Handling Compositional Time Series with Varying Number of Parts

When different polling organisations conduct political party preference polls at different times, different parties might be reported. If the estimated voter shares of these polls are combined into a time series we obtain a compositional time series, but with varying number of parts, thus prohibiting the use of standard compositional time series analysis tools. We discuss the problem and suggest a solution by imputing the unreported parts. The method is applied to a short compositional time series of party preference polls from Sweden.