On 2016 Apr 19, S. Hong Lee commented:

Many thanks for your comments. The two models (response variables) do not simply represent association between genome-wide markers, previously associated with schizophrenia, and woman’s age at first birth. The two models (response variables) are the functions derived from the relationship between the schizophrenia risk in children and their mother’s age, studied in a large national-wide study (please see reference #11 in the paper). So what we tested was that the relationship between the schizophrenia risk in children and their mother’s age found from the previous (totally independent) study was associated with the relationship between schizophrenia risk profile score (inferred from SNP effects estimated in another independent SCZ GWAS data set) and age at first birth. Although we used SNP effects to infer risk profile score, we did not think the test should be corrected for the number of SNPs. We used a response variable and one set of risk profile score (N x 1) in a single test (not multiple sets of risk profile score).

In addition, I would like to explain a bit more about the variance explained by the predictor in the target data set (R2 in Table 2). The variance explained by the predictor in the target data set is not the actual variance of the underlying effects. It means it can increase more when there is less sampling error (i.e. with a larger samples size). For example, if you estimate SNP effects in a discovery (or training) data set, and the estimated SNP effects are projected into an independent target data set, the R2 explained by the predictor (from the estimated SNPs effects) depends on the estimation error of the SNP effects. So, if the sample size in the discovery data set (SCZ GWAS in our case) is increased, the R2 can be increased (hence lower p value).

On 2016 Mar 31, Anastasia Levchenko commented:

Reading the paper, I did not see a clear mention that the calculated p-values (Table 2) were corrected for multiple testing (and if they were, it is not clear which statistical method was used). For instance, “statistically significant” association between a genome-wide marker and a phenotype is represented by a p-value ≤ 5E-08, if the Bonferroni correction is used. Don’t the two models (response variables), used in the present study, represent association between genome-wide markers, previously associated with schizophrenia, and woman’s age at first birth? If they do, then it is not clear based on what grounds the authors claim that the p-value = 4.1E-04 is “statistically significant” in the present study setting. Near the end of Discussion the authors state that “A caveat of our findings is that the genetic association between risk of SCZ and delayed AFB was only marginally significant given the number of analyses performed, perhaps reflecting smaller sample size and correspondingly larger standard errors for women with delayed AFB, and so require replication in a larger sample”, probably referring to the p-value = 1.4E-02. Again, the reader sees no support for the statement that p = 1.4E-02 is “significant”, although “marginally”.

On 2016 Mar 31, Anastasia Levchenko commented:

Reading the paper, I did not see a clear mention that the calculated p-values (Table 2) were corrected for multiple testing (and if they were, it is not clear which statistical method was used). For instance, “statistically significant” association between a genome-wide marker and a phenotype is represented by a p-value ≤ 5E-08, if the Bonferroni correction is used. Don’t the two models (response variables), used in the present study, represent association between genome-wide markers, previously associated with schizophrenia, and woman’s age at first birth? If they do, then it is not clear based on what grounds the authors claim that the p-value = 4.1E-04 is “statistically significant” in the present study setting. Near the end of Discussion the authors state that “A caveat of our findings is that the genetic association between risk of SCZ and delayed AFB was only marginally significant given the number of analyses performed, perhaps reflecting smaller sample size and correspondingly larger standard errors for women with delayed AFB, and so require replication in a larger sample”, probably referring to the p-value = 1.4E-02. Again, the reader sees no support for the statement that p = 1.4E-02 is “significant”, although “marginally”.