Multiple imputation with missing indicators as proxies for unmeasured variables: simulation study.

Area of Science:

• Epidemiology
• Biostatistics
• Health Data Science

Background:

• Missing data in routinely collected health information, such as electronic health records, can be informative.
• The presence or absence of data can hold significant value.
• Naive use of missing indicators can introduce bias, but strategic application may mitigate this.

Purpose of the Study:

• To evaluate the effectiveness of using missing indicators combined with multiple imputation for causal effect estimation.
• To determine conditions under which this approach reduces bias, especially in missing not at random scenarios.
• To explore the impact of unmeasured confounding on bias reduction.

Main Methods:

• A simulation study was conducted to assess various missing data mechanisms, including missing not at random and missing at random.
• Directed acyclic graphs and structural models were employed to represent causal structures.
• Missing data was handled using complete case analysis and multiple imputation, with and without missing indicator terms.

Main Results:

• Multiple imputation with missing indicators demonstrated minimal bias in most causal effect estimation scenarios.
• The method showed no introduced bias in missing completely at random situations.
• Bias reduction was observed in missing not at random scenarios where the missingness depended on the variable itself.
• In cases of unmeasured confounding, the approach could either reduce or increase bias.

Conclusions:

• Strategic use of missing indicators alongside multiple imputation can enhance causal effect estimation when missingness is informative.
• This combined approach is not detrimental when data is missing at random.
• The findings highlight the potential of leveraging informative missingness to improve the accuracy of causal inference from health data.