Overall survival comparing laparoscopic to open surgery for right‐sided colon cancer: propensity score inverse probability weighting population study

This retrospective cohort study reports on overall survival and short‐term complications, comparing laparoscopic to open resection for right‐sided colon cancers. It is one of the largest studies in the field with generalizable population‐level results.


Introduction Background
Colorectal cancer is a leading cause of cancer death in developed countries and surgical resection is the mainstay of treatment. 1 Laparoscopic surgery is a well-established technique for the resection of colonic cancers. The first series of laparoscopic colorectal surgery was published in 1991. 2 Laparoscopic resection for colon cancer has short-term advantages of decreased blood loss, decreased postoperative length of stay, quicker return of gut function and better quality of life scores 3-5 but preliminary randomized controlled trials did not show a difference in overall survival between the two modalities. [6][7][8][9] The rationale for re-examining this topic is that the seminal randomized controlled trials in this area are at least 15 years old and had small sample sizes resulting in a lack of precision in effect estimates. 6,[8][9][10][11][12][13][14] The technique of laparoscopic colorectal surgery has been broadly adopted in Australia and is now a mature technique. Therefore, re-examination of this area to provide contemporary data with real-world, generalizable outcomes and large numbers, will help reinforce the randomized literature on this topic. 6,8,[14][15][16][17][18][19][20] The gold standard for addressing bias in comparative studies is randomisation. This maximizes the likelihood that all known and unknown baseline characteristics are exchangeable between the exposure and the control groups. It is extremely unlikely that the resource-use and time could be justified for further randomized controlled trials on the topic of laparoscopic compared with open surgery for colon cancer but a retrospective cohort study provides a practical alternative. However, multiple sources of bias within a cohort study can result in different baseline characteristics between the treatment groups which could affect the result. In this study, to minimize bias due to confounding, the technique of multivariable regression analysis was used to adjust for the potential confounding effects of differing levels of co-variables between the treatment groups. Furthermore, the analysis was repeated using the newer technique of propensity score weighting to achieve more even distribution of baseline characteristics between treatment groups, in a manner similar to the effect of randomisation (at least in-terms of measurable co-variables). Neither analysis method eliminates all bias. Both techniques have been presented to allow the reader more familiar with regression analysis to compare the results found with propensity score weighting.

Objectives
This study reports on overall survival (OS) and short-term outcomes comparing laparoscopic to open resection of right-sided colonic adenocarcinoma in public hospital patients over a 5-year period in the state of Victoria, Australia. It uses the technique of inverse probability weighting to balance cohorts and novel algorithms of administrative data (AD) linked to a death database to capture detailed population level data.

Study design
This retrospective cohort study used prospectively-gathered AD. Multivariable regression analysis was used to adjust for the potentially confounding effects of co-variables. The analysis was repeated employing a propensity score model with Inverse probability of treatment weighting to help minimize selection bias relating to measurable co-variables that may have determined whether patients were allocated to either treatment arm. This analysis method balanced the co-variables between study groups by creating pseudo-populations of slightly different sizes than the original data.
In this study, we have used hospital AD to gain access to population-level clinical outcomes data which captures all inpatient admissions with detail concerning patient demographics, co-morbidities, operation type, post-surgery complications, histopathology and some information concerning staging. Although hospital AD is primarily used for the purpose of billing, it is collected in a systematic prospective manner and is a potential source of individual covariate information relevant to CRC surgery. We have used algorithms of code combinations, validated against clinical notes to increase accuracy and applicability of AD for clinical outcomes research. 21,22 Because AD only records inpatient data and thus only inpatient mortality, linkage to a state-wide death registry was also utilized for the purpose of long-term survival analysis.
We examined only right-sided colon cancer resections because techniques of laparoscopic right-sided resections are more standardized among surgeons and therefore may represent less potential for technical heterogeneity than left or rectal resections. We previously found that coding for laparoscopic resection of right-sided cancers at our institution had an accuracy of 87% (sensitivity 80% and specificity 100%). 23 Victoria is the second most populated state in Australia with 6.5 million people, mostly urban dwellers. 24 All Australian residents have access to public healthcare with about half the population being additionally privately insured. Most surgeons work in both public and private sectors.
This study was performed in line with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement 25 and utilized a guideline for propensity score studies recommended by the EQUATOR network. 26 The Human Research and Ethics Committee, National Health and Medical Research Council of Australia (reference number HREC/52415/MH-2019), granted ethics approval.

Participants
This study included all public hospital patients having surgical resection for right-sided cancers from 2014 to 2018 within the state of Victoria, Australia. Cases were found from the Victoria Admitted Episodes Dataset (VAED), being the central repository of all Victorian hospitals' AD. This data was extracted from the VAED by applying algorithms created as part of a previous study to increase the accuracy of the data. 22 The algorithms stratified patients based on the location of the tumour (right colon was defined as all right and transverse colon including splenic flexure, left colon was defined as descending and sigmoid colon and rectum included recto-sigmoid and rectum). Illogical combinations were eradicated by combining the code for 'tumour location' with the code for 'type of operation'. On comparison with clinical notes, the accuracy of these validated algorithms of AD for right, left and rectum was 93%, 89% and 88% respectively. 22 Coding patterns across the state are homogenous, as demonstrated in previous studies performed by our group. [27][28][29][30] Exclusion criteria were private patients (algorithms were only validated for public hospitals), patients under 10 years, synchronous and metachronous cancers. Duplicate data, rectal and left sided cancers were excluded.

Exposure and control
Principal exposure of interest was laparoscopic surgery and control was open surgery.
Surgery type was defined as laparoscopic if it was coded as such. From mid-2013 the coding used for AD in Australia (the International Classification of Diseases Tenth Revision, Australian Modification (ICD-10-AM)) released unique codes for laparoscopic colorectal surgery, for this reason 2014 was chosen as the study starting date. Prior to this, to code for a laparoscopic operation, the open code was combined with a generic code for laparoscopy. For this study, patients were defined as having a laparoscopic operation if a unique laparoscopic colorectal code was used or a code for open surgery was combined with the generic laparoscopic code. Cases converted from laparoscopic to open surgery were recorded as laparoscopic to maintain the intention to treat analysis. Patients were defined as having an open operation if they only had a code for an open CRC resection.

Outcomes
The principal outcome of interest was overall survival. Death data (outcome and date) was obtained by linkage to the Victorian Death Index (VDI) and censoring time for survivors was the date of discharge of their last inpatient stay subsequent to their index admission (almost all patient had subsequent inpatient admissions which could include day-case endoscopy and chemotherapy). Starting date was defined as date of admission for surgical procedure.
Secondary outcomes of interest (short-term outcomes) were: return to theatre, prolonged length of stay (LOS) (>14 days) defined as discharge date subtracted from admission date, inpatient mortality and discharge destination (home/other).

Potential confounders
Patient demographics, comorbidity scores (American Society of Anaesthesiologists (ASA) score and Charlson Comorbidity Index score(CCI)) 31 and pathological details were recorded. AD records lymph node metastases (92% accuracy) and distant metastases (88% accuracy) 22 but does not record formal TNM staging. AD does not capture immunohistochemistry data.
Hospital case volume (high volume ≥50 CRC resections per year, low volume <50 CRC resections per year) and hospitals with specialist CR surgery training status (Australian and New Zealand Training Board in Colon and Rectal Surgery (ANZTBCRS)) were recorded. Victorian ANZTBCRS hospitals have been defined previously. 28 Although Peter MacCallum Cancer Centre is not an ANZTBCRS hospital, it was included given its specialist workload.
Specifically for regression analysis but not for propensity scoring, ASA was defined with the following categories: ASA 1-2, low; ASA 3-5, high; ASA not recorded, NR.
In a similar fashion, CCI score was defined as 0, 1 or ≥2 (with ≥2 conferring higher mortality) to aid regression analysis. The variables used to calculate CCI have been validated for use in the ICD-10-AM. 31 This categorisation has been used previously. [31][32][33][34] HDU+ was used to categorize those patients requiring more than standard ward care. It included all patients having postoperative mechanical ventilation and/or a monitored bed including intensive care unit (ICU), high dependency unit (HDU) or coronary care unit (CCU).
The hospital geographic location was defined using the Accessibility/Remoteness Index of Australia (ARIA+). ARIA+ is a measure for defining geographical remoteness based on accessibility to services that is nationally recognized, dividing Australia into five    geographical zones (S1). 28 All public hospital in Victoria are located in the 3 ARIA+ zones of metropolitan, inner regional and outer regional.

Statistical analysis
Stata version 15.0 (College Station, TX, USA) was used for data analysis. This study used propensity score inverse probability of treatment weighting (IPTW) to create a pseudo-population whereby the action of balancing covariables produced changes in cohort size that had balanced baseline characteristics between the exposure and treatment group. To maintain the power of the study, we elected to use IPTW which excludes fewer patients than propensity score matching. 35 Moreover, IPTW estimates average treatment effect (ATE) for the whole study population as distinct from matching which only estimates the average treatment effect on the treated (ATT). Pre-and perioperative characteristics used for calculating the propensity score were age, sex, ASA score, CCI, year of operation, ARIA+, elective versus emergency admission, hospital case volume, ANZTBCRS status, lymph node metastases and distant metastases. Weights were trimmed at the 1st and 99th centiles to both avoid extreme weights and to meet the positivity assumption (complete overlap in propensity score distributions). Comparison of unweighted and weighted baseline variables using standardized mean differences (SMD) and associated p-values was used to assess exposure group balance. Covariables were deemed to be balanced if the SMD was less than 0.2 Using the weighted sample, the hazard of death was estimated using Cox proportional hazards (PH) regression analysis. The robust method was used to estimate variance to account for the weighted nature of the pseudopopulation. 36 As a supportive analysis, univariable Cox PH regression was performed on unweighted data to assess the hazard of death for the principal exposure of interest and all baseline variables. All Covariables with a ≥10% change from the univariable analysis were included in a multivariable analysis, with mode of access retained as principal exposure.
For comparison, results were presented as the unweighted univariable hazard of death and the unweighted multivariable hazard of death and finally the hazard of death using the IPTW pseudo-population.
Schoenfeld residuals and visual assessment of cumulative hazard graphs were used to test the proportional hazards assumption.
A similar structure of analysis utilizing propensity score weighted analysis supplemented with an unweighted univariable and multivariable logistic regression analysis was performed for the secondary outcomes with results presented as odds ratios (OR).

Results
There were 3603 patients identified for the study after previously described algorithms were applied: 1439 open patients and 2164 laparoscopic patients. Included in the laparoscopic cohort were 383 patients who had conversion to open (Fig. 1).
For the IPTW analysis, 118 patients were excluded when the weights were trimmed below the 1st and above the 99th centile (Fig. 2).
The baseline characteristics of the unweighted and weighted populations can be seen in Table 1. The original raw (unweighted) data of the study is presented in the left hand column. The right hand column shows that the process of propensity score weighting achieved an even balance of co-variables between the treatment groups based on the low SMDs and high p-values, albeit by producing pseudo-populations of different sizes to the original treatment groups.
Median clinical follow-up time (excluding time to death) of the weighted data was 1. During the total follow-up period, 28% of patients undergoing open surgery and 14% of patients undergoing laparoscopic surgery died (Fig. 3).  Univariable Cox PH regression analysis of the unweighted data showed a 47% increased overall survival for patients in the laparoscopic cohort compared to the open cohort (HR 0.53, 95% CI 0.45-0.61, P < 0.001). Results of Cox PH univariable regression analysis for all co-variables are displayed in supplementary Tables 1-5. However, in the unweighted data, multivariable Cox regression did not show a significant effect of laparoscopic surgery compared with open surgery on OS when adjusted for the effects of geographical location, CCI and ASA scores, age, type of admission, lymph node involvement and presence of distant metastases, (HR 0.90, 95% CI 0.76-1.07, P = 0.22). On Cox PH regression analysis of the IPTW weighted data, there was a similar hazard ratio to the unweighted multivariable result, again providing no evidence of effect of laparoscopic surgery on OS (HR 0.87, 95% CI 0.72-1.04, P = 0.12) ( Table 2).
In terms of secondary outcomes, the weighted data showed lower odds of prolonged length of stay (OR 0.58, 95% CI 0.46-0.72, P = <0.001) and discharge destination (not home) (OR 0.69, 95% CI 0.55-0.88, P = <0.001) in the laparoscopic cohort compared to the open cohort. There was no difference in inpatient mortality (OR 0.88, 95% CI 0.41-1.90, P = 0.75) and return to theatre (OR 0.76, 95% CI 0.53-1.10, P = 0.15). These findings were similar to the multivariable logistic regression analysis of the unweighted data ( Table 3). Results of univariable logistic regression analysis of secondary outcomes for unweighted data of all covariables are presented in Supplementary Tables S1-S5.

Discussion
This study used IPTW propensity scoring to convert observational data into pseudorandomised data. The strength of this study lies in its size, which is greater than any RCT comparing laparoscopic to open resection for colon cancer and its up-to-date and generalizable population-level data.
Propensity score analysis found no difference in long-term OS comparing laparoscopic to open resection for right-sided colon cancers. In terms of short-term outcomes, laparoscopic resection resulted in reduced odds of prolonged LOS, and discharge destination other than home compared to open resection. There was no difference between laparoscopic resection and open resection in regards to inpatient mortality and return to theatre. When the unweighted data were analysed, the initial strong effect, seen on univariable regression, of laparoscopic surgery on improved survival was no longer present when confounding variables were adjusted for on multivariable regression. This study used data from the last 6 years to investigate the effects of laparoscopy in colon cancer resection in an era when it had become an established and commonly used surgical technique. Since being validated by the COST trial (2002) 6 Five-year follow-up did not find any difference in OS in the subset of colonic patients (percentage difference À7.0%, 95% CI À3.4% to 18.3%). 9  In terms of secondary outcomes, LOS was shorter in the laparoscopic cohort of most of the major RCTs 4,5,15,18,37,38 similar to our findings. Consistent with our study, previous studies found no difference between laparoscopic and open cohorts in terms of inpatient mortality. 15,18,[37][38][39] Most RCTs involved single centers or, if multi-centered, were highly selective. This potentially decreases generalisability. This is not the case in this study which pooled all available data within a population and reflects real-world activity. It is also the largest study to use propensity scoring in this setting. 40 This study was deliberately restricted to right-sided cancers in an attempt to investigate the effect of laparoscopic resection utilizing the least heterogenous population. Right-sided laparoscopic resections were also more accurately coded in AD than other colorectal resections.
It is a limitation of this study that right-sided resections included transverse colon resections.
Another limitation of this study is that of AD. AD does not record every complication type; of relevance to colorectal surgeons, there is not a code for anastomotic leak. However, in this study it is likely that the sequalae of this complication were captured in our primary and secondary outcomes. The other notable omission in AD is outpatient chemotherapy and radiation treatment, although this was unlikely to be differential between open and laparoscopic surgery.
A further limitation of this study is bias, as with any observational study. It was not possible to adjust for all elements of the likely allocation bias that determined whether a patient had laparoscopic or open surgery (e.g., anatomically, the tumour may not lend itself to laparoscopic resection). However, in this study measurable confounding variables were adjusted for in two ways, at the outcome level through logistic regression analysis and at exposure level through propensity scoring using IPTW. It Is reassuring that our results and baseline characteristics of our cohorts are similar to large RCTs in the literature, indicating that balance may have been achieved between important baseline variables and most major confounders adjusted for. Although not measurable in this study, allocation bias might select simpler cases for laparoscopic surgery, resulting in a bias towards better survival outcomes in the laparoscopic cohort. Despite this potential bias, this study showed no evidence of a survival benefit from laparoscopic surgery.
In this study, follow-up was based on a death index and inpatient admissions and therefore may seem short. However, follow-up data is complete and detailed and extended over 5 years. The only deaths that were potentially missed were those patients who died outside the state of Victoria. We could have made follow-up time of this study seem longer by assuming that the end date of the study was the date of last follow-up for patients with no death record. However, this would have ignored the available detailed follow-up data from subsequent in-patient admissions, and ignored the principle of censoring at last clinical follow-up which is an important aspect of a survival analysis.
IPTW creates pseudo-populations where the numbers of patients in the exposure and control cohorts can differ significantly from the unweighted data. This can appear disconcerting, particularly if there were small numbers in one cohort to begin with. Propensity scoring is a balancing procedure in which all covariates, apart from the exposure, are balanced between cohorts. This balancing procedure aims to mitigate the effects of treatment allocation bias present in observational data (at least for measurable co-variables) to allow the study to become more akin to a randomized trial and is becoming more commonly used in the surgical literature. 26

Conclusion
This study validates the use of laparoscopic surgery for right-sided colon cancer, showing similar long-term overall survival and inpatient mortality compared to open surgery. Laparoscopic surgery is superior to open surgery for the short-term outcomes of return to theatre, LOS, and discharge destination other than home.