Dataset Information

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

ABSTRACT:

Background

Within routinely collected health data, missing data for an individual might provide useful information in itself. This occurs, for example, in the case of electronic health records, where the presence or absence of data is informative. While the naive use of missing indicators to try to exploit such information can introduce bias, its use in conjunction with multiple imputation may unlock the potential value of missingness to reduce bias in causal effect estimation, particularly in missing not at random scenarios and where missingness might be associated with unmeasured confounders.

Methods

We conducted a simulation study to determine when the use of a missing indicator, combined with multiple imputation, would reduce bias for causal effect estimation, under a range of scenarios including unmeasured variables, missing not at random, and missing at random mechanisms. We use directed acyclic graphs and structural models to elucidate a variety of causal structures of interest. We handled missing data using complete case analysis, and multiple imputation with and without missing indicator terms.

Results

We find that multiple imputation combined with a missing indicator gives minimal bias for causal effect estimation in most scenarios. In particular the approach: 1) does not introduce bias in missing (completely) at random scenarios; 2) reduces bias in missing not at random scenarios where the missing mechanism depends on the missing variable itself; and 3) may reduce or increase bias when unmeasured confounding is present.

Conclusion

In the presence of missing data, careful use of missing indicators, combined with multiple imputation, can improve causal effect estimation when missingness is informative, and is not detrimental when missingness is at random.

SUBMITTER: Sperrin M

PROVIDER: S-EPMC7346454 | biostudies-literature | 2020 Jul

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

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

Sperrin Matthew M Martin Glen P GP

BMC medical research methodology 20200708 1

<h4>Background</h4>Within routinely collected health data, missing data for an individual might provide useful information in itself. This occurs, for example, in the case of electronic health records, where the presence or absence of data is informative. While the naive use of missing indicators to try to exploit such information can introduce bias, its use in conjunction with multiple imputation may unlock the potential value of missingness to reduce bias in causal effect estimation, particula ...[more]

PMID: 32640992

Similar Datasets

Project description:BackgroundQuestionnaires used in longitudinal studies may have questions added or removed over time for numerous reasons. Data missing completely at a follow-up survey is a unique issue for longitudinal studies. While such excluded questions lack information at one follow-up survey, they are collected at other follow-up surveys, and covariances observed at other follow-up surveys may allow for the recovery of the missing data. This study utilized data from a large longitudinal cohort study to assess the efficiency and feasibility of using multiple imputation (MI) to recover this type of information.MethodsMillennium Cohort Study participants completed the 9-item Patient Health Questionnaire (PHQ) depression module at 2 time points (2004, 2007). The suicidal ideation item in the module was set to missing for the 2007 assessment. Several single-level MI models using different sets of predictors and forms of suicidal ideation were used to compare self-reported values and imputed values for this item in 2007. Additionally, associations with sleep duration and smoking status, which are related constructs, were compared between self-reported and imputed values of suicidal ideation.ResultsAmong 63,028 participants eligible for imputation analysis, 4.05% reported suicidal ideation on the 2007 survey. The imputation models successfully identified suicidal ideation, with a sensitivity ranging between 34 and 66% and a positive predictive value between 36 and 42%. Specificity remained above 96% and negative predictive value above 97% for all imputed models. Similar associations were found for all imputation models on related constructs, though the dichotomous suicidal ideation imputed from the model using only PHQ depression items yielded estimates that were closest with the self-reported associations for all adjusted analyses.ConclusionsAlthough sensitivity and positive predictive value were relatively low, applying MI techniques allowed for inclusion of an otherwise missing variable. Additionally, correlations with related constructs were estimated near self-reported values. Therefore, the other 8 depression items can be used to estimate suicidal ideation that was completely missing from a survey using MI. However, these imputed values should not be used to estimate population prevalence.

Project description:BackgroundLongitudinal categorical variables are sometimes restricted in terms of how individuals transition between categories over time. For example, with a time-dependent measure of smoking categorised as never-smoker, ex-smoker, and current-smoker, current-smokers or ex-smokers cannot transition to a never-smoker at a subsequent wave. These longitudinal variables often contain missing values, however, there is little guidance on whether these restrictions need to be accommodated when using multiple imputation methods. Multiply imputing such missing values, ignoring the restrictions, could lead to implausible transitions.MethodsWe designed a simulation study based on the Longitudinal Study of Australian Children, where the target analysis was the association between (incomplete) maternal smoking and childhood obesity. We set varying proportions of data on maternal smoking to missing completely at random or missing at random. We compared the performance of fully conditional specification with multinomial and ordinal logistic imputation, and predictive mean matching, two-fold fully conditional specification, indicator based imputation under multivariate normal imputation with projected distance-based rounding, and continuous imputation under multivariate normal imputation with calibration, where each of these multiple imputation methods were applied, accounting for the restrictions using a semi-deterministic imputation procedure.ResultsOverall, we observed reduced bias when applying multiple imputation methods with restrictions, and fully conditional specification with predictive mean matching performed the best. Applying fully conditional specification and two-fold fully conditional specification for imputing nominal variables based on multinomial logistic regression had severe convergence issues. Both imputation methods under multivariate normal imputation produced biased estimates when restrictions were not accommodated, however, we observed substantial reductions in bias when restrictions were applied with continuous imputation under multivariate normal imputation with calibration.ConclusionIn a similar longitudinal setting we recommend the use of fully conditional specification with predictive mean matching, with restrictions applied during the imputation stage.

Project description:BackgroundIncomplete categorical variables with more than two categories are common in public health data. However, most of the existing missing-data methods do not use the information from nonresponse (missingness) probabilities.MethodsWe propose a nearest-neighbour multiple imputation approach to impute a missing at random categorical outcome and to estimate the proportion of each category. The donor set for imputation is formed by measuring distances between each missing value with other non-missing values. The distance function is calculated based on a predictive score, which is derived from two working models: one fits a multinomial logistic regression for predicting the missing categorical outcome (the outcome model) and the other fits a logistic regression for predicting missingness probabilities (the missingness model). A weighting scheme is used to accommodate contributions from two working models when generating the predictive score. A missing value is imputed by randomly selecting one of the non-missing values with the smallest distances. We conduct a simulation to evaluate the performance of the proposed method and compare it with several alternative methods. A real-data application is also presented.ResultsThe simulation study suggests that the proposed method performs well when missingness probabilities are not extreme under some misspecifications of the working models. However, the calibration estimator, which is also based on two working models, can be highly unstable when missingness probabilities for some observations are extremely high. In this scenario, the proposed method produces more stable and better estimates. In addition, proper weights need to be chosen to balance the contributions from the two working models and achieve optimal results for the proposed method.ConclusionsWe conclude that the proposed multiple imputation method is a reasonable approach to dealing with missing categorical outcome data with more than two levels for assessing the distribution of the outcome. In terms of the choices for the working models, we suggest a multinomial logistic regression for predicting the missing outcome and a binary logistic regression for predicting the missingness probability.

Project description:Whole brain fMRI analyses rarely include the entire brain because of missing data that result from data acquisition limits and susceptibility artifact, in particular. This missing data problem is typically addressed by omitting voxels from analysis, which may exclude brain regions that are of theoretical interest and increase the potential for Type II error at cortical boundaries or Type I error when spatial thresholds are used to establish significance. Imputation could significantly expand statistical map coverage, increase power, and enhance interpretations of fMRI results. We examined multiple imputation for group level analyses of missing fMRI data using methods that leverage the spatial information in fMRI datasets for both real and simulated data. Available case analysis, neighbor replacement, and regression based imputation approaches were compared in a general linear model framework to determine the extent to which these methods quantitatively (effect size) and qualitatively (spatial coverage) increased the sensitivity of group analyses. In both real and simulated data analysis, multiple imputation provided 1) variance that was most similar to estimates for voxels with no missing data, 2) fewer false positive errors in comparison to mean replacement, and 3) fewer false negative errors in comparison to available case analysis. Compared to the standard analysis approach of omitting voxels with missing data, imputation methods increased brain coverage in this study by 35% (from 33,323 to 45,071 voxels). In addition, multiple imputation increased the size of significant clusters by 58% and number of significant clusters across statistical thresholds, compared to the standard voxel omission approach. While neighbor replacement produced similar results, we recommend multiple imputation because it uses an informed sampling distribution to deal with missing data across subjects that can include neighbor values and other predictors. Multiple imputation is anticipated to be particularly useful for 1) large fMRI data sets with inconsistent missing voxels across subjects and 2) addressing the problem of increased artifact at ultra-high field, which significantly limit the extent of whole brain coverage and interpretations of results.

Dataset Information

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

Background

Methods

Results

Conclusion

Publications

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

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets