Modeling the effect of days and road type on peak period travels using structural equation modeling and big data from radio frequency identification for private cars and taxis

This study employed weekday, weekend, road type, type of car that is taxi or private car variables to explain their effect on peak hour travels. In this section, we explain the similarities of variable choice with others studies and state how our study is different from other studies. Van Wee et al. [35] stated in their study the importance of temporal factors on travel behavior, emphasizing that there are opportunities available depending on the time of day chosen. This goes to explain the relevance of weekday, weekend and time of day factors in analyzing travel patterns. Yu et al. [36] used weekday peak hour, weekday non-peak hour and weekend non-peak hour to investigate different crash rate circumstances. Even though this was used to investigate crash rates, whereas our study uses it for vehicular movement, it still serves as valuable variables for peak hour traffic studies. Loudon et al. [37] argued in their research the need to specifically model peak hour travels instead of the traditional way of making assumptions with percentages based on twenty-four-hour estimates. They used 6 am to 9 am peak period and 3 pm to 6 pm off-peak period travel for their research. Our research descriptive statistics, however, confirmed the 6 am to 9 am morning peak period but had a 4 pm to 7 pm evening peak period instead. Chen et al. [38] also employed peak and off-peak hour variables to investigate the influence of congestion and passenger load on bus emissions compared with passenger cars but they only collected data for weekdays. Chiou, Fu [39],Seedam et al. [6] only used 7 am to 9 am morning peak period variable in their study but the former used it to study behaviors of drivers with respect to congestion charges while the latter used it to investigate motorcycles impact on fuel consumption and emissions. Holguin-Veras et al. [40] in their study to report the results of a project that evaluated the effects of a pricing initiative on the behavior of passengers, used variables of weekday and weekend peak and off-peak period travels, and also employed a morning peak period of 6 am to 9 am and evening peak period of 4 pm to 7 pm. Their study was however focused on passenger behavioral changes. Ben-Elia, Ettema [41], also used peak periods in their research but used it to investigate a reward system for not commuting during peak hours of 6.30 am to 9.30 am. Yazici et al. [42] analyzed temporal travel variances between urban roads and highways and concluded that there is a variability of travel time patterns between urban roads and highways. They also used different vehicles for their analysis but did not specifically conclude on the relationship between these vehicle types and peak or off-peak hour travels. These previous studies have used only one or two road types in their investigations whereas our paper employs the choice of four (4) road types (secondary, primary, tertiary and motorway). Tarko, Perez-Cartagena [43] utilized temporal and spatial variables for their investigation on peak hour factor but used site to site analysis instead of the type of road we are using for the analysis. They concluded on the significance of spatial factor on peak hour factor variability, thus further endorsing the importance of our choice for road type as a variable for peak hour investigations.

Nanjing which used to be the capital of China but now the capital of the Jiangsu Province has inherent urban traffic issues of peak time traffic congestion. The Nanjing RFID project which started in 2010 was originally intended to meet the needs of testing vehicle emission standards, provide smart parking management, conduct urban traffic management and analyze and offer traffic monitoring services. Each vehicle registered in Nanjing was installed with an RFID electronic tag and RFID readers were installed with video capturing devices along designated streets in the town. The data recorded by the readers of the vehicles is transmitted to the Nanjing Information Center. By 2015 Nanjing had over 518 base station readers and electronic tags on vehicles.

The RFID data for this research covers the month of May in 2014. The RFID vehicle movement data covers over 881,103 vehicles with a record of over 122,000,000 delivered as a text file. The initial data consist of three datasets: The vehicle path data, the vehicle type data, and the RFID reader data. The vehicle path data is made up of the license number of the vehicles (Lic_No), the date and time the vehicle was spotted by the RFID reader, the vehicle type code, the number of lanes of the road and the RFID reader ID. The vehicle type data is made up of the vehicle type code and the vehicle type. The RFID reader data is made up of the RFID reader ID, the road name that the RFID reader has been installed, the longitude and latitude of the RFID reader and the number of road lanes on which the RFID reader is installed. Samples of the three datasets and their attributes are shown in Tables 1, 2 and 3. The license numbers of vehicles for this paper are scrambled to protect the identity of the vehicle owners. This study utilizes RFID readers covering 496 base stations installed in Nanjing city and used to capture the data.

There are five groups of variables namely weekdays, weekends, road type, vehicle type and peak and off-peak hours that was extracted from the raw data and used for the study. The weekdays have 5 variables (Mondays, Tuesdays, Wednesdays, Thursdays, Fridays) while the weekend has 2 variables (Saturdays and Sundays). The vehicle types used for the study are private cars and taxis. The road types consist of four classes: motorway, primary, secondary and tertiary roads. The road classification was adapted from OpenStreetMap (OSM). The OSM data for Nanjing city was downloaded using BBBike online service (extract.​bbbike.​org). The downloaded data was further converted to ESRI ArcGIS Network data set using the ArcGIS OSM Editor Toolbox. The road types used are a modified version of classification of roads by OSM, where living street and trunk roads are omitted. Readers falling on living street and trunk roads were assigned road types in our classification closest to roads connecting these road types. The RFID readers are installed along roads, thus assigning a road type to each RFID reader. A vehicle tracked by an RFID reader therefore also automatically determines the classification of road that vehicle is using. The peak and off-peak period variables group consist of three variables: morning and evening peak-periods from 6 am to 9 am and between 4 pm to 7 pm respectively and off-peak hours from 9 am to 12 noon.

The vehicle type codes columns in Tables 1 and Table 2 make it possible for each vehicle type in Table 2 to be known. The RFID_ID columns in Tables 1 and 3 also serve as a primary key allowing Tables 1 and 3 to be joined, thus combining vehicle path data to the RFID reader data. Joining Tables 1 and 3 is beneficial for retrieving vehicles’ locations at specific dates and times and thus determining the movement of the vehicles. For example, in Table 4, for vehicle with Lic_No. AV4688, the path of the vehicle from the time 12.37 to 12.50 can be traced and the number of times the vehicle was detected by RFID readers can be determined within this time period.

The study focused on private cars and taxis both of which have the vehicle type code 103 and are designated as small buses as shown in Table 2. Pre-processing and initial viewing of the data was done using PostgreSQL 9.4 software. The data was next exported to SPSS for further processing. The descriptive statistics revealed that private cars and taxis make up over 90% of vehicles in Nanjing. The rest of the vehicle types were therefore filtered out and deleted leaving only private cars and taxis, amounting to a total record of 111,147,589 made up of 830,193 vehicles. The distribution of the total number of records, which represents the number of times the RFID reader detected vehicles for each day of the week is shown in Table 5. To differentiate between which vehicles are taxis and which are private cars, the number of times the RFID reader detected a vehicle and the total counts for that particular vehicle was used. The number of times a vehicle was counted by RFID readers in a day was tried out until an optimal number of counts per day were assigned to private cars and taxis. Counts greater than 25 per day were taken to be taxis and counts lower than 25 per day were considered as private cars. This procedure gave the total number of taxis as 10,228 and private cars as 819,965 which corroborated with the Nanjing statistical yearbook of 2014 and SAP [20].

The variable values were determined using the number of times each vehicle is detected by the RFID reader. The number of detections for each vehicle for Mondays through to Sundays was determined using SPSS for the month of May 2014. Similarly, the total number of detections for peak hours 6 am to 9 am and 4 pm to 7 pm and off-peak hours of 9 am to 12 noon were also determined. Each vehicle is detected on a particular road by the RFID reader with the reader ID used as a reference. Knowing the longitudes and latitudes for each of the RFID readers, the positions of the readers were plotted in a GIS software. The OSM data was draped on the plotted RFID positions after which each reader was labeled with a road type. The roads were inspected with respect to each RFID reader and each reader assigned to the road on which it coincided or with the road closest to it. Thus, each vehicle movement was assigned as motorway, secondary, primary or tertiary roads as shown in Table 4. The total number of times each vehicle used a road type was computed. Peak and off-peak-hour times were chosen based on previous literature as explained earlier in the variables section and also revealed from the one month histogram of May 2014 as shown in Fig. 1. These peak and off-peak periods were also chosen because various literature and investigations used similar peak hours [37, 40, 41]. The rest of the hours displayed similar characteristics as the hours chosen and presumed therefore that they will exhibit same results as hours chosen for study. A sample of values computed and used for the model is displayed in Table 6.

Based on previous study recommendations, a two-step approach was used for the analyses [44, 45]. Before testing of the structural model, the indicators were subjected to Confirmatory Factor Analysis (CFA) to evaluate how each item represents their corresponding latent variables and also determine the validity and reliability of the scale. Because the construct of weekday and weekend were common, the second order CFA was employed. In all, 57 parameters were estimated consisting of 32 regression weights, 6 covariances, and 19 variances.

According to previous studies, goodness-of-fit index (GFI), adjusted goodness-of-fit index (AGFI), comparative fit index (CFI) and normed fit index (NFI) values of 0.9 or above and root mean square error of approximation (RMSEA) of 0.08 or less indicate a good fit between the model and the data [45, 46]. As displayed in Fig. 2, the model fitted the observed variance-covariance matrix very well and was within the acceptable range as the comparisons fit indices were all above 0.09 and RMSEA of less than 0.08 was achieved. The standardized regression weights of the observed variables range from 0.65 to 0.89 with a significant level of p < 0.001, indicating that, the observed variables significantly represent their respective latent constructs and permitted for further analysis. The squared multiple correlations (SMCs) range from 0.42 to 0.79 indicating the explanatory power of the research model.

In Table 7, the result shows that all the composite reliabilities (CRs) were over the recommended value of 0.7, [46], indicating that the scales were reliable. The average variance extracted (AVE) for each construct ranged from 0.68 to 0.77 and the square roots of AVE of each construct were greater than the correlation coefficients with other constructs, suggesting that the scales have good convergent and discriminant validity.

From Fig. 3, it is clear that the independent variables (weekdays) positively predicted dependent variables (the peak and off-peak hours) from 6 am to 9 am, 9 am to 12 noon and 4 pm to 7 pm. On Mondays, the movement of the vehicles for morning peak hours, from 6 am to 9 am (β = 0.20, t-value =97.29, p < 0.001) and evening peak hours from 4 pm to 7 pm (β = 0.21, t-value =135.35, p < 0.001) are more than off-peak hours from 9 am to 12 noon (β = 0.13, t-value =88.95, p < 0.001). Concerning the movement of vehicles on Tuesdays, evening peak hours, 4 pm to 7 pm was higher followed by off-peak hours, 9 am to 12 noon with morning peak, 6 am to 9 am showing the least vehicle movement. The result further show that on Wednesdays the movement of the vehicles for morning peak,6 am to 9 am and off-peak hours 9 am to 12 noon were greater (β = 0.16, t-value = 79.82, p < 0.001; β = 0.15, t-value = 98.06, p < 0.001 respectively) than evening peak hours 4 pm to 7 pm. On Thursdays, the movement of the vehicles for off-peak hours 9 am to 12 noon (β = 0.18, t-value =105.67, p < 0.001) is more, followed by morning peak hours, 6 am to 9 am (β = 0.16, t-value = 70.68, p < 0.001). The least movements among the Thursdays are the evening peak, 4 pm to 7 pm. On Fridays, the movement of vehicles in the morning peak, from 6 am to 9 am (β = 0.21, t-value = 93.78, p < 0.001) and the evening peak hour, from 4 pm to 7 pm (β = 0.24, t-value =137.60, p < 0.001) are higher than off-peak hours 9 am to 12 noon. The Friday evening peak hours are however slightly higher than morning peak hours as also corroborated by findings from Viti et al. [47],Yazici et al. [42]. Generally, weekday movement has a significant influence on peak hours. However, from the result, it is obvious that the coefficient weight of Mondays and Fridays appears to be greater compared to other weekdays.

On the other hand, again from Fig. 3, during weekends, except the path from Sundays morning’s peak from 6 am to 9 am, which is not significant, all the paths coefficients were significant. Nevertheless, the weight of the significance was not strong, suggesting less possibility of traffic congestion occurring during weekends. The movement of the vehicles on weekends appears to be less on all the peak hours compared to the weekdays.

The result further shows that when the weekdays alone were included in the model, the model explained 63%, 80% and 78% of the variance at 6 am to 9 am, 9 am to 12 noon and 4 pm to 7 pm peak hours respectively. When the weekends were included in the model, 1% of the variance was added to the 9 am to 12-noon off-peak hour as shown in Fig. 3. Specifically, considering both the weekends and weekdays in the model, the total amount of the variance explained in the morning, noon off-peak and the evening peak hours were 63%, 81%, and 78% respectively. This also further confirms that weekdays made more significant influence than weekends on peak hours.

From all these results, it essentially means that movements of vehicles during morning peak hours 6 am to 9 am are less dispersed but more significant connoting that morning peak hour movements have similar patterns as compared to evening off-peak from 4 pm to 7 pm. The hours with least patterns are morning off-peak from 9 am to 12 noon, even though day of the week and time of the day have more significance on morning peaks than off-peak hours as also verified by [48]. This is understandable in that for morning peak hours most people are either moving to their workplace, school or regular activity for the day, whereas with off-peak hours, the movement is less predictable as no regular pattern is discerned. For weekends, it is reasonable that there will be more vehicular movement during off-peak hours since most people do not need to meet the morning hour rush time but can decide to travel after 9 am.

In this research, which includes identifying the significance of car types contributing to peak hour traffic in Nanjing, it was important to compare the movement between taxis and private cars which forms the majority of cars in Nanjing. From Table 8, the result shows that on Mondays and Wednesdays, the movement of the private cars is more than the taxis on all the peak hours. On Tuesdays, the movement of the taxis in the morning peak hours, 6 am to 9 am was not significant (β = 0.03, t-value = 1.67) whereas, the movement of private cars (β = 0.15, t-value =96.00, p < 0.001) is significant. However, from 9 am to 12 noon, the movement of taxis appears to be slightly more than that of the private cars. On the other hand, from 4 pm to 7 pm, private cars move more than the taxis. These findings imply that in Nanjing, most people use their private cars for commuting to work during the morning peak hours and going back home during the evening peak hours from 4 pm to 7 pm. However, in between work hours, that is off-peak hours of 9 am to 12 noon most people use public transport or taxis.

The result further indicated that on Thursdays, from 6 am through to 12 noon, the taxis seem to move more than the private cars. However, on the same day, from 4 pm to 7 pm, the private cars (β = 0.14, t-value = 124.71, p < 0.001) move more compared to the taxis (β = 0.12, t-value = 7.91, p < 0.001). Additionally, on Friday mornings, the movement of both private and taxis appears to be in equilibrium, whereas as from 4 pm to 7 pm, the taxis move more than the private cars. Finally, on weekends, the movement of taxis seems to be more than the private cars in Nanjing. Nonetheless, from 4 pm to 7 pm the private cars move more than the taxis during weekends.

From the above findings, it is clear that from 4 pm to 7 pm on both weekdays and weekends, there is substantial movement of vehicles thus causing traffic congestion and this could be attributed to the more use of private cars. In addition, from 6 am to 9 am morning peak hours, in the weekdays, the movements of private cars are more than the taxis, except on Thursdays where the taxis move more than the private cars from 6 am through to 12 noon. This also implies that private cars movement during morning peak hours from 6 am to 9 am is significant and this means that private car usage is still the preferred choice for many travelers due to reasons like comfort, reliability, convenience, privacy [49] and accessibility [50]. Generally speaking, this postulation is in agreement with the hypothesis that private cars contribute to peak hour traffic more than taxis and also conforms with other studies that have proved the preferred use of private cars to public transport [51‐53]. In Table 8, the variance explained in peak and off-peak hours for taxis and private cars are displayed.

From Fig. 4, the movement of both taxi and private cars when combined with all the road types (Motorway, primary, secondary and tertiary roads) had a significant relationship with both peak and off-peak hours. Therefore, hypothesis 3 (The choice of road type by cars (taxi and private cars) will significantly influence the movement in peak and off period) is supported. However, on the motorway, the significant coefficient weight is small for both peak and off-peak hour period compared to primary, secondary and tertiary roads. This means that choice of the motorway by vehicles does not influence vehicular movement for both the peak and off-peak hours. However, choice of road types primary, secondary and tertiary have a significant influence on peak hour travels. This postulation is supported by Yazici et al. [42] who exposed in their study that urban road choice is affected by time of the day.

Furthermore, together with the independent variables (Motorway, primary, secondary and tertiary roads), the model explained 62%, 80% and 76% of the variance in dependent variables, 6 am to 9 am, 9 am to 12 and 4 pm to 7 pm respectively. Based on previous evidence by Cohen [54] it is obvious that the variance explained by dependent variables are high to support the validity of the current study model using road types in predicting various peak hours.

Comparing choice of road types by taxis and private cars, the choice of motorway by private cars was more significant to the various peak hours than the taxis as shown in Table 9, whereas on a tertiary road, the taxis have high significant beta weight than the private cars.

Regarding the choice of primary road, from 6 am to 9 am and 4 pm to 7 pm, the private cars had a highly significant coefficient compared to taxis. This indicates that at the morning and evening peak hours, private cars prefer using the primary roads than the other road types. This could also suggest that most private car owners think the congestion on the primary road will be less thus influencing their choice. However, from 9 am to 12 noon, the choice of primary roads by taxis appears to be slightly higher than that of the private cars. This choice could further be substantiated by earlier findings in this study suggesting that there is more movement of taxis during off-peak hours from 9 am to 12 noon than peak periods. The results further show that the choice of the secondary road by private cars highly predicted the morning peak hour 6 am to 9 am, off-peak 9 am to 12 noon and evening peak 4 pm to 7 pm, than the taxis, suggesting that in Nanjing, the private cars may frequently choose secondary roads compared to taxis. This further suggests that secondary roads are likely to be congested during these peak periods. The choice of secondary roads by private cars could be because of the large population in the study area or prior knowledge of all alternative secondary roads.