Brian K. Lee

Improving propensity score weighting using machine learning

Machine learning techniques such as classification and regression trees (CART) have been suggested as promising alternatives to logistic regression for the estimation of propensity scores. The authors examined the performance of various CART-based propensity score models using simulated data. Hypothetical studies of varying sample sizes (n=500, 1000, 2000) with a binary exposure, continuous outcome, and 10 covariates were simulated under seven scenarios differing by degree of non-linear and non-additive associations between covariates and the exposure. Propensity score weights were estimated using logistic regression (all main effects), CART, pruned CART, and the ensemble methods of bagged CART, random forests, and boosted CART. Performance metrics included covariate balance, standard error, per cent absolute bias, and 95 per cent confidence interval (CI) coverage. All methods displayed generally acceptable performance under conditions of either non-linearity or non-additivity alone. However, under conditions of both moderate non-additivity and moderate non-linearity, logistic regression had subpar performance, whereas ensemble methods provided substantially better bias reduction and more consistent 95 per cent CI coverage. The results suggest that ensemble methods, especially boosted CART, may be useful for propensity score weighting.

Parental depression, maternal antidepressant use during pregnancy, and risk of autism spectrum disorders: population based case-control study

Objective To study the association between parental depression and maternal antidepressant use during pregnancy with autism spectrum disorders in offspring. Design Population based nested case-control study. Setting Stockholm County, Sweden, 2001-07. Participants 4429 cases of autism spectrum disorder (1828 with and 2601 without intellectual disability) and 43 277 age and sex matched controls in the full sample (1679 cases of autism spectrum disorder and 16 845 controls with data on maternal antidepressant use nested within a cohort (n=589 114) of young people aged 0-17 years. Main outcome measure A diagnosis of autism spectrum disorder, with or without intellectual disability. Exposures Parental depression and other characteristics prospectively recorded in administrative registers before the birth of the child. Maternal antidepressant use, recorded at the first antenatal interview, was available for children born from 1995 onwards. Results A history of maternal (adjusted odds ratio 1.49, 95% confidence interval 1.08 to 2.08) but not paternal depression was associated with an increased risk of autism spectrum disorders in offspring. In the subsample with available data on drugs, this association was confined to women reporting antidepressant use during pregnancy (3.34, 1.50 to 7.47, P=0.003), irrespective of whether selective serotonin reuptake inhibitors (SSRIs) or non-selective monoamine reuptake inhibitors were reported. All associations were higher in cases of autism without intellectual disability, there being no evidence of an increased risk of autism with intellectual disability. Assuming an unconfounded, causal association, antidepressant use during pregnancy explained 0.6% of the cases of autism spectrum disorder. Conclusions In utero exposure to both SSRIs and non-selective monoamine reuptake inhibitors (tricyclic antidepressants) was associated with an increased risk of autism spectrum disorders, particularly without intellectual disability. Whether this association is causal or reflects the risk of autism with severe depression during pregnancy requires further research. However, assuming causality, antidepressant use during pregnancy is unlikely to have contributed significantly towards the dramatic increase in observed prevalence of autism spectrum disorders as it explained less than 1% of cases.

Weight Trimming and Propensity Score Weighting

Propensity score weighting is sensitive to model misspecification and outlying weights that can unduly influence results. The authors investigated whether trimming large weights downward can improve the performance of propensity score weighting and whether the benefits of trimming differ by propensity score estimation method. In a simulation study, the authors examined the performance of weight trimming following logistic regression, classification and regression trees (CART), boosted CART, and random forests to estimate propensity score weights. Results indicate that although misspecified logistic regression propensity score models yield increased bias and standard errors, weight trimming following logistic regression can improve the accuracy and precision of final parameter estimates. In contrast, weight trimming did not improve the performance of boosted CART and random forests. The performance of boosted CART and random forests without weight trimming was similar to the best performance obtainable by weight trimmed logistic regression estimated propensity scores. While trimming may be used to optimize propensity score weights estimated using logistic regression, the optimal level of trimming is difficult to determine. These results indicate that although trimming can improve inferences in some settings, in order to consistently improve the performance of propensity score weighting, analysts should focus on the procedures leading to the generation of weights (i.e., proper specification of the propensity score model) rather than relying on ad-hoc methods such as weight trimming.

Autism Spectrum Disorders in the Stockholm Youth Cohort: Design, Prevalence and Validity

Objective Reports of rising prevalence of autism spectrum disorders (ASD), along with their profound personal and societal burden, emphasize the need of methodologically sound studies to explore their causes and consequences. We here present the design of a large intergenerational resource for ASD research, along with population-based prevalence estimates of ASD and their diagnostic validity. Method The Stockholm Youth Cohort is a record-linkage study comprising all individuals aged 0–17 years, ever resident in Stockholm County in 2001–2007 (N = 589,114). ASD cases (N = 5,100) were identified using a multisource approach, involving registers covering all pathways to ASD diagnosis and care, and categorized according to co-morbid intellectual disability. Prospectively recorded information on potential determinants and consequences of ASD were retrieved from national and regional health and administrative registers. Case ascertainment was validated through case-note review, and cross validation with co-existing cases in a national twin study. Results The 2007 year prevalence of ASD in all children and young people was 11.5 per 1,000 (95% confidence interval 11.2–11.8), with a co-morbid intellectual disability recorded in 42.6% (41.0–44.2) of cases. We found 96.0% (92.0–98.4) of reviewed case-notes being consistent with a diagnosis of ASD, and confirmed ASD in 85.2% (66.2–95.8) of affected twins. Conclusions Findings from this contemporary study accords with recently reported prevalence estimates from Western countries at around 1%, based on valid case ascertainment. The Stockholm Youth Cohort, in light of the availability of extensive information from Swedens registers, constitutes an important resource for ASD research. On-going work, including collection of biological samples, will enrich the study further.

Maternal hospitalization with infection during pregnancy and risk of autism spectrum disorders.

Animal models indicate that maternal infection during pregnancy can result in behavioral abnormalities and neuropathologies in offspring. We examined the association between maternal inpatient diagnosis with infection during pregnancy and risk of ASD in a Swedish nationwide register-based birth cohort born 1984-2007 with follow-up through 2011. In total, the sample consisted of 2,371,403 persons with 24,414 ASD cases. Infection during pregnancy was defined from ICD codes. In the sample, 903 mothers of ASD cases (3.7%) had an inpatient diagnosis of infection during pregnancy. Logistic regression models adjusted for a number of covariates yielded odds ratios indicating approximately a 30% increase in ASD risk associated with any inpatient diagnosis of infection. Timing of infection did not appear to influence risk in the total Swedish population, since elevated risk of ASD was associated with infection in all trimesters. In a subsample analysis, infections were associated with greater risk of ASD with intellectual disability than for ASD without intellectual disability. The present study adds to the growing body of evidence, encompassing both animal and human studies, that supports possible immune-mediated mechanisms underlying the etiology of ASD.

Apolipoprotein e genotype, cortisol, and cognitive function in community-dwelling older adults.

BACKGROUND Elevated cortisol indicates stress and may be a risk factor for cognitive decline in aging. Genetic factors may influence individual vulnerability to the adverse effects of stress on cognitive function in aging. METHOD The authors investigated whether the gene-environment interaction between the genotype for the apolipoprotein E gene (APOE) and cortisol predicted cognitive performance in older urban adults. Cross-sectional data were analyzed from a population-based sample of 50-70-year-old men and women. Cognitive performance, salivary cortisol levels, and APOE genotype were assessed in 962 subjects. Scores on 20 standard cognitive tests were combined into seven domain scores (language, processing speed, eye-hand coordination, executive functioning, verbal memory and learning, visual memory and learning, visuoconstruction). RESULTS In adjusted models, while a higher cortisol level was associated with worse cognitive scores, the slopes of the adverse relations were steeper in persons with at least one APOE-epsilon4 allele. Effect sizes were large: for example, the effect of having one epsilon4 allele plus a cortisol area under the curve greater than the 75th percentile was equivalent to a decrease in language score expected from an age increase of 8.0 years (95% confidence interval=1.7-14.4), while having two epsilon4 alleles and a cortisol area under the curve greater than the 75th percentile was equivalent to an age increase of 33.4 years (95% CI=14.8-52.0). CONCLUSIONS These data suggest that APOE genotype modifies cortisols relations with cognitive functioning and that the epsilon4 allele increases vulnerability of the aging brain to adverse effects of stress.

Joint Effects of Obesity and Vitamin D Insufficiency on Insulin Resistance and Type 2 Diabetes: Results from the NHANES 2001–2006

OBJECTIVE The possible interaction of serum 25-hydroxyvitamin D [25(OH)D] and obesity in regard to type 2 diabetes and insulin resistance has not been well studied. To explore the effect modification of obesity on the association between 25(OH)D and insulin resistance/type 2 diabetes, data were examined from a nationally representative sample. RESEARCH DESIGN AND METHODS The analytic sample for the type 2 diabetes analysis (n = 12,900) was limited to participants from the National Health and Nutrition Examination Survey (NHANES) 2001–2006 over 20 years of age. Participants >20 years of age assigned to the morning session and free of diabetes were limited to the insulin resistance analysis (n = 5,806). Multiplicative interaction was assessed through a cross-product interaction term in a multiple logistic regression model. The presence of additive interaction between insufficient 25(OH)D and obesity (indicated by BMI or waist circumference) was evaluated by calculation of the relative excess risk due to interaction (RERI) and attributable proportion due to interaction (AP). RESULTS There was no multiplicative interaction of insufficient 25(OH)D and obesity on type 2 diabetes or insulin resistance. Furthermore, none of the RERI or AP values were statistically significant in the diabetes analysis. However, there was strong additive interaction between abdominal obesity and insufficient 25(OH)D (RERI 6.45 [95% CI 1.03–11.52]) in regard to insulin resistance. In addition, 47% of the increased odds of insulin resistance can be explained by interaction between insufficient 25(OH)D and high BMI (AP 0.47 [95% CI 0.08–0.87]). CONCLUSIONS Within a cross-sectional, nationally representative sample, abdominal obesity and insufficient 25(OH)D interact to synergistically influence the risk of insulin resistance.

Parental age and the risk of autism spectrum disorders: findings from a Swedish population-based cohort

BACKGROUND The objectives of this study were to examine the independent and dependent associations of maternal and paternal age and risk of offspring autism spectrum disorders (ASD), with and without intellectual disability (ID). METHODS The sample consisted of 417 303 Swedish children born 1984-2003. ASD case status (N = 4746) was ascertained using national and regional registers. Smoothing splines in generalized additive models were used to estimate associations of parental age with ASD. RESULTS Whereas advancing parental age increased the risk of child ASD, maternal age effects were non-linear and paternal age effects were linear. Compared with mothers at the median age 29 years, those <29 had similar risk, whereas risk increased after age 30, with an odds ratio (OR) of 1.75 [95% (CI): 1.63-1.89] at ages 40-45. For fathers, compared with the median age of 32 years, the OR for ages 55-59 was 1.39 (1.29-1.50). The risk of ASD was greater for older mothers as compared with older fathers. For example, mothers aged 40-45 (≥97.2th percentile) had an estimated 18.63 (95% CI: 17.25-20.01) ASD cases per 1000 births, whereas fathers aged 55-59 (≥99.7th percentile) had 16.35 (95% CI: 15.11-17.58) ASD cases per 1000 births. In analyses stratified by co-parental age, increased risk due to advancing paternal age was evident only with mothers ≤35 years. In contrast, advancing maternal age increased risk regardless of paternal age. Advancing parental age was more strongly associated with ASD with ID, compared with ASD without ID. CONCLUSIONS We confirm prior findings that advancing parental age increases risk of ASD, particularly for ASD with ID, in a manner dependent on co-parental age. Although recent attention has emphasized the effects of older fathers on ASD risk, an increase of n years in maternal age has greater implications for ASD risk than a similar increase in paternal age.

The Changing Epidemiology of Autism Spectrum Disorders

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition with lifelong impacts. Genetic and environmental factors contribute to ASD etiology, which remains incompletely understood. Research on ASD epidemiology has made significant advances in the past decade. Current prevalence is estimated to be at least 1.5% in developed countries, with recent increases primarily among those without comorbid intellectual disability. Genetic studies have identified a number of rare de novo mutations and gained footing in the areas of polygenic risk, epigenetics, and gene-by-environment interaction. Epidemiologic investigations focused on nongenetic factors have established advanced parental age and preterm birth as ASD risk factors, indicated that prenatal exposure to air pollution and short interpregnancy interval are potential risk factors, and suggested the need for further exploration of certain prenatal nutrients, metabolic conditions, and exposure to endocrine-disrupting chemicals. We discuss future challenges and goals for ASD epidemiology as well as public health implications.

Prognostic score–based balance measures can be a useful diagnostic for propensity score methods in comparative effectiveness research

OBJECTIVE Examining covariate balance is the prescribed method for determining the degree to which propensity score methods should be successful at reducing bias. This study assessed the performance of various balance measures, including a proposed balance measure based on the prognostic score (similar to a disease risk score), to determine which balance measures best correlate with bias in the treatment effect estimate. STUDY DESIGN AND SETTING The correlations of multiple common balance measures with bias in the treatment effect estimate produced by weighting by the odds, subclassification on the propensity score, and full matching on the propensity score were calculated. Simulated data were used, based on realistic data settings. Settings included both continuous and binary covariates and continuous covariates only. RESULTS The absolute standardized mean difference (ASMD) in prognostic scores, the mean ASMD (in covariates), and the mean t-statistic all had high correlations with bias in the effect estimate. Overall, prognostic scores displayed the highest correlations with bias of all the balance measures considered. Prognostic score measure performance was generally not affected by model misspecification, and the prognostic score measure performed well under a variety of scenarios. CONCLUSION Researchers should consider using prognostic score-based balance measures for assessing the performance of propensity score methods for reducing bias in nonexperimental studies.