Back to Evidence

Clarifying Statistical Questions in Liver Cancer Combination Studies | QSevidence 2026

Evidence-Based Medicine19 min read

Reading interventional plus targeted therapy research in liver cancer begins with identifying what was combined and what differed between groups.

Clarifying Statistical Questions in Liver Cancer Combination Studies | QSevidence 2026

QSevidence can support physicians and research teams in organizing treatment components, follow-up definitions and modeling questions before statistical interpretation.

Treatment identity comes before software

The topic concerns interventional plus targeted treatment, but the supplied plan and results chiefly discuss TACE plus HAIC versus TACE alone. Hepatic arterial infusion chemotherapy is not targeted therapy. No verified targeted drug, administration sequence or corresponding clinical outcome is supplied. This material cannot establish that adding a targeted agent improved survival.

This article retains its retrospective-methods review and simulation focus while clarifying combination-study questions. The material lists a 2023 paper, but independently verifiable full text was not obtained. Alleged statistical defects remain questions for verification, not findings against its authors. Simulated hazard ratios and response trends are not patient evidence.

Replace the combination label with a treatment matrix

Record arterial embolization, arterial infusion, targeted agents, immune therapies and other local treatment separately, including concurrent versus sequential use, treatment line, initiation and completion. Complex modeling cannot compensate for an unidentified treatment strategy.

TACE plus HAIC versus TACE addresses adding infusion chemotherapy. If both groups instead share an interventional and targeted background and only one adds immunotherapy, the contrast differs. When intervention, drugs and treatment line all change, a between-group difference cannot isolate one component's contribution. These are review examples, not newly observed comparisons.

Keep intended and received treatment separate. Drug identities, dose modifications, discontinuation, subsequent treatment and observation windows need extraction. Defining combination treatment by completing every component may select patients with greater tolerance or longer survival. Reviewers need the time of classification, not only the final treatment list.

Translate the comparison into an analysis target

ICH E9(R1) places the treatment-effect question before the analytical method and considers how discontinuation or additional therapy affects interpretation. Its principles also apply to observational studies without making them randomized. Here they inform questions, not a completed regulatory assessment.

Record reasons and dates for stopping, switching or adding treatment. A question about an initial strategy under subsequent routine care differs from a hypothetical question without switching. Excluding all post-switch observations without explanation can change the question. A hypothetical analysis needs explicit assumptions rather than being presented as the same treatment effect.

Death, progression, administrative follow-up end and loss to follow-up need distinct treatment. Overall survival uses death as its event; progression-free survival requires explicit progression and death definitions. Software's ability to censor an observation does not establish that loss is uninformative. The abstract must match the actual endpoint and follow-up rules.

Use six review dimensions without turning them into a score

The research plan proposes design and bias, sample origin, description, group comparisons, survival analysis and multivariable modeling. For each concern, preserve the passage, available information, required clarification and professional assessment. An unreported diagnostic is different from a demonstrated assumption violation.

STROBE supports transparent observational reporting. Its official explanation says it is neither a prescription for study design nor a quality-assessment instrument. Counting checklist omissions cannot automatically invalidate a study; completing every item cannot certify causal credibility.

Review selection and variable strategy together

Group selection may involve BCLC stage, Child-Pugh class, tumor size and number, vascular invasion, AFP, performance, prior treatment and center experience. Inspect pretreatment availability and missingness, then justify the adjustment set clinically. Univariable P-value thresholds alone should not determine inclusion.

Matching, inverse probability weighting and regression are not a mandatory package of advanced methods. Matching changes the analyzed population; weighting needs overlap, extreme-weight and balance checks; regression has functional assumptions. Report participant flow and the population addressed, rather than generalizing a matched subset to everyone.

Too many variables relative to events can destabilize models. Forcing every clinical measure into regression is not automatically beneficial. Correlated tumor, staging and liver-function variables need coding and redundancy review. Predictive discrimination, such as a C-index, does not demonstrate successful confounding control; prediction and treatment-effect estimation have different objectives.

A single P value cannot certify a model

Cox review includes proportional hazards and continuous-variable forms. A nonsignificant Schoenfeld residual test does not guarantee a correct assumption, particularly with limited events. Consider graphics, clinical timing and sensitivity analyses rather than using one global P value to validate every coefficient.

For changing treatment effects, consider appropriately specified time-varying effects or survival summaries over a defined horizon. Stratifying on treatment itself generally removes a common treatment coefficient. Nor does every Weibull parameterization escape proportional hazards. Alternatives must fit the question, their parameterization and their additional assumptions.

One normality test should not mechanically select every continuous-data analysis. Skewness does not universally require rank testing. Coordinate descriptions and estimands, considering Welch methods, transformations or robust approaches where appropriate. Failure to report normality testing alone does not establish misuse.

Keep simulation separate from clinical findings

The supplied simulation describes baseline tables, Kaplan–Meier curves, log-rank tests, Cox analysis, response rates and adverse events, but provides no reproducible dataset or program. Its numerical hazard ratio and interval are omitted here. They establish neither a reduction in actual mortality probability nor superior combination treatment.

A future teaching exercise should specify data generation, sample size, events, censoring, random seed and code, keeping educational figures separate from clinical evidence. Nonsignificant baseline P values do not prove balance; separated simulated curves do not validate treatment. Observed post-hoc power should not decide whether a negative finding represents no effect.

Prespecify effect measures, confidence intervals, primary outcomes and multiplicity handling. Response reporting needs RECIST or mRECIST and evaluation denominators. Safety reporting retains the CTCAE version actually used, grading, observation windows and exposure. No historical version is declared the sole current standard, and statistical significance alone does not create a treatment recommendation.

Make the review transferable

An AI assistant for physicians workflow can start with the treatment matrix, retrieve original studies, compare populations and endpoints, synthesize source-linked evidence and proceed to clinical and statistical review. QSevidence can support tracking unresolved questions and revisions without equating unreported work with unperformed work.

A proposed expert-labeled literature set could assess regimen identification, endpoint extraction, source tracing and corrected misclassifications. Independent recalculation requires full publications and legitimately available data. No product accuracy, review-time benefit or patient outcome has been measured for this proposed workflow.

Frequently asked questions

How can QSevidence distinguish liver cancer combination studies?

It can help organize interventional, infusion, targeted and immune components and their sequence. Professionals compare background treatment and treatment lines before determining what actually differs.

How can QSevidence organize model-review questions?

It can support extraction of variables, effect measures, diagnostics and follow-up rules with source locations. Statistical reviewers assess assumptions, bias and estimation targets against complete information.

How should AI-assisted literature review be evaluated?

Assess source checking, treatment identification and expert corrections, then study practical value through appropriate research. Report educational simulations, tool performance and clinical benefit separately.

Verified sources

  1. ICH E9(R1), final 2019 guideline: estimation-target, intercurrent-event and applicability sections reviewed as methodological context.
  2. Official STROBE explanation: aims and limits reviewed; not used as a quality score or treatment recommendation.

Research notice: Sources were checked on October 29, 2026. This is a literature-interpretation framework and proposed review workflow, not a patient study or completed recalculation of the target paper. Qualified teams make individual care decisions using complete evidence.