The paper ‘Bayesian time series analysis of terrestrial impact cratering’ was published in Mon. Not. R. Astron. Soc. 416, 1163–1180 (2011). There is an error in the write-up of the method in sections 3.2 and 3.4 of the article. It is essentially just an error in the notation used, but this may interfere with a complete understanding of the method. There is no mistake in the concept or in the code, so this does not affect the results, discussion or conclusions in any way.
The issue is with equation (5) (section 3.2, Bailer-Jones 2011), the equation for the measurement model. Roughly speaking the measurement model gives the probability of getting the measured data given the true value. More precisely, it is the probability of measuring the event at time given both the true time of the event, , and the uncertainty, , in the measurement . In other words, rather than being , the measurement model is(and not as was written in the article). Comparing the above to equation (5) in Bailer-Jones (2011), the right-hand-sides are identical to what we had before (the order of and is irrelevant for a Gaussian). In fact, the left-hand-sides are also identical if we consider the data to be rather than , and implicitly conditioned on the measurement uncertainty, . The correct perspective, which was not clear in Bailer-Jones (2011), is that we should think of the measurement model as a model of conditioned not only on the (unknown) true time, , but also on a fixed (measured/estimated) noise model parameter, . Yet as we neither infer nor marginalize over in this work, it has no practical consequence whether we regard this as measured data or a fixed parameter.
This changes the notation in section 3.4 of Bailer-Jones (2011), but not the actual content of the equations or the calculations. One can just replace with and remember that there is an implicit conditioning on . None the less, for completeness, the second, third and fourth paragraphs of this section with the corrected notation are given below.
The consequence of this more precise notation is that it reminds us that the evidence, , is actually conditioned on the fixed measurement uncertainties. This was not explicit in the original article, but the results, conclusions and discussion are entirely unaffected.
AMENDED PARAGRAPHS OF SECTION 3.4 (BAILER-JONES 2011)
…The probability of observing data from model with parameters is , the likelihood for one event. The time series model predicts the true age of an event, which is unknown. Applying the rules of probability we marginalize over this to getis specified becomes conditionally independent of and ; in the second term – the time series model – is independent of . As the data are fixed, we consider both terms as functions of . Both must be properly normalized probability density functions. . The principle of this calculation is illustrated in Fig. 3: the likelihood of an event for a given model is the integral of the probability distribution for the event (equation 5) over the time series model, . Specific cases for the latter are introduced in Section 3.6 below. drops out of the second term due to conditional independence). For a given set of data (crater time series), we calculate this evidence for the different models we wish to compare, each parametrized by some parameters . The parameter prior encapsulates our prior knowledge (i.e. independent of the data) of the probabilities of different parameters, normally established from the context of the problem (see Section 3.7)…