Impact of data gaps on the accuracy of annual and monthly average daily bicycle volume calculation at permanent count stations

Abstract

This research explores the impact of missing rate of cycling count data on the accuracy of monthly and annual average daily bicycle volume estimates (MADB and AADB). The study made use of a full year of daily bicycle counts at six count stations in Vancouver, Canada. Two missing data patterns were simulated in this study: Completely at Random (MCR) and Not Missing at Random (NMR), also known as the systematic pattern. In the first pattern, repeated random samples of daily bicycle count of different missing rates were drawn from the full data set and used to calculate MADBs and AADB at each count station. In the second pattern, long period data gaps were assumed for periods of one week to four months and MADTs and AADBs were calculated. The estimates calculated from incomplete data were compared to the actual estimates and the errors for each scenario were determined. The results showed that the impact of missing counts on the estimation accuracy of the AADB is minimal where the errors did not exceed 5%, even for high missing rates. This is conditional on that the data is missing randomly and there are a few samples that cover each month of the year. On the other hand, the estimation errors of MADBs were found to be relatively high when the missing rates were high. These results indicated that even if half of the permanent counter data is missing at random, the maximum estimation error would not exceed 14%. The combined impact of AADB and MADB estimation was captured by comparing the MFs calculated using full data versus those calculated by incomplete data. The results showed maximum errors of 94% and 34% for missing rates of 90% and 70%. For the scenario of long period data gaps, the maximum estimation error of AADB ranged between 1.5% and 21.1% when data was missing for one week to four months. Disaggregate error analysis showed that missing data of July would have the most negative impact on the estimation accuracy of AADB. Finally, a Multiple Imputation (MI) method was applied to fill in data gaps for high missing rates. The method led to a maximum AADB estimation error of <3% even if four months of data were continuously missing at one count station.