How we calculate your chances

What the tool claims, and what it does not

The tool estimates what happened to a large group of women with similar recorded characteristics, tracked to whether they had a baby. It is not a prediction about any one person, not a diagnosis, and not advice on whether to pursue treatment at all.

The model's own authors advise against using it to decide whether to have IVF in the first place. It estimates outcomes under treatment, not the benefit of treating versus not treating, which would require data on comparable untreated couples that does not exist at a national level. See our medical disclaimer for how this, and every other estimate on the site, should be read.

The model

Developed by McLernon and colleagues, published in the BMJ in 2016, using 113,873 women in the HFEA national registry between 1999 and 2008. Externally validated and updated by Ratna, Bhattacharya and McLernon in Human Reproduction in 2023, using a further 91,035 women treated between 2010 and 2016. Full citations are in the references below.

We use the 2023 updated version. The original 2016 coefficients calibrated poorly against the newer cohort, and several changed direction: relationships that held in the 1999 to 2008 data did not hold in the newer one.

The model predicts live birth, not pregnancy, cumulatively across up to six complete cycles, meaning one egg collection plus every fresh and frozen transfer arising from it. We use the pre-treatment Method 3 (model revision) coefficient column before a cycle starts, and the post-treatment Method 2 (logistic recalibration) column once egg and embryo details are known, so a reader can check our arithmetic directly against the paper's own tables.

Both papers are open access, and the 2023 paper is CC BY 4.0.

The three places our numbers differ from the published paper

A clinician checking our work needs to know precisely where we did not simply copy the published tables. There are three places, stated plainly.

One: a corrected coefficient. The published table gives 0.040 for a double blastocyst transfer in the post-treatment model. Logistic recalibration multiplies every coefficient in that column by a single factor, stated in the paper's supplementary material as 0.684, and 23 of the 24 values in the column match it exactly. This one value is off by a factor of ten. We use 0.398, the value the column's own pattern implies. We believe this is an uncorrected typographical error in the source table; the 2024 correction to this paper explicitly does not alter any coefficient, so the error was not caught there either.

Two: the treatment year. The model includes year of treatment as an offset from the most recent year in the underlying data, not as a calendar year. We pass zero, which the paper's supplementary material states gives the most current prediction the model supports. There is no extrapolation to a future year.

Three: values for details not yet given. When a field is left blank, we use the average across the study population rather than assuming the single most common profile: 4 years of infertility, 63% with no previous pregnancy, and the recorded prevalence of each diagnosis. Setting the diagnosis flags to zero instead would assert that none of the four named causes applies, which describes a minority of patients and would push every incomplete estimate slightly downward.

The range, and why it is not a confidence interval

The range shown in the tool is not a confidence interval. It is the spread the answer could take across the details not yet supplied, so it narrows as fields are completed and disappears once everything is given. At age 34 with no other detail supplied, the range is roughly 31 to 41 out of 100, because duration of infertility, diagnosis, and pregnancy history alone can move the figure by about that much.

We cannot produce a genuine individual confidence interval for a specific answer. That would require the model's full variance-covariance matrix, which is not published.

How accurate it is

The updated pre-treatment model has a c-statistic of 0.67; the post-treatment model, 0.75. A c-statistic measures how well the model separates women with a better prognosis from women with a worse one. It does not mean the model predicts any individual outcome correctly 67% or 75% of the time.

The authors report this as being among the highest discrimination achieved for IVF prediction models to date. Calibration, whether the predicted numbers match reality across the population, matters more than discrimination for a tool like this, because the population being predicted for is relatively homogeneous: broadly comparable women undergoing broadly comparable treatment.

What the model does not account for

The model does not use AMH, antral follicle count, BMI, ethnicity, smoking, or alcohol use. Endometriosis is not a predictor either, despite affecting about 7% of the study population.

A patient with an endometriosis diagnosis is treated as having none of the four causes the model does measure. That does move her estimate slightly, not because endometriosis itself carries any weight in the model, but because ruling out tubal factor, male factor, anovulation and unexplained infertility is itself information the model uses. Our tool's own copy stated this incorrectly at launch, claiming endometriosis had no effect on the number at all. That was wrong, and is corrected here and in the tool itself.

The AMH point is worth restating precisely: the largest externally validated model of cumulative live birth contains no measure of ovarian reserve at all, and still predicts live birth to a useful standard. See what actually decides whether IVF works for what that means in practice.

Where the data is weaker than it looks

The diagnosis effects are small, and not all of them are stable across model versions. Male factor infertility carried a coefficient of −0.101 in the 2016 model and +0.051 in the 2023 revalidation: it reversed direction entirely. Treat the four diagnosis causes as modest adjustments that have shifted between versions, not as settled facts. It is also why the tool warns that adding a diagnosis can move an estimate either way.

In the 2010 to 2016 validation cohort, duration of infertility was 97% missing and previous pregnancy 100% missing. Both were imputed from 1998 to 2007 records, after the HFEA removed those questions from its own forms. Two of the eight model inputs therefore rest on imputation in the very cohort the coefficients come from.

The cohort's mean age was 35. Below about 25 and above about 43, the estimate rests on comparatively few women, and the tool flags this at those ages.

The underlying data is British, and it is now roughly a decade old.

Why we cannot give you an Indian figure

The statutory National ART and Surrogacy Registry publishes clinic registration, not treatment outcomes. The last multi-centre Indian dataset with real reach reports pregnancy rate rather than live birth and could not stratify by age, because clinics did not submit it. One Bangalore hospital cohort clears a real quality bar but is one hospital's results, not a national figure. What actually decides whether IVF works covers this in more detail.

Our figures describe British patients, and we do not know precisely how they transfer to an Indian population.

Your data

Every calculation runs in your browser. Nothing you enter is sent to us, stored, or logged. We do not receive it and could not produce it if asked. See our privacy policy for how this applies across the rest of the site.

Who reviewed this

This page has not yet been reviewed by an independent clinician. We will name them here, with their credentials and the date reviewed, once that review has happened.

Corrections, and how to tell us we are wrong

If you find an error, including in the maths, the citations, or how we have represented the source papers, email contact@fertilitydecoded.org. We correct mistakes and list them here, dated.

No corrections have been made since this page was published.

When we will update this

We will revisit this page when the model is further validated or updated, when HFEA publishes materially newer figures than the 2010 to 2016 cohort, or when an Indian dataset appears that meets the bar set out above: live birth as the outcome, and a real age breakdown.

References

• McLernon DJ, Steyerberg EW, Te Velde ER, Lee AJ, Bhattacharya S. Predicting the chances of a live birth after one or more complete cycles of in vitro fertilisation: population based study of linked cycle data from 113,873 women. BMJ 2016;355:i5735. DOI: 10.1136/bmj.i5735.
• Ratna MB, Bhattacharya S, McLernon DJ. External validation of models for predicting cumulative live birth over multiple complete cycles of IVF treatment. Human Reproduction 2023;38(10):1998-2010. DOI: 10.1093/humrep/dead165.
• Correction to the above. Human Reproduction 2024;39(7):1580-1581. DOI: 10.1093/humrep/deae099. This correction affects worked examples in the original paper only, not the model coefficients used here.

The model's authors run their own calculator at opis.asf.abdn.ac.uk, implementing the same models described on this page. If you want a second opinion on our arithmetic independent of this site, that is the place to check it.