Cerebrospinal fluid cytotoxicity does not affect survival in amyotrophic lateral sclerosis; Methodological issues
We have read the article written by Galán et al.,1 published in the Acta Neurologica Scandinavica in December 29, 2016, meticulously. The study purposed to examine the potential clinical consequences related with the cytotoxicity of cerebrospinal fluid (CSF) in a cohort of patients with amyotrophic lateral sclerosis (ALS). The authors collected CSF from 31 patients with ALS, in which 21 patients (67.7%) presented a cytotoxic CSF. Then, some variables including cytotoxicity, gender, ALS form of onset, and age at diagnosis were entered into multivariate Cox proportional hazard (PH) regression model to determine which one of them independently predict the survival from diagnosis to death. Although the present investigation made noteworthy contribution to the area, some methodological issues must be noted to avoid misinterpretations.
First, the association between independent variables with survival rate was assessed using the Cox PH models, while PH assumption as one of the major assumptions has not been evaluated graphically or using the appropriate methods such as time‐dependent covariates methods. This assumption has been violated in the study conducted by Galán and colleagues which is shown in their article as Figure 1.1 Therefore, extended Cox regression models is proposed to be applied when this assumption is not established to avoid any misleading results.2
Second, a valuable cohort study was conducted on ALS patients, whereas the sample size is not determined. The sample size must be calculated in epidemiologic investigations, and it is prerequisite to found an association and to achieve a requested precision in parameter estimates.3
Third, in the study conducted by Galán et al., forward stepwise regression model was applied to build multivariate Cox regression model, whereas the study sample was not enough and significant associations would be changed to non‐significant association as the penalty of low sample size is power reduction of statistical tests. As rule of thumb, at least 10 observations are proposed to be added to initially calculated sample size per each variable is included to multivariate model.5 In the aforementioned study, not only initial sample size is not calculated but also the calculated sample size has not been increased per each variable included.
Finally, authors indicated that the presence of cytotoxicity in CSF does not predict studied outcome which is optimistic interpretation. The study should be performed with enough sample size, and the prediction models need to be validated through the appropriate methods.