Bureaucratic training, consumer knowledge, and work performance: A smartphone field experiment

I introduce two interventions. The first is the kind of training materials that bureaucrats typically receive in China. The second is the kind of consumer knowledge that the bureaucrats receive from dominant consumer products. In addition, I also have a control group that does not receive any intervention.

2023-11-06

2023-11-20

My primary outcomes will be metrics related to the work performances of the bureaucrats. I will also measure whether the bureaucrats have used our mini-program on a daily basis.

Specifically, the metrics will include binary variables, such as whether bureaucrats arrive at their offices on time in the morning and whether they leave their offices early in the evening. The metrics will also include other continuous variables, for example, how many hours bureaucrats have worked on a single day and how many extra hours they work on a single day. I will also measure how often they visit work-related sites and how many sites they visit on a single day.

I will also measure how much carbon footprint do bureaucrats have on a single day. I will also measure how long they spend on reading my interventions, which are articles of different types. Also, I will measure how often they answer questions related to the first intervention correctly. Furthermore, I will record self-reported behaviors that are related to individual carbon footprint.

The self-reported behaviors that are related to individual carbon footprint include check boxes where I ask if the bureaucrats to upload photos that show eco-friendly behaviors. I will use image annotation and hand-coding to identify these behaviors and create variables.

I use a WeChat miniprogram that I have built to conduct this experiment.

My mini-program uses GPS trajectory data on smartphones to estimate the carbon emissions from transportation. The estimation can be broken down into two functionalities. The first function retrieves location trajectory data frequently from the users’ smartphones to estimate the distance and duration of travel and the means of transportation. Studies find that smartphone GPS is accurate in locating individuals in an urban setting because the assisted GPS module has a horizontal location accuracy of fewer than 15 meters most times. When individuals start to use our mini-program, they will choose a means of transportation from a short list that includes walking, biking, taking a bus, taking a subway, taking a taxi, or driving. Once they start a trip, we retrieve their GPS trajectories every 10 seconds so that we can record the speed and travel distance. Additionally, motion sensors, such as accelerators, on smartphones can also record the individuals’ step count during a certain period, which yields acceptable results compared to medical pedometers. My mini-program also accesses the step count every minute through WeChat, which is then used to calibrate the individuals’ carbon footprint other than transportation.

In cases where the user travels using multiple modes of transportation or mistakenly selects the wrong mode, we have developed a machine learning algorithm to automatically label and correct the transportation mode. To test the algorithm, we recruited 43 volunteers who used our mini-program for two months and recorded 121 valid entries of trip data. Using these data, we computed several features such as trip distances, durations, average speeds, and standard deviations of distances per ten seconds, and used them to build a decision tree classifier. We randomly selected 70% of the trip data as our training set and used the remaining data as the test set. The trained decision tree classifier was then used to predict the mode of transportation, and our estimate suggests that the algorithm achieved an overall accuracy of 72.7% in determining the means of transportation.

The second function is to estimate the amount of carbon emissions based on the individual’s mode of transportation that we have estimated and the distance they traveled. We calculate the carbon footprint segment by segment through individuals’ transportation, using the Greenhouse Gases Equivalencies Calculator provided by the United States Environmental Protection Agency. For example, if a user travels from Duke Kunshan University to Shanghai Jiao Tong University, our mini-program will enable us to estimate, using her or his route and speed, as well as the distance she or he traveled by taxi, train, metro, or walking respectively.

I use a computer to randomize which group a participant will belong. The chance of the participant to join (1) bureaucratic training group, (2) consumer knowledge group, or (3) control group is the same.

Individual

No

I plan to have at least 90 individuals. But I will try to recruit all employees at the bureau

90 bureaucrats or more

The sample is divided equally across three groups. For example, if the bureau has 90 individuals, I will assign 30 for bureaucratic training, 30 for consumer knowledge learning, and 30 for control

My method can detect very small differences across groups. However, I plan to detect whether the difference of working hours per day is greater than 0.5 hours between groups. I estimate that I will need at least 29 individuals for my experiment.