Thirteen Questions About Using Machine Learning in Causal Research (You Won't Believe the Answer to Number 10!).

Machine learning is gaining prominence in the health sciences, where much of its use has focused on data-driven prediction. However, machine learning can also be embedded within causal analyses, potentially reducing biases arising from model misspecification. Using a question-and-answer format, we provide an introduction and orientation for epidemiologists interested in using machine learning but concerned about potential bias or loss of rigor due to use of "black box" models. We conclude with sample software code that may lower the barrier to entry to using these techniques.

Use of Intrauterine Devices and Risk of Human Immunodeficiency Virus Acquisition Among Insured Women in the United States.

Concerns have been raised about progestin-containing contraceptives and the risk of human immunodeficiency virus (HIV) acquisition. Based on health insurance data from women in the United States with intrauterine device (IUD) insertions during 2011-2018, there was no increased risk of incident HIV diagnosis for levonorgestrel-releasing IUDs versus copper IUDs.

Defining and Identifying Per-protocol Effects in Randomized Trials.

In trials with noncompliance to assigned treatment, researchers might be interested in estimating a per-protocol effect-a comparison of two counterfactual outcomes defined by treatment assignment and (often time-varying) compliance with a well-defined treatment protocol. Here, we provide a general counterfactual definition of a per-protocol effect and discuss examples of per-protocol effects that are of either substantive or methodologic interest. In doing so, we seek to make more concrete what per-protocol effects are and highlight that one can estimate per-protocol effects that are more than just a comparison of always taking treatment in two distinct treatment arms. We then discuss one set of identifiability conditions that allow for identification of a causal per-protocol effect, highlighting some potential violations of those conditions that might arise when estimating per-protocol effects.

Brief Report: Estimating Differences and Ratios in Median Times to Event.

Time differences and time ratios are often more interpretable estimates of effect than hazard ratios for time-to-event data, especially for common outcomes. We developed a SAS macro for estimating time differences and time ratios between baseline-fixed binary exposure groups based on inverse probability-weighted Kaplan-Meier curves. The macro uses pooled logistic regression to calculate inverse probability of censoring and exposure weights, draws Kaplan-Meier curves based on the weighted data, and estimates the time difference and time ratio at a user-defined survival proportion. The macro also calculates the risk difference and risk ratio at a user-specified time. Confidence intervals are constructed by bootstrap. We provide an example assessing the effect of exclusive breastfeeding during diarrhea on the incidence of subsequent diarrhea in children followed from birth to 3 years in Vellore, India. The SAS macro provided here should facilitate the wider reporting of time differences and time ratios.

Epidemiology's continuing contribution to public health: The power of "Then and Now".

The 47th annual meeting of the Society for Epidemiologic Research hosted 17 invited speakers charged by the Executive Committee with presenting some of the many ways that epidemiologists have improved the health of the general population. There were 9 "Then and Now" sessions that were structured to focus on how early epidemiologists overcame research hurdles and advanced health through innovative strategies. For most topics, a longstanding expert was paired with an excellent contemporary epidemiologist working in the area, and both were given the freedom to deliver an integrated story about epidemiology's temporal role in protecting and promoting public health. Epidemiologic discoveries in cardiovascular, cancer, and perinatal epidemiology were discussed on day 1, followed by discussions of accomplishments in reducing exposures that adversely impact health (nutrition, environment/occupation, and tobacco use) on day 2. Topics with relevancy for many aspects of epidemiology were presented on day 3, including infectious diseases, social forces, and causal thinking in epidemiologic research. Given the large number of outstanding senior and junior epidemiologists that attended the meeting, choosing speakers was a unique challenge. What became evident from all sessions was the passion that epidemiologists have for population health, tempered with concerns for remaining true to epidemiologic principles, the timely adoption of innovative methods, and the responsible interpretation of research findings.

Early Life Antibiotic Exposure Is Not Associated with Growth in Young Children of Vellore, India.

OBJECTIVES: To estimate the effects of antibiotic exposures in the first 6 months of life on short- and long-term growth. STUDY DESIGN: In a prospective observational cohort study of 497 children from Vellore, India, we estimated short-term effects of antibiotics during the first 6 months using longitudinal general linear regression to model weight-for-age, height-for-age, and weight-for-height z-scores in monthly intervals. To estimate long-term effects, we modeled growth from 6 months to 3 years as a function of antibiotic use in the first 6 months. We also estimated the effects of antibiotics on the monthly relative risks of underweight, stunting, and wasting in the first 6 months and to 3 years. RESULTS: Underweight, stunting, and wasting were common in this population: 31%, 32%, and 15% on average after 6 months of age, respectively. There was no association between antibiotic exposures before 6 months and growth during that period. From 6 months to 3 years, adjusted absolute differences in weight and height were small (approximately -100 g and no more than -2 mm overall, respectively) and not statistically significant. CONCLUSIONS: Antibiotic exposures early in life were not associated with increased or decreased growth. The combination of malnutrition and recurrent illness likely complicate the relationship between antibiotic exposures and growth among children in low and middle-income countries.

Commentary: Epidemiology in the era of big data.

Big Data has increasingly been promoted as a revolutionary development in the future of science, including epidemiology. However, the definition and implications of Big Data for epidemiology remain unclear. We here provide a working definition of Big Data predicated on the so-called "three V's": variety, volume, and velocity. From this definition, we argue that Big Data has evolutionary and revolutionary implications for identifying and intervening on the determinants of population health. We suggest that as more sources of diverse data become publicly available, the ability to combine and refine these data to yield valid answers to epidemiologic questions will be invaluable. We conclude that while epidemiology as practiced today will continue to be practiced in the Big Data future, a component of our field's future value lies in integrating subject matter knowledge with increased technical savvy. Our training programs and our visions for future public health interventions should reflect this future.

Postnatal cytomegalovirus exposure in infants of antiretroviral-treated and untreated HIV-infected mothers.

HIV-1 and CMV are important pathogens transmitted via breastfeeding. Furthermore, perinatal CMV transmission may impact growth and disease progression in HIV-exposed infants. Although maternal antiretroviral therapy reduces milk HIV-1 RNA load and postnatal transmission, its impact on milk CMV load is unclear. We examined the relationship between milk CMV and HIV-1 load (4-6 weeks postpartum) and the impact of antiretroviral treatment in 69 HIV-infected, lactating Malawian women and assessed the relationship between milk CMV load and postnatal growth in HIV-exposed, breastfed infants through six months of age. Despite an association between milk HIV-1 RNA and CMV DNA load (0.39 log(10) rise CMV load per log(10) rise HIV-1 RNA load, 95% CI 0.13-0.66), milk CMV load was similar in antiretroviral-treated and untreated women. Higher milk CMV load was associated with lower length-for-age (-0.53, 95% CI: -0.96, -0.10) and weight-for-age (-0.40, 95% CI: -0.67, -0.13) Z-score at six months in exposed, uninfected infants. As the impact of maternal antiretroviral therapy on the magnitude of postnatal CMV exposure may be limited, our findings of an inverse relationship between infant growth and milk CMV load highlight the importance of defining the role of perinatal CMV exposure on growth faltering of HIV-exposed infants.

Association Between HIV and Prevalence and Manifestations of Asthma: Analysis of the Multicenter AIDS Cohort Study and Women's Interagency HIV Study.

BACKGROUND: The association between HIV and asthma prevalence and manifestations remains unclear, with few studies including women. SETTING: A retrospective observational cohort study from the Multicenter AIDS Cohort Study and Women's Interagency HIV Study. METHODS: Asthma was defined in 2 ways: (1) self-report and (2) robust criteria requiring all the following: lack of fixed airflow obstruction, presence of wheeze on the St. George's Respiratory Questionnaire (SGRQ), and report of asthma therapies. Estimates of asthma prevalence and asthma-related manifestations were compared by HIV serostatus. RESULTS: A total of 1815 men and 2122 women were included. Asthma prevalence did not differ between people with HIV (PWH) and people without HIV regardless of definition: self-report (men, 12.0% vs. 11.2%; women, 24.3% vs. 27.5%) and robust criteria (men, 5.0% vs. 3.4%; women, 12.8% vs. 13.2%). Among men with asthma, worse respiratory symptom burden was reported among those with HIV, regardless of asthma definition. Among women with self-reported asthma, those with HIV had less respiratory symptom burden. Regardless of serostatus, women with robust-defined asthma had similar respiratory symptoms across SGRQ domains and similar frequencies of phlegm, shortness of breath, and wheezing. CONCLUSIONS: Among PWH and people without HIV, asthma prevalence was 2-fold to 3-fold higher using self-reported definition rather than robust definition. In men and women, HIV was not associated with increased asthma prevalence. In men, HIV was associated with more respiratory symptoms when asthma was self-reported; the relationship was attenuated with the robust criteria. Further studies are needed to explore asthma phenotypes among PWH.

A comparison of methods to estimate the hazard ratio under conditions of time-varying confounding and nonpositivity.

In occupational epidemiologic studies, the healthy worker survivor effect refers to a process that leads to bias in the estimates of an association between cumulative exposure and a health outcome. In these settings, work status acts both as an intermediate and confounding variable and may violate the positivity assumption (the presence of exposed and unexposed observations in all strata of the confounder). Using Monte Carlo simulation, we assessed the degree to which crude, work-status adjusted, and weighted (marginal structural) Cox proportional hazards models are biased in the presence of time-varying confounding and nonpositivity. We simulated the data representing time-varying occupational exposure, work status, and mortality. Bias, coverage, and root mean squared error (MSE) were calculated relative to the true marginal exposure effect in a range of scenarios. For a base-case scenario, using crude, adjusted, and weighted Cox models, respectively, the hazard ratio was biased downward 19%, 9%, and 6%; 95% confidence interval coverage was 48%, 85%, and 91%; and root MSE was 0.20, 0.13, and 0.11. Although marginal structural models were less biased in most scenarios studied, neither standard nor marginal structural Cox proportional hazards models fully resolve the bias encountered under conditions of time-varying confounding and nonpositivity.

Adverse maternal and neonatal outcomes among women with preeclampsia with severe features <34 weeks gestation with versus without comorbidity.

OBJECTIVES: To determine adverse maternal and neonatal outcomes among women with preeclampsia with severe features who delivered <34 weeks comparing those with versus without a comorbid condition. STUDY DESIGN: A retrospective analysis from the U.S. Consortium on Safe Labor Study of deliveries <34 weeks with preeclampsia with severe features. We examined the association of each comorbid condition versus none with adverse maternal and neonatal outcomes. The comorbidities (not mutually exclusive) were chronic hypertension, pregestational diabetes, gestational diabetes, twin gestation, and fetal growth restriction. MAIN OUTCOMES: Maternal outcome: eclampsia, thromboembolism, ICU admission, and/or death; and neonatal outcome: intracranial/periventricular hemorrhage, hypoxic-ischemic encephalopathy/periventricular leukomalacia, stillbirth, and/or perinatal death. RESULTS: Among 2217 deliveries, 50% had a comorbidity, namely chronic hypertension (30%), pregestational diabetes (8%), gestational diabetes (8%), twin gestation (10%), and fetal growth restriction (7%). Adverse maternal and neonatal outcomes occurred in 10% and 12% of pregnancies, respectively. Pregnancies with preeclampsia with severe features delivered <34 weeks complicated by gestational diabetes (adjusted risk difference, aRD: -4.9%, 95%CI: -9.11 to -0.71), twin gestation (aRD: -5.1%, 95%CI: -8.63 to -1.73), and fetal growth restriction (aRD: -4.7%, 95%CI: -7.96 to -1.62) were less likely to result in adverse maternal outcome compared to pregnancies without comorbidity, but not chronic hypertension and pregestational diabetes. A pregnancy complicated by fetal growth restriction (aRD: 12.2%, 95%CI: 5.48 to 19.03) was more likely to result in adverse neonatal outcome, but not other comorbid conditions. CONCLUSIONS: Preeclampsia with severe features <34 weeks complicated by comorbidity was generally not associated with an increased risk of adverse maternal and neonatal outcomes, with the exception of fetal growth restriction.

Tuberculosis treatment and risk of stavudine substitution in first-line antiretroviral therapy.

BACKGROUND: Treatment for tuberculosis (TB) is common among individuals receiving stavudine-containing highly active antiretroviral therapy (HAART), but the effect of TB treatment on stavudine toxicity has received little attention. We estimated the effect of TB treatment on risk of stavudine substitution among individuals receiving first-line HAART. METHODS: We evaluated a cohort of 7066 patients who initiated HAART from April 2004 through March 2007 in Johannesburg, South Africa. Three exposure categories were considered: ongoing TB treatment at HAART initiation, concurrent initiation of TB treatment and HAART, and incident TB treatment after HAART initiation. The outcome was single-drug stavudine substitution. Adjusted hazard ratios (aHRs) were estimated using marginal structural models to control for confounding, loss to follow-up, and competing risks. RESULTS: Individuals with ongoing and concurrent TB treatment were at increased risk of stavudine substitution, irrespective of stavudine dosage. For ongoing TB treatment, aHR was 3.18 (95% confidence interval [CI], 1.82-5.56) in the first 2 months of HAART, 2.51 (95% CI, 1.77-3.54) in months 3-6, and 1.19 (95% CI, 0.94-1.52) thereafter. For concurrent TB treatment, aHR was 6.60 (95% CI, 3.03-14.37) in the first 2 months, 1.88 (95% CI, 0.87-4.09) in months 3-6, and 1.07 (95% CI, 0.65-1.76) thereafter. There was no effect of incident TB on stavudine substitution risk. CONCLUSIONS: Risk of stavudine substitution was increased among patients who received TB treatment and was especially elevated during the period soon after HAART initiation. In settings in which alternative antiretroviral drugs are available, initiation of stavudine therapy in patients receiving TB treatment may need to be reconsidered.

Optimizing screening for acute human immunodeficiency virus infection with pooled nucleic acid amplification tests.

Recent studies have shown the public health importance of identifying individuals with acute human immunodeficiency virus infection (AHI); however, the cost of nucleic acid amplification testing (NAAT) makes individual testing of at-risk individuals prohibitively expensive in many settings. Pooled NAAT (or group testing) can improve efficiency and test performance of testing for AHI, but optimizing the pooling algorithm can be difficult. We developed simple, flexible biostatistical models of specimen pooling with NAAT for the identification of AHI cases; these models incorporate group testing theory, operating characteristics of biological assays, and a model of viral dynamics during AHI. Pooling algorithm sensitivity, efficiency (test kits used per individual specimen evaluated), and positive predictive value (PPV) were modeled and compared for three simple pooling algorithms: two-stage minipools (D2), three-stage hierarchical pools (D3), and square arrays with master pools (A2m). We confirmed the results by stochastic simulation and produced reference tables and a Web calculator to facilitate pooling by investigators without specific biostatistical expertise. All three pooling strategies demonstrated improved efficiency and PPV for AHI case detection compared to individual NAAT. D3 and A2m algorithms generally provided better efficiency and PPV than D2; additionally, A2m generally exhibited better PPV than D3. Used selectively and carefully, the simple models developed here can guide the selection of a pooling algorithm for the detection of AHI cases in a wide variety of settings.

Past-month cannabis use among U.S. individuals from 2002-2015: An age-period-cohort analysis.

BACKGROUND: Cannabis is the most commonly used illicit drug among U.S. adolescents and adults, but little is known about factors that drive trends in cannabis use prevalence. To better understand drivers of these trends, we aimed to estimate age, period, and cohort effects on past-month cannabis use among U.S. individuals age 12 and older from 2002 to 2015. METHODS: We conducted an age-period-cohort analysis on past-month cannabis use among participants ages 12 and older using the National Survey on Drug Use and Health (NSDUH), an annual cross-sectional nationally-representative survey of drug use. Additionally, we examined how age, period, and cohort effects differed across gender. Participants (n = 779,799) self-reported cannabis patterns using a computer-assisted telephone interview (CATI). RESULTS: Past-month cannabis use in this population increased from 6.0% in 2002 to 8.1% in 2015. Distinct age, period, and cohort effects were observed. Compared to participants ages 12-13, participants ages 18-21 (PR: 16.8, 95% CI: 15.6, 18.1) and 22-25 (PR: 13.2, 95% CI: 12.2, 14.4) had dramatically higher prevalence of past-month cannabis use. Compared to participants in 2002, participants in 2014 (PR: 1.2, 95% CI: 1.1, 1.4) and 2014 (PR: 1.2, 95% CI: 1.1, 1.4) had slightly higher prevalence of past-month cannabis use. Compared to the 1940s birth cohort, the 1950s birth cohort (PR: 1.8, 95% CI: 1.5, 2.2) had a higher prevalence of past-month cannabis use. CONCLUSIONS: Past-month cannabis use is prevalent and increasing among U.S. adults. Distinct age, period, and cohort effects are at play, though age effects are strongest.

Association of Household Opioid Availability and Prescription Opioid Initiation Among Household Members.

Importance: Increases in prescription opioid use in the United States have been attributed to changing prescribing guidelines and attitudes toward pain relief; however, the spread of opioid use within households through drug diversion may also be a contributing factor. Objective: To investigate whether individuals living in a household with a prescription opioid user are more likely to initiate prescription opioids themselves, compared with individuals in households with a prescription nonsteroidal anti-inflammatory drug (NSAID) user. Design, Setting, and Participants: This was a retrospective cohort study using administrative health care claims data from 2000 to 2014 of commercial insurance beneficiaries sharing a health plan with continuous prescription drug coverage, without opioid or NSAID use in the prior year. Enrollees were followed from the date of the first prescription filled by a household member for an eligible opioid or NSAID until initiation of prescription opioids, disenrollment, or administrative censoring after 1 year or the end of follow-up on December 31, 2014. Risk of opioid initiation was derived from inverse probability-weighted (IPW) Kaplan-Meier estimators that adjusted for potential confounders, prognostic factors, and predictors of censoring. Exposure: Outpatient pharmacy dispensing of a prescription opioid or prescription NSAID. Main Outcomes and Measures: Incident outpatient pharmacy fill for a prescription opioid by a household member. Results: From 2000 to 2014, 12 695 280 individuals were exposed to prescription opioids and 6 359 639 to prescription NSAIDS through an index prescription to a household member. The IPW estimated risk of opioid initiation in the subsequent year was 11.83% (95% CI, 11.81%-11.85%) among individuals exposed to prescription opioids in the household, compared with 11.11% (95% CI, 11.09%-11.14%) among individuals exposed to prescription NSAIDs, resulting in a risk difference of 0.71% (95% CI, 0.68%-0.74%). An unmeasured confounder that is modestly associated with the exposure (eg, prevalence difference = 0.9%) and the outcome (eg, risk difference = 0.9) after adjustment for all measured variables could explain our observed estimate of the overall risk difference (0.71%). Conclusions and Relevance: Living in a household with a prescription opioid user may increase risk of prescription opioid use, which may reflect both increased access to these products as well as shared risk factors, such as prescriber preference and prescription drug monitoring.

Reduction in diarrhoeal rates through interventions that prevent unnecessary antibiotic exposure early in life in an observational birth cohort.

BACKGROUND: Antibiotic treatment early in life is often not needed and has been associated with increased rates of subsequent diarrhoea. We estimated the impact of realistic interventions, which would prevent unnecessary antibiotic exposures before 6 months of age, on reducing childhood diarrhoeal rates. METHODS: In data from a prospective observational cohort study conducted in Vellore, India, we used the parametric g-formula to model diarrhoeal incidence rate differences contrasting the observed incidence of diarrhoea to the incidence expected under hypothetical interventions. The interventions prevented unnecessary antibiotic treatments for non-bloody diarrhoea, vomiting and upper respiratory infections before 6 months of age. We also modelled targeted interventions, in which unnecessary antibiotic use was prevented only among children who had already stopped exclusive breast feeding. RESULTS: More than half of all antibiotic exposures before 6 months (58.9%) were likely unnecessary. The incidence rate difference associated with removing unnecessary antibiotic use before 6 months of age was -0.28 (95% CI -0.46 to -0.08) episodes per 30 child-months. This implies that preventing unnecessary antibiotic exposures in just 4 children would reduce the incidence of diarrhoea by 1 from 6 months to 3 years of age. CONCLUSIONS: Interventions to reduce unnecessary antibiotic use among young children could result in an important reduction in diarrhoeal rates. This work provides an example application of statistical methods which can further the aim of presenting epidemiological findings that are relevant to public health practice.

An updated systematic review of epidemiological evidence on hormonal contraceptive methods and HIV acquisition in women.

OBJECTIVE AND DESIGN: Some studies suggest that specific hormonal contraceptive methods [particularly depot medroxyprogesterone acetate (DMPA)] may increase women's HIV acquisition risk. We updated a systematic review to incorporate recent epidemiological data. METHODS: We searched for articles published between 15 January 2014 and 15 January 2016 and hand-searched reference lists. We identified longitudinal studies comparing users of a specific hormonal contraceptive method against either nonusers of hormonal contraception or users of another specific hormonal contraceptive method. We added newly identified studies to those in the previous review, assessed study quality, created forest plots to display results, and conducted a meta-analysis for data on DMPA versus non-use of hormonal contraception. RESULTS: We identified 10 new reports of which five were considered 'unlikely to inform the primary question'. We focus on the other five reports, along with nine from the previous review, which were considered 'informative but with important limitations'. The preponderance of data for oral contraceptive pills, injectable norethisterone enanthate, and levonorgestrel implants do not suggest an association with HIV acquisition, though data for implants are limited. The new, higher quality studies on DMPA (or nondisaggregated injectables), which had mixed results in terms of statistical significance, had hazard ratios between 1.2 and 1.7, consistent with our meta-analytic estimate for all higher quality studies of hazard ratio 1.4. CONCLUSION: Although confounding in these observational data cannot be excluded, new information increases concerns about DMPA and HIV acquisition risk in women. If the association is causal, the magnitude of effect is likely hazard ratio 1.5 or less. Data for other hormonal contraceptive methods, including norethisterone enanthate, are largely reassuring.