2007 HSR&D National Meeting Abstract
1014 — Multiple Imputation of Right-Truncated Laboratory Data
Cheng C (CHERP) , Stone RA
(CHERP), Obrosky DS
(CHERP), DeRubertis FR
(VA Pittsburgh Healthcare System)
To impute baseline hemoglobin A1c (HbA1c) levels in a longitudinal study where some laboratory values are reported only as exceeding a cut-point.
The Diabetes Telemonitoring Study compares home telemonitoring-based (HT) care management with less intense care coordination (CC) to help veterans with diabetes better manage their disease. The primary outcome, HbA1c, was measured for each of the 138 enrolled subjects at baseline and 3 months. At baseline, finger-stick HbA1c was performed to ascertain study eligibility (7.5%); a separate laboratory HbA1c by venipuncture also was assessed. These finger-stick values are complete while lab values are missing for 10 veterans, including three CC veterans. Seven HbA1c laboratory values were right-truncated at 11.5%, 11.8%, or 12.3%. Multiple imputation based on finger-stick values was done using the Imputation by Chained Equations algorithm in Stata. From a large number of imputations generated, we used the first 10 sets for which the imputed values for all seven truncated observations satisfied the corresponding range restrictions. We compared the multiple imputation approach, complete case analysis and simple replacement methods (substituting truncation cut-points or finger-stick values) with respect to (i) the estimated slope of lab vs. finger-stick HbA1c values and (ii) the estimated mean HbA1c in the two treatment groups.
The regression coefficient for the finger-stick is 1.00 (s.e. 0.030), based on 10 imputations. The corresponding coefficients are (0.99, 0.031) for complete case analysis; (0.91, 0.028) substituting truncation points; and (0.97, 0.025) substituting finger-stick values. The estimated HbA1c means for the HT and CC groups were 9.60 (s.e. 0.20) and 9.44 (s.e. 0.16) based on 10 imputations; 9.35 (s.e. 0.15) and 9.31 (s.e. 0.17) for complete cases; 9.43 (s.e. 0.16) and 9.51 (s.e. 0.18) substituting the truncation points and 9.43 (s.e. 0.16) and 9.56 (s.e. 0.19) substituting the finger-stick values.
Complete case analysis and simple substitutions produced downwardly-biased estimators. Restricting multiple imputations to satisfy the truncation constraints yields unbiased estimates with variances that appropriately reflect uncertainty.
A standard imputation algorithm can be readily modified to accommodate truncated reporting of laboratory data.