Comment on: Risk of Pancreatic Cancer in Relation to ABO Blood Group and Hepatitis C Virus Infection in Korea: A Case-Control Study

To the Editor: The quality of each analysis is dependent on the background knowledge, and to be bound to our knowledge on the same time when we are applying our awareness for example during analysis and reporting the results. Maybe authors of the paper about the association between pancreatic cancer and ABO blood group (1) are aware from the below mentioned comments; but, usually a third person may observe some hints from outside which persons involved in the work cannot see. This paper (1) consolidate previous literature in this filed. However, I am interested to know why authors did not use conditional logistic regression while they have paired match their controls for age, gender and date of admission or visit? The analysis should be matched (Paired t-test and conditional logistic regression) when our controls are paired matched with our cases. Authors have only studied 34% of eligible cases who had available data about ABO blood group. For preventing selection bias, it is recommended to compare cases with missing data (66%) and the others (34%) specifically when the percentage of missing data is high. Any results can be distorted in these situations. We are not aware are there any differences between participants and non-participants in this study? According to the mentioned information about data gathering, it seems that such data are available but have not analyzed or reported. Usually, we have not easy access to such data. About control group, were there selected from the base of the cohort (case-cohort) or from the persons with similar time of exposure which are enrolling at the time of selecting cases (nested case-control)? They have mentioned that cases and controls were matched for date of admission/visit; but, it is not determined if they are matched for duration of follow up (nested case-control) or not. It can determine type of study, the strength of association and many other details. As mentioned in methods, "multivariate logistic regression analyses, including age, gender, smoking history and diabetes, were performed to estimate the adjusted odds ratios (AOR)." Such sentence dictates me that they have entered all these variables in the model and have run the logistic regression once. So, ORs are adjusted not only for age and gender; but also, for other variables which are included in the model. For example, in table 2, Non-O blood group OR in comparison with O blood group has an OR which is adjusted for age, gender, smoking history and diabetes. However, they have mentioned it is age- and gender-adjusted OR. They do not mean that these are variables that they have analyzed separately in different logistic models. If they want to mention which variables are evaluated in multivariable logistic regression, they should also express HBsAg and anti-HCV, as well. Therefore, according to the text, we should consider these ORs adjusted for all mentioned variables entered the model. Results show that AORs for pancreatic cancer in subjects with blood types A is 1.36 (1.09-1.71; P = 0.08). It should be significant based on confidence interval, because it has not covered the number one. However, according to P value, it seems non-significant (higher than usual cut off point for P value which is 0.05). Is there any explanation by authors? It is not unusual for borderline values. However, it is some more far to be considered borderline. When we pooled data, subgroup differences may be obscured. The log-rank test can detect differences between two different groups like AB and O. Here, the authors have compared all four groups (A, B, AB, and O) at once which may hide the underlying significant difference.