Will admit to having never used a Cox model (or perhaps I did in graduate school and have since forgotten), but just reading the basics, I'm not sure it would be appropriate in this case. My understanding is that the model makes the assumption that the probability of the "hazard" (in this case, a pregnancy) is does not change over time (i.e. the probability of getting pregnant in month n+1 is the same as the probability of getting pregnant in month n, controlling for the other variables). We know that births are seasonal, so the likelihood of getting pregnant in a given month is not constant. Given that vaccination rates were quickly changing during the enrollment period, it could be that the sample sizes of unvaccinated/vaccinated varied greatly between the low pregnancy months and the high pregnancy months.
Very good point. I also suspect that bias of Cox regression increases if the time individuals spend in the study decreases. For this study, it is around 3 cycles on average. I work a lot with time series of credit ratings data (estimation of probabilities of default and such) but I would never use estimators that intransparent.
Looking at the supplemental data from the paper, there are clues as to why the adjustments changed the direction of the signal. In particular, the unvaccinated women reported more frequent sex (smaller % < 1/wk, larger % >4x per wk), and a larger % of unvaxxed reported reported "doing something to increase chances of conception". As always, how we arrive at these "adjustments" can make all the difference. This is why we need more of an open source system so others can confirm the analyses, as the bad cat recommends.
Looking even deeper, I'm quite certain the adjustments were responsible. If you look at page 6 of the supplemental, my interpretation is that they made very large adjustments in the factors listed above.
I understood this paper, and all the others, as really working with unadjusted case counts to produce the "unadjusted" figures, and all strange effects in these figures being due to the estimator chosen. Don't the adjustments (propensity scores and such) only go into the "adjusted" columns?
I don't understand your charts as they seem to be missing the grouping characteristic of each row. For example, in the age chart, there are a bunch of rows and I assume that a row is an age range. In the alcohol chart, I assume the rows are how much alcohol one drinks.
Perhaps your simple point is that the raw fr vs cox fr ratio should be close to 1 so band information doesn't matter. But the band information matters to me so I can verify your numbers from the references.
I omitted the grouping characteristics in order not to distract from the technical aspect (checking if ratios are close to 1, as you said). When the characteristics are present, we can't help but go into narrative mode. Here they are:
Will admit to having never used a Cox model (or perhaps I did in graduate school and have since forgotten), but just reading the basics, I'm not sure it would be appropriate in this case. My understanding is that the model makes the assumption that the probability of the "hazard" (in this case, a pregnancy) is does not change over time (i.e. the probability of getting pregnant in month n+1 is the same as the probability of getting pregnant in month n, controlling for the other variables). We know that births are seasonal, so the likelihood of getting pregnant in a given month is not constant. Given that vaccination rates were quickly changing during the enrollment period, it could be that the sample sizes of unvaccinated/vaccinated varied greatly between the low pregnancy months and the high pregnancy months.
Very good point. I also suspect that bias of Cox regression increases if the time individuals spend in the study decreases. For this study, it is around 3 cycles on average. I work a lot with time series of credit ratings data (estimation of probabilities of default and such) but I would never use estimators that intransparent.
Looking at the supplemental data from the paper, there are clues as to why the adjustments changed the direction of the signal. In particular, the unvaccinated women reported more frequent sex (smaller % < 1/wk, larger % >4x per wk), and a larger % of unvaxxed reported reported "doing something to increase chances of conception". As always, how we arrive at these "adjustments" can make all the difference. This is why we need more of an open source system so others can confirm the analyses, as the bad cat recommends.
Looking even deeper, I'm quite certain the adjustments were responsible. If you look at page 6 of the supplemental, my interpretation is that they made very large adjustments in the factors listed above.
I understood this paper, and all the others, as really working with unadjusted case counts to produce the "unadjusted" figures, and all strange effects in these figures being due to the estimator chosen. Don't the adjustments (propensity scores and such) only go into the "adjusted" columns?
When do we get to list Matthew 2:13-16?
I hope that at some point I can safely quote Matthew 2:20...
I don't understand your charts as they seem to be missing the grouping characteristic of each row. For example, in the age chart, there are a bunch of rows and I assume that a row is an age range. In the alcohol chart, I assume the rows are how much alcohol one drinks.
Perhaps your simple point is that the raw fr vs cox fr ratio should be close to 1 so band information doesn't matter. But the band information matters to me so I can verify your numbers from the references.
Did I miss something?
I omitted the grouping characteristics in order not to distract from the technical aspect (checking if ratios are close to 1, as you said). When the characteristics are present, we can't help but go into narrative mode. Here they are:
https://www.file-upload.net/download-15093297/Fecundity.xlsx.html