Missing Covariate Research Articles

A flexible approach for variable selection in large-scale healthcare database studies with missing covariate and outcome data

BackgroundPrior work has shown that combining bootstrap imputation with tree-based machine learning variable selection methods can provide good performances achievable on fully observed data when covariate and outcome data are missing at random (MAR). This approach however is computationally expensive, especially on large-scale datasets.MethodsWe propose an inference-based method, called RR-BART, which leverages the likelihood-based Bayesian machine learning technique, Bayesian additive regression trees, and uses Rubin’s rule to combine the estimates and variances of the variable importance measures on multiply imputed datasets for variable selection in the presence of MAR data. We conduct a representative simulation study to investigate the practical operating characteristics of RR-BART, and compare it with the bootstrap imputation based methods. We further demonstrate the methods via a case study of risk factors for 3-year incidence of metabolic syndrome among middle-aged women using data from the Study of Women’s Health Across the Nation (SWAN).ResultsThe simulation study suggests that even in complex conditions of nonlinearity and nonadditivity with a large percentage of missingness, RR-BART can reasonably recover both prediction and variable selection performances, achievable on the fully observed data. RR-BART provides the best performance that the bootstrap imputation based methods can achieve with the optimal selection threshold value. In addition, RR-BART demonstrates a substantially stronger ability of detecting discrete predictors. Furthermore, RR-BART offers substantial computational savings. When implemented on the SWAN data, RR-BART adds to the literature by selecting a set of predictors that had been less commonly identified as risk factors but had substantial biological justifications.ConclusionThe proposed variable selection method for MAR data, RR-BART, offers both computational efficiency and good operating characteristics and is utilitarian in large-scale healthcare database studies.

Planning a method for covariate adjustment in individually randomised trials: a practical guide.

BackgroundIt has long been advised to account for baseline covariates in the analysis of confirmatory randomised trials, with the main statistical justifications being that this increases power and, when a randomisation scheme balanced covariates, permits a valid estimate of experimental error. There are various methods available to account for covariates but it is not clear how to choose among them.MethodsTaking the perspective of writing a statistical analysis plan, we consider how to choose between the three most promising broad approaches: direct adjustment, standardisation and inverse-probability-of-treatment weighting.ResultsThe three approaches are similar in being asymptotically efficient, in losing efficiency with mis-specified covariate functions and in handling designed balance. If a marginal estimand is targeted (for example, a risk difference or survival difference), then direct adjustment should be avoided because it involves fitting non-standard models that are subject to convergence issues. Convergence is most likely with IPTW. Robust standard errors used by IPTW are anti-conservative at small sample sizes. All approaches can use similar methods to handle missing covariate data. With missing outcome data, each method has its own way to estimate a treatment effect in the all-randomised population. We illustrate some issues in a reanalysis of GetTested, a randomised trial designed to assess the effectiveness of an electonic sexually transmitted infection testing and results service.ConclusionsNo single approach is always best: the choice will depend on the trial context. We encourage trialists to consider all three methods more routinely.

Distribution‐free model selection for longitudinal zero‐inflated count data with missing responses and covariates

In many medical and social science studies, count responses with excess zeros are very common and often the primary outcome of interest. Such count responses are usually generated under some clustered correlation structures due to longitudinal observations of subjects. To model such longitudinal count data with excess zeros, the zero-inflated binomial (ZIB) models for bounded outcomes, and the zero-inflated negative binomial (ZINB) and zero-inflated poisson (ZIP) models for unbounded outcomes all are popular methods. To alleviate the effects of deviations from model assumptions, a semiparametric (or, distribution-free) weighted generalized estimating equations has been proposed to estimate model parameters when data are subject to missingness. In this article, we further explore important covariates for the response variable. Without assumptions on the data distribution, a model selection criterion based on the expected weighted quadratic loss is proposed to select an appropriate subset of covariates, especially when count responses have excess zeros and data are subject to nonmonotone missingness in both responses and covariates. To understand the selection effects of the percentages of excess zeros and missingness, we design various scenarios for covariate selection in the mean model via simulation studies and a real data example regarding the study of cardiovascular disease is also presented for illustration.

Association between serum β2-microglobulin and mortality in Japanese peritoneal dialysis patients: A cohort study.

Higher serum β2-microglobulin (B2M) concentrations are associated with higher mortality in the general population, non-dialyzed chronic kidney disease patients and patients receiving hemodialysis (HD). However, this relationship among patients on peritoneal dialysis (PD) has not been validated. We collected baseline data for 3,011 prevalent PD patients from a nationwide dialysis registry in Japan at the end of 2010. Clinical outcomes for 9 years were then evaluated using the registry at the end of 2011 to 2019. All-cause and cardiovascular (CV) mortality was assessed using Cox regression analysis and competing-risks regression analysis, respectively. We used multiple imputation to deal with missing covariate data. During a median follow-up of 87 months, 2,054 patients transferred to combined therapy with PD and HD or HD directly. A total of 3,011 patients, 1,235 (41.0%) died, including 437 patients (14.5%) from CV causes. Among them, 612 patients died after transfer to other dialysis modalities. Univariate analyses revealed no significant association between serum B2M and mortality, whereas higher serum B2M was independently associated with both all-cause and CV mortalities in adjusted models. However, the significant association between serum B2M and CV mortality disappeared in analysis treating serum B2M as a categorical variable. The effect of serum B2M on all-cause mortality was significantly higher among patients with higher urinary volume and a significant interaction was evident. Using a large-scale registry, we found that serum B2M contributes tenuously but significantly to worse outcome and residual kidney function significantly affects this relationship. On the contrary, serum B2M per se had no predictive value for patient outcome in prevalent PD patients.

On the bias of complete‐ and shifting‐case meta‐regressions with missing covariates

Missing covariates is a common issue when fitting meta‐regression models. Standard practice for handling missing covariates tends to involve one of two approaches. In a complete‐case analysis, effect sizes for which relevant covariates are missing are omitted from model estimation. Alternatively, researchers have employed the so‐called "shifting units of analysis" wherein complete‐case analyses are conducted on only certain subsets of relevant covariates. In this article, we clarify conditions under which these approaches generate unbiased estimates of regression coefficients. We find that unbiased estimates are possible when the probability of observing a covariate is completely independent of effect sizes. When that does not hold, regression coefficient estimates may be biased. We study the potential magnitude of that bias assuming a log‐linear model of missingness and find that the bias can be substantial, as large as Cohen's d = 0.4–0.8 depending on the missingness mechanism.

Evaluation of multiple imputation approaches for handling missing covariate information in a case-cohort study with a binary outcome

BackgroundIn case-cohort studies a random subcohort is selected from the inception cohort and acts as the sample of controls for several outcome investigations. Analysis is conducted using only the cases and the subcohort, with inverse probability weighting (IPW) used to account for the unequal sampling probabilities resulting from the study design. Like all epidemiological studies, case-cohort studies are susceptible to missing data. Multiple imputation (MI) has become increasingly popular for addressing missing data in epidemiological studies. It is currently unclear how best to incorporate the weights from a case-cohort analysis in MI procedures used to address missing covariate data.MethodA simulation study was conducted with missingness in two covariates, motivated by a case study within the Barwon Infant Study. MI methods considered were: using the outcome, a proxy for weights in the simple case-cohort design considered, as a predictor in the imputation model, with and without exposure and covariate interactions; imputing separately within each weight category; and using a weighted imputation model. These methods were compared to a complete case analysis (CCA) within the context of a standard IPW analysis model estimating either the risk or odds ratio. The strength of associations, missing data mechanism, proportion of observations with incomplete covariate data, and subcohort selection probability varied across the simulation scenarios. Methods were also applied to the case study.ResultsThere was similar performance in terms of relative bias and precision with all MI methods across the scenarios considered, with expected improvements compared with the CCA. Slight underestimation of the standard error was seen throughout but the nominal level of coverage (95%) was generally achieved. All MI methods showed a similar increase in precision as the subcohort selection probability increased, irrespective of the scenario. A similar pattern of results was seen in the case study.ConclusionsHow weights were incorporated into the imputation model had minimal effect on the performance of MI; this may be due to case-cohort studies only having two weight categories. In this context, inclusion of the outcome in the imputation model was sufficient to account for the unequal sampling probabilities in the analysis model.

Association between chronic conditions and urinary incontinence in females: a cross-sectional study using national survey data.

Background:Urinary incontinence affects up to half of women, yet few speak to their health care provider about or receive treatment for the condition. To aid with identifying subpopulations at risk for urinary incontinence, we examined the associations between 10 chronic health conditions and urinary incontinence among Canadian adult females.Methods:We conducted a cross-sectional analysis of survey data from the Canadian Community Health Survey (2013–2014) involving female respondents aged 25 years or older living in a private dwelling. Presence of chronic conditions and urinary incontinence were measured by self-report. We used logistic regression modelling with sampling weights, controlling for age, income, ethnicity, body mass index and smoking. Multiple imputation and probabilistic bias analysis were used to address missing covariate data and unmeasured confounding from parity.Results:Our analysis included 60 186 respondents representing more than 12 million Canadian females, of whom 45.8% (95% confidence interval [CI] 45.0%–46.6%) reported at least 1 chronic condition. Chronic conditions were associated with more than twice the odds of urinary incontinence (adjusted odds ratio [OR] 2.42, 95% CI 2.02–2.89). Associations were largest for bowel disorders (adjusted OR 2.92, 95% CI 2.44–3.49); modest for chronic obstructive pulmonary disease (adjusted OR 2.00, 95% CI 1.63–2.45), asthma (adjusted OR 1.82, 95% CI 1.52–2.19), arthritis (adjusted OR 1.98, 95% CI 1.74–2.24) and heart disease (adjusted OR 1.73, 95% CI 1.48–2.02); and smallest for diabetes (adjusted OR 1.20, 95% CI 1.02–1.41) and high blood pressure (adjusted OR 1.27, 95% CI 1.12–1.44). Results slightly attenuated but did not substantively change after imputation and bias analysis.Interpretation:We found that chronic conditions are associated with significantly higher odds of comorbid urinary incontinence among Canadian adult females, which is consistent with previous research. Our findings support routine inquiry regarding urinary incontinence symptoms among women accessing health care for chronic conditions.

A comparison of two approaches to implementing propensity score methods following multiple imputation


 Background. In observational research on causal effects, missing data and confounding are very common problems. Multiple imputation and propensity score methods have gained increasing interest as methods to deal with these, but despite their popularity methodologists have mainly focused on how they perform in isolation.
 
 Methods. We studied two approaches to implementing propensity score methods following multiple imputation, both of which have been used in applied research, and compared their performance by way of Monte Carlo simulation for a continuous outcome and partially unobserved covariate, treatment or outcome data. In the first, so-called Within, approach, propensity score analysis is performed within each of m imputed datasets, and the resulting m effect estimates are averaged. In the Across approach, for each subject the m estimated propensity scores are averaged first, after which the propensity score method is implemented based on each subject’s average propensity score. Because of its common use, complete case analysis was also implemented. Five propensity score estimators were studied, including regression, matching, and inverse probability weighting.
 
 Results. The Within approach was found to be superior to the Across approach in terms of bias as well as variance in settings with missing covariate data, when missing data were missing at random as well as when they were missing completely at random. In settings with incomplete treatment or outcome values only, the Within and Across approaches yielded similar results. Complete case analysis was generally least efficient and unbiased only in scenarios where missing data were missing completely at random.
 
 Conclusion. We advise researchers not to use the Across approach as the default method, because even when data are missing completely at random, this may yield biased effect estimates. Instead, the Within is the preferred approach when implementing propensity score methods following multiple imputation.

Pairwise joint modeling of clustered and high-dimensional outcomes with covariate missingness in pediatric pneumonia care.

Multiple outcomes reflecting different aspects of routine care are a common phenomenon in health care research. A common approach of handling such outcomes is multiple univariate analyses, an approach which does not allow for answering research questions pertaining to joint inference. In this study, we sought to study associations among nine pediatric pneumonia care outcomes spanning assessment, diagnosis and treatment domains of care, while circumventing the computational challenge posed by their clustered and high-dimensional nature and incompletely recorded covariates. We analyzed data from a cluster randomized trial conducted in 12 Kenyan hospitals. There were varying degrees of missingness in the covariates of interest, and these were multiply imputed using latent normal joint modeling. We used the pairwise joint modeling strategy to fit a correlated random effects joint model for the nine outcomes. This entailed fitting 36 bivariate generalized linear mixed models and deriving inference for the joint model using pseudo-likelihood theory. We also analyzed the nine outcomes separately before and after multiple imputation. We observed joint effects of patient-, clinician- and hospital-level factors on pneumonia care indicators before and after multiple imputation of missing covariates. In both pairwise joint modeling and separate univariate analysis methods, enhanced audit and feedback improved documentation and adherence to recommended clinical guidelines over time in six and five pneumonia care indicators, respectively. Additionally, multiple imputation improved precision of parameter estimates compared to complete case analysis. The strength and direction of association among pneumonia outcomes varied within and across the three domains of pneumonia care

Imputation of Missing Covariates in Randomized Controlled Trials with Continuous Outcomes: Simple, Unbiased and Efficient Methods

ABSTRACT The literature on dealing with missing covariates in nonrandomized studies advocates the use of sophisticated methods like multiple imputation (MI) and maximum likelihood (ML)-based approaches over simple methods. However, these methods are not necessarily optimal in terms of bias and efficiency of treatment effect estimation in randomized studies, where the covariate of interest (treatment group) is independent of all baseline (pre-randomization) covariates due to randomization. This has been shown in the literature, but only for missingness on a single baseline covariate. Here, we extend the situation to multiple baseline covariates with missingness and evaluate the performance of MI and ML compared with simple alternative methods under various missingness scenarios in RCTs with a quantitative outcome. We first derive asymptotic relative efficiencies of the simple methods under the missing completely at random (MCAR) scenario and then perform a simulation study for non-MCAR scenarios. Finally, a trial on chronic low back pain is used to illustrate the implementation of the methods. The results show that all simple methods give unbiased treatment effect estimation but with increased mean squared residual. It also turns out that mean imputation and the missing-indicator method are most efficient under all covariate missingness scenarios and perform at least as well as MI and LM in each scenario.

Clustering and Prediction With Variable Dimension Covariates

In many applied fields incomplete covariate vectors are commonly encountered. It is well known that this can be problematic when making inference on model parameters, but its impact on prediction performance is less understood. We develop a method based on covariate dependent random partition models that seamlessly handles missing covariates while completely avoiding any type of imputation. The method we develop allows in-sample as well as out-of-sample predictions, even if the missing pattern in the new subjects’ incomplete covariate vector was not seen in the training data. Any data type, including categorical or continuous covariates are permitted. In simulation studies, the proposed method compares favorably. We illustrate the method in two application examples. Supplementary materials for this article are available here.

Association of lower dementia risk with improved air quality in older women

AbstractBackgroundLate‐life ambient air pollution is a novel risk factor for brain aging. Air quality improvement has been associated with improved cardiopulmonary health and reduced mortality; however, whether it is associated with reduced dementia risk is unknown.MethodWe studied a geographically‐diverse cohort from the Women’s Health Initiative Memory Study‐Epidemiology of Cognitive Health Outcomes (WHIMS‐ECHO). Participants included 2239 older women (aged 74‐92) without dementia at WHIMS‐ECHO enrollment in 2008‐12. Incident dementia cases, based on DSM‐IV criteria, were centrally‐adjudicated annually (2008‐18). Yearly mean concentrations of ambient PM2.5 (fine particulate matter) and NO2 at residential locations were estimated using validated regionalized universal kriging models. These annual estimates were then averaged over 3‐year periods at both immediately preceding (recent exposure) and 10 years prior to (remote exposure) WHIMS‐ECHO enrollment. Individual‐level improved air quality was defined as reduction from remote to recent exposures. We used Cox proportional hazard models to estimate the hazard ratios (HR) for dementia risk associated with air quality measures, adjusting for sociodemographic, lifestyle, and clinical characteristics. We incorporated inverse‐probability weighting to account for differential attrition during the follow up. Models were replicated using multiple imputation to include those with missing air pollution and covariate data as sensitivity analyses.ResultsWe identified 398 incident dementia cases during median 6.1 years of follow up (interquartile range [IQR]=5.2 years). PM2.5 and NO2 reduced significantly over the 10 years before WHIMS‐ECHO enrollment. Greater improvement in air quality was significantly associated with reduced dementia risks (HRPM2.5=0.80 per IQR=1.78µg/m3, 95% confidence interval [CI]=[0.71, 0.91]; HRNO2=0.80 per IQR=3.91ppb, 95%CI=[0.71, 0.90]), equivalent to the lower risk observed in women 2.4 years younger at baseline. There was some evidence that recent PM2.5 exposure was associated with higher dementia risk (HRPM2.5=1.16 per IQR=2.90µg/m3, 95%CI=[0.98, 1.38]), but the lower dementia risk associated with improved air quality remained after further adjusting for the recent exposure effects. The observed associations did not substantially differ by age, education, geographic region, APOE e4 genotypes, or cardiovascular risk factors. Similar results were found in the sensitivity analyses using multiple imputation.ConclusionsLong‐term air quality improvement in late‐life was associated with lower dementia risk in older women.

Multimorbidity and the risk of incident dementia in UK Biobank

AbstractBackgroundThere is strong evidence that certain individual conditions, such as diabetes, hypertension and stroke, increase the risk of dementia. However, there is limited longitudinal research investigating the role of multimorbidity as a risk factor for developing dementia.MethodParticipants were selected from UK Biobank, a population‐based cohort of half a million women and men aged between 40‐73 years at recruitment. Between 2006‐2010, participants completed a range of assessments at baseline, which included self‐reporting pre‐existing health conditions during a nurse‐led verbal interview. This was used to determine multimorbidity, defined as having two or more, out of a possible 42, chronic physical and mental health conditions used in recent large‐scale epidemiology studies to establish multimorbidity. Incident all‐cause dementia was ascertained using cohort‐wide hospital inpatient and death records collected up to 2020.ResultAfter excluding participants less than 60 years old at baseline (n=285,016), those with prevalent dementia (n=230), and those with missing covariate data (n=10,350) a total of 206,974 participants remained, of whom, 89,210 (43.1%) had multimorbidity. The five most common individual conditions in the analytic sample were hypertension (36.2%), pain (21.5%), cancer (10.7%), asthma (10.7%) and dyspepsia (10.1%). Over 2.28 million person years of follow‐up (mean = 11.0, standard deviation = 1.96), 4,813 participants developed dementia. In Cox proportional‐hazards models adjusted for age, sex, ethnicity, education, socioeconomic status and APOE‐e4, multimorbidity was associated with a 70% increased risk of developing dementia (Hazard Ratio [HR] = 1.70, 95% Confidence Interval [CI] 1.60‐1.80), compared to ≤1 conditions. The association remained consistent when restricting to incident dementia diagnosed after 3, 6 and 9 years respectively. There was evidence of a dose‐response effect, with HRs of 1.44 (95% CI 1.34‐1.55), 1.68 (95% CI 1.55‐1.82), 2.33 (95% CI 2.10‐2.59), 2.80 (95% CI 2.41‐3.24) and 3.43 (95% CI 2.84‐4.15) observed for 2, 3, 4, 5 and ≥6 conditions, respectively, compared to ≤1 conditions.ConclusionIn a population of more than 200,000 older adults, multimorbidity was strongly associated with an increased risk of incident dementia. Further research investigating which particular combinations of conditions contribute to dementia risk is warranted.

Long-term exposures to PM2.5, black carbon and NO2 and prevalence of current rhinitis in French adults: The Constances Cohort.

Rhinitis is one of the most common disease worldwide with a high and increasing prevalence. There is limited knowledge on the link between long-term exposure to air pollution and rhinitis. We aim to study the associations between long-term exposure to air pollutants and self-reported current rhinitis among adults from Constances, a large French population-based cohort. Current rhinitis was defined at inclusion from questionnaire by the presence of sneezing, runny or blocked nose in the last 12months without a cold or the flu. Annual concentrations of nitrogen dioxide (NO2), particulate matter≤2.5µm (PM2.5) and black carbon (BC) were estimated at the participants' residential address by European land-use regression models. Cross-sectional associations between annual exposure to each air pollutant and current rhinitis were estimated using logistic models adjusted for age, sex, smoking, education level and French deprivation index. The health prevention centers were taken into account by marginal models with generalized estimating equations. Several supplementary analyses were performed. Analyses were performed on 127,108 participants (47years old on average, 54% women, 19% current smokers). The prevalence of current rhinitis, allergic (AR) and non-allergic (NAR) rhinitis were 36%, 25% and 11% respectively. Adjusted ORs for current rhinitis were 1.13 (1.08, 1.17), 1.12 (1.07, 1.17), and 1.11 (1.06, 1.17) per one interquartile range increase of PM2.5 (4.85µg/m3), BC (0.88·10-5 m-1) and NO2 (17.3µg/m3) respectively. Significant and positive associations were also found for both AR and NAR. Results were similar in men and women, and in the different smoking strata, and were consistent with meta-analysis or after imputing missing covariates. An increase of modeled annual average residential exposure to PM2.5, BC, and NO2 was significantly associated with an increase of prevalence of current rhinitis in adults in the French general population. The results suggest that among air pollutants, BC may be of special interest.

Psychosocial determinants of cardiovascular events among black Americans with chronic kidney disease or associated risk factors in the Jackson heart study

BackgroundIndividuals with chronic kidney disease (CKD), hypertension (HTN), or diabetes mellitus (DM) are at increased risk for cardiovascular disease (CVD). The extent to which psychosocial factors are associated with increased CVD risk within these individuals is unclear. Black individuals experience a high degree of psychosocial stressors due to socioeconomic factors, environment, racism, and discrimination. We examined the association between psychosocial factors and risk of CVD events among Black men and women with CKD and CKD risk factors in the Jackson Heart Study.Methods and ResultsWe identified 1919 participants with prevalent CKD or CKD risk factors at baseline. We used rotated principal component analysis - a form of unsupervised machine learning that may identify constructs not intuitively identified by a person - to describe five groups of psychosocial components (including negative moods, religiosity, discrimination, negative outlooks, and negative coping resources) based on a battery of questionnaires. Multiple imputation by chained equation (MICE) was used to impute missing covariate data. Cox models were used to quantify the association between psychosocial components and incident CVD, defined as a fatal coronary heart disease event, myocardial infarction, cardiac procedure (angiography or revascularization procedure), or stroke. Of the 929 participants in the analysis, 67% were female, 28% were current/former smokers with mean age of 56 years and mean BMI of 33 kg/m2. Over a median follow-up of 8 years, 6% had an incident CVD event. In multivariable models, each standard deviation (SD) increase in the religiosity component was associated with an increased hazard for CVD event (hazard ratio [HR] = 1.52, 95% CI: 1.09–2.13).ConclusionsReligiosity was associated with CVD among participants with prevalent CKD or CKD risk factors. Studies to better understand the mechanisms of this relationship are needed.

Real World Survival Outcomes of CPX-351 Versus Venetoclax and Azacitadine for Initial Therapy in Adult Acute Myeloid Leukemia

Real World Survival Outcomes of CPX-351 Versus Venetoclax and Azacitadine for Initial Therapy in Adult Acute Myeloid Leukemia

Response to cardiac resynchronisation therapy in men and women: a secondary analysis of the SMART-AV randomised controlled trial

ObjectivesThere is a controversy about whether both sexes’ response to cardiac resynchronisation therapy (CRT) is similar. We aimed to assess a causal effect of sex on CRT response.DesignSecondary analysis of...

Individual Patient Data Meta-Analysis and Network Meta-Analysis.

Meta-analyses are often conducted using trial-level summary data. However, when individual patient data (IPD ) is available, there is greater flexibility in the analysis and a wider range of statistical models that can be fitted. There are two approaches to fitting IPD models. The traditional two-stage approach involves analyzing each trial individually in the first stage and then combining trial estimates of treatment effectiveness in the second stage using methods developed for aggregate data meta-analysis. Growing in popularity is the one-stage approach in which trials are analyzed and synthesized within one statistical model whilst the clustering of patients within trials is accounted for. This chapter outlines both fixed effect and random effects one- and two-stage meta-analysis models for continuous, binary, and time-to-event outcomes. The meta-analysis framework is then extended to the scenario where there are more than two treatments and network meta-analysis models are described.The availability of IPD provides greater statistical power for investigating interactions between treatments and covariates. Treatment-covariate interactions contain both within- and across-trial information where the across-trial information may be subject to ecological bias. This chapter presents network meta-analysis models separating out the within- and across-trial information and finishes by considering practical solutions for dealing with missing covariate data, assessing the consistency assumption, combining IPD and aggregate data and specific considerations for time-to-event outcomes.

Maternal occupational exposure to chemicals and neurocognitive development at 4-5 years of age

BACKGROUND AND AIM: Despite evidence that suggests fetal brain development is particularly sensitive to endocrine disrupting chemicals (EDCs), limited data exist regarding childhood neurodevelopment in relation to maternal exposure to EDCs at levels relevant to occupational settings. This study’s objective was to examine the association between maternal occupational EDC exposures and childhood neurocognitive function at 4-5 years of age. METHODS: This analysis included 1,058 mother-child pairs in the {INfancia y Medio Ambiente} (INMA) project (2003-2008). Likelihood of exposure to several EDC groups was estimated using a job-exposure matrix and assigned to each mother based on her self-reported longest held job during her index pregnancy. At her child’s 4-5-year follow-up visit, the McCarthy Scales of Children’s Abilities was administered by a trained psychologist, yielding a general cognitive index and sub-scales of verbal, perceptive-performance and quantitative reasoning. We imputed missing covariate data and analyzed associations separately for each cohort using linear regression, weighting for inverse probability of attrition and adjusting for covariates informed by a directed acyclic graph: maternal age, education, country of birth, parity, gestational weight gain, pre-pregnancy body mass index, smoking and alcohol use during pregnancy, verbal intelligence, child sex and age at assessment, and the psychologist who administered the assessment. Finally, we combined cohort-specific effect estimates using random-effects meta-analysis. RESULTS:We observed an association between maternal occupational exposure to organic solvents and decreased quantitative reasoning scores at age 4-5 (-5.8, 95% confidence interval: -11.0, -0.5). Associations of similar magnitude and direction were observed for the same domain in relation to exposure to alkylphenolic compounds, phthalates and miscellaneous chemicals (i.e., benzophenones, parabens and siloxanes), though these were not statistically significant. CONCLUSIONS:This study found limited evidence for a role of maternal occupational EDC exposures on childhood cognitive function. Further research should consider the combined impact of multiple workplace exposures. KEYWORDS: neurodevelopmental outcomes, chemical exposures, occupational exposures, endocrine disrupting chemicals, solvents, occupational epidemiology

The development and validation of prognostic models for overall survival in the presence of missing data in the training dataset: a strategy with a detailed example

BackgroundThe United Kingdom Myeloma Research Alliance (UK-MRA) Myeloma Risk Profile is a prognostic model for overall survival. It was trained and tested on clinical trial data, aiming to improve the stratification of transplant ineligible (TNE) patients with newly diagnosed multiple myeloma. Missing data is a common problem which affects the development and validation of prognostic models, where decisions on how to address missingness have implications on the choice of methodology.MethodsModel buildingThe training and test datasets were the TNE pathways from two large randomised multicentre, phase III clinical trials. Potential prognostic factors were identified by expert opinion. Missing data in the training dataset was imputed using multiple imputation by chained equations. Univariate analysis fitted Cox proportional hazards models in each imputed dataset with the estimates combined by Rubin’s rules. Multivariable analysis applied penalised Cox regression models, with a fixed penalty term across the imputed datasets. The estimates from each imputed dataset and bootstrap standard errors were combined by Rubin’s rules to define the prognostic model.Model assessmentCalibration was assessed by visualising the observed and predicted probabilities across the imputed datasets. Discrimination was assessed by combining the prognostic separation D-statistic from each imputed dataset by Rubin’s rules.Model validationThe D-statistic was applied in a bootstrap internal validation process in the training dataset and an external validation process in the test dataset, where acceptable performance was pre-specified.Development of risk groupsRisk groups were defined using the tertiles of the combined prognostic index, obtained by combining the prognostic index from each imputed dataset by Rubin’s rules.ResultsThe training dataset included 1852 patients, 1268 (68.47%) with complete case data. Ten imputed datasets were generated. Five hundred twenty patients were included in the test dataset. The D-statistic for the prognostic model was 0.840 (95% CI 0.716–0.964) in the training dataset and 0.654 (95% CI 0.497–0.811) in the test dataset and the corrected D-Statistic was 0.801.ConclusionThe decision to impute missing covariate data in the training dataset influenced the methods implemented to train and test the model. To extend current literature and aid future researchers, we have presented a detailed example of one approach. Whilst our example is not without limitations, a benefit is that all of the patient information available in the training dataset was utilised to develop the model.Trial registrationBoth trials were registered; Myeloma IX-ISRCTN68454111, registered 21 September 2000. Myeloma XI-ISRCTN49407852, registered 24 June 2009.