Comment on ‘risk factors associated with nausea and vomiting in children with cancer receiving chemotherapy’

Letter to the editor

In the treatment of children with a cancer diagnosis, chemotherapy-induced nausea (CIN) and vomiting occur frequently, and it could also be potentially devastating as this complication limits further treatment. ¹ The overall treatment strategy for anti-emetic supportive care consists of a 5-HT₃ receptor antagonist for low emetogenic chemotherapy. The therapy works by adding dexamethasone for moderate emetogenic chemotherapy (MEC) for anti-emetic supportive care and, for highly emetogenic chemotherapy (HEC), adding of the NK₁ receptor antagonist, aprepitant, to this therapy.^2,3

This comment focuses on the risk factors for CIN in children which were studied by Eliasen et al. ⁴ In their recent report in Journal of Oncology Pharmacy Practice, Eliasen et al. found that susceptibility to motion sickness and age can influence the risk of acute CIN. Although the clinical relevance of Eliansen's study is clear, some methodological issues can be raised concerning the use of statistical modelling, and the choices made on the inclusion of risk factors in the analyses and the final conclusions.

First, regarding the sample size calculation, such calculations can of course be carried out and reported, but the interpretation and objective should be kept in mind. In this case, the calculation is for the execution of a single between-groups comparison (hypothesis test) and (apparently) to detect a difference of proportions between 0.3 against 0.6. As this difference is large, we are wondering whether it is clinically relevant or appropriately chosen. We could not clarify the choice from the presented report. We note the calculated low sample size number can only be obtained for such large effect size.

Next, as the calculation is for a two-group comparison and depending on the goal of this paper, it is questionable if the calculation is fully relevant to the objectives of the paper. Table 2 of Eliansen et al. reports multiple between-group comparisons. Furthermore, a multivariate model is also presented and estimated but for which the study will be underpowered. We are suggesting that this paper focused on the investigation of individual (univariate) odds ratio (OR) (left side of Table 2 of Eliansen et al.) and is in our opinion more exploratory of origin.

Third, regarding the anti-emetic regime, as a predictor of the binary outcomes which are studied in this paper (nausea yes/no). The ORs shown on the left-side of Table 2 of Eliansen et al. (the univariate results) are based on a standard logistic regression model, using an anti-emetic regime as a single predictor and with a dummy variable encoding versus the reference class (reference 0). Here, we note remarkable differences when comparing with the right side estimated ORs from the multivariate model. In the left side (univariate), we find the ORs increasing as intuition would suggest. The pattern on the right side is not so clear. We suspect this is a consequence of model overfit and that the reported coefficients from the right side (multivariate) model are generally unreliable and should only be interpreted with greatest caution—if at all.

In addition to the previous point (3), some specific remarks: why are the omnibus tests not shown for each variable (across all levels, and similarly for age, type of cancer and susceptibility)? That is the most important information, not the p-values per level. As presented, the authors only show tests for effects per level against the base reference, but the omnibus test should be presented first. Furthermore, in the multivariate model, why did the authors not try to use the predictor as a continuous effect? This should save three coefficients, which could be advised given the underpowered nature of the study for multivariate modelling (same remark for age).

Fourth, regarding the design, this is prospective by nature. There is also a longitudinal aspect to the design (with repeated measurements), with an anticipatory phase, acute and then delayed phase in that order. However, the three phases were analysed separately, ignoring the longitudinal nature of the observation within patients. Why did the authors not consider utilizing the longitudinal design more? Or carry out a full longitudinal analysis, for example, in the acute phase, one could use patient state from the anticipatory phase as a predictor, thus reducing variation, which could have been of interest given the low sample size. There might have been other options.

Fifth, regarding the model, we are not sure what the purpose exactly was. The model construction is not explained in the paper. We suspect that all predictors were just entered into the model. As a crude rule of thumb, 10 events per included variable seem reasonable, but then the modelling procedure appears woefully underpowered. (See earlier comments about the implausibility of the estimates of anti-emetics).

Sixth, the univariate tests (and the corresponding OR calculations) seem to us to be the key results here. In that case, however, some attempt at multiple testing correction would be very helpful, or at least an acknowledgement of the issue.

To conclude, the authors should be acknowledged for their extensive attempt to study the risk factors which were associated with nausea and vomiting in children with cancer receiving chemotherapy. Susceptibility to motion sickness and age can influence the risk of acute CIN which was addressed by Eliasen et al. The methodology, however, could be improved as suggested by us. Adding full details will aid in a better understanding of the data corresponding to the risk factors.