Are deleted Tweets a source of bias and what can we learn from studying phantom tweets?

Social media analysis brings the promise of big data sets, new meta data and the unprecedented inclusion of direct public response. At the same time, social media is known for several biases when it comes to composition of its userbase and their respective experiences of for example echo chambers. Researchers survey social media platforms often by using special APIs to retroactively gather data on past to present activities. But how is deleted content represented in data gathered this way? Twitter for example will not present any data of deleted tweets, expect for the state of being deleted. For Twitter in particular, one can also observe its universe just in time. A special live API enables researches to follow posts of users, hashtags or other elements in real time. This study recorded a year of tweeting by MPs from German federal parliaments to construct a data set of 140539 Tweets. A year later, how many of these tweets are still present? What are the factors leading to Tweets being deleted? The inferential analysis of this contribution examines this question on level of MPs, as well as by studying meta data of Tweets directly. The exploratory part of the paper studies the differences between the corpus of preserved deleted Tweets and the corpus of present Tweets using computational text analysis. Finally, the contribution tries to formulate a framework for common biases that must be expected in political science studies using data collected a posteriori. For future analyses, mitigation strategies are discussed.