How much audiological data is needed for convergence? One year!

Authors: Niels H. Pontoppidan1, Jeppe H. Christensen1 , Kang Sun1
1Eriksholm Research Centre, Oticon A/S, Rørtangvej 20, 3070 Snekkersten, Denmark.

Background How representative are studies that analyze hearing aid users’ daily sound environments? Most audiological studies using ecological momentary assessment and data logging with hearing aids have been pilot studies with a few participants and a limited duration and only a few studies have included many participants for a long duration. In this study we investigate the stability of the sound environment estimate distribution in a sample comprising of 53 unique users and 120 unique days.

Methods We analyze the open access data set published by the EVOTION consortium to enable further and open analyzes of the representativity of other dimensions captured by the participants in the EVOTION project. The analyzed data is 3 dimensional having a user ID, the sound environment class (quiet, noise, speech, speech in noise) and the date. We sample days and users from the dataset to investigate how the distribution converges towards the distribution of the full dataset. With 100 repetitions for each sample, we vary the number of users and days and compare the distribution of the 4 classes to the distribution in the full dataset.

Results The results show that the mean square deviation between the distribution of the four classes in the samples and the full dataset follows decreases linearly on a log-log plot with quiet having a shallower slope than the noise class. Inspection of the outliers indicate that once the product of days and users approach 360 then the distribution of the four class only deviates slightly from the distribution in the total dataset.

Conclusion The present investigation indicates that studies of sound environments needs no less than one person-year of data before the sound environment distribution has converges.  Moreover, it suggests that the convergence can be achieved by many combinations of users and durations.