When Missing Data is Related to the Outcome of Interest
People can have missing data in surveys for a variety of reasons. In cross-sectional surveys, people may choose not to respond to some questions resulting in item missingness. In longitudinal studies, people may miss a wave of data collection or stop responding to the survey altogether. Conventional methods of handling missingness such as multiple imputation and maximum likelihood estimation assume that missingness is random after adjusting for variables of interest expected to be related to the outcome. I am interested in circumstances where missing values may be related directly to the outcome, when data is missing not at random. For example, in studies of older adults where health or frailty are the outcomes of interest, the individuals lost to follow up may also be those who are more likely to be in poor health and unable to respond to surveys. A body of my research considers how assumptions about missingness may shape our conclusions about health change and the social and demographic factors associated with changing health outcomes.
Related Studies:
Jackson, H. & Engelman, M. (2021) “Deaths, Disparities, and Cumulative (Dis)Advantage: How Social Inequities Produce an Impairment Paradox in Later Life.” Journal of Gerontology: Series A.
Engelman, M. & Jackson, H. (2019) “Gradual Change or Punctuated Equilibrium? Reconsidering Patterns of Health in Later Life.” Demography, 56(6), 2323-2347.
Jackson, H., Engelman, M., & Bandeen-Roche, K. (2017). “Robust Respondents and Lost Limitations: The Implications of Nonrandom Missingness for the Estimation of Health Trajectories.” Journal of Aging and Health, 31(4), 685-708.
Collecting Information about Sensitive Topics
Some survey topics can be sensitive to discuss. People may feel uncomfortable disclosing their prior experiences and worry that their “true” response may subject them to social stigma or legal repercussions. A recent area of my research considers whether surveys can indirectly collect information about sensitive topics. Using indirect approaches may both protect respondent privacy and improve reporting of sensitive events. My work has focused on evaluating the list experiment. The list experiment randomly divides respondents into two groups, a control group that is asked to report how many, but not which, relatively non-sensitive items apply to them, and a treatment group that is asked to report about the same list of non-sensitive items, plus a sensitive item. The difference in means between the two groups represents the incidence of the sensitive item. I am studying whether the list experiment may improve estimates of abortion incidence and how combined data approaches may reduce the variance around estimates.
Related Studies:
Kissling, A. & Jackson, H. (2022) “Estimating Prevalence of Abortion using List Experiments: Findings from a Survey of Women in Delaware and Maryland.” Women’s Health Issues. 33(1):33- 40.
Jackson, H. & Rendall, M. “Addressing Abortion Underreporting in Surveys with the List Experiment: Lifetime and Five-Year Abortion Incidence with Multivariate Estimation of Socio-demographic and Health Associations in two U.S. States.” Under Review.
Jackson, H. & Rendall, M. “Lifetime abortion incidence when abortion care is covered by Medicaid: Maryland versus five comparison states.” Health Services Research. Accepted.
How Survey Context and Design Influences Estimates
We analyze survey data to estimate a variety of outcomes. These estimates may be influenced by the information reported by respondents, the strategies for handling missing data, and how information is weighted. A body of my research has considered how survey redesigns may influence estimates of health outcomes. Specifically, I was part of a team that implemented and evaluated the redesign of the health content in two nationally representative household surveys, the Current Population Survey Annual Social and Economic Supplement (CPS ASEC) and the Survey of Income and Program Participation (SIPP).
Related Studies:
Jackson, H. & Starkey, K. (2023) “Out-of-Pocket Medical Expenditures in the Redesigned Current Population Survey: Evaluating Improvements to Data Processing.” Medical Care Research and Review.80(5): 548-557.
Jackson, H., Young, N., & Taylor, D. (2021) “Beyond Question Wording: How Survey Design and Administration Shape Estimates of Disability.” Disability and Health Journal, 101115.
Berchick, E. & Jackson, H. (2021) “Data Processing Improvements for Estimates of Health Insurance Coverage in the Current Population Survey Annual Social and Economic Supplement.” Medical Care Research and Review, 10775587211000812.
Jackson, H. & Berchick, E. (2020) “Improvements in Uninsurance Estimates for Fully Imputed Cases in the Current Population Survey Annual Social and Economic Supplement”. Inquiry: The Journal of Health Care Organization, Provision, and Financing, 57, 1-8.