I’ve run out of ideas for the regression table below, which is from a prediction model for liver transplantation. The paper is available here. The primary outcome was graft failure at 1-year, and the authors fitted a logistic model. So far, so good. Here are the results.

Looking at the first two columns, it appears that all the continuous variables have been dichotomized except the last one, which was not continuous. The next column is the coefficients \beta. As this is a logistic model, it must either be a log odd ratio or an odds ratio. Unfortunately, all signs are positive, which leaves us in the dark. Using common sense, maybe the OR is reported because exp(8.57) would be insane? It seems that for one of the variables with three categories, one coefficient was not reported. Also, one coefficient seems to be on the wrong line; no way Retransplantation=NO is associated with a higher risk. The whole paper’s point is that retransplantation is a big no-no (using a DCD liver). The other coefficients make more sense (e.g., younger donors, smaller BMI, shorter ischemia time all reduce the chances of graft failure).

However, in the next step, the authors derived a risk score system. For example, with Donor age and using a cuffoff of 60 years, they calculated the median for the younger and older group (46 and 66); they call them midpoints. The difference is 20 years, so they multiply the difference in the medians with the coefficient 20 \times 0.084 = 1.68. Thus, the \beta X is basically the effect when going from 46 to 66 years, and the Risk score is just the rounded version of this.

Given this procedure, I would expect that they did not dichotomize before fitting the regression model but subsequently for the midpoints. This would also explain that there is only one coefficient for functional donor warm ischemia time. But I would also assume that they did this step on the log OR scale; that would make more sense, wouldn’t it? Thus, I conclude:

- They did not dichotomize to fit the logistic model. All variables were entered into the model as continuous data, and the last was binary. Thus, you get one coefficient per variable. However, they did not report the intercept.
- The \beta are log ORs due to the way they constructed the risk score.
- Yes, the 8.571 is insane, reflecting a more than 5,000-fold increase in the chance of graft failure. I would love to see the uncertainty around this effect, but the authors failed to report standard errors or a 95% confidence interval. If the problem is complete separation, I would expect the p-value to be very high, not very small.
- The Table is terrible, or maybe it is me who is very, very confused.

I would be very grateful for tips and advice and whether I am on the right track. I am not a Sherlock Holmes.