In this case study, I describe how Dr.DataScience used specialized statistical analysis and data-science expertise to help a client with a serious problem they faced during their clinical research: a gap between the analysis results and their hands-on sense of the clinical setting. Finding that one’s own initial data analysis differs from the expected result—leaving one uneasy about publication or treatment decisions—is something many researchers experience.

In response to this concern, Dr.DataScience delved deeply into the characteristics of the data and selected and applied the most appropriate analytical method, providing reliable scientific evidence that allowed the client to move on to the next step with confidence. In accordance with our confidentiality agreement, no specific disease names, treatment names, figures, or individual patient information are disclosed; however, the statistical methods used, the process, and the type of findings obtained are the same as in the actual analysis.

Background and objective

To evaluate the effect of a particular treatment procedure on patients’ recurrence-free survival, the client had already performed survival analysis using a Cox proportional hazards model. However, these results simply would not match the intuition and felt sense, built over many years of rich clinical experience, that “this treatment procedure should not differ that much from the alternative procedure.”

A situation in which the figures obtained do not align with a clinician’s “felt sense” raised doubts about the direction of the research and the interpretation of the results, becoming a major barrier that kept the client from moving ahead with a conference presentation or manuscript submission. The client had a strong wish to identify the root cause of this “sense of unease,” obtain reliable analysis results, and advance their clinical research.

Data and variables

This analysis used anonymized medical and clinical data. The subjects were data extracted from a patient cohort meeting specific conditions. Specifically, the following main types of variables were included.

1. Response (outcome) variables
   • Time until a particular event occurs: time information from a given intervention until a clinically important event (e.g., recurrence, progression of a particular disease, occurrence of a complication, etc.) occurs.
   • Whether the event occurred: information indicating whether the event was actually observed, or whether the data were censored (the event did not occur during the period).
2. Explanatory variables (covariates)
   • Principal comparison factor: the variable indicating the main intervention or status to be evaluated (e.g., different treatments, classification into particular patient groups, etc.).
   • Patient background factors: a wide variety of data such as sex, age group, disease stage, presence of comorbidities, lifestyle, particular laboratory values, and genetic factors. These variables were considered as confounders that could affect the analysis results.

Analytical methods

1. Data-quality check and preprocessing
   • From the medical data to be analyzed, I identified and carefully excluded or imputed inappropriate data points that could undermine the reliability of the analysis.
   • For variables that needed to be converted into a form suitable for statistical analysis, I applied appropriate preprocessing and built the analysis dataset.
2. Verifying the assumptions of the standard survival-analysis model
   • To verify the applicability of the Cox proportional hazards model—the common survival-analysis model the client had used—I examined in detail, using the statistical test Schoenfeld residuals test, whether its key assumptions (e.g., proportional hazards) fit the data.
   • This rigorous verification clearly revealed that, for the principal comparison factor, the model’s assumption was statistically significantly not satisfied. This finding was the very core of the client’s “sense of unease.”
3. Selecting the optimal survival-analysis method and covariates for the data’s characteristics
   • Taking into account the data characteristic of non-proportional hazards, Dr.DataScience adopted the RMST (Restricted Mean Survival Time) model as the central method for the re-analysis, as it does not rely on that assumption. This model directly compares the difference in mean event-free time (time until an event) over a specified observation period, enabling accurate and interpretable comparative evaluation even under non-proportional hazards.
   • To enhance the model’s reliability, in selecting the explanatory variables (covariates) for the covariate-adjusted analysis I carefully chose the optimal combination of covariates so as to avoid the problem of multicollinearity.
   (Figure: Example of the RMST model / Group A [left] vs. Group B [right])
4. Multifaceted analysis using the adopted model
   • ・Univariate analysis
     • First, to evaluate the difference in time-to-event between the principal comparison factors, I performed univariate analysis using the adopted model.
     • The analysis found no statistically significant difference between the principal comparison factors.
   • ・Covariate-adjusted analysis
     • To account for the influence of patient background factors and the like, I performed covariate-adjusted analysis using the adopted model.
     • Even with the multiple selected covariates included, no statistically significant difference was found between the principal comparison factors. Furthermore, I re-ran the adjusted analysis using only the covariates statistically suggested to have an influence, but the result was the same: no statistically significant difference was found.

Overview of the main results and clinical considerations

This advanced data analysis produced an objective and solid result: there was no statistically significant difference in “time-to-event” between the principal comparison factors with respect to the particular clinical outcome. This result matched beautifully the client’s intuition, built over many years of clinical experience, that “there probably isn’t that much difference between the treatment approaches.”

This finding suggests that a particular intervention may not have a clear advantage over the other, which carries great significance for decision-making in the clinical setting. Dr.DataScience’s statistical considerations also touched on the points on which these results might be debated at the time of academic presentation (papers and conferences), contributing further to the credibility of the client’s research.

Dr.DataScience’s contribution

This case demonstrated how deeply and practically Dr.DataScience’s statistical-analysis expertise can contribute to a serious problem in a client’s medical and clinical research.

• ・Scientific clarification of the “sense of unease” and identification of the problem
• For the client’s vaguely felt question—”the gap between my clinical felt sense and the analysis results”—I identified the root cause through a statistical approach of rigorously verifying the assumptions of the existing statistical model, and provided objective scientific evidence.
• ・Choosing the optimal method for the data’s characteristics
• In a situation where applying a common statistical model was difficult, I deeply understood the inherent characteristics of the data (non-proportional hazards) and applied a corresponding advanced alternative method (the RMST model), thereby deriving more robust and reliable results.
• ・Multifaceted verification and ensuring robustness of the results
• Going beyond univariate analysis, I performed detailed covariate-adjusted analyses in multiple patterns, thoroughly confirming the stability and reliability of the results. This process gave the client firm confidence in their findings.
• ・Contribution to academic presentation and provision of practical implications
• The results and their statistical considerations identified in advance the points that might be raised from a statistical standpoint when the client gives future conference presentations or submits peer-reviewed papers, enabling appropriate responses. I also provided practical implications useful for decision-making in the clinical setting.

In the statistical analysis of complex medical data, Dr.DataScience engages sincerely with the client’s research challenges and, with advanced expertise and thorough, side-by-side support, contributes to obtaining reliable insights and improving the quality of clinical research.