in Table 15. In the first 4 years, fewer than five patients underwent surgery per year. This increased from 2001 to a maximum of 18 (out of a total of 156 surgical patients) in 2008. More men (83%) were diagnosed with a primary shoulder dislocation than women, and more men underwent surgery within 6 months (40 men vs. 16 women). The majority of patients were aged 17–21 years at the time of their first-time TASD, but similar numbers were operated on among those aged 18–25 years.

Only 8% of patients were overweight or obese, with missing data on BMI for 40% of all patients. There was a pattern of decreasing numbers of patients by deprivation quintile, but this pattern was not evident for surgical patients. Only three patients did not have a recorded IMD score. Most patients were non-smokers (54%) and there were missing data on smoking for 7% of all patients. Most patients drank alcohol (49%) and there were missing data on alcohol consumption for 41% of all patients. Most patients had no comorbid conditions (93%). Many patients were from the North West (n = 636) and South Central regions (n = 543), and fewer were from the North East region (n = 96). A small proportion of patients had been diagnosed with epilepsy (3%), but a higher proportion of all surgical patients had been diagnosed with epilepsy (8%). A small proportion (6%) of patients were prescribed painkillers in the 3 months prior to their first-time TASD, which increased to 12% in the 1 month following TASD. Of the patients prescribed painkillers, both 3 months prior to and 1 month after their TASD, only 1% of patients were prescribed a different painkiller. Only 22 patients died during the study period.

Multiply imputed data analyses

Multiple imputation by chained equations was conducted for patients with missing data on BMI, smoking, drinking and IMD. Appendix 7 presents the HRs, 95% CIs and p-values for each of these variables using all available data, a complete-case analysis and the multiply imputed data. The HRs are similar for each group, indicating that the multiple imputation process was successful.

Cox survival estimates for complete cases and the multiply imputed data, with the factors that had an impact on re-dislocations, are presented in Appendix 8. In total, only 1790 cases had complete data on all factors. For these patients, a protective but non-significant adjusted effect of surgery was observed (HR 0.55, 95% CI 0.29 to 1.04; p = 0.068). Year of first-time TASD, younger age at first-time TASD and an epilepsy diagnosis were significant risk factors for a re-dislocation.

Using the multiply imputed data resulted in similar HRs, 95% CIs and the same risk factors. The protective effect of surgery was less marked and remained non-significant (HR 0.76, 95% CI 0.52 to 1.11; p = 0.151).

Propensity score analysis

Logistic regression was used to calculate the propensity scores. The distribution of propensity scores in patients receiving surgery within 6 months and in those who had non-operative treatment are presented in Figure 10. Prior to matching, the propensity scores tended to be higher in the surgery group, but there was a substantial amount of overlap, indicating that it is possible to proceed with propensity score matching to estimate the treatment effect. Following matching, the propensity score density plots are very similar between the surgery and non-surgery patient groups. The propensity score matching does not include the non-surgical patients in the left-hand peak of the distribution.

FIGURE 10.

Propensity scores for first-time TASD patients diagnosed during 1 April 1997–26 April 2014, who had surgery within 6 months or no surgery within CPRD-HES, in England. (a) Prior to PS matching; and (b) following PS matching. PS, propensity score; SD, standard deviation.

Table 16 presents the baseline characteristics of first-time TASD patients who had surgery or non-operative treatment within 6 months. The SMDs, prior to propensity score matching, show that the two groups of patients do differ for most characteristics (SMD > 0.1). During propensity score matching, each surgical patient (n = 156) was matched to three non-surgical patients (n = 468). In addition to this, the time between the date of the first-time TASD and the date of surgery was allocated to the non-surgical patients for 102 surgical patients, making a total of 570 non-surgical patients. The remaining 54 surgical patients had no time to be re-allocated because the date of their first-time TASD was the same as their date of surgery. Following propensity score matching, the SMDs were much smaller than before (all < 0.1), suggesting that the balancing was successful.

Figure 11 presents Kaplan–Meier estimates of probability of survival in the surgical and non-surgical groups; the hazards were proportional (Schoenfeld Residuals Test; p = 0.39). There appears to be a small survival advantage for surgical patients that is not statistically significant.

FIGURE 11.

Kaplan–Meier survival estimates for first-time TASD patients diagnosed during 1 April 1997–26 April 2014 who had surgery within 6 months or no surgery within CPRD-HES, in England.

Table 17 presents the final results of the effect of surgery within 6 months compared with no surgery among first-time TASD patients. The median follow-up in both groups was similar and the rate of shoulder re-dislocations was slightly higher in the non-surgical group [0.36 per 1000 person-years (95% CI 0.29 to 0.43 per 1000 person-years) in the non-surgical group compared with 0.30 per 1000 person-years (95% CI 0.21 to 0.43 per 1000 person-years) in the surgical group], although this was not statistically significant as demonstrated by the wide CIs. Overall, the effect of surgery within 6 months appeared slightly protective, at HR 0.88 (95% CI 0.58 to 1.35; p = 0.565), but this was not statistically significant.

Study design

A population-based, propensity-score-matched cohort study to control for confounding at baseline has been conducted. Young adults (aged 16–35 years) who presented with a first-time occurrence of TASD were selected from two NHS computerised databases: the CPRD and HES. Figure 12 shows a detailed illustration of the study plan that is described in more detail in the following sections.

FIGURE 12.

The UK TASH-D phase 2 plan design for patients who received surgery or no surgical treatment within 12 months of diagnosis of a first-time TASD during 1 April 1997–31 March 2015, in England.

This study used methodology identical to that described in the internal validation study in Chapter 2 and the 6-month analysis in Chapter 4; all participants had to have 2 years of clinical data within the CPRD before their first-time TASD to ensure that only ‘primary’ dislocations had been captured. The same rules were followed and the first Read code entry in the CPRD for shoulder dislocation was defined as the ‘first’ dislocation. All events were collected using the pre-agreed, validated list of CPRD Read codes (see Appendix 1). Patients experienced their first-time TASD between 1 April 1997 and 31 March 2015, allowing up to 3 years of follow-up for each patient to the end of the study on 26 April 2016. As in the previous analysis, a 6-week washout period for re-dislocation codes within the CPRD was applied to all patients following their TASD to avoid duplicate records.

The key changes from the previous analysis described in Chapter 4 were to include:

• patients having surgery up to 12 months from the date of their first-time TASD

• patients with any length of follow-up, rather than at least 2 full years

• follow-up up to 3 years, rather than 2 years.

Outcome

The outcome was time to a shoulder re-dislocation as defined by the CPRD Read codes given in Appendix 1. For the surgical group, the shoulder re-dislocation can occur between 6 weeks and 2 years following the date of surgery. For the non-surgical group, the shoulder re-dislocation could occur between 6 weeks and 2 years from the date of the first-time TASD.

Any patients who died during the study were censored on their date of death.

Data sources

The analysis utilised two computerised NHS databases: the CPRD and HES, as described in Chapter 4. The linked databases were managed by a senior data manager using Python and SQL, and statistical analyses were conducted using Stata and R.

Sample size

The sample size calculation in the study plan, described fully in the previous chapter, required a total sample size of 3065 patients with 656 re-dislocations, with 90% power at the 5% significance level and allowing for a 10% loss to follow-up, to detect an absolute difference of 5% between the surgical and non-surgical groups. It was assumed that 1073 (35% of the total) patients would receive surgery. The 6-month analysis was underpowered on this basis.

A senior statistician reran the sample size calculation, reducing the power to 80%, having unequal groups (1 : 10 ratio), extending follow-up to 3 years from 2 years and looking at surgical intervention up to 12 months (rather than 6 months). To detect a HR of 0.73 (26% surgical re-dislocations vs. 19.5% non-surgical re-dislocations) would require a total of 3456 patients, of whom 314 were surgical patients and 3142 were non-surgical patients, and at least 695 re-dislocations.

Participants

Exclusion criteria

The following patients were excluded.

• those aged 16–35 years with a first-time TASD who could not be linked by CPRD-HES

• those with prior shoulder surgery for a shoulder dislocation

• those whose instability was treated with rotator cuff repair surgery or fracture surgery prior to or following a TASD.

Exclusions were made in accordance with the criteria above and these are described in Table 18. Following linkage to HES data, there was a linkage loss of 1234 patients and six duplicates were identified and removed. A very small proportion of patients (n = 193) was excluded for having shoulder surgery for prior dislocations, rotator cuff tears or fractures.

TABLE 18 - Shoulder dislocation exclusions for patients aged 16–35 years during 1 April 1997–26 April 2014 within CPRD in England with the linkage to HES records

Exclusions	Excluded (n)	Total (n)	%
Eligible patients in the CPRD cohort for the internal validation study following exclusions made in Table 1		6046	100
HES data linkage loss	1234		20
Duplicate removal	6		< 1
Prior shoulder surgery for a shoulder dislocation	155		3
Surgery for rotator cuff tear prior to TASD	1		< 1
Surgery for rotator cuff tear 6 months following TASD	4		< 1
Surgery for a shoulder fracture prior to TASD	14		< 1
Surgery for shoulder fracture in the 6 months following TASD	19		< 1
Total exclusions	1433		24
Patients included in subsequent analyses		4613	76

In total, 4613 patients remained available for analysis, which was greater than the minimum number of patients (n = 3456) required for a sufficient sample size at 80% power (Figure 13). Of these, 342 were surgical patients (slightly more than the minimum required, n = 314) and 4271 were non-surgical patients (much greater than the minimum required, n = 3142). Re-dislocations were observed in 912 patients (much greater than the minimum required, n = 695). Among the surgical group, 18% of patients suffered a re-dislocation, and 20% of patients suffered a re-dislocation in the non-surgical group. The difference in re-dislocation proportions between the surgical and non-surgical groups was only 2%, rather than the 5% difference used in the power calculation, which means that the power is reduced. However, overall, this study of surgical intervention up to 12 months on shoulder re-dislocations should have had sufficient power.

FIGURE 13.

Shoulder dislocation exclusions flow chart for patients aged 16–35 years during 1 April 1997–31 March 2016 within CPRD in England, with linkage to HES records.

Descriptive characteristics

A cohort of 4613 patients diagnosed with a first-time TASD during 1 April 1997–31 March 2015 was identified in the CPRD. The CPRD Read codes used to identify these patients are listed in Appendix 9. Of these patients, 342 underwent shoulder surgery within 6 months. The OPCS 4.7 surgical codes identified within HES are given in Appendix 10.

The descriptive characteristics of patients categorised by receiving surgery within 12 months or no surgery are described in Table 19. In the first 4 years, < 10 patients underwent surgery per year, but this increased from 2001 to 2010 to a maximum of 34 patients out of a total of 342 surgical patients. More men (82%) were diagnosed with a primary shoulder dislocation than women, and more men underwent surgery within 12 months (n = 302 men vs. n = 40 women). The majority of patients were aged 17–21 years at the time of their first-time TASD, but similar numbers were operated on among those aged 18–25 years.

Only 8% of patients were overweight or obese, with missing data on BMI for 42% of all patients. There was a pattern of decreasing numbers of patients by deprivation quintile, but this pattern was not evident for surgical patients. Only three patients did not have a recorded IMD score. Most patients were non-smokers (53%) and there were missing data on smoking for 9% of all patients. Most patients drank alcohol (47%) and there were missing data on alcohol consumption for 44% of all patients. Most patients had no comorbid conditions (93%). Many patients were from the North West (n = 743) and South Central regions (n = 669), and fewer were from the North East region (n = 113). A small proportion of patients had been diagnosed with epilepsy (3%), but a higher proportion of all surgical patients had been diagnosed with epilepsy (6%). A small proportion (6%) of patients were prescribed painkillers in the 3 months prior to their first-time TASD, which increased to 12% in the 1 month following a TASD. Of the patients prescribed painkillers, both 3 months prior to and 1 month after their TASD, only 1% were prescribed a different painkiller. Only 22 patients died during the study period.

Multiply imputed data analyses

Multiple imputation by chained equations was conducted for patients with missing data on BMI, smoking, drinking and IMD. Appendix 11 presents the HRs, 95% CIs and p-values for each of these variables using all available data, a complete-case analysis and the multiply imputed data. The HRs are similar for each group, indicating that the multiple imputation process was successful.

Cox survival estimates for complete cases and the multiply imputed data with the factors that had an impact on re-dislocations are presented in Appendix 12. Only 2100 cases had complete data on all factors. For these patients, a protective but non-significant adjusted effect of surgery was observed (HR 0.63, 95% CI 0.40 to 1.00; p = 0.052). A younger age at first-time TASD and an epilepsy diagnosis were significant risk factors for a re-dislocation.

Using the multiply imputed data resulted in similar HRs, 95% CIs and the same risk factors. The protective effect of surgery was less marked but was significant (HR 0.73, 95% CI 0.55 to 0.95; p = 0.022).

Propensity score analysis

Logistic regression was used to calculate the propensity scores. The distributions of propensity scores in the patients receiving surgery within 12 months and in those patients who had non-operative treatment are presented in Figure 14. Prior to matching, the propensity scores tended to be higher in the surgery group, but there is a substantial amount of overlap, indicating that it is possible to proceed with propensity score matching to estimate the treatment effect. Following matching, the propensity score density plots are very similar between the surgery and non-surgery patient groups. The propensity score matching will not include the non-surgical patients in the left-hand peak of the distribution.

FIGURE 14.

Propensity scores for first-time TASD patients diagnosed during 1 April 1997–31 March 2015 who had surgery within 12 months (black) or no surgery (green) within CPRD-HES, in England. (a) Prior to PS matching; and (b) following PS matching. PS, propensity score; SD, standard deviation.

Table 20 presents the baseline characteristics of first-time TASD patients who had surgery within 12 months or non-operative treatment. The SMDs, prior to propensity score matching, show that the two groups of patients do differ for most characteristics (SMD > 0.1). During propensity score matching, 295 surgical patients were each matched to 10 non-surgical patients (n = 2950). In addition to this, the time between the date of the first-time TASD and the date of surgery was allocated to the non-surgical patients for 233 surgical patients, making a total of 3183 non-surgical patients. The remaining 62 surgical patients had no time to be re-allocated because the date of their first-time TASD was the same date as their surgery. Following propensity score matching, the SMDs were much smaller than before (all < 0.1), which suggested that the balancing was successful.

Figure 15 presents Kaplan–Meier estimates of probability of survival in the surgical and non-surgical groups. The hazards were not proportional as shown by the two lines crossing at approximately 750 days and by the Schoenfeld’s Residuals Test (p = 0.003). We estimated time-varying hazards, which showed that there was no effect in the first year of follow-up (p = 0.458). This means that, initially, surgical patients had a similar rate of re-dislocations to non-surgical patients. Between 1 and 3 years of follow-up, the non-surgical group was more likely to re-dislocate (p = 0.022). Overall, there appears to be a small survival advantage for surgical patients, but this is not statistically significant.

FIGURE 15.

Kaplan–Meier survival estimates for first-time TASD patients diagnosed during 1 April 1997–31 March 2015 who had surgery within 12 months (green) or no surgery (black) within CPRD-HES, in England.

Table 21 presents the final results of the effect of surgery within 12 months compared with no surgery among first-time TASD patients. The median follow-up was less in the non-surgical group. The rate of shoulder re-dislocations was similar between the surgical and non-surgical groups (0.26 per 1000 person-years, 95% CI 0.20 to 0.33 per 100,000 person-years, compared with 0.26 per 1000 person-years, 95% CI 0.24 to 0.28 per 100,000 person-years, respectively). Overall, there was no difference between the effect of surgery and non-surgery within 12 months on shoulder re-dislocations (HR 1.17, 95% CI 0.88 to 1.55; p = 0.274).

To test for residual confounding, the patients were split into quintiles, based on the value of the propensity score, and HRs were produced for each quintile (see Appendix 13). The numbers of surgeries were equal in each quintile. HRs were found to differ between quintiles, suggesting that there was unmeasured confounding in the study consistent with the a priori risk factors not available in the CPRD or HES. Quintile 5 was found to have a twofold increase in the risk of dislocation after surgery (HR 2.07, 95% CI 1.23 to 3.46). Some key differences between the characteristics of patients within each stratum are shown in Appendix 14. Quintile 5, for instance, included more men, people who had a more recent first-time TASD, fewer alcohol drinkers and more people with epilepsy than the other quintiles.

Sample

The collated sample comprised patients aged 16–35 years with a first-time TASD who had at least 2 years of data (washout period) before a first-time entry Read code for shoulder dislocation and with up to 3 years of follow-up coding, who were registered at ‘active’ general practices in England (see Figure 11).

Definition of the primary outcome

The primary outcome was re-dislocation following a first-time TASD. We determined an entry date for each participant, which was the date of surgery for the surgical cohort and the date of first shoulder dislocation for the non-surgical cohort. Observation time was calculated from the entry date to an exit date, which was defined as the earliest date of recorded re-dislocation or 3 years after the index date.

Candidate predictors

Potential predictors of re-dislocation were defined a priori by expert consensus and informed by the validation study (see Appendix 4). Only eight of these predictors were available in the CPRD and were used as candidate predictors in the multivariable prediction model (Table 22). Owing to the sparseness of the CCI score in the surgical cohort, it was not considered in the model building in this cohort.

Continuous predictors

Fractional polynomials were used to explore the presence of non-linear relationships of continuous predictors (e.g. age, BMI); however, a linear relationship was found to be a good approximation. 34

Missing data

We assumed that missing data occurred at random and we carried out multiple imputation. 24 Missing values were predicted on the basis of all other predictors as well as the outcome. One hundred imputed data sets were generated with imputed values, reflecting the uncertainty associated with the imputations. Models were fitted on each imputed data set and coefficients combined using Rubin’s rules.

Model development

All candidate predictors in Table 22 were included in the multivariable Cox regression models for predicting re-dislocation. Because predictors were chosen a priori based on clinical consensus, and only a small number of candidate predictors were available in the CPRD, no reduction of predictors was considered.

Assessment of model performance and internal validation

The predictive ability of the model was assessed in terms of discrimination. 35 Discrimination is the ability of the model to differentiate between individuals who have a re-dislocation and those who do not. Discrimination was assessed by calculating the concordance (c)-index; a value of 0.5 indicates no discrimination (equivalent to tossing a coin) and a value of 1 indicates perfect discrimination.

Optimism in the performance was assessed by bootstrap resampling. 35 We drew 200 samples with replacement from the original data, with the same size as the original derivation data. In each bootstrap sample the entire modelling process was repeated. This process was repeated over each of the 100 imputed data sets, and an averaged, optimism-corrected c-index was taken.

The R software environment (version 3.5.0) was used for all analyses. We followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis statement for reporting our analyses. 36,37

Prediction model: surgical cohort

Eight predictors were included in the model to predict re-dislocation in the surgery cohort (Table 24). With an effective sample size of 61 re-dislocation events, this yields an events-per-variable (EPV) number of 5.1 (61 events/12 regression coefficients), which is much smaller than the widely recommended EPV number of 10, indicating the likelihood of overfitting because of a small sample size.

Age and epilepsy were the only statistically significant predictors (at the p < 0.05 level). The apparent predictive performance of the model, as measured by the c-index, was moderate, with a c-index of 0.72 (95% CI 0.65 to 0.80), which dropped slightly to 0.67 after correcting for optimism (because of overfitting).

Prediction model: non-surgical cohort

Nine predictors were included in the model to predict re-dislocation in the non-surgery cohort (Table 25). With an effective sample size of 851 re-dislocation events, this yields an EPV number of 56.7 (851 events/15 regression coefficients), which is higher than the widely recommended EPV number of 10, indicating a sufficient sample size for model development and the minimal likelihood of overfitting.

Age, sex and epilepsy were statistically significant predictors (at the p < 0.05 level). The apparent predictive performance of the model, as measured by the c-index, was low (c-index 0.58, 95% CI 0.56 to 0.60), which dropped to 0.56 after correcting for optimism (because of overfitting).

Main findings

The use of primary care data to predict the risk of re-dislocation following a first-time TASD is limited. However, the performances of the two models were markedly different, with better performance in the surgery cohort. Both models identified age and epilepsy as statistically significant predictors of re-dislocation, but in the non-surgery cohort sex was also identified as a statistically significant predictor.

The model developed in the surgery cohort showed moderate performance, with a c-index of 0.67, suggesting that the information collected has some predictive capacity but that additional information (risk factors) is needed to allow better predictions of re-dislocation. For the model developed in the non-surgery cohort, the predictive ability of the model was poor, with a c-index of only 0.57, and, therefore, has no use for the risk factors available in predicting re-dislocation in this cohort of patients.

Strengths and limitations

This study has several strengths. The models were developed using data routinely collected in electronic health-care records in primary care. Widely recommended statistical methodology was followed to develop and evaluate the models, including the exploration of complex relationships (e.g. non-linearity) in continuous measurements (e.g. age and BMI). Bootstrapping techniques were used to internally validate the models.

There are also some limitations. We were only able to include a small number of predictors, fewer than half of those identified in the clinical consensus. This had a considerable impact on our ability to develop models to accurately predict the risk of re-dislocation. There was also a considerable number of missing data (e.g. BMI, alcohol consumption). However, we followed recommended guidance and increased the number of imputations to 100 to account for the large number of missing data and any uncertainty in the imputation. 24 Despite using a large electronic health records database (CPRD linked to HES), the number of eligible individuals, notably in the surgical cohort, was surprisingly small, with only 61 re-dislocation events during the study period. A small sample size can lead to overfitting. However, to counter the risk of overfitting, internal validation was carried out using bootstrapping to obtain unbiased estimates of model performance. Another limitation was the lack of a separate data set to carry out external validation, particularly to evaluate the model in the surgery cohort, which showed some predictive accuracy in the internal validation.