Evidence in incomplete neutralisation: tonal data and Bayesian inference

I take incomplete neutralisation (IN) to be phonological alternations that result in seemingly indistinguishable outputs which nevertheless leave instrumentally detectable differences. German final devoicing is an example. Traditionally, voiced coda obstruents (e.g. in rad) are thought to be devoiced and neutralised with voiceless coda obstruents (e.g. in rat) in a " classic example of a phonological rule " (Wiese, 1996). However, numerous experimental studies have challenged this view, showing fine-grained acoustic and even perceptual difference between allegedly neutralised pairs like rad and rat, the field has yet to reach anything resembling a consensus as to whether IN effects are genuine (Roettger et al., 2014). This is partly due to persistent scepticism of IN studies on methodological grounds. These sceptics suspect a case of " bad data " , and they have good reasons to do so. IN presents a challenging task for experimentalists: we have to test for an effect which will be small (largely imperceptible) by definition, in the face of many potential confounds. These confounds cited by critics primarily include laboratory speech effects and Fourakis and Iverson find that IN in German final obstruents occur only if subjects read word lists that shows minimal pairs. When the experiment looks like morphological paradigm elicitation (e.g. conjugate strong verbs given the infinitives), IN effect becomes statistically insignificant. In another line of objection, Warner et al. (2006) attribute IN in Dutch final devoicing to " spelling pronunciation " , as only minimal pairs that differ in spelling (e.g. vs ) elicit acoustic differences. Another " bad data problem " in the IN literature, surprisingly under-acknowledged, concerns the use of frequentist statistical inference (usually resulting in a p value) for data evaluation. Frequentist tools are fundamentally unable to reject the alternative hypotheses or positively establish equivalence (Gallistel, 2009). For these models, a small p may indicate statistical significance, but a large p is uninformative – the absence of evidence does not constitute evidence of absence. The dominance of frequentist statistics in linguistics and in the IN literature creates a persistent bias towards rejecting the null hypothesis in favour of IN. Studies which do not find significant difference can simply be explained as having insufficient statistical power (e.g. not enough items and subjects; Roettger et al., 2014). I argue that the study of tone sandhi and the use of Bayesian statistical inference provide alternatives that will expand the typology of IN (good …