Challenging the Two-for-One Assumption: A Novel Implementation of the Two-Method Planned Missing Data Design

Researchers are often faced with the dilemma of choosing between a cost-effective measure that is not as reliable but is able to be administered to larger sample, or a cost-prohibitive measure that is considered very reliable but is not suitable to be administered to a large sample. The current two-method missing data design uses modern methods of treating missing data to allow researchers to use both types of measurements and enjoy the benefits of both types of assessments, granted they meet the assumptions that the two types of measurements assess a common underlying construct. This study aims to use Monte Carlo simulation demonstrate that this assumption is not necessary for all applications of this measurement design by testing the hypothesis that having correlated constructs is sufficient to recover the information lost through the imposition of planned missing data, as well as a set of conditions that may influence the viability of this alternative approach, including the reliability of the item indicators, the strength of the association between the latent constructs, and the amount of planned missing data. Results indicated that having at least two cost-prohibitive items and a correlation of .100 or greater resulted in absolute relative bias (ARB) estimates from 0.00% to 3.45%, and when a single, reliable indicator (λ = .950) resulted in ARB estimates from 0.99% to 6.49%. Conditions with a lower correlation or a lower reliability for the cost-prohibitive single indicator resulted in ARB values well above the 10% threshold. Relative efficiency was also evaluated, and was strongly related to the amount of planned missing data, with 20% missingness resulting in an average 79.6% efficiency, 50% missingness resulting in an average 50.2% efficiency, and 80% missingness resulting in an average 20% efficiency. These results successfully demonstrated the viability of this alternative design’s ability to return unbiased parameter estimates under a majority of the 810 conditions evaluated while highlighting areas where this method did not perform well.