ML Versus MI for Missing Data With Violation of Distribution Conditions

Yuan, Ke‐Hai; Yang-Wallentin, Fan; Bentler, Peter M.

doi:10.1177/0049124112460373

Cited by 102 publications

(82 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Our prior simulation study (Hallgren & Witkiewitz, 2013) also found that the MI and FIML methods produced less biased estimates of alcohol treatment outcomes regardless of sample size and that a larger sample size does not protect against bias when using single imputation methods (e.g., LOCF; missing = heavy drinking). Importantly, we found both in the current study and in our prior work (Hallgren & Witkiewitz, 2013) that MI and FIML produced similar standard errors and seemed to perform equally well in recovering the truth, which is inconsistent with prior studies that have found MI to produce biased estimates when distributional assumptions are violated, particularly for smaller sample sizes (Demirtas et al, 2008; Yuan et al, 2012). …”

Section: Discussioncontrasting

confidence: 99%

Methods to Analyze Treatment Effects in the Presence of Missing Data for a Continuous Heavy Drinking Outcome Measure When Participants Drop Out from Treatment in Alcohol Clinical Trials

Witkiewitz

Falk

Kranzler

et al. 2014

Alcoholism Clin & Exp Res

View full text Add to dashboard Cite

Background Attrition is common in alcohol clinical trials and the resultant loss of data represents an important methodological problem. In the absence of a simulation study, the drinking outcomes among those who are lost to follow-up are not known. Individuals who drop out of treatment and continue to provide drinking data, however, may be a reasonable proxy group for making inferences about the drinking outcomes of those lost to follow-up. Methods We used data from the COMBINE study, a multisite, randomized clinical trial, to examine drinking during the four months of treatment among individuals who dropped out of treatment but continued to provide drinking data (i.e., “treatment dropouts;” n=185). First, we estimated the observed treatment effect size for naltrexone versus placebo in a sample that included both treatment completers (n=961) and treatment dropouts (n=185; total N=1146), as well as the observed treatment effect size among just those who dropped out of treatment (n=185). In both the total sample (N=1146) and the dropout sample (n=185), we then deleted the drinking data after treatment dropout from those 185 individuals to simulate missing data. Using the deleted datasets we then estimated the effect of naltrexone on the continuous outcome percent heavy drinking days using six methods to handle missing data (last observation carried forward, baseline observation carried forward, placebo mean imputation, missing=heavy drinking days, multiple imputation (MI), and full information maximum likelihood (FIML)). Results MI and FIML produced effect size estimates that were most similar to the true effects observed in the full data set in all analyses, while missing=heavy drinking days performed the worst. Conclusions Although missing drinking data should be avoided whenever possible, MI and FIML yield the best estimates of the treatment effect for a continuous outcome measure of heavy drinking when there is dropout in an alcohol clinical trial.

show abstract

Section: Discussioncontrasting

confidence: 99%

Methods to Analyze Treatment Effects in the Presence of Missing Data for a Continuous Heavy Drinking Outcome Measure When Participants Drop Out from Treatment in Alcohol Clinical Trials

Witkiewitz

Falk

Kranzler

et al. 2014

Alcoholism Clin & Exp Res

View full text Add to dashboard Cite

show abstract

“…For the bias and RMSE of the intercept mean, the non-normal degree had no impact on the two methods. For the slope mean, the bias and RMSE of MAR-based ML method was not affected by the degree of non-normality, and the conclusions agree with previous studies (Enders, 2001; Tacksoo et al, 2009; Yuan et al, 2012). But the MNAR-based DK method was more greatly impacted by the non-normal degree.…”

Section: Discussionsupporting

confidence: 90%

“…Four levels of skewness and kurtosis were considered. In light of the degrees of non-normal distribution adopted by previous simulation studies (Yuan and Bentler, 2000; Enders, 2001; Tacksoo et al, 2009; Yuan et al, 2012), the skewness and kurtosis were set at 0 and 3 for normal data, 0.5 and 6 for the slightly non-normal case, 2 and 15 for the moderately non-normal case, and 3 and 33 for the extremely non-normal case, respectively.Two missingness mechanisms, i.e., MNAR and MAR, were involved.…”

Section: Methodsmentioning

confidence: 99%

Comparison of Different LGM-Based Methods with MAR and MNAR Dropout Data

Chen

Cui

et al. 2017

Front. Psychol.

View full text Add to dashboard Cite

The missing not at random (MNAR) mechanism may bias parameter estimates and even distort study results. This study compared the maximum likelihood (ML) selection model based on missing at random (MAR) mechanism and the Diggle–Kenward selection model based on MNAR mechanism for handling missing data through a Monte Carlo simulation study. Four factors were considered, including the missingness mechanism, the dropout rate, the distribution shape (i.e., skewness and kurtosis), and the sample size. The results indicated that: (1) Under the MAR mechanism, the Diggle–Kenward selection model yielded similar estimation results with the ML approach; Under the MNAR mechanism, the results of ML approach were underestimated, especially for the intercept mean and intercept slope (μi and μs). (2) Under the MAR mechanism, the 95% CP of the Diggle–Kenward selection model was lower than that of the ML method; Under the MNAR mechanism, the 95% CP for the two methods were both under the desired level of 95%, but the Diggle–Kenward selection model yielded much higher coverage probabilities than the ML method. (3) The Diggle–Kenward selection model was easier to be influenced by the non-normal degree of target variable's distribution than the ML approach. The level of dropout rate was the major factor affecting the parameter estimation precision, and the dropout rate-induced difference of two methods can be ignored only when the dropout rate falls below 10%.

show abstract

“…Multiple imputation is among the optimal options for handling missing data (Kline, 2005), yielding relatively unbiased estimates and performing particularly well relative to other approaches in small samples (Graham & Schafer, 1999; Yuan et al, 2012) and when flexibility in analytic strategies is required (Allison, 2003; Schafer, 1999). To capture changes in the mediator and outcome variables, residualized gain scores (Tucker, Damarin, & Messick, 1966) were calculated for pre- and post-treatment or -waitlist scores on the measures of emotion dysregulation, DSH, and BPD symptoms.…”

Section: Resultsmentioning

confidence: 99%

Mechanisms of change in an emotion regulation group therapy for deliberate self-harm among women with borderline personality disorder

Gratz

Bardeen

Levy

et al. 2015

Behaviour Research and Therapy

View full text Add to dashboard Cite

Despite growing evidence for the efficacy of Gratz and colleagues’ emotion regulation group therapy (ERGT) for deliberate self-harm (DSH) among women with borderline personality disorder (BPD), the proposed mechanism of change in this treatment (i.e., emotion regulation) remains largely unexamined. This study examined change in emotion dysregulation as a mediator of the effects of this ERGT on DSH and BPD symptoms, as well as the extent to which change in emotion dysregulation during treatment predicts further improvements in DSH during a 9-month follow-up. Participants included 61 female outpatients with BPD and recent DSH who were randomly assigned to receive this ERGT in addition to their ongoing outpatient therapy immediately (n = 31) or after 14 weeks (n = 30). Measures of emotion dysregulation, DSH, and BPD symptoms were administered pre- and post-treatment or -waitlist, and at 9-months post-treatment (for participants in both conditions who received ERGT). Results from a series of mediation analyses provide further support for emotion regulation as a mechanism of change in this treatment. Specifically, results revealed that improvements in emotion dysregulation over the course of treatment mediated the observed reductions in BPD cognitive and affective symptoms during treatment and predicted further improvements in DSH during follow-up.

show abstract

ML Versus MI for Missing Data With Violation of Distribution Conditions

Cited by 102 publications

References 46 publications

Methods to Analyze Treatment Effects in the Presence of Missing Data for a Continuous Heavy Drinking Outcome Measure When Participants Drop Out from Treatment in Alcohol Clinical Trials

Methods to Analyze Treatment Effects in the Presence of Missing Data for a Continuous Heavy Drinking Outcome Measure When Participants Drop Out from Treatment in Alcohol Clinical Trials

Comparison of Different LGM-Based Methods with MAR and MNAR Dropout Data

Mechanisms of change in an emotion regulation group therapy for deliberate self-harm among women with borderline personality disorder

Contact Info

Product

Resources

About