China Family Panel Studies

• 1. Introduction

  ›

  The China Family Panel Studies set out to do something enormously ambitious: follow tens of thousands of Chinese households across time, returning every two years to ask about their money, their health, their children, and their inner lives. Every child over the age of nine in a sampled household sits for an interview. Every adult is tracked. And the survey doesn't stop when families move or grow; biological and adopted children born after the baseline round are permanently added to the roster, designated "genetic members" and followed for the rest of the project's life.

  The survey launched in the spring of 2010, run by the Institute of Social Science Survey at Peking University. Its goal was explicit: document how Chinese society, economy, population, education, and health are changing, and put that evidence in the hands of researchers and policymakers. What makes this project so unusual isn't just its scale. It's the particular combination of questions it asks, reaching from household debt and land distribution to mental health scales, locus of control scores for children, and inter-personal trust measures borrowed from the World Values Survey.

  How do you actually sample a population of that size fairly? Why do five provinces get far more households than the rest? And what happened when the 2010 field teams went out to meet their 16,000-household target? Those questions run through every layer of how this survey was designed and executed.

• 2. Sampling Architecture

  ›

  The original target for the 2010 baseline was 16,000 households, drawn from 25 provinces, municipalities, and autonomous regions. That figure excluded Hong Kong, Macao, Taiwan, Xinjiang, Tibet, Qinghai, Inner Mongolia, Ningxia, and Hainan. The remaining coverage was designed to represent roughly 95% of China's population.

  The method used was Probability-Proportional-to-Size Sampling, layered with implicit stratification. Administrative units and socioeconomic status served as the main stratification variables. Where local GDP per capita data were available, that figure was used to rank communities within an administrative unit. Where GDP per capita wasn't on hand, the proportion of nonagricultural population or raw population density stood in as a substitute.

  The sample was split into two halves. Eight thousand households came from five provinces selected for oversampling: Shanghai, Liaoning, Henan, Gansu, and Guangdong. Each of those five sub-samples held 1,600 households. The logic was that each province would be representative at the regional level, enabling cross-region comparisons independent of the national frame. The remaining 8,000 households were drawn from the other 20 provinces, which the designers called the "small provinces." Combined and weighted, the two halves produced a frame representative of the national population.

  The sampling ran in three stages. Administrative districts or counties served as the Primary Sampling Unit. Villages or neighborhood communities formed the second stage. At the third stage, housing units were selected from street listings using a random starting point and an equal-probability method. The result was a nested design capable of producing estimates at the individual, household, and community level simultaneously.

• 3. What the 2010 Baseline Found

  ›

  Field collection for the first wave ran from the 1st of April 2010 to the 15th of September 2010. When the interviewers finished, the baseline had reached 14,960 households, falling short of the 16,000-household target. Inside those households, teams had completed interviews with 33,600 adults and 8,990 youths, for a total of 42,590 individuals.

  Response rates from that first wave were documented in detail. At the household level, the response rate was 81.25%. The cooperation rate among households that were contacted reached 96.58%. The contact rate was 84.13%, and the refusal rate was just 2.67%. At the individual level, the response rate was 84.14%, the cooperation rate 87.01%, the contact rate 96.7%, and the refusal rate 8.47%.

  Keeping those numbers honest required a dense set of quality controls. The 2010 team applied telephone checks, field checks, audio record checks, interview reviews, and statistical analyses to guard against problems ranging from poorly designed questionnaires to irregular interviewer behavior to data entry errors. Consistency checking, wild code checking, and text recoding were applied as cleaning operations after collection. At baseline, the software platform running the survey was Blaise 4.8, alongside SAS 9.1 and 9.2.

  All members of those baseline households, together with any biological and adopted children they would have in later years, were locked in as permanent participants. The project's term for them is "genetic members," and the design calls for following them across every future wave.

• 4. What Gets Asked and Measured

  ›

  The topical scope of CFPS is unusually wide for a single survey. Economic activities, educational attainment, family relationships and dynamics, migration, and physical and mental health all fall within its frame. Subject terms catalogued in the study documentation run from attitudes, children, and cognitive functioning through debt, employment, and family businesses to pensions, public transportation, real estate, and time utilization.

  The survey includes a set of established psychological and developmental scales. Mental health is measured using the K6 and the CESD20. Children are assessed on the Big Five Questionnaire and the HOME Scale, as well as the Nowicki-Strickland Locus of Control Scale and the Rosenberg Self Esteem Scale. Parental attitudes are captured through a scale measuring the value of children to parents. Inter-personal trust comes from an instrument borrowed directly from the World Values Survey.

  Collection modes have included face-to-face interviews, telephone interviews, computer-assisted personal interview (CAPI), and computer-assisted telephone interview (CATI), with mixed-mode approaches used across waves. To hold attrition down, the team employed follow-up visits, supervisory checks, and historical matching.

  The data are available in both Chinese and English. They are not distributed through ICPSR or other third-party archives; access requires a data-use application submitted directly to the Institute of Social Science Survey at Peking University. The Peking University Open Research Data Dataverse mirrors the baseline release under the DOI 10.18170/DVN/45LCSO.

• 5. The Team and Its Backers

  ›

  Four principal investigators are credited on the project, all affiliated with Peking University. Yu Xie holds appointments at both Peking University and Princeton University. Xiaobo Zhang, Ping Tu, and Qiang Ren round out the team, each based at Peking University.

  Funding has come from two sources: the National Natural Science Foundation of China and Peking University itself. That combination of a national science foundation and the home institution of the principal investigators is a familiar model for large longitudinal surveys, placing the study at the intersection of government scientific priorities and university-based research infrastructure.

  The survey became longitudinal in the formal sense starting with the 2012 wave, running every two years from that point. Subsequent waves ran from July 2012 through March 2013, from July 2014 through June 2015, and then continued in 2016, 2018, and 2020. The 2020 wave was fielded from July through the end of that year. A 2022 wave followed. The design calls for follow-up of all core members on a yearly basis, with the biennial full-wave collections as the primary data-gathering events.