Web appendix: Supplementary information OPCS coding (Office of Population, Censuses and Surveys Classification of Surgical Operations and Procedures) (4th revision) Procedure OPCS code/name Cholecystectomy J18.1 Total cholecystectomy and excision of surrounding tissue J18.2 Total cholecystectomy and exploration of common bile duct J18.3 Total cholecystectomy NEC J18.4 Partial cholecystectomy and exploration of common bile duct J18.5 Partial cholecystectomy NEC J18.8 Other specified excision of gall bladder J18.9 Unspecified excision of gall bladder Laparoscopic approach Y75.1 Laparoscopically assisted approach to abdominal cavity Y75.2 Laparoscopic approach to abdominal cavity NEC Y75.3 Robotic minimal access approach to abdominal cavity Y75.4 Hand assisted minimal access approach to abdominal cavity Y75.5 Laparoscopic ultrasonic approach to abdominal cavity Y75.8 Other specified minimal access to abdominal cavity Y75.9 Unspecified minimal access to abdominal cavity OPCS4.2 (until 142006) Y50.8 Other Specified Approach Through Abdominal Cavity Open conversion Y71.4* Failed minimal access approach converted to open Y71.8* Other specified * inconsistently applied in earlier years and not used in analysis Operative cholangiography J37.2 Operative cholangiography through cystic duct J37.3 Direct puncture operative cholangiography Includes: Operative cholangiography NEC ERCP J38.x Endoscopic incision of sphincter of Oddi J39.x Other therapeutic endoscopic operations on ampulla of Vater J40.x Endoscopic retrograde placement of prosthesis in bile duct J41.x Other therapeutic endoscopic retrograde operations on bile duct J42.x Therapeutic endoscopic retrograde operations on pancreatic duct J43.x Diagnostic endoscopic retrograde examination of bile duct and pancreatic duct J44.x Diagnostic endoscopic retrograde examination of bile duct J45.x Diagnostic endoscopic retrograde examination of pancreatic duct Bile duct exploration J18.2 Total cholecystectomy and exploration of common bile duct J18.4 Partial cholecystectomy and exploration of common bile duct In combination with cholecystectomy code: J31.x Open introduction of prosthesis into bile duct (ALL) J33.x Incision of bile duct (ALL) J35.x Incision of sphincter of Oddi using duodenal approach (ALL) 1
ICD coding of excluded patients Diagnosis name ICD10 code N Carcinoma in situ of other and unspecified digestive organs D015 1 Diffuse nonhodgkin's lymphoma C833 1 Leiomyoma of uterus D259 1 Lymphoid leukaemia C911 1 Malignant neoplasm of breast C504 6 Malignant neoplasm of bronchus and lung C349 5 Malignant neoplasm of cervix uteri C539 1 Malignant neoplasm of colon C189 11 Malignant neoplasm of gallbladder C23X 99 Malignant neoplasm of kidney, except renal pelvis C64X 5 Malignant neoplasm of liver and intrahepatic bile ducts C229 17 Malignant neoplasm of oesophagus C159 3 Malignant neoplasm of other and illdefined digestive organs C269 1 Malignant neoplasm of pancreas C250 38 Malignant neoplasm of rectosigmoid junction C19X 2 Malignant neoplasm of rectum C20X 2 Malignant neoplasm of small intestine C170 3 Malignant neoplasm of stomach C169 11 Malignant neoplasm without specification of site C80X 6 Maligt neoplasm of other and unspec parts biliary tract C249 7 Neoplasm of uncertain/unknown behaviour of endocrine glands D441 1 Neoplasm uncert or unkn behaviour oral cav and diges organs D377 6 Neuromuscular dysfunction of bladder NEC N318 1 Open wound of abdomen, lower back and pelvis S318 2 Other and unspecified types of nonhodgkin's lymphoma C851 8 Other malignant neoplasms of skin C444 2 Sec malignant neoplasm of respiratory and digestive organs C787 18 Secondary and unspecified malignant neoplasm of lymph nodes C772 1 Secondary malignant neoplasm of other sites C796 3 Injury of intraabdominal organs S3610 12 Total 275 2
Note on the exclusion of patients treated in private hospitals The patients in the dataset from private hospitals were exclusively from NHS waitinglist initiatives. We did not have access to data of the standard patients from these hospitals. These numbers therefore do not reflect total hospital volume and so we had no choice but to exclude them. Definition of categorical volume groups The definition of categorical volume groups was considered at length prior to starting the study. Firstly and most importantly, hospital volume is considered as a continuous variable in all models, i.e. mean annual hospital procedure counts are used without conversion to a threelevel categorical variable. In the mortality and length of stay models, there is a significant relationship between the continuous hospital volume variable and the outcome variable. Reoperation and readmission are nonlinear and the continuous hospital volume variable does not correlate particularly well when considered in a linear model. The continuous variable is difficult to report as it is not intuitive being in logunits. To provide an easily interpretable result (at the cost of throwing away some information), we report hospital volume as a categorical variable with threelevels (low, medium and high volume). We have used a method whereby the entire cohort is equally divided into three groups. While this does afford maximum statistical, our prime motivation for using this method was to avoid accusations of arbitrarily setting cutoff points to maximise volume group differences. In general, other methods require dataanalysis to be performed and cutoff points defined on the basis of that analysis. Choice of morbidity score Different models of comorbidity were explored all models were run with both Charlson and Elixhauser scores. The choice made almost no difference to model performance. The 2 loglikelihoods are presented below (the lower the value, the better the goodness of fit of the model). As can be seen in the table below, there is at most a 0.3% difference between Charlson and Elixhauser for all outcome variables. Charlson Elixhauser mortality 2612.4 2622.9 reoperation 18276 18271 readmission 32692 32658 length of stay 1178428 1178259 The table below shows how morbidity model choice does not significantly alter parameter estimates and standard errors for mortality. Given all this, we have used the simpler method (Charlson). Hosp vol* Charlson Elixhauser mortality Low 1.512 (1.1092.086) 1.487 (1.0912.050) Medium 1.487 (1.0912.050) 1.434 (1.0471.984) *compared with high volume group 3
Simulation procedure and R script. A simulation procedure has been used to model absolute risk differences for patients of varying baseline risk. Rather than present the mathematical basis of this (which can be found in the references) we have included what we hope is a lay explanation. Consider a hypothetical survey of adult height. An investigator randomly samples a population and determines a sample mean (a quantity of interest) for adult height. The greater the number in the sample, the closer the sample mean is to that of the population and the more information we have about the probability distribution of height in that population. Now consider a logistic regression model of mortality after cholecystectomy against a number of explanatory variables (say age, sex and deprivation score). In the normal manner, coefficients (β) are estimated for explanatory variables together with a standard error reflecting the uncertainty of that point estimate. By setting explanatory variables at a chosen level (e.g. age = 50, sex = female, SIMD = 4), the absolute probability of death (a quantity of interest) can be determined from the model function. In a manner akin to the population sampling example above, the simulation procedure involves taking random draws from the probability distributions underlying the model coefficients. This is repeated many times (we have used 1000 000) and these simulated parameter estimates used to generate 1000 0000 estimates of the quantity of interest, in this example the absolute probability of death. The mean (expected value) of these estimates can be presented, together with SE/95% CI for the simulated distribution. 27 R is an opensource statistical environment which can be downloaded here http://www.rproject.org/. Below is a script which generates sample data and demonstrates a simple example of an absolute risk difference analysis. #### Example using randomly generated variables #### ## Define variables ## #Randomly define two sets of hospitals with 10000 cases in total hosp.group<factor(rbinom(10000,1,0.5)) #Define mortality. 5% in hosp.group = 1, 3% in hosp.group = 0 mortality<ifelse(hosp.group==1,rbinom(sum(hosp.group==1),1,0.05),rbinom(sum(hosp.group==0),1, 0.03)) #Age: random normal distribution with mean 50 SD 20. age<rnorm(10000,50,20) #Morbidity score. Randomly assigned. simd<rbinom(10000,4,0.5) #Bring variables together in dataframe data<data.frame(mortality, age, hosp.group, simd) ##Use package Zelig to specify logit model (install in normal manner) #install.packages("zelig") library(zelig) z.mortality<zelig(mortality ~ age + hosp.group + simd, model="logit", data=data) summary(z.mortality) #What follows is a standard logit model #The absolute numbers will differ depending on the randomly generated variables Call: zelig(formula = mortality ~ age + hosp.group + simd, model = "logit", data = data) Deviance Residuals: Min 1Q Median 3Q Max 0.4176 0.3309 0.2896 0.2610 2.7035 4
Coefficients: Estimate Std. Error z value Pr(> z ) (Intercept) 3.74543 0.17770 21.077 < 2e16 *** age 0.00477 0.00243 1.963 0.0496 * hosp.group1 0.47142 0.09956 4.735 2.19e06 *** simd 0.08433 0.04853 1.737 0.0823. Signif. codes: 0 *** 0.001 ** 0.01 * 0.05. 0.1 1 (Dispersion parameter for binomial family taken to be 1) Null deviance: 3627.5 on 9999 degrees of freedom Residual deviance: 3597.8 on 9996 degrees of freedom AIC: 3605.8 Number of Fisher Scoring iterations: 6 #Model coefficients expressed as odds ratios exp(z.mortality$coefficients) > exp(z.mortality$coefficients) (Intercept) age hosp.group1 simd 0.02362553 1.00478153 1.60226195 1.08798506 #Want to find the absolute risk difference for mortality (first difference in expected values) between hospital groups, holding age and simd at mean #Set covariates at chosen level x.mortality_0 <setx(z.mortality, hosp.group=0) x.mortality_1 <setx(z.mortality, hosp.group=1) #Simulate quantity of interest s.mortality<sim(z.mortality, x=x.mortality_0, x1=x.mortality_1) summary(s.mortality) Model: logit Number of simulations: 1000 Values of X (Intercept) age hosp.group1 simd 1 1 50.02982 0 1.9867 Values of X1 (Intercept) age hosp.group1 simd 1 1 50.02982 1 1.9867 Expected Values: E(Y X) #expected values for hosp.group=1 mean sd 2.5% 97.5% 1 0.03448615 0.002514324 0.02981267 0.03978032 Predicted Values: Y X 0 1 1 0.978 0.022 First Differences in Expected Values: E(Y X1)E(Y X) #difference in mortality about 2% as would be expected given the variable definitions mean sd 2.5% 97.5% 1 0.01936902 0.004039653 0.01124012 0.02697273 Risk Ratios: P(Y=1 X1)/P(Y=1 X) mean sd 2.5% 97.5% 1 1.569786 0.1468423 1.299199 1.872838 5
Upper panel, total cholecystectomy count by volume group 19982007. Lower panel, proportion of cholecystectomies completed laparoscopically by admission type and hospital volume group. 6
Additional multivariable models Mortality (standard logit model) Reoperation (fixedeffects model) Readmission (fixedeffects model) Odds ratio (95% CI) p value Odds ratio (95% CI) p value Odds ratio (95% CI) p value Length of stay (generalised linear model)* Coefficient (logscale; 95% CI) p value Operation year 0.93 (0.890.97) 1.03 (1.011.04) 1.01 (1.001.02) 0.0720 0.039 (0.0420.037) Age, y <40 4054 5569 70 Gender Male Female Admission type Elective Nonelective Diagnosis Cholelithiasis Cholecystitis Acute pancreatitis Other 3.50 (1.329.32) 10.74 (4.3226.70) 38.86 (15.8595.28) 1.46 (1.151.87) 0.012 5.64 (4.267.47) 2.07 (1.522.83) 1.61 (0.823.15) 8.55 (6.3411.54) 0.92 (0.811.04) 1.20 (1.061.35) 1.53 (1.331.75) 0.002 1.34 (1.221.47) 0.166 0.175 0.003 2.60 (2.382.85) 0.84 (0.740.94) 1.30 (1.011.66) 1.31 (1.111.55) 0.92 (0.851.00) 0.92 (0.841.00) 1.09 (0.991.21) 1.22 (1.141.31) 0.002 0.041 0.002 0.046 0.038 0.070 1.69 (1.581.81) 1.00 (0.931.08) 1.06 (0.851.31) 1.27 (1.121.43) 0.165 (0.1470.183) 0.419 (0.4010.437) 0.964 (0.9420.987) 0.184 (0.1680.200) 0.098 0.613 0.762 (0.7450.778) 0.084 (0.0670.101) 0.156 (0.1020.210) 0.336 (0.3050.367) Deprivation 0.82 (0.750.91) 0.96 (1.041.17) 0.001 0.96 (0.940.98) 0.039 (0.0420.037) Morbidity 1.39 (1.261.54) 1.10 (1.1041.17) 0.001 1.27 (1.22 1.32) 0.144 (0.1320.155) Hospital volume High Medium Low 1.45 (1.062.00) 1.52 (1.112.08) 0.022 0.010 1.77 (1.591.97) 1.24 (1.101.39) 1.17 (1.091.26) 1.10 (1.021.19) 0.015 0.254 (0.2380.271)) 0.227 (0.2100.243) * Generalised linear model of length of stay following cholecystectomy. The LOS data were first fitted to different distributions with varying link functions using generalised linear modelling: linear (with identity or log link functions), gamma (with log link), Poisson (with log link) and negative binomial (with log link). Model performance was assessed using likelihood ratio tests and the Akaike information criterion. The gamma distribution with loglink function fitted the data best and was used in the final analysis. As errors in estimates can occur if backexponentiation is performed, the results are left on the logscale. 7
Subset analysis of outcomes after laparoscopic cholecystectomy, excluding primary open and open conversion procedures (N=43 407) Mortality (relogit model) Reoperation (mixed effects multilevel model) Readmission (mixed effects multilevel model) Odds ratio (95% CI) p value Odds ratio (95% CI) p value Odds ratio (95% CI) p value Length of stay (Cox proportional hazards) Hazard ratio for discharge (95% CI) p value Operation year 0.92 (0.841.00) 0.056 1.03 (1.011.05) 0.004 1.02 (1.011.03) 0.005 1.04 (1.031.04) Age, y <40 4054 5569 70 Gender Male Female Admission type Elective Nonelective Diagnosis Cholelithiasis Cholecystitis Acute pancreatitis Other 2.26 (0.7211.49) 6.51 (2,6032.02) 23.29 (9.63114.83) 2.03 (1.253.28) 0.229 0.002 4.28 (2.637.04) 1.30 (0.662.30) 7.49 (4.0013.25) 0.83 (0.710.97) 1.00 (0.861.16) 1.21 (1.001.46) 0.004 1.31 (1.151.50) 0.401 0.018 0.963 0.052 2.39 (2.092.72) 0.85 (0.721.01) 1.25 (0.891.74) 1.49 (1.151.94) 0.058 0.196 0.003 0.90 (0.821.00) 0.88 (0.800.97) 1.04 (0.911.18) 1.19 (1.091.30) 0.042 0.014 0.597 1.72 (1.581.87) 1.02 (0.921.13) 1.00 (0.771.30) 1.30 (1.091.56) 0.724 0.988 0.004 0.91 (0.890.94) 0.75 0.730.76) 0.50 (0.480.51) 0.93 (0.910.96) 0.48 (0.470.50) 0.94 (0.920.97) 0.91 (0.840.98) 0.94 (0.900.99) Deprivation 0.72 (0.580.86) 0.001 0.96 (0.921.00) 0.053 0.94 (0.910.96) 1.04 (1.041.05) Morbidity 1.31 (1.011.56) 0.016 1.14 (1.041.24) 0.004 1.30 (1.231.37) 0.89 (0.870.90) Hospital volume High Medium Low 2.14 (1.184.18) 1.88 (1.003.75) 0.018 0.060 1.94 (1.542.43) 1.28 (1.021.61) 0.036 1.23 (1.131.35) 1.14 (1.041.26) 0.008 0.72 (0.700.74) 0.79 (0.770.80) 0.013 0.027 8
Logistic regression models for probability of completing cholecystectomy using an open vs. laparoscopic approach, and of performing an operative cholangiogram Odds of cholecystectomy completed open vs. laparosopic (fixedeffects logit model) Odds of operative cholangiogram being performed (fixedeffects logit model) Odds ratio (95% CI) p value Odds ratio (95% CI) p value Operation year 1.09 (1.081.10) 0.94 (0.930.95) Age, y <40 4054 5569 70 Gender Male Female Admission type Elective Nonelective Diagnosis Cholelithiasis Cholecystitis Acute pancreatitis Other 1.20 (1.141.27) 1.57 (1.491.65) 2.31 (2.162.46) 1.44 (1.421.56) 1.56 (1.491.63) 1.49 (1.421.56) 0.72 (0.620.85) 2.95 (2.723.20) 1.04 (0.971.11) 1.15 (1.071.23) 1.19 (1.091.29) 1.07 (1.011.14) 0.285 0.022 1.65 (1.561.75) 0.56 (0.520.61) 2.30 (1.992.67) 0.74 (0.660.83) Deprivation 0.93 (0.920.94) 0.98 (0.960.99) 0.013 Morbidity 1.14 (1.101.17) 1.02 (0.981.06) 0.385 Hospital volume High Medium Low 1.72 (1.631.81) 2.86 (2.723.00) 0.27 (0.250.29) 0.50 (0.470.53) 9