In reply: Predicting abusive head trauma.

The first issue relates to the author’s comments that ‘the authors derived likelihoods even when one or more features were unknown (usually rib fractures and retinal haemorrhage at this stage of care). As this data is not missing at random, multiple imputation may have introduced bias’.1 We suspect that ‘likelihood’ is used within the editorial as a synonym for ‘probability’, however in statistical parlance these are distinct concepts. We did not derive likelihoods but derived predicted probabilities of abusive head trauma (AHT) given particular combinations of clinical features. Second, as the author infers,1 multiple imputation is statistically valid providing the data are missing at random (MAR).2 Importantly, the MAR assumption is just that an assumption rather than a property of the data.3 We believe that the MAR assumption is reasonable in this context as clinical decisions about whether to perform a skeletal survey or ophthalmology exam are usually determined by the ‘measured’ presence or absence of other features, that …