Tuning multiple imputation by predictive mean matching and local residual draws.
Morris TP, White IR, Royston P. · 2014
Read Tuning multiple imputation by predictive mean matching and local residual draws. by Morris TP, White IR, Royston P. (2014) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4051964.txt.
Introduction
Tim P. Morris, Ian R. White, and Patrick Royston tune two shortcut methods for filling gaps in data, in BMC Medical Research Methodology in 2014. Multiple imputation guesses missing values from a model. When the model is off, the guesses can mislead. Two fixes borrow from real records instead: one takes a whole observed value from a similar donor case, the other takes only the donor's leftover error.
The team reviewed the variants of both methods and tested them in simulations, first with a correct model and then with a wrong one. One warning stands out: never draw from a single donor, a default in some software. A pool of about ten donors behaved far better. The shortcuts help when the model is slightly wrong. Nothing replaces building it with care.
Catalog record
| CREATOR | Morris TP, White IR, Royston P. |
|---|---|
| YEAR | 2014 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Tuning multiple imputation by predictive mean matching and local residual draws By Tim P Morris, Ian R White, Patrick Royston Published in: BMC Medical Research Methodology DOI: 10.1186/1471-2288-14-75 ABSTRACT Abstract Background Multiple imputation is a commonly used method for handling incomplete covariates as it can provide valid inference when data are missing at random. This depends on being able to correctly specify the parametric model used to impute missing values, which may be difficult in many realistic settings. Imputation by predictive mean matching (PMM) borrows an observed value from a donor with a similar predictive mean; imputation by local residual draws (LRD) instead borrows the donor’s residual. Both methods relax some assumptions of parametric imputation, promising greater robustness when the imputation model is misspecified. Methods We review development of PMM and LRD and outline the various forms available, and aim to clarify some choices about how and when they should be used. We compare performance to fully parametric imputation in simulation studies, first when the imputation model is correctly specified and then when it is misspecified. Results In using PMM or LRD we strongly caution against using a single donor, the default value in some implementations, and instead advocate sampling from a pool of around 10 donors. We also clarify which matching metric is best. Among the current MI software there are several poor implementations. Conclusions PMM and LRD may have a role for imputing covariates (i) which are not strongly associated with outcome, and (ii) when the imputation model is thought to be slightly but not grossly misspecified. Researchers should spend efforts on specifying the imputation model correctly, rather than expecting predictive mean matching or local residual draws to do the work. Keywords: Multiple imputation, Imputation model, Predictive mean matching, Local residual draws, Missing data BACKGROUND The presence of missing data is a common issue in medical research, leading to reduced precision and sometimes bias in parameter estimates. Multiple imputation (MI) can alleviate these issues and is popular approach to dealing with missing data [1-3]. It is impossible to know for certain how data went missing. In thinking about the process there are three important scenarios [4]: 1. Missing completely at random (MCAR). The probability of data being missing does not depend on observed or unobserved data. 2. Missing at random (MAR). Conditional on observed data, the probability of data being missing does not depend on unobserved data. MCAR is a special case of MAR. 3. Missing not at random (MNAR). Conditional on observed data, the probability of data being missing still depends on unobserved data. Researchers analysing incomplete datasets should consider the process by which data may have gone missing, and perform analyses that are valid given this assumption. MI involves specifying a parametric model for the missing data given the observed data and drawing missing values from the posterior predictive distribution M>1 times. This model is henceforth referred to as the imputation model. The M filled-in datasets are analysed identically according to the model that would have been used in the absence of missing data. We term this model the analysis model. The M parameter estimates are then combined using ‘Rubin’s rules’ [5]. Multiple imputation can provide valid inference given any of the above mechanisms, although standard software implementations impute assuming MAR (MCAR) by default. If the imputation model is specified correctly, Rubin’s rules lead to consistent parameter estimation and confidence intervals that fully incorporate uncertainty due to missing data [6]. For imputing a covariate it is advisable to include in the imputation model (i) variables thought to predict missingness, (ii) variables associated with the variable being imputed, and (iii) the outcome variable of the analysis model [3,7]. One of the biggest challenges for users of MI is specifying the imputation model correctly. This is not always easy to do, even for seemingly simple analyses: for instance when the analysis model contains nonlinear functions of incomplete covariates [8]. Predictive mean matching (PMM) [9] and local residual draws (LRD) [10] are methods for drawing imputations that relax some of the assumptions of parametric imputation. In doing so they may improve robustness of inference with missing data to misspecification of the imputation model. These methods are outlined briefly below and described further in the Methods section. For an incomplete variable x, an imputation model is fitted with parameters α and covariates z . Parametric imputation proceeds by drawing α from its posterior distribution, before drawing missing values of x from the posterior predictive distribution conditional on the draw α ∗. The draws of the imputation model parameters make parametric imputation ‘proper’ [6] and may be taken parametrically or by the approximate Bayesian bootstrap [11]. PMM and LRD differ from parametric imputation as follows. Let h index observations with x observed and j index observations with x missing. For all h, the linear predictor α obs z h is calculated, and for all j, the linear predictor α mis z j is calculated ( α obs and α mis will be defined in the Methods section). Observed values close to the linear-predicted value are selected as the donor pool. Often, but not always, the donor pool is fixed as containing k candidate donors. One of these is selected at random to ‘donate’. PMM imputes the donor’s x h . LRD adds the donor’s residual to the recipient’s linear predictor. In the remainder of this article, we give technical details of these methods reviewing their development and the various forms available, along with the rationale for their use. Two simulation studies on PMM and LRD are then described and reported: in the first, the imputation model is correct; in the second, the imputation model is mis-specified. We illustrate various approaches to imputing a missing covariate for a cohort study in ovarian cancer. We finish with a discussion and some conclusions. This article describes the rationale for PMM and LRD, and their development and evaluation in previous work. They are evaluated further in some simple and then more challenging settings. Our focus is on incomplete continuous covariates, though in principle both methods may be used to impute ordinal or categorical covariates. We aim to clarify some choices about how PMM and LRD should be implemented and when they should be used. METHODS The development of predictive mean matching and local residual draws In this section, we provide a technical description of PMM and LRD, review the development of the various flavours available – of which there are several – and clarify some details. Table 1 summarises software implementations of PMM and LRD, as of February 2014, and provides some details on options for changing the default values, if available. [Table: Summary of existing software implementations of PMM and LRD] Both PMM and LRD begin by calculating a predictive distance δ hj , which can be thought of as a measure of match quality. For all j the k observations minimising |δ hj | are identified where [formula: (1)δhj=αmiszj-αobszh,] and one of these is selected at random. For PMM [9] the imputed value xj∗ is taken as x h . For LRD [19] the imputed value xj∗ is [formula: (2)xj∗=αmiszj+xh-αobszh.] Defining the matching distance Little initially introduced PMM, suggesting the calculation of δ