I understand how Gibbs sampling works to allow estimations of joint probability distributions in a simple sense. For example, using a simple two dimensional example (A and B are two different tennis players whose results are not independent, and 0=losing, 1 = winning), you might start with the following conditional probabilities:

P(A=0| B=0) = 0.5,

P(A=0| B=1) = 0.125

P(B=0|A=0) = 0.5

P(B=0| A =1) = 0.125

To get things moving, you would then suppose a starting value for A (let it be 0). We can then use that value of A (A=0) for the next iteration for B, where we therefore look at P(B=0|A=0). As shown above, there is a 0.5 probability that B=0 in that case. Let's run a random number (0-1) generator, which happens to yield 0.67. As this is greater than the probability of 0.5, we take B=1 (using the rule that if the random number is lower than the conditional probability, 0 is yielded, and if the random number is greater than the conditional probability, 1 is yielded). This gives us the first pair of joint values: A=0 and B=1 [which can be written as 0,1]. We now run the next iteration for A using that last value of B (B=1). P(A=0|B=1) is 0.125. The random number generator yields 0.28, so we take A =1. We then look at P(B=0|A=1) [as the last value of A yielded was 1], which is 0.125. The random number generator yields 0.34, which means we take B to be 1. So our second pair of values is: A=1, B=1 [or, 1,1]. We can repeat this process for a very large number of iterations, and if we then count the numbers of paired values that are 0,0; 0,1; 1,0; and 1,1, we should be able to estimate the joint probability distribution. I have attached a simple excel program that carries out such a Gibbs sampling. It can be seen that the estimation of the joint probability distribution is very close to the actual joint probability distribution from which the conditional probabilities were calculated (in practice, of course you wouldn't have access to the true joint probabilities as then you'd have no reason to do the Gibbs sampling). I have largely based my example on an excellent YouTube video by Ben Lambert.

However, this is where I need advice and help. I do not understand how the above example relates to network meta-analyses (NMAs). For example, imagine a network meta-analysis of three treatments A, B and C. How do the data from these studies relate to conditional probabilities? For example, if the odds ratio of outcome X is 0.5 for the comparison of A vs B, the odds ratio of outcome X is 0.2 for the comparison of B vs C, and the odds ratio of outcome X is 0.1 for the comparison of A versus C (clearly no incoherence here!), how do we proceed? I have a vague idea that we could use the odds ratios to calculate conditional probabilities, but can't quite grasp exactly what should happen. I have looked at most of the relevant documents (like the DSU document TSD2) but these don't explain exactly what occurs in the Gibbs sampling itself. Can anyone describe, in simple terms, how the sampling would proceed in an NMA, in relation to the model of Gibbs sampling I have given earlier?

More Mark Christopher Perry's questions See All
Similar questions and discussions