Injection drug use (IDU) is a primary contributor to the syndemic of overdose,^1-3 human immunodeficiency virus (HIV),^4-6 hepatitis C,^7-9 and other associated conditions.^10,11 An estimated 3.7 million people (1.5% of the US adult population) injected drugs in 2018, about 5 times the 2011 estimate.^12,13 The incidence rate of acute hepatitis C virus (HCV) infection in the United States has quadrupled from 2011 to 2021 (0.4-1.6 cases per 100 000), and IDU is the most commonly reported risk factor.¹⁴ There were 69 800 estimated acute HCV infections during 2021, which was more than 15 times the 2030 viral hepatitis national strategic goal of 4400.¹⁵

To identify potential outbreaks, understand transmission, and better address the syndemic, it is crucial to monitor and track IDU through national viral hepatitis surveillance data. Obtaining information on IDU is challenging due to limited resources to investigate acute cases and reluctance to disclose potentially stigmatizing behaviors. In the Centers for Disease Control and Prevention’s (CDC’s) National Notifiable Diseases Surveillance System (NNDSS), more than 60% of reported acute hepatitis C cases are missing IDU data.¹⁴ The large number of missing IDU data limits the usefulness of surveillance data to inform public health prevention efforts, such as linkage to treatment and harm reduction services.

Choosing the appropriate approach to handle missing data depends on the pattern of missingness: missing completely at random (MCAR, missingness is independent of any variable), missing at random (MAR, missingness is associated with other variables), and missing not at random (MNAR, missingness is associated with the missing variable itself).¹⁶ Listwise deletion (LD) relies on the MCAR assumption, but this assumption is often invalid leading to biased estimates. Imputation of missing data generally starts with the MAR assumption and then addresses the possibility of MNAR through sensitivity analysis.¹⁷

Multiple imputation by chained equations (MICE) is a principled and flexible approach to impute missing data.^18,19 MICE generates multiple imputed datasets by iteratively predicting missing values for each variable through an imputation model based on other variables. To our knowledge, MICE has not been applied to impute missing IDU information for acute viral hepatitis cases from the national surveillance system though it has been applied to other data elements for other conditions.^20,21 One study used a logistic regression (LR)-based imputation model to impute missing substance use information but was limited to a small cohort.²² As national surveillance data usually have a large number of variables with complex relationship patterns such as nonlinearity, non-normality, and hierarchical interactions, specifying the appropriate parametric model for imputing missing IDU data could be challenging and labor-intensive.

Recent studies have shown that machine learning algorithms such as random forest (RF) could be powerful tools when they are incorporated with MICE.^23-25 RF combines the outputs from multiple decision trees,²⁶ does not require determining an imputation model, and allows inclusion of many predictors for automatic screening while accounting for complex relationships among variables. RF-based imputation performs well under moderate to high missingness, even when data are MNAR in certain cases.²⁷

To better understand the contribution of IDU among acute hepatitis C cases in the United States, we aimed to establish a flexible and robust MICE approach to impute missing IDU information in the context of a large national surveillance dataset. Specifically, using NNDSS, we compared the performance of LD and 3 classes of models to impute missing IDU data through MICE: parametric LR, semi-parametric predictive mean matching (PMM),²⁸ and nonparametric RF model (standard RF [sRF] through R “mice” package,¹⁹ and fast implementation of RF [fRF] through R package “miceRanger”²⁹). Evaluation and sensitivity analyses were conducted to assess the performance of each imputation model.

To address the high missing values of IDU in national surveillance data, this study assessed 3 imputation models through MICE. We found that the fRF-based MICE imputation produced the most accurate estimation of IDU prevalence while remaining robust to the MNAR assumption. The estimated IDU prevalence among acute hepatitis C cases in the United States during 2019-2021 was as high as 85.1% as compared to 63.5% from the LD approach.

Among the different approaches we evaluated for handling missing IDU data, LD had the most bias compared to other MICE imputations. This was expected as LD excluded a large amount of data which reduced statistical power, and the assumption of MCAR did not hold. In addition, we found that in certain circumstances (such as using Dataset 2 with smaller sampler size), LR showed worse performance than LD, which might be explained by the less optimized imputation model. In all the tested scenarios, RF-based MICE imputation, especially fRF, showed better performance and efficiency than LD and other models. These differences might reflect the different levels of robustness of imputation modeling when dealing with large surveillance data with complex patterns such as interactions and nonlinearities. A high-quality imputation model is necessary to allow MICE to have accurate and efficient estimation.³² Our analysis suggested that the data-driven machine learning such as RF was advantageous because it can automatically capture valuable underlying patterns from many data elements to increase estimation accuracy and efficiency. This approach is necessary and advantageous to improve the inference and interpretation of viral hepatitis surveillance data. Many jurisdictions do not currently have sufficient resources to obtain nor impute IDU information on all acute hepatitis C cases; some might lack the resources to conduct imputation themselves. A national-level imputation approach could generate information to inform jurisdiction hepatitis C surveillance and prevention program efforts, improving imputation by use of data from multiple jurisdictions.

While we have demonstrated significant associations for missing IDU information with the observed variables (ie, MAR), we cannot exclude the possibility that IDU missingness may be associated with other unobserved variables or IDU itself due to potential stigma or other concerns (ie, MNAR). To address this possibility, it is recommended to include as many related variables as possible in the imputation model.- ¹⁶ In our analysis, in addition to the case-level surveillance data, we added 108 county-level measures with diverse themes from public data sources. However, this also caused computational speed issues with sRF, but not for fRF, which is designed for high-dimensional data. Simulation studies showed that RF outperformed other machine learning mechanisms such as k-nearest neighbors and was robust when data were MNAR.²⁷ This finding is consistent with our sensitivity analysis.

The fRF-MICE may help us to better understand the prevalence and burdens of IDU-associated acute hepatitis infections in the United States. A previous study had used acute hepatitis C infection as a proxy measure for IDU, assuming IDU is the major risk for the infection.³³ However, the true extent to which IDU contributed to acute hepatitis C has not been investigated from national surveillance system data. Another study limited to acute hepatitis C cases with available risk data estimated that the prevalence of IDU among these cases ranged from 58.5% in 2004 to 84.0% in 2014.⁹ Due to small samples used, these estimates were likely to be unstable and not nationally representative. Compared to the proxy-based approach, the multiple imputation-based approach from fRF-MICE, which considers the entire study population, might provide a more robust estimation on the IDU-associated infections.

Through the application of fRF-MICE, we conducted an analysis to compare the prevalence of IDU among different sociodemographic subgroups (Table 2). For example, the highest IDU prevalence of 93.0% was estimated for non-Hispanic Native American/Alaska Native persons compared to the lower value of 72.9% for non-Hispanic Asian/Pacific Islander persons, which might indicate potential racial disparities among IDU-associated infections. However, further studies are warranted due to the limited sample size at the subpopulation level by race/ethnicity. Another application of fRF-MICE is to study trends of IDU-associated infections over person, place, and time. Given the robustness of fRF-MICE, future research may explore the performance of this approach at a more refined geographic level, such as the county or city level, or within a specific subpopulation. This exploration would enhance the ability to detect clusters or outbreaks associated with IDU.

Application of fRF-MICE to address missing data and enhance understanding of surveillance data may also help public officials and health providers prioritize populations with a high prevalence of IDU, which can be used to support hepatitis prevention and elimination activities. For example, public officials and health providers may offer harm reduction services, including syringe services, education on safe injection and non-injection drug alternatives, medication for opioid use disorder, HIV and viral hepatitis testing and treatment, and hepatitis A and hepatitis B vaccination to sociodemographic populations and geographic areas with high prevalence of IDU where health care is not readily accessible. The fRF-MICE-based IDU estimation may help guide evidence-based strategies for targeted outreach and resource allocation, which can reduce morbidity, mortality, and further transmission of viral hepatitis.

This study has several limitations. First, RF does not have model specifications and may be more difficult to interpret than parametric models. However, RF can rank the importance of variables used for imputation, which may be helpful for understanding the model. Second, RF is usually computationally intensive when dealing with high-dimensional large datasets; this limitation can be alleviated by using more computationally efficient methods such as fRF-MICE. Third, like all other imputations, fRF-MICE assumes the missing data follow MAR. Sensitivity analyses should be conducted when MNAR is possible. Fourth, this analysis does not fully explore the selection bias introduced by variables or data sources used in the model. Further analysis is needed to explore potential bias and refine the model accordingly. Lastly, caution should be exercised when interpreting differences at the subpopulation level, especially when sample sizes are small. Potential bias might exist with the fRF-MICE due to limited information available to capture the characteristics and variability of small subpopulations. Further work is warranted to assess the validity of fRF-MICE in addressing the issue of missing IDU data at the state or local, such as comparing the imputation analysis results with those obtained from enhanced surveillance studies that have more complete IDU data.

IDU is the most significant driver for acute hepatitis C in the United States; accurate and complete data are crucial for guiding hepatitis C elimination efforts. While efforts are underway to improve the completeness of surveillance data, imputation could serve as a supplemental approach to handle missing data appropriately when obtaining complete data is challenging. This analysis demonstrated that RF-based MICE imputation is an effective, flexible, and straightforward approach to address missing IDU data when analyzing high-dimensional national surveillance data with complex patterns like NNDSS. The utilization of advanced imputation methods, such as RF-MICE, has the potential to inform public health interventions aimed at addressing the syndemic of viral hepatitis, HIV, and drug overdose driven by IDU.