Data-Driven Studies of Interviewers Productivity for Outbound Call Centers

以數據所驅動的外撥電話中心訪問員生產力研究

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date22 Dec 2020

Abstract

In recent years, outbound call centers play an increasingly critical role in business services, market research, telemarketing, and other areas. Outbound call centers enable proactivity in contacting potential respondents, but create more operational challenges than inbound call centers. Emerging telephone deceptions have recently caused a significant drop in response rate on outbound survey calls, which severely affects the sustainability and operations of outbound call centers. This study is driven by the challenges confronted by outbound call centers, which motivates us to build a useful database from an operations-oriented perspective and identify the impact factors of interviewer productivity to improve the operational efficiency of outbound call centers.

Call centers employ millions of agents all over the world. Workforce management remains the most important issue in call center operations. From the perspective of return on investment, the workforce accounts for a vast proportion of variable costs in outbound call centers. Therefore, it is worth enhancing operational efficiency by improving interviewer productivity in order to reduce the variable costs. As a basic and principal performance indicator, the number of completed interviews conducted by each interviewer per shift is used as a productivity measurement to evaluate call center workforce performance in this study.

During the past decade, we have accumulated extensive data on the experience of conducting telephone surveys in the Statistical Consulting Unit of City University of Hong Kong. In this study, we design and build a primary database on all identified potential impact factors on interviewer productivity, filtered from fragmentary raw data that includes structured interviewers' behaviors data, the demographic information of interviewers, the logbook information, and the unstructured conversational audio files from the Computer Assisted Telephone Interviewing (CATI) system. In particular, distinguishing this study from previous studies, we used factor analysis technique to transform a series of voice features, the high-dimensional auditory spectrum bands from 0 to 8Hz, into five latent factors representing different characteristics of the human voice. To the best of our knowledge, this is the first study attempting to utilize structured data information together with voice features to predict interviewer productivity.

Based on the constructed database, we focused on identifying the impact factors affecting interviewer productivity with different data analysis methods. Due to the phenomenon of repeated observations from the same interviewer attending more than one shift, the methodology of longitudinal data analysis is required to contend with the correlation between responses from the same individual. We use a popular longitudinal data analysis methodology, the marginal model with Generalized Estimating Equations (GEE) estimation method, together with backward elimination to identify significant impact factors on interviewer productivity from all the candidate factors in our database. In analyzing the model results, significant variables reflect the factors affecting interviewer productivity and provide guidance for the operations of outbound call centers. Moreover, we assess model predictive performance of the marginal model compared with the Poisson regression model on call center's data.

Furthermore, we adopted an alternative approach, using the generalized linear mixed-effects model (GLMM), to analyze call center's data. We proposed a two-stage method on the penalized GLMM via Lasso to select impact factors on interviewer productivity, where the re-estimation approach helps to improve estimation accuracy in GLMM modeling. Cross-validation was used to evaluate the forecasting performance of the penalized GLMM, compared with the Poisson regression model and the marginal model with GEE based on three scoring rules. The results showed that the GLMM performs better than the other two models to some extent in 1000 CV trails.

Our research findings can teach outbound call center managers how to select interviewers during recruitment, how to provide technical guidance in agents' training, and how to give evaluation suggestions to retain talent. In particular, when recruiting interviewers for a survey project, the manager can ask candidates to read the survey introduction and record their voice. Through vocal analysis, men with a lower average pitch, women with a higher average pitch, and candidates with lower energy in the bass region could be prioritized in hiring. When training new agents before the survey, supervisors can guide agents to use a suitably loud voice and moderate speech rate to introduce the survey in detail when respondents answer the phone. Supervisors can also recommend agents wait about 30-40 seconds for a call attempt. In addition, managers and supervisors can pay attention to outstanding agents who work hard, log more total working time, and have more conversational skill with higher Completes to First Refusals Ratio (CFR) and shorter average lengths for interviews. Supervisors should consider retaining these talents for future survey projects.