On the relation between likelihood ratios and p-values for testing success probabilities

Abstract: We investigate the asymptotic relation between likelihood ratios and p-values. We do that in a setting in which exact computations are possible: a coin-tossing context where the hypotheses of interest address the success probability of the coin. We obtain exact asymptotic results and conclude that the p-value scales very differently than the likelihood ratio. We also investigate the p-value of the likelihood ratio, that is, the probability of finding a more extreme likelihood ratio under the various hypotheses. Here we also find explicit asymptotic relations, with similar conclusions. Finally, we study the expected value of the likelihood ratio in an optional stopping context. Our results imply, for instance, that in a coin-tossing context, a p-value of 0.05 cannot correspond to an actual likelihood ratio larger than 6.8.