Paperonium

← Back to the plain-language case

Abstract reports univariate hazard ratios as adjusted estimates and attaches them to the wrong treatments

Lancet Reg Health Eur (2025) Β· doi:10.1016/j.lanepe.2025.101402 Β· PMC12859599

Summary

The Abstract states that "in the adjusted multivariate model" immunotherapy and targeted therapy had hazard ratios (HRs) for second primary cancer (SPC), vs chemotherapy, of 0.470 (95% CI 0.338–0.654) and 0.701 (0.574–0.855). In the Results these same two HR/CI pairs are reported as the univariate Cox estimates, with the treatments the other way round: 0.470 for targeted therapy and 0.701 for immunotherapy. The adjusted multivariate estimates in the Results, repeated in the Discussion, are 0.768 (0.622–0.947) for immunotherapy and 0.556 (0.385–0.803) for targeted therapy. The paper's own event counts and incidence rates support the Results' assignment: the width of each interval matches the number of events in the group the Results name, and the crude rate ratios rank targeted therapy as the more protective treatment. As a result, the Abstract overstates the adjusted risk reduction for immunotherapy (53% vs 23%), understates it for targeted therapy (30% vs 44%), and reverses which treatment looks more protective after adjustment. Both treatments remain associated with lower SPC risk in every version.

What the paper reports

In the adjusted multivariate model, both immunotherapy and targeted therapy were associated with a reduced risk of SPCs: HR = 0.470 (95% CI 0.338–0.654) and HR = 0.701 (95% CI 0.574–0.855), respectively. Additional associated factors with SPCs included metastatic disease, HR = 0.456 (95% CI 0.377–0.552); history of tobacco use, HR = 1.599 (95% CI 1.127–2.269); and current smoking, HR = 1.319 (95% CI 0.923–1.884) p < 0.001.

Abstract, Findings

In the univariate Cox regression, the following factors were explored for their association with the development of SPCs: type of treatment received, age at diagnosis, sex, tumour stage, histological type, and smoking habits were explored. Both targeted therapy and immunotherapy, when compared to chemotherapy, were identified as protective factors with HR = 0.470 (95% CI 0.338–0.654) and HR = 0.701 (95% CI 0.574–0.855), respectively.

Results, "Factors associated with second primary cancer", paragraph 1

Other factors associated with the development of SPCs included metastatic disease HR = 0.456 (95% CI 0.377–0.552), history of tobacco use HR = 1.599 (95% CI 1.127–2.269), and active tobacco use HR = 1.319 (95% CI 0.923–1.884); p < 0.001.

Results, "Factors associated with second primary cancer", paragraph 3 (between the univariate paragraph and the adjusted-model paragraph)

In the adjusted multivariate model, the effect of treatment was confirmed with immunotherapy and targeted therapy, HR = 0.768 (95% CI 0.622–0.947) and HR = 0.556 (95% CI 0.385–0.803); p < 0.001, respectively. Other associated factors such as age HR = 1.028 (95% CI 1.018–1.039), small cell histology HR = 0.704 (95% CI 0.513–0.965), and metastatic disease HR = 0.518 (95% CI 0.428–0.626); p < 0.001, were also found.

Results, "Factors associated with second primary cancer", paragraph 4

Multivariate Cox regression analysis confirmed that both immunotherapy and targeted therapy were protective for SPC compared to chemotherapy with HR = 0.768 (95% CI 0.622–0.947) and HR = 0.556 (95% CI 0.385–0.803); p < 0.001, respectively.

Discussion, paragraph 1

The inputs used for the checks below:

A total of 480 patients (2.3%) developed a SPC. The incidence rate per 100 person-years were 1.059 for the immunotherapy group, 0.661 for the targeted therapy group, and 1.791 for the chemotherapy group. […] By treatment group, 298 patients (62.1%) had received chemotherapy, 142 (29.6%) immunotherapy, and 40 (8.3%) targeted therapy. The proportion of patients who developed an SPC was lowest in the targeted therapy group (1.5%) followed by the immunotherapy group (2.1%) and the chemotherapy group (2.9%) (Table 3).

Results, "Characteristics of the population with a second primary cancer by treatment type", paragraph 1

How the Abstract uses the result:

Patients treated with immunotherapy and targeted therapy exhibited a significantly lower incidence of SPCs. Tobacco use was a key associated factor.

Abstract, Interpretation

What the paper's own numbers show

Method. The Abstract, Results and Discussion values were compared as printed. Two checks show which treatment each HR/CI pair belongs to. (1) The standard error of a log HR implied by a printed 95% CI is ln(upper/lower)/(2 Γ— 1.96). For a Cox HR comparing two groups, this SE is approximately √(1/d₁ + 1/dβ‚€), where d₁ and dβ‚€ are the event counts in the two groups (142 immunotherapy, 40 targeted therapy and 298 chemotherapy events; Results and Table 3). The approximation is close for an unadjusted comparison and somewhat larger after adjustment. (2) The crude incidence-rate ratio of each group vs chemotherapy was computed from the reported rates per 100 person-years. Each CI was also checked for internal consistency: its log-scale midpoint should equal the printed HR. All numbers come from recompute.py.txt.

Quantity Location Reported Recomputed from the paper's own numbers
Immunotherapy vs chemotherapy, "adjusted multivariate" Abstract, Findings 0.470 (0.338–0.654) Results and Discussion give the adjusted HR as 0.768 (0.622–0.947). 0.470 (0.338–0.654) is the Results' univariate value for targeted therapy
Targeted therapy vs chemotherapy, "adjusted multivariate" Abstract, Findings 0.701 (0.574–0.855) Results and Discussion give the adjusted HR as 0.556 (0.385–0.803). 0.701 (0.574–0.855) is the Results' univariate value for immunotherapy
SE(log HR) implied by 0.470 (0.338–0.654) Abstract; Results ΒΆ1 β€” 0.1684. Expected from events: targeted therapy (40 vs 298) 0.1684, immunotherapy (142 vs 298) 0.1020. Ratio implied/expected 1.000 for targeted therapy, 1.651 for immunotherapy
SE(log HR) implied by 0.701 (0.574–0.855) Abstract; Results ΒΆ1 β€” 0.1017. Matches immunotherapy (ratio 0.997), not targeted therapy (0.604)
SE(log HR) implied by adjusted 0.768 (0.622–0.947) Results ΒΆ4; Discussion immunotherapy 0.1072, 1.052 Γ— the immunotherapy event-based value. Consistent
SE(log HR) implied by adjusted 0.556 (0.385–0.803) Results ΒΆ4; Discussion targeted therapy 0.1875, 1.114 Γ— the targeted-therapy event-based value. Consistent
Log-scale midpoint of each CI all four pairs 0.470, 0.701, 0.768, 0.556 0.470, 0.701, 0.767, 0.556. Each pair is internally consistent, so no digit has been garbled
Crude SPC rate ratio vs chemotherapy from rates per 100 person-years (1.059, 0.661, 1.791) β€” Immunotherapy 0.591 (0.484–0.722), targeted therapy 0.369 (0.265–0.513). Targeted therapy is the more protective group, as in the Results' univariate assignment
12-month cumulative incidence ratio vs chemotherapy Abstract; Results (1.22%, 0.85%, 0.67%) β€” Immunotherapy 0.697, targeted therapy 0.549. Same ordering
Hazard reduction, immunotherapy (100 Γ— (1 βˆ’ HR)) Abstract vs Results ΒΆ4 53.0% (abstract) 23.2% (adjusted model). The abstract's figure is 2.28 times the adjusted one
Hazard reduction, targeted therapy Abstract vs Results ΒΆ4 29.9% (abstract) 44.4% (adjusted model)
Which treatment is more protective after adjustment Abstract vs Results ΒΆ4 immunotherapy (0.470 < 0.701) targeted therapy (0.556 < 0.768)
Metastatic disease, listed in the abstract after the "adjusted" treatment HRs Abstract vs Results ΒΆ3, ΒΆ4 0.456 (0.377–0.552) 0.456 is the Results ΒΆ3 value (between the univariate and adjusted paragraphs). The adjusted-model value is 0.518 (0.428–0.626)
Current smoking, "p < 0.001" Abstract; Results ΒΆ3 1.319 (0.923–1.884) The CI includes 1. The Wald p implied by the CI is 0.13 (z = 1.52), so the "p < 0.001" cannot refer to this term on its own
Wald p implied by the adjusted CIs Results ΒΆ4; Discussion ("p < 0.001") β€” Immunotherapy z = βˆ’2.46, p = 0.014; targeted therapy z = βˆ’3.13, p = 0.0017. These are significant at 0.05, but neither is below 0.001; the printed p may refer to a joint test of the treatment variable
Crude SPC proportions Abstract; Results; Table 3 2.9%, 2.1%, 1.5% 298/10,137 = 2.94%, 142/6,820 = 2.08%, 40/2,628 = 1.52% with the followed-up denominators of Supplementary Figure S1 (sum 19,585). These are consistent and not part of the error

So the Abstract's two "adjusted" treatment HRs are the univariate pair, with the treatment labels swapped. The event-count check fits the Results' labelling exactly (ratios 0.997 and 1.000) and misfits the Abstract's labelling by factors of 1.65 and 0.60. The crude rate ratios point the same way. The adjusted estimates in the body are consistent with the event counts and with the expected modest SE inflation after covariate adjustment.

Figure

Abstract vs Results: the same HR/CI pairs appear as adjusted estimates in the Abstract and as univariate estimates, with treatments swapped, in the Results

Figure 1. Where each treatment HR appears in the paper and how it is labelled. Orange text gives the values as printed. The left panel is the Abstract's "adjusted multivariate model". The upper right panel is the Results' univariate Cox paragraph, which carries the same two HR/CI pairs with the treatment labels reversed. The lower right panel is the Results' adjusted multivariate paragraph, repeated in the Discussion, which carries different values under the same model label as the Abstract. The blue text below summarises the checks recomputed from the paper's own event counts and incidence rates (see recompute.py.txt). Both checks agree with the treatment labels in the Results.

Standard error implied by each printed CI vs the standard error the event counts imply

Figure 2. For each of the four printed treatment HRs, the orange diamond is the standard error of the log HR implied by its 95% CI. The blue vertical lines are the approximate standard errors expected from the event counts, √(1/d₁ + 1/dβ‚€): 0.102 for immunotherapy vs chemotherapy (142 vs 298 events, solid) and 0.168 for targeted therapy vs chemotherapy (40 vs 298 events, dashed). The interval of HR 0.470, which the Abstract assigns to immunotherapy, sits exactly on the targeted-therapy line. The interval of HR 0.701, which the Abstract assigns to targeted therapy, sits on the immunotherapy line. The two adjusted intervals from the Results lie just to the right of their own group's line, as expected after covariate adjustment. The axis starts at zero.

Why it matters

Explanations considered

Status

Suggested correction

  1. In the Abstract, Findings, replace the treatment sentence with the adjusted estimates the Results and Discussion report: immunotherapy HR 0.768 (95% CI 0.622–0.947) and targeted therapy HR 0.556 (95% CI 0.385–0.803), vs chemotherapy. If the univariate values are kept as well, label them as univariate and attach them to the right treatments: targeted therapy 0.470 (0.338–0.654), immunotherapy 0.701 (0.574–0.855).
  2. Make the Abstract's "additional associated factors" match the model it describes. For the adjusted model, give metastatic disease 0.518 (0.428–0.626), age 1.028 (1.018–1.039) and small cell histology 0.704 (0.513–0.965). Report the adjusted tobacco HRs if the model included smoking, as the Methods list it among the covariates.
  3. Give a p value for each term, or say which test "p < 0.001" refers to. As printed, it is attached to current smoking, whose CI includes 1, and to two adjusted treatment HRs whose CIs imply p = 0.014 and p = 0.0017.
  4. Consider adding a table of the univariate and multivariate SPC models (Cox and, as the Methods describe, Fine–Gray) so that each estimate's model and covariate set is explicit. If needed, revise the Interpretation sentence on tobacco use to match the adjusted results.