Discussion about this post

User's avatar
David Mears's avatar

Thank you so much for this write up. I also scraped the data from HFEA, with some generous radius around London ie excluding far-away clinics.

I then asked Claude Opus 4.8 to calculate the correlation between the success metrics. The correlation is so low that I wonder if it would be a mistake to try and optimise for these metrics - I infer from the low correlation that, rather than indicating a 'general factor' of clinic competence, the success rates might not be much better than random, at which point you should use different signals to choose clinics, such as the surgeon's relevant experience level (especially, as in my case, when I'm looking to get an operation done that doesn't map to either of the metrics, namely a surgery on me, and I'm not female).

Does that seem like a fair inference to you? NB I do not know how correlations are calculated.

From Claude:

The two metrics are only weakly correlated:

┌────────────────┬───────┐

│ Measure │ Value │

├────────────────┼───────┤

│ Pearson r │ 0.28 │

├────────────────┼───────┤

│ R² │ 0.08 │

├────────────────┼───────┤

│ Paired clinics │ 26 │

└────────────────┴───────┘

Plain-English answer: if a clinic scores well on births per embryo transferred, that tells you very little about how it scores on births per egg collection. The correlation is positive but weak — knowing one metric explains only about 8% of the variation in the other. The remaining ~92% is driven by something else.

★ Insight ─────────────────────────────────────

- Why these two diverge: "per embryo transferred" measures what happens after an embryo is in place, while "per egg collection" folds in everything upstream — how many eggs fertilise, how many become viable embryos, lab quality, freezing strategy. A clinic can be great at one stage and mediocre at another, which is exactly why the correlation is low.

- R² vs r: R² (0.08) is the more honest "predictive power" number here — it's literally the fraction of variance in one metric explained by the other. The eye-catching r=0.28 sounds less trivial than the 8% it really represents.

- Small-n caveat: with only 26 paired clinics, an r of 0.28 is not statistically significant (the ~95% threshold for n=26 is around r≈0.39), so the safest reading is "no reliable linear relationship detected."

─────────────────────────────────────────────────

Phoebe Arslanagić-Little's avatar

Thank you for writing this Luzia! I imagine it will be of help and use to many.

5 more comments...

No posts

Ready for more?