Thanks for the great article and I largely agree with your overall analysis and claims regarding miscarriage rates. I hope you don't mind a bit of quibbling over a statistical point, offered to strengthen and refine your case and head-off would-be critics: Survival analysis (including nonparametric and Cox proportional hazards) is a gene…
Thanks for the great article and I largely agree with your overall analysis and claims regarding miscarriage rates. I hope you don't mind a bit of quibbling over a statistical point, offered to strengthen and refine your case and head-off would-be critics:
Survival analysis (including nonparametric and Cox proportional hazards) is a general set of methods for any kind of time-to-event data (not just cancer) designed to handle left and/or right censoring and provide a more refined breakdown of risks over time beyond raw rates. Censoring just means we do not know a data value exactly, but only an upper or lower bound, and so I find the statement "left censoring to 0 weeks is nonsensical" to be nonsensical itself.
Well-conducted survival analysis should enhance and complement comparisons of raw rates, and it seems perfectly reasonable to apply them here given the nearly complete uncertainty of gestational age values less than 6 weeks. Your own brief R analysis supports this and effectively compares survival curves over at-risk time windows for the two cohorts. As you have nicely said, we could do much more with full data, and I would contend this would include well-formed survival models with a left-censoring indicator at 6 weeks.
A lot of "normal" people have derived great value from Kaplan-Meier plots and Cox models for decades. So it is not survival analysis per se that is the problem here, nor Xu's sample size calculation; it is failure to apply survival analysis properly and make unbiased comparisons with raw rates, including adjustments for all potential confounders.
Thanks for your addition. I'd agree that, as long as comparisons were equivalent they must be valid. The point of the article is to show how the people defending the product chop and change how they assess that data point. I think this is the first ever article to address it.
Thanks for the great article and I largely agree with your overall analysis and claims regarding miscarriage rates. I hope you don't mind a bit of quibbling over a statistical point, offered to strengthen and refine your case and head-off would-be critics:
Survival analysis (including nonparametric and Cox proportional hazards) is a general set of methods for any kind of time-to-event data (not just cancer) designed to handle left and/or right censoring and provide a more refined breakdown of risks over time beyond raw rates. Censoring just means we do not know a data value exactly, but only an upper or lower bound, and so I find the statement "left censoring to 0 weeks is nonsensical" to be nonsensical itself.
Well-conducted survival analysis should enhance and complement comparisons of raw rates, and it seems perfectly reasonable to apply them here given the nearly complete uncertainty of gestational age values less than 6 weeks. Your own brief R analysis supports this and effectively compares survival curves over at-risk time windows for the two cohorts. As you have nicely said, we could do much more with full data, and I would contend this would include well-formed survival models with a left-censoring indicator at 6 weeks.
A lot of "normal" people have derived great value from Kaplan-Meier plots and Cox models for decades. So it is not survival analysis per se that is the problem here, nor Xu's sample size calculation; it is failure to apply survival analysis properly and make unbiased comparisons with raw rates, including adjustments for all potential confounders.
Thanks for your addition. I'd agree that, as long as comparisons were equivalent they must be valid. The point of the article is to show how the people defending the product chop and change how they assess that data point. I think this is the first ever article to address it.