Secondary Logo

Journal Logo

Original Articles

An evaluation of treatments and survival rates for pancreatic adenocarcinoma through survival analysis with inverse probability of treatment weighting: a population-based study

Wang, Suzhena; Wang, Chenb; Shi, Fuyana; Tao, Enxuec; Zhu, Gaopeia; Li, Juana; Feng, Jianinga; Wang, Xiaoxuana; Guo, Jingd; Zheng, Qingfenge; Zhang, Bof

Author Information
doi: 10.1097/JP9.0000000000000060
  • Open

Abstract

Introduction

Pancreatic cancer accounts for about 3% of all cancers in the United States and about 7% of all cancer deaths, according to the American Cancer Society's estimate. Mortality rates of pancreatic cancer continue to be very high, with an estimated 45,750 (23,800 men and 21,950 women) deaths in the United States in 2019.[1] Global Cancer Observatory estimates predicted 458,918 diagnoses and 432,242 deaths from pancreatic cancer globally in 2018.[2] Due to pancreatic cancer's high mortality, investigating survival rates of various treatments to postulate prognosis is of interest to both of clinicians and patients. The mainstays of current curative treatments are surgery (pancreatectomy), radiation therapy, and surgery (pancreatectomy) plus adjuvant radiotherapy. Surgical resection is generally considered as the effective curative treatment for pancreatic cancer, but <20% of patients are candidates for surgery.[3] Radiotherapy can destroy the cancer cells rapidly, shrink the tumors before surgery, and eliminate most cancer cells that remain in the treated area after surgery. Thus, surgical resection combined with radiotherapy is often considered a treatment method with better survival.[4]

Although treatment techniques have advanced significantly in recent years, prognoses of pancreatic cancer have not much improved. The 5-year relative survival rate of pancreatic cancer for all stages combined is 8%.[5] Even for the small percentage of patients diagnosed with localized disease, the 5-year survival rate is only 28% to 32%. About 52% of patients are diagnosed with distant disease, for which 5-year survival rate is 3%.[5] Multimodality treatment approaches offer better survival outcomes, but they remain controversial because of the absence of conclusive evidence.[4]

The Surveillance, Epidemiology, and End Results (SEER)[6] database from the National Cancer Institute of the United States provides national cancer registries for researchers to evaluate cancer outcomes. In this study, we attempt to minimize the influence of confounding covariates and evaluate survival outcomes of the 3 treatments (surgery, radiation, and surgery plus adjuvant radiotherapy) for pancreatic adenocarcinoma. Inverse probability of treatment weighting (IPTW)[7–10] survival analysis was performed to assess the benefits of the 3 different therapies for pancreatic adenocarcinoma patients.[11,12]

Materials and Methods

Data source

The SEER dataset from 2004 to 2014 was the data source for this population-based study. Patients with pathologically confirmed pancreatic adenocarcinoma were identified by the International Classification of Disease for Oncology, Third Edition (ICD-O-3) morphology codes of 8140/3 (adenocarcinoma, not otherwise specified), as well as 8500/3 (duct carcinoma, not otherwise specified), and topography codes of C25.0, C25.1, C25.2, C25.3, C25.7, C25.8, and C25.9. A total of 37,393 pancreatic adenocarcinoma patients were identified, then limited to those who received surgery, radiation, or surgery plus adjuvant radiotherapy. Exclusion criteria included (i) patients who had other malignant tumor prior to pancreatic adenocarcinoma or whose pancreatic adenocarcinoma was diagnosed after 2014, (ii) patients who died prior to date of diagnosis, (iii) patients in whom the clinical stages were missed, (iv) patients who had missing/unknown vital status, (v) patients who had no evidence of a primary tumor or in situ cancer, and (vi) patients who were diagnosed by autopsy or death certificate only. Patients were also excluded if any of the following essential variables were listed as missing or unknown: survival time, age, tumor size, race, numbers of primary site, number of positive nodes, diagnostic confirmation, location of primary site, tumor differentiation, and derived tumor stage from American Joint Committee on Cancer staging (AJCC). A total of 8191 patients were included in our study: 3409 patients received surgery, 2865 received radiotherapy, and 1917 received surgery plus adjuvant radiotherapy. This study was approved by the Ethics Review Board of Weifang Medical University and conformed to the provisions of the Declaration of Helsinki. This study was a secondary analysis of the existing SEER dataset, and posed no more than minimal risk to patients. Therefore, the requirement to obtain informed consent from patients was waived.

Statistical analysis

In this study, we adopted IPTW survival analysis[9,13–15] assisted by the generalized boosted models (GBMs) as the primary analysis approach to assess survival rates of the pancreatic adenocarcinoma patients who received surgery, radiation, or surgery plus adjuvant radiotherapy. GBM is an iterative fitting algorithm that can measure the nonlinear relationship between an outcome and pretreatment covariates by fitting a piecewise constant model and outputting a binary outcome through multivariate nonparametric regression tree.[9,13] There is only a single simple regression tree at the initial step in GBM, but a new tree is added when the new iteration comes on the rules that the tree provides the best fit to the residuals of the model from the previous iteration. Iterations are repeated until a presetting value is reached or the absolute standardized bias is <0.2 or the Kolmogorov–Smirnov (KS) <0.1.[13] In this study, the covariates potentially associated with treatment selection, including age, gender, marital status, race, tumor size, tumor grade, tumor primary sites, disease extent condition, and tumor T, N, M stages were included in the GBMs for generation of propensity scores. GBMs with these covariates were used to generate a continuous propensity score through a multinomial logistic regression model and to estimate the probability that a patient would undergo surgery, radiation, or surgery plus adjuvant radiotherapy. Then, the IPTW survival analysis was conducted with these generated propensity scores.

Baseline characteristics of the different treatment groups were compared and evaluated by chi-square test for categorical variables and F test for continuous variables in the original sample. Here, the “original sample” refers to the study sample collected as described in “Data source” section. The IPTW created a “pseudo-sample” consisting of the identical subjects in the original sample, but each subject in the pseudo-sample was assigned a weight derived from the IPTW to minimize the impact of treatment selection bias and other potential confounders.[15,16] As previously stated, the propensity scores were estimated by the GBMs to predict the probability of a patient undergoing different treatments. The following methods were employed to estimate the overall and cancer-specific survivals: the Kaplan–Meier estimator weighted by IPTW method was carried out to estimate survival curves; IPTW log-rank test was conducted to compare the survival curves; and IPTW Cox proportional hazards analysis and their 95% confidence intervals (CIs) were performed to estimate mortality hazard ratios.[17–20] All significance tests were two-tailed, and P values <.05 were considered statistically significant. All statistical analyses were carried out using R version 4.0 (R Foundation for Statistical Computing, Vienna, Austria). The GBM algorithms were calculated through the R package “twang” (Toolkit for Weighting and Analysis of Nonequivalent Groups).[19] The R package “IPWsurvival” was applied to estimate log-rank test and adjusted survival curves based on the IPTW.[21] The R package “coxphw” was applied to perform IPTW Cox proportional hazard regression.[18]

Results

Demographic and clinical characteristics of the original sample and the pseudo-sample at baseline

Among the 8191 pancreatic adenocarcinoma patients included in this study, the median (25% quantile and 75% quantile) study follow-up times were 13.0 (6.0 and 24.0) months, 15.0 (6.0 and 32.0) months, 9.0 (5.0 and 15.0) months, and 18.0 (11.0 and 33.0) months in the entire sample, surgery group, radiation group, and surgery plus adjuvant radiotherapy group, respectively. The mean (standard deviation) ages of the entire sample, group, radiation group, and surgery plus adjuvant radiotherapy group were 65.71 (11.69) years, 65.61 (12.64) years, 66.98 (11.46) years, and 64.00 (9.90) years, respectively.

Baseline demographic and clinical characteristics of the pancreatic adenocarcinoma patients are listed in Tables 1 and 2. In the original sample, before IPTW balancing, the 3 groups took significantly different proportions in the categories of all covariates except for gender (P = .571). However, after the IPTW balancing in the pseudo-sample, the 3 groups took similar proportions (no significant difference was detected in proportions) in the categories of all covariates except for tumor primary site (P = .024) and T stage (P = .017) (Table 2).

Table 1 - Clinical characteristics of patients among 3 groups before inverse probability of treatment weighting (IPTW)
Variable All patients Surgery(n = 3409) Radiotherapy(n = 2865) Surgery plus adjuvant radiotherapy (n = 1917) P value
Age (yr) <.001
 20–39 152 (1.86) 109 (3.20) 27 (0.94) 16 (0.83)
 40–59 2201 (26.87) 874 (25.64) 733 (25.58) 594 (30.99)
 60–79 4863 (59.37) 1997 (58.58) 1647 (57.49) 1219 (63.59)
 ≥80 975 (11.90) 429 (12.58) 458 (15.99) 88 (4.59)
Gender .571
 Female 3988 (48.69) 1680 (49.28) 1392 (48.59) 916 (47.78)
 Male 4203 (51.31) 1729 (50.72) 1473 (51.41) 1001 (52.22)
Marital status <.001
 Married 5091 (62.15) 2067 (60.63) 1721 (60.07) 1303 (67.97)
 Single 918 (11.21) 424 (12.44) 297 (10.37) 197 (10.28)
 Separated 53 (0.65) 15 (0.44) 26 (0.91) 12 (0.63)
 Divorced 857 (10.46) 344 (10.09) 319 (11.13) 194 (10.12)
 Widowed 1016 (12.40) 435 (12.76) 415 (14.49) 166 (8.66)
 Unmarried or domestic partner 13 (0.16) 9 (0.26) 2 (0.07) 2 (0.10)
 Unknown 243 (2.97) 115 (3.37) 85 (2.97) 43 (2.24)
Race <.001
 White 6498 (79.33) 2752 (80.73) 2200 (76.79) 1546 (80.65)
 Black 889 (10.85) 327 (9.59) 357 (12.46) 205 (10.69)
 American Indian 46 (0.56) 15 (0.44) 20 (0.70) 11 (0.57)
 Asian 745 (9.10) 306 (8.98) 287 (10.02) 152 (7.93)
 Other and unknown 13 (0.16) 9 (0.26) 1 (0.03) 3 (0.16)
Primary site <.001
 Pancreatic tail 996 (12.16) 602 (17.66) 185 (6.46) 209 (10.90)
 Pancreatic head 5351 (65.33) 2129 (62.45) 1809 (63.14) 1413 (73.71)
 Pancreatic body 868 (10.60) 293 (8.59) 437 (15.25) 138 (7.20)
 Other 976 (11.92) 385 (11.29) 434 (15.15) 157 (8.19)
Tumor size (mm) <.001
 <50 6429 (78.49) 2696 (79.08) 2129 (74.31) 1604 (83.67)
 50–100 1552 (18.95) 587 (17.22) 682 (23.80) 283 (14.76)
 >100 128 (1.56) 85 (2.49) 31 (1.08) 12 (0.63)
 Unknown or uncertain 82 (1.00) 41 (1.20) 23 (0.80) 18 (0.94)
Grade <.001
 Well 956 (11.67) 643 (18.86) 120 (4.19) 193 (10.07)
 Moderately 2452 (29.94) 1314 (38.55) 286 (9.98) 852 (44.44)
 Poorly 1919 (23.43) 973 (28.54) 333 (11.62) 613 (31.98)
 Undifferentiated 126 (1.54) 53 (1.55) 44 (1.54) 29 (1.51)
 Not applicable 2738 (33.43) 426 (12.50) 2082 (72.67) 230 (12.00)
Disease extent condition <.001
 Localized 1200 (14.65) 723 (21.21) 314 (10.96) 163 (8.50)
 Regional 5298 (64.68) 2104 (61.72) 1634 (57.03) 1560 (81.38)
 Distant 1693 (20.67) 582 (17.07) 917 (32.01) 194 (10.12)
T stage <.001
 T1 595 (7.26) 427 (12.53) 85 (2.97) 83 (4.33)
 T2 1420 (17.34) 663 (19.45) 488 (17.03) 269 (14.03)
 T3 4729 (57.73) 2163 (63.45) 1122 (39.16) 1444 (75.33)
 T4 1343 (16.40) 145 (4.25) 1077 (37.59) 121 (6.31)
 TX 104 (1.27) 11 (0.32) 93 (3.25) 0 (0.00)
N stage <.001
 N0 3957 (48.31) 1518 (44.53) 1786 (62.34) 653 (34.06)
 N1 4004 (48.88) 1876 (55.03) 869 (30.33) 1259 (65.68)
 NX 230 (2.81) 15 (0.44) 210 (7.33) 5 (0.26)
M stage <.001
 M0 7084 (86.49) 3062 (89.82) 2148 (74.97) 1874 (97.76)
 M1 1107 (13.51) 347 (10.18) 717 (25.03) 43 (2.24)

Table 2 - Clinical characteristics of patients among 3 groups after inverse probability of treatment weighting (IPTW)
Variable Surgery (%) Radiotherapy (%) Surgery plus adjuvant radiotherapy (%) Minimum P value
Age (yr) .721
 20–39 3.2 0.9 0.8
 40–59 25.6 25.6 31.0
 60–79 58.6 57.5 63.6
 ≥80 12.6 16.0 4.6
Gender .858
 Female 49.3 48.6 47.8
 Male 50.7 51.4 52.2
Marital status .633
 Married 60.6 60.1 68.0
 Single 12.4 10.4 10.3
 Separated 0.4 0.9 0.6
 Divorced 10.1 11.1 10.1
 Widowed 12.8 14.5 8.7
 Unmarried or domestic partner 0.3 0.1 0.1
 Unknown 3.4 3.0 2.2
Race .550
 White 80.7 76.8 80.6
 Black 9.6 12.5 10.7
 American Indian 0.4 0.7 0.6
 Asian 9.0 10.0 7.9
 Other and unknown 0.3 0.0 0.2
Primary site .024
 Pancreatic tail 17.7 6.5 10.9
 Pancreatic head 62.5 63.1 73.7
 Pancreatic body 8.6 15.3 7.2
 Other 11.3 15.1 8.2
Tumor size (mm) .639
 <50 79.1 74.3 83.7
 50–100 17.2 23.8 14.8
 >100 2.5 1.1 0.6
 Unknown or uncertain 1.2 0.8 0.9
Grade .232
 Well 18.9 4.2 10.1
 Moderately 38.5 10.0 44.4
 Poorly 28.5 11.6 32.0
 Undifferentiated 1.6 1.5 1.5
 Not applicable 12.5 72.7 12.0
Disease extent condition .417
 Localized 21.2 11.0 8.5
 Regional 61.7 57.0 81.4
 Distant 17.1 32.0 10.1
T stage .017
 T1 12.5 3.0 4.3
 T2 19.4 17.0 14.0
 T3 63.4 39.2 75.3
 T4 4.3 37.6 6.3
 TX 0.3 3.2 0.0
N stage .157
 N0 44.5 62.3 34.1
 N1 55.0 30.3 65.7
 NX 0.4 7.3 0.3
M stage .189
 M0 89.8 75.0 97.8
 M1 10.2 25.0 2.2

Survival in the original sample

Figures 1 and 2 show both the estimated 5-year overall Kaplan–Meier survival curves and cancer-specific Kaplan–Meier survival curves for the 3 groups in the original sample and in the pseudo-sample. Table 3 shows the analysis results derived from Cox proportional hazard regression (hazard ratios, their 95% confidence intervals or CIs, and P values) for all-cause mortality and cancer-specific mortality in the original sample; Table 4 displays these results in the pseudo-sample.

Figure 1
Figure 1:
Overall Kaplan–Meier survival curves for the 3 groups in the original sample and in the pseudo-sample.
Figure 2
Figure 2:
Cancer-specific Kaplan–Meier survival curves for the 3 groups in the original sample and in the pseudo-sample.
Table 3 - Hazard ratios, confidence internals, and P values obtained from Cox proportional hazard models for all-cause mortality and cancer-specific mortality in the original sample (CI, 95% confidence interval)
All-cause mortality Cancer-specific mortality
Variable Hazard ratio (CI) P value Hazard ratio (CI) P value
Treatment
 Surgery plus radiotherapy Reference Reference
 Surgery 1.17 (1.08–1.26) <.001 1.10 (1.01–1.20) .030
 Radiotherapy 2.77 (2.54–3.03) <.001 3.02 (2.72–3.34) <.001
Age (yr)
 20–39 Reference Reference
 40–59 2.39 (1.78–3.21) <.001 2.16 (1.58–2.96) <.001
 60–79 3.19 (2.38–4.28) <.001 2.46 (1.80–3.36) <.001
 ≥80 4.56 (3.37–6.16) <.001 2.88 (2.08–3.98) <.001
Gender
 Female Reference Reference
 Male 1.10 (1.04–1.16) .001 1.07 (1.01–1.14) .033
Marital status
 Married Reference Reference
 Single 1.15 (1.05–1.26) .002 1.15 (1.04–1.27) .006
 Separated 1.10 (0.79–1.52) .578 1.24 (0.88–1.75) .220
 Divorced 1.21 (1.10–1.32) <.001 1.24 (1.12–1.37) <.001
 Widowed 1.18 (1.09–1.29) <.001 1.18 (1.07–1.30) .001
 Unmarried or domestic partner 1.27 (0.57–2.83) .559 1.66 (0.74–3.70) .218
 Unknown 1.04 (0.89–1.21) .656 1.08 (0.91–1.29) .383
Race
 White Reference Reference
 Black 1.13 (1.04–1.23) .005 1.06 (0.96–1.17) .228
 American Indian 1.00 (0.71–1.42) .979 1.12 (0.77–1.62) .557
 Asian 1.13 (1.04–1.24) .007 1.23 (1.11–1.36) <.001
 Other and unknown 0.55 (0.21–1.48) .240 0.55 (0.18–1.72) .308
Tumor primary site
 Pancreatic tail Reference Reference
 Pancreatic head 1.20 (1.09–1.32) <.001 1.30 (1.16–1.46) <.001
 Pancreatic body 1.02 (0.91–1.15) .719 1.08 (0.93–1.24) .322
 Other 1.08 (0.96–1.21) .188 1.18 (1.03–1.36) .017
Tumor size (mm)
 <50 Reference Reference
 50-100 1.06 (0.99–1.13) .095 1.05 (0.97–1.14) .216
 >100 1.10 (0.87–1.38) .417 1.14 (0.87–1.49) .339
 Unknown or uncertain 1.11 (0.84–1.47) .458 1.18 (0.85–1.63) .332
Tumor grade
 Well Reference Reference
 Moderately 1.76 (1.58–1.97) <.001 1.93 (1.69–2.21) <.001
 Poorly 2.23 (1.99–2.49) <.001 2.49 (2.17–2.85) <.001
 Undifferentiated 2.48 (1.99–3.10) <.001 2.85 (2.21–3.68) <.001
 Not applicable 1.76 (1.57–1.97) <.001 1.94 (1.69–2.23) <.001
Disease extent condition
 Localized Reference Reference
 Regional 1.27 (1.11–1.45) <.001 1.35 (1.16–1.58) <.001
 Distant 1.23 (1.04–1.46) .017 1.29 (1.06–1.58) .012
T stage
 T1 Reference Reference
 T2 1.49 (1.28–1.72) <.001 1.66 (1.39–1.99) <.001
 T3 1.56 (1.33–1.82) <.001 1.71 (1.42–2.06) <.001
 T4 1.58 (1.33–1.86) <.001 1.65 (1.35–2.02) <.001
 TX 2.00 (1.54–2.58) <.001 1.93 (1.43–2.62) <.001
N stage
 N0 Reference
 N1 1.32 (1.25–1.41) <.001 1.38 (1.29–1.48) <.001
 NX 1.53 (1.33–1.78) <.001 1.57 (1.34–1.85) <.001
M stage
 M0 Reference Reference
 M1 1.74 (1.54–1.98) <.001 1.77 (1.53–2.05) <.001

Table 4 - Hazard ratios, confidence internals, and P values obtained from Cox proportional hazard models for all-cause mortality and cancer-specific mortality in the pseudo-sample (CI: 95% confidence interval)
All-cause mortality Cancer-specific mortality
Variable Hazard ratio (CI) P value Hazard ratio (CI) P value
Treatment
 Surgery plus radiotherapy Reference Reference
 Surgery 1.15 (1.11–1.20) <.001 1.11 (1.06–1.17) <.001
 Radiotherapy 2.53 (2.44–2.63) <.001 2.73 (2.61–2.86) <.001
Age (yr)
 20–39 Reference Reference
 40–59 1.87 (1.62–2.15) <.001 1.54 (1.33–1.79) <.001
 60–79 2.38 (2.07–2.74) <.001 1.62 (1.40–1.87) <.001
 ≥80 3.18 (2.75–3.69) <.001 1.84 (1.58–2.15) <.001
Gender
 Female Reference Reference
 Male 1.16 (1.13–1.20) <.001 1.13 (1.08–1.17) <.001
Marital status
 Married Reference Reference
 Single 1.12 (1.06–1.18) <.001 1.12 (1.06–1.19) <.001
 Separated 1.25 (1.01–1.56) .042 1.49 (1.19–1.87) .001
 Divorced 1.09 (1.04–1.15) .001 1.18 (1.11–1.25) <.001
 Widowed 1.37 (1.30–1.44) <.001 1.39 (1.31–1.47) <.001
 Unmarried or domestic partner 0.89 (0.44–1.77) .730 1.20 (0.60–2.40) .602
 Unknown 1.05 (0.96–1.16) .287 1.05 (0.93–1.17) .442
Race
 White Reference Reference
 Black 1.16 (1.10–1.22) <.001 1.10 (1.04–1.17) .001
 American Indian 1.10 (0.87–1.39) .432 1.25 (0.97–1.61) .086
 Asian 1.13 (1.07–1.20) <.001 1.28 (1.20–1.36) <.001
 Other and unknown 0.16 (0.07–0.37) <.001 0.15 (0.06–0.40) <.001
Tumor primary site
 Pancreatic tail Reference Reference
 Pancreatic head 1.13 (1.07–1.20) <.001 1.17 (1.10–1.25) <.001
 Pancreatic body 1.01 (0.94–1.09) .704 0.98 (0.90–1.07) .706
 Other 0.95 (0.88–1.01) .119 1.05 (0.97–1.14) .251
Tumor size (mm)
 <50 Reference Reference
 50–100 0.98 (0.94–1.02) .361 0.93 (0.89–0.97) .002
 >100 1.36 (1.18–1.56) <.001 1.16 (0.98–1.39) .088
 Unknown or uncertain 0.93 (0.78–1.12) .452 1.13 (0.93–1.39) .225
Tumor grade
 Well Reference Reference
 Moderately 1.57 (1.48–1.66) <.001 1.69 (1.58–1.82) <.001
 Poorly 1.72 (1.61–1.82) <.001 1.97 (1.83–2.12) <.001
 Undifferentiated 2.12 (1.87–2.41) <.001 2.44 (2.10–2.83) <.001
 Not applicable 1.42 (1.34–1.51) <.001 1.61 (1.50–1.73) <.001
Disease extent condition
 Localized Reference Reference
 Regional 1.09 (1.01–1.17) .033 1.13 (1.03–1.24) .009
 Distant 1.02 (0.92–1.13) .703 1.13 (1.01–1.27) .034
T stage
 T1 Reference Reference
 T2 1.29 (1.19–1.40) <.001 1.66 (1.50–1.84) <.001
 T3 1.44 (1.32–1.57) <.001 1.80 (1.62–2.01) <.001
 T4 1.99 (1.82–2.19) <.001 2.23 (1.99–2.51) <.001
 TX 2.34 (1.96–2.78) <.001 2.31 (1.87–2.86) <.001
N stage
 N0 Reference Reference
 N1 1.22 (1.18–1.26) <.001 1.26 (1.21–1.31) <.001
 NX 1.74 (1.57–1.92) <.001 2.00 (1.79–2.24) <.001
M stage
 M0 Reference Reference
 M1 1.74 (1.61–1.88) <.001 1.59 (1.45–1.73) <.001

In the 3 groups in the original sample, the mortality rates were 40.4% in the surgery group (1378 patients died during the study period), 69.1% in the radiation group (1981 patients died), and 49.3% in the surgery plus adjuvant radiotherapy group (946 patients died). The overall survival rates at the end of the 1st, 3rd, and 5th years were 64.4%, 35.8%, and 28.2%, respectively, in the surgery group; 36.3%, 5.2%, and 2.5%, respectively, in the radiation group; and 78.7%, 34.4%, and 22.0%, respectively, in the surgery plus adjuvant radiotherapy group (Fig. 1A). The corresponding median survival times were 22 months, 9 months, and 23 months in the surgery group, the radiation group, and the surgery plus adjuvant radiotherapy group, respectively. The log-rank test gave a P value of <.001 when comparing the overall and cancer-specific Kaplan–Meier survival curves of the 3 groups in the original sample. The results obtained from Cox proportional hazard regression revealed that patients who received surgery had lower survival rates compared to the patients who received surgery plus adjuvant radiotherapy (all-cause mortality hazard ratio, 1.17; CI 1.08–1.26; P < .001). Similarly, patients who received radiation had lower survival rates compared to those who received surgery plus adjuvant radiotherapy (all-cause mortality hazard ratio, 2.77; CI 2.54–3.03; P < .001).

The cancer-specific survival rates at the 1st, 3rd, and 5th year were 74.0%, 48.3%, and 41.1%, respectively, in the surgery group; 45.3%, 9.2%, and 5.2%, respectively, in the radiation group; and 83.3%, 43.6%, and 31.2%, respectively, in the surgery plus adjuvant radiotherapy group (Fig. 1B). The corresponding cancer-specific median survival times were 36, 11, and 29 months in the surgery, radiation, and surgery plus adjuvant radiotherapy groups, respectively. Analysis of the Cox proportional hazard regression revealed that the cancer-specific survival rate in the surgery group was lower than the surgery plus adjuvant radiotherapy group (cancer-specific mortality hazard ratio, 1.10; CI 1.01–1.20; P = .030). The cancer-specific survival rate was also lower for the radiation group compared to the surgery plus adjuvant radiotherapy group (HCC-specific mortality hazard ratio, 3.02; CI 2.72–3.34; P < .001).

The following covariates were the risk factors identified by the Cox proportional hazard regression for cancer-specific mortality: age 40 to 59 years (vs age 20–39 years; hazard ratio, 2.16; CI, 1.58–2.96; P < .001), age 60 to 79 years (vs age 20–39 years; hazard ratio, 2.46; CI, 1.80–3.36; P < .001), age >80 years (vs age 20–39 years; hazard ratio, 2.88; CI, 2.08–3.98; P < .001), marital status “single” (vs marital status “married”; hazard ratio, 1.15; CI, 1.04–1.27; P = .006), marital status “divorced” (vs “married”; hazard ratio, 1.24; CI, 1.12–1.37; P < .001), marital status “windowed” (vs “married”; hazard ratio, 1.18; CI, 1.07–1.30; P = .001), male (vs female; hazard ratio, 1.07; CI, 1.01–1.14; P = .033), Asian (vs White; hazard ratio, 1.23; CI, 1.11–1.36; P < .001), tumor primary site pancreatic head (vs pancreatic tail; hazard ratio, 1.30; CI, 1.16–1.46; P < .001), other primary site (vs pancreatic tail; hazard ratio, 1.18; CI, 1.03–1.36; P = .017), tumor grade “moderately” (vs tumor grade “well”; hazard ratio, 1.93; CI, 1.69–2.21; P < .001), tumor grade “poorly” (vs tumor grade “well”; hazard ratio, 2.49; CI, 2.17–2.85; P < .001), tumor grade “undifferentiated” (vs tumor grade “well”; hazard ratio, 2.85 CI, 2.21–3.68; P < .001), regional disease extent condition (vs localized disease extent condition; hazard ratio, 1.35; CI, 1.16–1.58; P < .001), distant disease extent condition (vs localized disease extent condition; hazard ratio, 1.29; CI, 1.06–1.58; P = .012), T2 (vs T1; hazard ratio, 1.66 CI, 1.39–1.99; P < .001),T3 (vs T1; hazard ratio, 1.71; CI, 1.42–2.06; P < .001), T4 (vs T1; hazard ratio, 1.65; CI, 1.35–2.02; P < .001), N1 (vs N0; hazard ratio, 1.38 CI, 1.29–1.48; P < .001), and M1 (vs M0; hazard ratio, 1.77; CI, 1.53–2.05; P < .001). The risk factors identified by the Cox proportional hazard regression for all-cause mortality in the original sample are also reported in Table 3.

Survival in the pseudo-sample

Baseline demographic and clinical characteristics of the patients in the pseudo-sample after the IPTW were are displayed in Table 2. The overall survival rates at the end of the 1st, 3rd, and 5th year were 61.2%, 31.1%, and 23.4%, respectively, in the surgery group; 40.0%, 7.4%, and 2.8%, respectively, in the radiation group; and 77.0%, 33.0%, and 22.4%, respectively, in the surgery plus adjuvant radiotherapy group (Fig. 1B). The corresponding median survival times were 19 months, 10 months, and 21 months in the surgery group, the radiation group, and surgery plus adjuvant radiotherapy group, respectively. The log-rank test gave a P value of <.001 when comparing the overall and cancer-specific Kaplan–Meier survival curves of the 3 groups in the pseudo-sample. Analysis of the IPTW Cox proportional hazard regression demonstrated that both the radiation and surgery groups had a higher all-cause mortality rate and higher cancer-specific mortality rates than the surgery plus adjuvant radiotherapy group in the pseudo-sample (Table 4).

The cancer-specific survival rates at the 1st, 3rd, and 5th year were 71.3%, 43.2%, and 36.1%, respectively, in the surgery group; 48.8%, 12.6%, and 5.3%, respectively, in the radiation group; and 82.6%, 44.2%, and 33.0%, respectively, in the surgery plus adjuvant radiotherapy group (Fig. 2B). The corresponding cancer-specific median survival times were 28 months, 12 months, and 29 months in the surgery group, the radiation group, and surgery plus adjuvant radiotherapy groups, respectively. The following covariates were risk factors for cancer-specific mortality: age 40 to 59 years (hazard ratio, 1.54; CI, 1.33–1.79; P < .001), age 60 to 79 years (hazard ratio, 1.62; CI, 1.40–1.87; P < .001), age >80 years (hazard ratio, 1.84; CI, 1.58–2.15; P < .001), male (vs female; hazard ratio, 1.13; CI, 1.08–1.17; P < .001), marital status “single” (vs marital status “married”; hazard ratio, 1.12; CI, 1.06–1.19; P < .001), marital status “separated” (vs “married”; hazard ratio, 1.49; CI, 1.19–1.87; P = .001), marital status “divorced” (vs “married”; hazard ratio, 1.18; CI, 1.11–1.25; < 0.001), marital status “divorced” (vs “married”; hazard ratio, 1.39; CI, 1.31–1.47; P < .001), Black (vs White; hazard ratio, 1.10; CI, 1.04–1.17; P = .001), Asian (vs White; hazard ratio, 1.28; CI, 1.20–1.36; P < .001), tumor primary site pancreatic head (vs pancreatic tail; hazard ratio, 1.17; CI, 1.10–1.25; P < .001), tumor size level 50 to100 mm (vs <50 mm; hazard ratio, 0.93; CI, 0.89–0.97; P = .002), tumor grade “moderately” (vs tumor grade “well”; hazard ratio, 1.69; CI, 1.58–1.82; P < .001), tumor grade “poorly” (vs tumor grade “well”; hazard ratio, 1.97; CI, 1.83–2.12; P < .001), tumor grade “undifferentiated” (vs tumor grade “well”; hazard ratio, 2.44; CI, 2.10–2.83; P < .001), regional disease extent condition (vs localized disease extent condition; hazard ratio, 1.13; CI, 1.03–1.24; P = .009), distant disease extent condition (vs localized disease extent condition; hazard ratio, 1.13; CI, 1.01–1.27; P = .034), T2 (versus T1; hazard ratio, 1.66 CI, 1.50–1.84; P < .001), T3 (vs T1; hazard ratio, 1.80; CI, 1.62–2.01; P < .001), T4 (vs T1; hazard ratio, 2.23; CI, 1.99–2.51; P < .001), N1 (vs N0; hazard ratio, 1.26 CI, 1.21–1.31; P < .001), and M1 (vs M0; hazard ratio, 1.59; CI, 1.45–1.73; P < .001). The risk factors identified by the Cox proportional hazard regression for all-cause mortality in the pseudo-sample are also reported in Table 4.

Discussion

Covariate confounding exists in observational studies and may bring bias to the evaluation of cancer outcomes. In this study, we applied the IPTW survival analysis with propensity scores to alleviate confounding among 3 therapeutic groups of pancreatic adenocarcinoma patients. We made attempts to minimize the estimation bias and evaluate the long-term survival of 3 groups.

The results of this study demonstrated both surgery and surgery plus adjuvant radiation as pancreatic cancer therapies were statistically associated with reduced overall and cancer-specific mortality in SEER database after the IPTW. The survival curves were significantly different among the 3 groups. The patients with pancreatic adenocarcinoma may not expect a better outcome in overall and cancer-specific survival. This result was validated by previous studies.[22] A few other retrospective studies using the SEER database had demonstrated the survival benefit of adjuvant radiotherapy among patients with resected pancreatic cancer.[23] In our study, the Kaplan–Meier survival curves for the surgery group and the surgery with adjuvant radiotherapy group cross each other in the middle of 5 years, indicating that adjuvant radiotherapy may provide only short-term survival benefit for patients with resected pancreatic cancer. Although patients who received adjuvant radiotherapy may be more likely to achieve local control than those without radiotherapy, patients in both groups may develop distant metastases.

There are some limitations in the present study. One limitation is that the SEER program does not provide information regarding which patients received chemotherapy. Treatment details and timing of chemotherapy are also unknown. These factors must affect the survival outcomes. In addition, the SEER program does not collect information on surgical margin status or functional status of patients, which may influence treatment selection and outcomes. Details of radiotherapy including radiation dose and field design are also mostly missing in the SEER program. This limited our clinical investigation.

Conclusions

The present study has shown that surgery with adjuvant radiotherapy is significantly associated with improved overall and cancer-specific survival among the patients with pancreatic adenocarcinoma in a large national database. This was demonstrated through propensity score IPTW survival analysis. However, the choice of pancreatic adenocarcinoma treatment may largely depend on patients’ clinical characteristics, and therefore clinicians should not determine treatment for pancreatic adenocarcinoma based on the conclusions drawn from this observational study.

Acknowledgments

None.

Author contributions

Dr. Suzhen Wang, Chen Wang, Dr. Fuyan Shi, Dr. Enxue Tao made equal contribution to this research article. Dr. Bo Zhang and Dr. Suzhen Wang are co-senior authors of this research article.

Financial support

Dr. Suzhen Wang's research was partially supported by the National Natural Science Foundation of China (No. 81872719), the National Bureau of Statistics Foundation Project (No. 2018LY79), the Natural Science Foundation of Shandong Province (No. 2019MH034), and the Poverty Alleviation Fund project of Weifang Medical University (No. FP1801001). Dr. Fuyan Shi's research was partially supported by the National Natural Science Foundation of China (No. 81803337), the Shandong Provincial Youth Innovation Team Development Plan of Colleges and Universities (No. 2019-6-156, Lu-Jiao), the Shandong Provincial Government Fund for Overseas Study (No. 27, 2019, Lu-Jiao), the Shandong Science and Technology Development Plan Project (No. 2015 WS0067), and the Weifang Medical University Doctoral Foundation Project (No. 2017BSQD51).

Conflicts of interest

The authors declare that they have no conflict of interests.

Ethics approval

This study was approved by the Ethics Review Board of Weifang Medical University and conformed to the principles of the Declaration of Helsinki. This study was a secondary analysis of the existing SEER dataset, and posed no more than minimal risk to patients. Therefore, the requirement to obtain informed consent from patients was waived.

References

1. American Cancer Society. Cancer Facts & Figures 2019. Atlanta: American Cancer Society; 2019. Available from: https://www.cancer.org/cancer/pancreatic-cancer/about/key-statistics.html. Accessed March 1, 2020
2. World Health Organization. The Global Cancer Observatory 2018. Available from: https://gco.iarc.fr/today/fact-sheets-cancers. Accessed March 1, 2020
3. Mahmood J, Shukla HD, Soman S, et al. Immunotherapy, radiotherapy, and hyperthermia: a combined therapeutic approach in pancreatic adenocarcinoma treatment. Cancers (Basel) 2018; 10:469.
4. Sugawara A, Kunieda E. Effect of adjuvant radiotherapy on survival in resected pancreatic adenocarcinoma: a propensity score surveillance, epidemiology, and end results database analysis. J Surg Oncol 2014; 110:960–966.
5. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2018. CA Cancer J Clin 2018; 68:7–30.
6. Surveillance, Epidemiology, and End Results (SEER) Program, SEER∗Stat Database. Available from: http://www.seer.cancer.gov. Accessed March 1, 2020
7. Imbens GW. National Bureau of Economic Research. The Role of the Propensity Score in Estimating Dose-Response Functions. In NBER Technical Working Paper Series No. t0237. Cambridge, MA: National Bureau of Economic Research; 1999. Electronic resource.
8. Gutman R, Rubin DB. Robust estimation of causal effects of binary treatments in unconfounded studies with dichotomous outcomes. Stat Med 2013; 32:1795–1814.
9. McCaffrey DF, Ridgeway G, Morral AR. Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychol Methods 2004; 9:403–425.
10. Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika 1983; 70:41–55.
11. Holland PW. Statistics and causal inference. J Am Stat Assoc 1986; 81:945–960.
12. McCaffrey DF, Griffin BA, Almirall D, et al. A tutorial on propensity score estimation for multiple treatments using generalized boosted models. Stat Med 2013; 32:3388–3414.
13. Austin PC, Schuster T. The performance of different propensity score methods for estimating absolute effects of treatments on survival outcomes: a simulation study. Stat Methods Med Res 2016; 25:2214–2237.
14. Lee BK, Lessler J, Stuart EA. Improving propensity score weighting using machine learning. Stat Med 2010; 29:337–346.
15. Pezzi A, Cavo M, Biggeri A, et al. Inverse probability weighting to estimate causal effect of a singular phase in a multiphase randomized clinical trial for multiple myeloma. BMC Med Res Methodol 2016; 16:150.
16. Cole SR, Hernan MA. Adjusted survival curves with inverse probability weights. Comput Methods Programs Biomed 2004; 75:45–49.
17. Lunceford JK, Davidian M. Stratification and weighting via the propensity score in estimation of causal treatment effects: a comparative study. Stat Med 2004; 23:2937–2960.
18. Dunkler D, Ploner M, Schemper M, et al. Weighted Cox regression using the R package coxphw. J Stat Softw 2018; 84:1–26.
19. Ridgeway G, McCarey D, Morral A, Burgette L, Griffin BA. Toolkit for weighting and analysis of nonequivalent groups: a tutorial for the twang package. Available from: https://www.rand.org/statistics/twang/downloads.html. Accessed March 1, 2020
20. Xie J, Liu C. Adjusted Kaplan–Meier estimator and log-rank test with inverse probability of treatment weighting for survival data. Stat Med 2005; 24:3089–3110.
21. Le Borgne F, Foucher Y. Package ‘IPWsurvival’: Propensity Score Based Adjusted Survival Curves and Corresponding Log-Rank Statistic. Available from: https://CRANR-projectorg/package=IPWsurvival 2017. Accessed March 1, 2020
22. Hazard L, Tward JD, Szabo A, et al. Radiation therapy is associated with improved survival in patients with pancreatic adenocarcinoma: results of a study from the Surveillance, Epidemiology, and End Results (SEER) registry data. Cancer 2007; 110:2191–2201.
23. Stessin AM, Meyer JE, Sherr DL. Neoadjuvant radiation is associated with improved survival in patients with resectable pancreatic adenocarcinoma: an analysis of data from the surveillance, epidemiology, and end results (SEER) registry. Int J Radiat Oncol Biol Phys 2008; 72:1128–1133.
Keywords:

Cox proportional hazard models, Generalized boosted models, Inverse probability of treatment weighting, Pancreatic adenocarcinoma, Propensity score, Survival analysis

Copyright © 2021 The Chinese Medical Association, Published by Wolters Kluwer Health, Inc. under the CCBY-NC-ND license.