Adjustment for Response Bias Via Two-phase Analysis

Abstract

Background Records-based studies often have limited covariate data, leading some researchers to collect survey data on a subset. Results for survey responders may be biased due to selective nonresponse and will be less precise due to the decreased responder sample size. We use data from a study of air pollution and birth outcomes to illustrate how a 2-phase analysis can yield less biased and more precise results. Methods Our phase 1 group was a cohort of Los Angeles births from which we obtained a phase 2 group of survey responders. We compared estimates for the odds ratio (OR) between entire pregnancy carbon monoxide (CO) exposure and low birth weight in the first- and second-phase groups, adjusting only for variables available for both groups. Results For CO exposure of 1 part per million or higher, the conventional adjusted ORs and 95% confidence intervals for low birth weight were 1.15 (1.06 –1.25) and 1.33 (1.06–1.68) for the phase 1 and 2 groups, suggesting a possible response bias and decreased precision in the latter estimate. We performed 2-phase analyses of the survey responders and found results similar to those for the cohort when we accounted for possible differential response by CO exposure. In our final analysis, we included both birth record and survey variables in a 2-phase model corrected for possible response bias. The results from weighted-, pseudo-, and maximum-likelihood were similar: 1.13 (1.03–1.25); 1.14 (1.01–1.29); and 1.10 (0.97–1.24), respectively. Conclusion Our approach provides a means of checking for response bias and adjusting both point and interval estimates to account for differential response.

The “Latina Epidemiologic Paradox” Revisited: The Role of Birthplace and Acculturation in Predicting Infant Low Birth Weight for Latinas in Los Angeles, CA

“…We did not adjust for gestational age, consistent with the assumption that since gestational age and LBW share etiologic pathways, adjustment is inappropriate in general [35, 36]; in a sensitivity analysis for our EPOS responder group, we added gestational age but found the inclusion of this variable had a negligible impact on our results. For analyses based on acculturation, we estimated ORs for risk factors for LBW within strata defined for high-acculturated USB and FB Latinas and low-acculturated FB Latinas.…”

“…In analyses for the EPOS responders and EPOS follow-up, both of which were taken from a nested case–control group, the study design resulted in an oversample of preterm normal weight births. To account for the sampling design and for possible bias due to survey non-response, we used sampling/response fractions and performed a weighted-likelihood two-phase analysis, as described in detail elsewhere [29, 34, 35]. The two-phase analysis was appropriate when considering exposures available for both the birth cohort and nested EPOS responders group (i.e., ethnicity and birthplace).…”

Generalizing Evidence From Randomized Clinical Trials to Target Populations: The ACTG 320 Trial

“…A derivation of the bias in the trial sample estimate of the population effect, defined as differences in means or proportions, is also given in the Appendix. Next, we describe the use of inverse probability-of-selection weights, which are an extension of Horvitz-Thompson weights (29) and have been used extensively in survey sampling (30)(31)(32), for confounder control (33), and have been discussed in the context of selection bias (7,19,34,35) or response bias in 2-phase studies (36).…”

What we don't know can hurt us: Nonresponse bias assessment in birth defects research

“…It is critical to consider nonresponse bias during the study design phase, and not just after analysis, so that necessary information, like the method of survey response or response wave, is collected and recorded. Even if there is evidence of nonresponse bias, analytical techniques (such as imputation or weighting) exist to estimate the impact of the bias, although they should be performed with caution, as additional, stronger assumptions must be made when performing these adjustments (Groves, ; Johnson et al, ; Hoggatt et al, ). It is strongly advised that when these methods are used, all assumptions are explicitly described for the readers.…”