As noted by @carlmc.bsky.social, one odd pattern in these results is that ALL the estimates and confidence interval bounds are multiples of 0.008
What could explain that? 🤔
I'm slightly baffled.
Could it be a funky scaling/units issue? But that would not apply to the CIs...
A very weird type of post-hoc rounding? But the p-values are not affected.
So I got Claude to extract estimates, CIs, and p values across all tables in the appendix.
eTable 6 is an outlier with 100% being multiples of 0.008. But all the tables have *some*.
For example, here is eTable 5 with all values that are multiples of 0.008 highlighted in red:
There are so many puzzling errors in these tables.
Here's an estimate that lies outside its confidence interval!
(Presumably it's a sign error: 0.035 would be the midpoint of *-*0.105 and 0.176)
But this type of error shouldn't happen! Tables should be automatically generated by statistical software.
If one estimate from the whole table has the wrong sign, it suggests that the tables must have been manually altered.
That's... not good.
Suggestions / explanations welcome!
paper: www-neurology-org.ezp-prod1.hul.harvard.edu/...
appendix: www-neurology-org.ezp-prod1.hul.harvard.edu/...
csv of appendix tables: github.com/...
(Note: csv extracted by Claude. Refer to Appendix pdf to check values.)
I found 5 cases where the estimate lies outside the reported CI.
Table 7: -0.016 (-0.001; 0.001) ➡️ no idea
Table 9: 0.035 (0.104; 0.176) ➡️ likely sign error, -0.104 0.001 (-0.002; 0.000) ➡️ rounding error? 0.098 (-0.222; 0.026) ➡️ likely sign error, -0.098 0.001 (-0.001; 0.000) ➡️ rounding error?
More strange discrepancies in Table 7.
🚩 Multiples of 0.008 🚩 Same estimate and CI for 2 different outcome vars 🚩 CI bound at 0.000 🚩 p-value inconsistent with reported CI
To get a p-value as high as those reported, the upper bound would have to be closer to 0.004 than 0.000.