Can't replicate kdm0

[drstrupf]

Dear Dayoon Kwon,

thank you for this nice package! I've successfully used the PhenoAge functionality to create custom models, and am now looking into KDM. As a first step, I repeated the training on NHANES 3 you did when preparing the data sets (lines 944 and following in nhanes_all.R). I.e. I used the same set of biomarkers and filter conditions for training:

train_fem_kdm0 = kdm_calc(data = NHANES3 %>% filter(gender==2 & age>=30 & age<=75 & pregnant==0), c("fev","sbp","totchol","hba1c","albumin","creat","lncrp","alp","bun"), fit = NULL, s_ba2 = NULL)

However, in train_fem_kdm0$data I then found that kdm and kdm0 are not equal, with kdm = (kdm0 - age) * x + age, where x is approximately 2.175 (min. 2.171, mean 2.175, max. 2.180). So this factor only almost constant between subjects.

Could you please help me understand what I'm doing wrong?

Thank you very much, and best regards
Ste

[dayoonkwon]

kdm0 and kdm are expected to differ. kdm0 was pre-computed using the original Levine 2013 algorithm with fixed published coefficients. kdm_calc re-estimates all parameters (k, q, s, s_ba2) from your training data. Even with the same biomarker set, the re-estimated parameters won't be identical to the original published values, which produces the scaling difference you're seeing. This is by design — the package allows you to train new versions of the algorithm rather than only reproducing the original.

[drstrupf]

Dear Dayoon Kwon,

Thank you for your reply, and sorry, I did not phrase my question properly. I was specifically asking about kdm0 in the NHANES data you provided. In these data files, you provide the columns phenoage0 and kdm0, which are described as computed using the original algorithms, i.e. Levine 2013 for KDM.

Now for PhenoAge, these published coefficients are hard-coded in nhanes_all.R from line 972 onwards and I can replicate the column phenoage0.

However, the calculation of kdm0 in nhanes_all.R from line 943 onwards is a re-training (no hard-coded coefficients), and if I follow these exact steps, I can not reproduce the kdm0 column in the provided NHANES datasets, even though it is supposed to be calculated by this block of code.

Thus I was wondering what the column kdm0 in the NHANES data files is, and I would be grateful if you could help me understand the dfference between the kdm0 in the data files and the kdm0 computed in nhanes_all.R from line 943 onwards.

Thank you & kind regards
Ste

[dayoonkwon]

The kdm0 values in the provided NHANES datasets were computed with an earlier version of kdm_calc that handled missing biomarkers differently. The current version of kdm_calc includes a rescaling step for missing biomarkers that was not present when kdm0 was originally computed. This is why re-running nhanes_all.R with the current code produces slightly different values. The underlying algorithm and biomarker set are the same — the difference is in how participants with incomplete data are handled during the variance (s_ba2) estimation.

[drstrupf]

Ah I see, thank you!