In this case study, I describe how Dr.DataScience used a wide range of statistical methods and rigorous data handling to help a client who wished to deeply evaluate, using their data on perioperative treatment for a certain cancer type (e.g., particular pre- and post-operative therapies A and B), how treatment effects and patient characteristics influence recurrence-free survival (RFS) and overall survival (OS).

In medical research, the analysis of events that occur over time (such as disease recurrence or death) is extremely important, but it is not uncommon to face complex challenges such as missing values, data bias, and violations of the assumptions of statistical models.

To address these challenges, Dr.DataScience systematically applied advanced statistical approaches: imputation of missing values by multiple imputation, correction of patient-characteristic bias by propensity score matching (PSM), and selection of the optimal method (RMST) when proportional hazards do not hold. This allowed me to extract reliable insights from the client’s data and provide objective evidence to inform clinical decision-making and future study design. In accordance with our confidentiality agreement, no specific figures or detailed clinical background are disclosed.

Background and objective

Using clinical data, the client wished to evaluate in detail the treatment effects of an existing treatment protocol and a particular new therapy, and to clarify their influence on recurrence-free survival (RFS) and overall survival (OS). In particular, they focused on the following themes.

• Theme I: Evaluating the influence of the magnitude of change in nutritional status on pathological response and RFS.
• Theme II: Evaluating the influence of the completion status of a particular perioperative therapy and postoperative adjuvant therapy on RFS and OS; and evaluating the influence of administering postoperative adjuvant therapy according to the presence or absence of lymph-node metastasis.

In these evaluations, the objective was to overcome the statistical challenges inherent in the data—missing values, bias in patient characteristics, and the assumptions of the survival-analysis model—and to perform reliable treatment-effect and factor analyses.

Data and variables

In this analysis, rigorous, multi-stage methods were adopted, from data-quality assurance to application of the optimal statistical model.

1. Handling of missing values
   • Missing values were confirmed in several patient-characteristic factors (e.g., particular tumor-related factors and multiple items concerning pathological status).
   • By using multiple imputation for the incomplete data containing these missing values, 50 imputed datasets were created. The analysis was then performed on each dataset, and the final results were pooled.
2. View on the normality of continuous patient-characteristic variables
   • For continuous variables such as age, body mass index (BMI), inflammation-related indices, nutritional indices, and electrolyte levels, normality tests indicated non-normality.
   • However, given the shape of the distributions and a sufficient sample size from the standpoint of the central limit theorem, these variables were treated as normally distributed for the analysis.
3. Correction of bias in patient characteristics
   • In the Theme II analysis, propensity score matching (PSM) was performed for a total of four patterns—a particular perioperative therapy (completed vs. not completed), a particular postoperative adjuvant therapy (completed vs. not completed), a particular adjuvant therapy in the lymph-node-positive group (administered vs. not administered), and a particular adjuvant therapy in the lymph-node-negative group (administered vs. not administered)—to correct bias in patient characteristics for each comparison pair.
4. Testing group differences in patient characteristics
   • For testing group differences in patient characteristics before and after PSM, evaluation was made mainly using the standardized mean difference (SMD). Supplementarily, Fisher’s exact test was adopted for categorical variables, and an appropriate test based on the view of normality was adopted for continuous variables.
5. Testing proportional hazards
   • Proportional hazards—the assumption of the log-rank test and the Cox proportional hazards model—was checked using Kaplan–Meier curves (KM plots) and Schoenfeld residuals.
   • Because there were scattered cases in this study where proportional hazards were not satisfied, analyses using survival rate and restricted mean survival time (RMST) as indices were also performed, so that the two approaches complemented each other.
6. Setting the time points for survival rate and RMST
   • The evaluation time point for survival rate and RMST was set at the clinically meaningful five years.
7. Survival-analysis results: Theme I
   • A significant difference for both RFS and OS was found only for the magnitude of change in the nutritional index (“maintained/increased” vs. “decreased”).
   • For this item, significant differences were detected in all of the following: the difference in survival time by the log-rank test, the hazard ratio by the Cox proportional hazards model, and the RMST difference by the RMST model; the five-year survival rate also tended to be higher for “maintained/increased.”
8. Covariate-analysis results: Theme I
   • The analysis centered on univariate and multivariate Cox proportional hazards models, with attention to the change in the p-value of each variable.
   • Variables whose p-values increased suggested the presence of background factors with a similar influence on RFS/OS, while variables whose p-values decreased suggested a possible role of complementing the factors influencing RFS/OS.
   • For RFS, a particular pathological T classification tended to be significant in univariate but non-significant in multivariate analysis, while weight change during preoperative chemotherapy tended to be non-significant in univariate but significant in multivariate analysis.
   • For OS, histological differentiation and the status of the resection margin tended to be significant in univariate but non-significant in multivariate analysis.
   • Four factors—presence of lymph-node metastasis, a particular pathological T classification, status of the resection margin, and presence of pathological complete response—significantly influenced the five-year RMST difference for both RFS and OS, while histological differentiation significantly influenced the RMST difference for OS only.
9. Differences in survival-analysis results before and after PSM: Theme II
   • Differences in survival-analysis results before and after PSM were compared in four patterns.
   • For the particular perioperative therapy (completed vs. not completed), before PSM both RFS and OS were non-significant, but after PSM significant results were obtained for OS in both the log-rank test and the Cox proportional hazards model. However, because this result did not satisfy proportional hazards and the p-value was close to the significance level, caution was deemed necessary regarding its reliability. In the RMST model, it was significant both before and after PSM.
   • For the particular postoperative adjuvant therapy (completed vs. not completed), the RMST difference for OS was significant both before and after PSM, and there were no significant differences in the other evaluation items, so it was judged that there was no change before and after PSM.
   • For adjuvant therapy in the lymph-node-positive group (administered vs. not administered), before PSM proportional hazards held for RFS, and significant differences were detected by the log-rank test and the proportional hazards model. However, after PSM the p-value for RFS rose above the significance level and was no longer significant. Considering that the pre-PSM p-value was close to the significance level, that proportional hazards were ensured after PSM, and that the sample size had shrunk, it was judged that the pre-PSM result should also be taken into account.
   • For adjuvant therapy in the lymph-node-negative group (administered vs. not administered), there was no change before and after PSM in any of the evaluation items, and no significant differences.
10. Covariate-analysis results: Theme II
    • As with Theme I, attention was paid to the change in each variable’s univariate and multivariate p-values.
    • For RFS, a particular pathological T classification tended to be significant in univariate but non-significant in multivariate analysis.
    • For OS, a particular pathological T classification / histological differentiation / status of the resection margin tended to be significant in univariate but non-significant in multivariate analysis, while the completion status of a particular perioperative therapy tended to be non-significant in univariate but significant in multivariate analysis.
    • Three factors—presence of lymph-node metastasis, a particular pathological T classification, and status of the resection margin—significantly influenced the three-year RMST difference for both RFS and OS, while histological differentiation and the completion status of a particular perioperative therapy significantly influenced the RMST difference for OS only.

Overview of the main results and clinical considerations

Through this analysis, multifaceted and detailed insights were obtained regarding the effects of perioperative treatment for the particular cancer type and the influence of patient characteristics.

• Importance of the magnitude of change in nutritional status: Multiple survival-analysis methods consistently suggested that the magnitude of change in the nutritional index (maintained/increased vs. decreased) is an independent prognostic factor with a significant influence on both RFS and OS. This shows that patients’ nutritional status is closely related to treatment effect and prognosis, underscoring the importance of nutritional management.
• Importance of bias adjustment by PSM: Applying PSM corrected bias in patient characteristics, and there were cases—such as the influence of a particular perioperative therapy on OS—where statistical significance emerged after PSM was applied. This means that more reliable treatment-effect evaluation, free of the influence of confounders, became possible.
• The challenge of proportional hazards and the use of RMST: When cases were confirmed in which the proportional-hazards assumption of the Cox model was not satisfied, using RMST as a complementary method showed that different insights can sometimes be obtained. This suggests that, rather than clinging to a single method, a flexible approach suited to the data’s characteristics is essential to enhancing the reliability of clinical research.
• Interpreting changes in p-values between univariate and multivariate analyses: The insight was obtained that cases where a factor significant in univariate analysis becomes non-significant in multivariate analysis suggest confounding by other background factors, while, conversely, cases where a non-significant factor becomes significant suggest a role of complementing that factor’s independent influence. This highlighted the complexity of identifying truly independent influencing factors.

These findings provide solid, data-based evidence for optimizing treatment strategies for the particular cancer type, stratifying patients, and determining the direction of future clinical-study design. In particular, overcoming realistic challenges such as missing values and imbalance in patient characteristics to derive reliable results is a significant step toward clinical application.

Dr.DataScience’s contribution

This case is a good example of how Dr.DataScience can contribute not only to supporting decision-making in the clinical setting but also to enhancing the scientific validity and visual persuasiveness of statistical insights when published as an academic paper. The analyses Dr.DataScience provides aim, through rigorous statistical methods and clear visualization of results, to ensure reliability in the peer-review process and to maximize the impact of the research.

For the data challenges the client faced (missing values, bias in patient characteristics, and the problem of proportional hazards), Dr.DataScience systematically applied multiple imputation, propensity score matching, and multiple survival-analysis methods including the RMST model. This made it possible to delve into the factors behind seemingly contradictory results and to support the scientific clarification of a clinical “sense of unease.” Beyond simply performing analyses, I clearly presented the limitations and strengths of the insights derived from the results, contributing from many angles so that the client could make solid, data-based decisions and so that their findings would be highly regarded academically.