This chapter begins the main analysis of the commissioned research (work package 2). The previous chapters described the internal and external validation of the data to be used in this and the following chapters.
Objectives of Chapter 4
To study the effect of surgical treatment on the 2-year recurrent shoulder dislocation rates in young adults in England when surgical treatment took place within 6 months of the first episode of TASD.
Methods
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. (Figure not reproduced here; see the original article.)
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.
Surgical group
The surgical group comprised patients in CPRD with a first-time TASD who underwent shoulder stabilisation surgery after their first dislocation. Early surgical repair in this NHS context means ‘a decision to treat surgically after the first TASD’ (as per the approved study protocol) and receiving surgery within 6 months of injury before any subsequent dislocations. This meant linking HES data to CPRD data in such a way that a HES surgical OPCS 4.7 code was seen to occur after a single first dislocation code in CPRD before that surgical date. The timelines between first dislocation codes and OPCS 4.7 codes were recorded. If a re-dislocation occurred prior to their surgery date, the patient was allocated to the non-surgical arm.
A further 6-week washout period for re-dislocation codes within CPRD was applied to all patients in the surgical group following their surgery date to avoid duplicate records because they would have been asked to immobilise their arm for this period and, thus, a re-dislocation would be highly unlikely to occur. During these 6 weeks, patients would most probably return to visit their GPs for painkillers or referral to physiotherapy. Surgery patients were followed up for at least 2 years from the date of their surgery. Patients who had surgery more than 6 months following their TASD were censored on their date of surgery. The surgery dates for these patients ranged from 25 December 1997 to 18 August 2014.
Non-surgical group
Although the most desirable control group would have been physiotherapy, the internal validation study identified that the referral codes for physiotherapy are lacking and unreliable in CPRD. Thus, conservative care has been defined as ‘non-surgical intervention’, with no linked OPCS 4.7 surgical shoulder codes, producing a control cohort of patients whose first-time shoulder dislocation has been treated non-operatively. Non-surgical patients were followed up for at least 2 years following the date of their first-time TASD, as illustrated in Figure 8.
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
Inclusion criteria
The CPRD records of 6046 patients, aged 16–35 years with 2 years of data in the CPRD before a first-time TASD, identified in the internal validation study (see Table 1) were linked to HES records.
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).
| 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. (Figure not reproduced here; see the original article.)
Statistical analysis
The aim of this study was to investigate the effect of surgical intervention within 6 months compared with non-surgical intervention on the rates of re-dislocation in young patients with a first-time TASD over a 2-year period. In addressing this research question, the exposure is whether or not a patient received surgery, and the primary outcome of interest is the time from date of surgery for surgical patients or date of the first-time TASD for non-surgical patients to having a subsequent re-dislocation within 2 years.
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
Results
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.
| Characteristic | Patient group, n (%) | ||
|---|---|---|---|
| Whole data set | No surgery | Surgery within 6 months of TASD | |
| Total | 3759 (100) | 3603 (96) | 156 (4) |
| Year of shoulder dislocation | |||
| 1997 | 62 (2) | 61 (2) | 1 (1) |
| 1998 | 106 (3) | 102 (3) | 4 (3) |
| 1999 | 120 (3) | 118 (3) | 2 (1) |
| 2000 | 143 (4) | 139 (4) | 4 (3) |
| 2001 | 192 (5) | 185 (5) | 7 (4) |
| 2002 | 226 (6) | 221 (6) | 5 (3) |
| 2003 | 270 (7) | 262 (7) | 8 (5) |
| 2004 | 276 (7) | 269 (7) | 7 (4) |
| 2005 | 258 (7) | 252 (7) | 6 (4) |
| 2006 | 280 (7) | 270 (7) | 10 (6) |
| 2007 | 303 (8) | 290 (8) | 13 (8) |
| 2008 | 297 (8) | 279 (8) | 18 (12) |
| 2009 | 275 (7) | 261 (7) | 14 (9) |
| 2010 | 294 (8) | 279 (8) | 5 (10) |
| 2011 | 265 (7) | 249 (7) | 16 (10) |
| 2012 | 212 (6) | 197 (5) | 15 (10) |
| 2013 | 159 (4) | 149 (4) | 10 (6) |
| 2014 | 21 (1) | 20 (1) | 1 (1) |
| Sex | |||
| Male | 3115 (83) | 2975 (83) | 140 (90) |
| Female | 644 (17) | 628 (17) | 16 (10) |
| Age at shoulder dislocation (years) | |||
| 16 | 204 (5) | 195 (5) | 9 (6) |
| 17 | 257 (7) | 251 (7) | 6 (4) |
| 18 | 241 (6) | 230 (6) | 11 (7) |
| 19 | 266 (7) | 255 (7) | 11 (7) |
| 20 | 254 (7) | 241 (7) | 13 (8) |
| 21 | 276 (7) | 266 (7) | 10 (6) |
| 22 | 242 (6) | 225 (6) | 17 (11) |
| 23 | 212 (6) | 203 (6) | 9 (6) |
| 24 | 222 (6) | 214 (6) | 8 (5) |
| 25 | 198 (5) | 187 (5) | 11 (7) |
| 26 | 180 (5) | 173 (5) | 7 (4) |
| 27 | 156 (4) | 148 (4) | 8 (5) |
| 28 | 128 (3) | 125 (3) | 3 (2) |
| 29 | 145 (4) | 138 (4) | 7 (4) |
| 30 | 147 (4) | 141 (4) | 6 (4) |
| 31 | 109 (3) | 105 (3) | 4 (3) |
| 32 | 137 (4) | 133 (4) | 4 (3) |
| 33 | 117 (3) | 113 (3) | 4 (3) |
| 34 | 131 (3) | 127 (4) | 4 (3) |
| 35 | 137 (4) | 133 (4) | 4 (3) |
| BMI (kg/m2) | |||
| < 25 | 1285 (34) | 1233 (34) | 52 (33) |
| 25.0–29.9 | 646 (17) | 617 (17) | 29 (19) |
| ≥ 30 | 309 (8) | 296 (8) | 13 (8) |
| Missing | 1519 (40) | 1457 (40) | 62 (40) |
| IMD 2004 (quintile of deprivation) | |||
| 1 (affluent) | 987 (26) | 959 (27) | 28 (18) |
| 2 | 840 (22) | 798 (22) | 42 (27) |
| 3 | 746 (20) | 723 (20) | 23 (15) |
| 4 | 691 (18) | 656 (18) | 35 (22) |
| 5 (deprived) | 492 (13) | 464 (13) | 28 (18) |
| Missing | 3 (< 1) | 3 (0) | 0 (0) |
| Smoking status | |||
| No | 2029 (54) | 1951 (54) | 78 (50) |
| Yes | 1137 (30) | 1086 (30) | 51 (33) |
| Ex-smoker | 314 (8) | 301 (8) | 13 (8) |
| Missing | 279 (7) | 265 (7) | 14 (9) |
| Drinking status | |||
| Yes | 1847 (49) | 1782 (49) | 65 (42) |
| No | 356 (9) | 340 (9) | 16 (10) |
| Missing | 1556 (41) | 1481 (41) | 75 (48) |
| CCI score | |||
| 0 | 3484 (93) | 3338 (93) | 146 (94) |
| 1 | 172 (5) | 166 (5) | 6 (4) |
| 2 | 66 (2) | 64 (2) | 2 (1) |
| ≥ 3 | 37 (1) | 35 (1) | 2 (1) |
| Region | |||
| East Midlands | 114 (3) | 108 (3) | 6 (4) |
| East of England | 444 (12) | 430 (12) | 14 (9) |
| London | 464 (12) | 445 (12) | 19 (12) |
| North East | 96 (3) | 89 (2) | 7 (4) |
| North West | 636 (17) | 603 (17) | 33 (21) |
| South Central | 543 (14) | 521 (14) | 22 (14) |
| South East Coast | 430 (11) | 420 (12) | 10 (6) |
| South West | 454 (12) | 430 (12) | 24 (15) |
| West Midlands | 441 (12) | 424 (12) | 17 (11) |
| Yorkshire and the Humber | 137 (4) | 133 (4) | 4 (3) |
| Epilepsy | 130 (3) | 118 (3) | 12 (8) |
| Painkiller prescriptions | |||
| Prescribed 3 months prior to TASD | 210 (6) | 198 (5) | 12 (8) |
| Prescribed 1 month following TASD | 446 (12) | 420 (12) | 26 (17) |
| Mortality | 22 (1) | 21 (< 1) | 1 (1) |
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. (Figure not reproduced here; see the original article.)
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.
| Characteristic | All eligible primary shoulder dislocation patients (N = 3784)a | Matched analysis (3 : 1): patients matched on PS (N = 726)a,b | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Surgery (n = 156) | Non-surgery (n = 3603) | SMD | Surgery (n = 156) | Non-surgery (n = 570) | SMD | |||||
| n (%) | Re-dislocation, n (%) | n (%) | Re-dislocation, n (%) | n (%) | Re-dislocation, n (%) | n (%) | Re-dislocation, n (%) | |||
| Re-dislocation | 156 (100) | 31 (20) | 3603 (100) | 731 (20) | 156 (100) | 31 (20) | 570 (100) | 102 (18) | ||
| Calendar year of shoulder dislocation | 156 (100) | 31 (20) | 3603 (100) | 731 (20) | 0.344 | 156 (100) | 31 (20) | 570 (100) | 102 (18) | –0.046 |
| Sex | ||||||||||
| Male | 140 (90) | 26 (19) | 2975 (83) | 618 (21) | –0.209 | 140 (90) | 26 (19) | 500 (88) | 89 (18) | –0.064 |
| Female | 16 (10) | 5 (31) | 628 (17) | 113 (18) | 16 (10) | 5 (31) | 70 (12) | 13 (19) | ||
| Age at shoulder dislocation (16- to 35-year-olds) | 156 (100) | 31 (20) | 3603 (100) | 731 (20) | –0.067 | 156 (100) | 31 (20) | 570 (100) | 102 (18) | 0.016 |
| BMI (kg/m2) | ||||||||||
| < 25 | 86 (55) | 20 (23) | 2159 (60) | 461 (21) | 0.114 | 86 (55) | 20 (23) | 320 (56) | 58 (18) | 0.030 |
| 25.0–29.9 | 44 (28) | 6 (14) | 972 (27) | 170 (17) | 44 (28) | 6 (14) | 162 (28) | 23 (14) | ||
| ≥ 30 | 26 (17) | 5 (19) | 472 (13) | 100 (21) | 26 (17) | 5 (19) | 88 (15) | 21 (24) | ||
| IMD 2004 (quintile of deprivation) | ||||||||||
| 1 (affluent) | 28 (18) | 2 (7) | 961 (27) | 189 (20) | 0.196 | 28 (18) | 2 (7) | 91 (16) | 18 (20) | –0.011 |
| 2 | 42 (27) | 7 (17) | 799 (22) | 168 (21) | 42 (27) | 7 (17) | 169 (30) | 33 (20) | ||
| 3 | 23 (15) | 8 (35) | 723 (20) | 139 (19) | 23 (15) | 8 (35) | 78 (14) | 8 (10) | ||
| 4 | 35 (22) | 8 (23) | 656 (18) | 133 (20) | 35 (22) | 8 (23) | 130 (23) | 21 (16) | ||
| 5 (deprived) | 28 (18) | 6 (21) | 464 (13) | 102 (22) | 28 (18) | 6 (21) | 102 (18) | 22 (22) | ||
| Smoking status | ||||||||||
| No | 85 (54) | 19 (22) | 2090 (58) | 446 (21) | 0.051 | 85 (54) | 19 (22) | 324 (57) | 59 (18) | 0.056 |
| Yes | 57 (37) | 10 (18) | 1184 (33) | 222 (19) | 57 (37) | 10 (18) | 202 (35) | 36 (18) | ||
| Ex-smoker | 14 (9) | 2 (14) | 329 (9) | 63 (19) | 14 (9) | 2 (14) | 44 (8) | 7 (16) | ||
| Drinking status | ||||||||||
| Yes | 128 (82) | 23 (18) | 2985 (83) | 597 (20) | 0.021 | 128 (82) | 23 (18) | 479 (84) | 80 (17) | 0.053 |
| No | 28 (18) | 8 (29) | 618 (17) | 134 (22) | 28 (18) | 8 (29) | 91 (16) | 22 (24) | ||
| CCI score | ||||||||||
| 0 | 146 (94) | 30 (21) | 3338 (93) | 682 (20) | –0.019 | 146 (94) | 30 (21) | 535 (94) | 96 (18) | 0.010 |
| 1 | 6 (4) | 0 (0) | 166 (5) | 31 (19) | 6 (4) | 0 (0) | 23 (4) | 5 (22) | ||
| 2 | 2 (1) | 0 (0) | 64 (2) | 11 (17) | 2 (1) | 0 (0) | 3 (1) | 0 (0) | ||
| ≥ 3 | 2 (1) | 1 (50) | 35 (1) | 7 (20) | 2 (1) | 1 (50) | 9 (2) | 1 (11) | ||
| Region | ||||||||||
| East Midlands | 6 (4) | 2 (33) | 108 (3) | 27 (25) | –0.024 | 6 (4) | 2 (33) | 24 (4) | 6 (25) | –0.003 |
| East of England | 14 (9) | 2 (14) | 430 (12) | 84 (20) | 14 (9) | 2 (14) | 44 (8) | 11 (25) | ||
| London | 19 (12) | 3 (16) | 445 (12) | 85 (19) | 19 (12) | 3 (16) | 76 (13) | 10 (13) | ||
| North East | 7 (4) | 0 (0) | 89 (2) | 24 (27) | 7 (4) | 0 (0) | 24 (4) | 5 (21) | ||
| North West | 33 (21) | 7 (21) | 603 (17) | 115 (19) | 33 (21) | 7 (21) | 126 (22) | 21 (17) | ||
| South Central | 22 (14) | 2 (9) | 521 (14) | 104 (20) | 22 (14) | 2 (9) | 74 (13) | 14 (19) | ||
| South East Coast | 10 (6) | 5 (50) | 420 (12) | 73 (17) | 10 (6) | 5 (50) | 42 (7) | 4 (10) | ||
| South West | 24 (15) | 7 (29) | 430 (12) | 93 (22) | 24 (15) | 7 (29) | 74 (13) | 10 (14) | ||
| West Midlands | 17 (11) | 3 (18) | 424 (12) | 100 (24) | 17 (11) | 3 (18) | 70 (12) | 18 (26) | ||
| Yorkshire and the Humber | 4 (3) | 0 (0) | 133 (4) | 26 (20) | 4 (3) | 0 (0) | 16 (3) | 3 (19) | ||
| Epilepsy | 12 (8) | 4 (33) | 118 (3) | 38 (32) | 0.195 | 12 (8) | 4 (33) | 42 (7) | 9 (21) | 0.012 |
| Painkiller prescriptions | ||||||||||
| 3 months prior to TASD | 12 (8) | 3 (25) | 198 (5) | 45 (23) | 0.088 | 12 (8) | 3 (25) | 48 (8) | 10 (21) | –0.027 |
| 1 month following TASD | 26 (17) | 7 (27) | 420 (12) | 75 (18) | 0.144 | 26 (17) | 7 (27) | 92 (16) | 18 (20) | 0.014 |
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. (Figure not reproduced here; see the original article.)
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.
| N | Patient group | HR (95% CI); p-value | HRa (95% CI); p-value | |||||
|---|---|---|---|---|---|---|---|---|
| Surgery | Non-surgery | |||||||
| Events (n) | Median follow-up (days) (IQR) | Rate (per 1000 person-years) (IQR) | Events | Median follow-up (days) (IQR) | Rate (per 1000 person-years) (IQR) | |||
| Shoulder re-dislocations | 31 | 731 (731–815) | 0.30 (0.21–0.43) | 102 | 731 (495–731) | 0.36 (0.29–0.43) | 0.84 (0.55 to 1.29); p = 0.429 | 0.88 (0.58 to 1.35); p = 0.565 |
Discussion
A population-based cohort of 3759 patients diagnosed with a first-time TASD during 1 April 1997–26 April 2014, with CPRD-HES linked records and 2 years of follow-up in England were identified. In total, only 156 patients in this data set received surgical treatment within 6 months of first-time TASD. Thus, despite the commissioned question, our findings conclude that early surgery after only one shoulder dislocation is uncommon in the NHS.
The overall finding from the propensity-score-matched analysis was that although surgery within 6 months appeared to be slightly protective, it was not a statistically significant deterrent for re-dislocations (HR 0.88, 95% CI 0.58 to 1.35; p = 0.565). The wide CI indicates that this study was underpowered. Owing to the unexpected small numbers receiving surgery after only one dislocation and the subsequent loss of study power, we are not able to confirm whether or not surgery within 6 months of a first-time anterior shoulder dislocation has any additional benefit on whether or not a patient suffers a re-dislocation.
This is the first time a large, primary care, national, observational data set has been used to examine the role of surgery on treating shoulder dislocations in England. The main strength of this study is that it uses a population-based cohort using real-world data from linked primary and secondary care databases and these databases are representative of the English population with respect to age and sex. The study also uses the latest statistical methods to account for missing data, confounding by indication and immortal time bias. There was a considerable number of missing data for BMI (40%) and for alcohol consumption (41%); fewer data were missing for smoking (7%) and IMD (n = 3). Data were successfully imputed for these covariates.
Although the number of patients (n = 3759) included in the analysis was greater than the minimum number required for statistical power (n = 3065), disappointingly, the main weakness of the study was the unexpected low number of NHS patients having surgery after one dislocation. In total, only 156 patients had undergone surgical treatment within 6 months of the date of their first-time TASD. The lack of surgical patients in the cohort resulted in an overall lack of statistical power. It was also observed that a substantial number of patients in this cohort of 16- to 35-year-olds had < 2 years of follow-up within the CPRD (n = 854), which may be related to moving location and changing GPs because of going to university or finding jobs.
The main disadvantage of using propensity score matching methods is that confounders for which no data are available result in a lack of adjustment. Other than age and sex, the risk factors recorded and available in the CPRD were considered less important risk factors. Many factors considered important by surgeons, including original cause/mechanism of shoulder dislocation, imaging findings of structural problems, anterior apprehension, occupation, sports played and level of sports, were not recorded in the observational data. This links with the findings later in Chapter 6, Prediction models. Finally, in this cohort, 20% of patients in the surgery and non-surgery group had suffered a shoulder re-dislocation. Responses to the GP questionnaire validation study (reported in Chapter 2) indicate that about one-third of patients suffered a re-dislocation within that CPRD cohort. It is possible that either shoulder re-dislocations have not been reported to GPs or the re-dislocations were not coded in the general practice’s computer system.
Conclusions
Overall, relatively few patients have surgery within 6 months of a first-time TASD in the NHS. This is probably a reflection of many GPs not referring patients with only one dislocation to secondary care and also because of NHS operative waiting times.
This study was underpowered, lacked sufficient follow-up data on many patients and did not include data on many of the important risk factors used by surgeons to make clinical decisions on the best care.
Based on these findings, and in an attempt to maximise the use of this data set to further examine the commissioned question of surgery after first-time shoulder dislocation, a further sensitivity analysis was planned and approved by the HTA programme and Independent Scientific Advisory Committee (ISAC) and is discussed in the following chapter.