Re: “Dynamic Risk Stratification for Predicting Recurrence in Patients With Differentiated Thyroid Cancer Treated Without Radioactive Iodine Remnant Ablation Therapy” by Park et al .: ( Thyroid 2017;27:524–530)

Dear Editor:

We read with great interest the article by Park et al. that was recently published in Thyroid (1). The authors aimed to validate a new dynamic risk-stratification system for the prediction of structural recurrent/persistent disease in patients with differentiated thyroid cancer who were treated without radioactive iodine remnant ablation therapy. As one of the main results, they reported that the hazard ratio (HR) of recurrent/persistent disease was significantly higher in the structural incomplete group compared to the excellent group (HR = 243.30 [confidence interval (CI) 46.53–1272.08]). This finding is questionable, as very large HRs with extremely wide CIs do not indicate a strong association between the exposure and outcome variables, but result from an insufficient number of observations in the exposure and outcome variables. This problem is called data sparsity and leads to what is known as sparse data bias. This bias inflates the HR and results in extremely wide CIs (2). Park et al. compared the hazard recurrent/persistent disease in the structural incomplete group with the excellent group. As there were only five patients in the structural incomplete group, data sparsity bias would be expected. Hence, it can be said that the aforementioned HR is inflated due to sparse data bias.

There have been a number of approaches proposed to reduce sparse data bias, but these are often difficult to implement and are not always effective. However, in 2015, Greenland et al. proposed a method called penalization, which more efficiently removes or reduces bias in logistic and related categorical and survival regressions (2). Hence, we respectfully suggest the authors apply the penalization method in their current and future work in order to obtain more unbiased measures of associations, with narrower and more precise CIs.

A take-home message for readers is that sparse data bias is a common problem in medical research (3,4), which should be considered and addressed by researchers using advanced statistical methods.