There were three main approaches to conducting the assessments: Composite Index Scores, Statistical Modeling, and Spatial Epidemiology.

The first main approach was the Composite Index Score (CIS) that uses methods similar to those developed for the Social Vulnerability Index (SVI) by the Geospatial Research, Analysis, and Services Program (GRASP). The SVI assesses the vulnerability of communities to disasters, such as earthquakes, hurricanes, and floods by ranking each census tract on fifteen factors and grouping them into four related themes. Each census tract receives a separate ranking for each of the four themes, as well as an overall ranking. The mechanics behind the CIS are similar. For this assessment, however, indicators relate to vulnerability, opioid overdose and/or to non-sterile IDU and are selected by subject matters experts. The rankings from the CIS then represent an indirect method of assessing risk for counties. An illustrative example is provided using a subset of the indicators identified in the National Vulnerability Assessment [1]: opioid prescriptions, drug arrests, drug overdose deaths, and per capita income. Similar to the SVI, if multiple variables on the same topic area are included, consider separate ranks for each topic or weighting the overall rank to account for the number of variables by topic area.

#### Missouri Vulnerability Assessment Project

Information provided by Becca Mickels, Chief, Bureau of Reportable Disease Informatics Missouri Department of Health and Senior Services

Missouri’s Opioid 2019 JVA project focused on the development and communication of an approach similar to the Social Vulnerability Index (SVI) by aggregating multiple factors related to opioid-related disease risk, and relating this risk across a geography including 114 counties (115 when including St. Louis). The primary intent of Missouri’s selected approach was to manage time and to offer simplicity in communication with the public. Overall, the results were a success and the methodology they employed seemed to be more transparent and easier to share.

#### Oregon Vulnerability Assessment Project

Information provided by Pickle, 2019. Viral Hepatitis Program Acute and Communicable Diseases.

Oregon’s County-Level JVA provided a thorough example of spatial epidemiology using a variety of choropleth map approaches, as introduced above. Oregon used multiple methodological approaches beginning with a high number of key indicator variables that were eventually reduced in a backward stepwise regression model until only significantly associated variables were included.

#### Illinois Vulnerability Assessment Project

Illinois undertook their JVA to inform work at the state level to conduct prevention activities, such as working with local health departments to develop jurisdictional response plans. The State leveraged national (e.g., Census Bureau) and local data sources (e.g., I-NEDSS; vital records) to derive 30 independent variables to test for association with the outcome of interest, HCV infections in individuals less than age 40.

#### Nebraska Vulnerability Assessment Project

Information provided by Felicia Quintana-Zinn

Nebraska’s Department of Health and Human Services (DHHS) sought to conduct a vulnerability assessment to both identify the high burden jurisdictions in the state, and to take a holistic approach to the entire state population. Nebraska has both rural and frontier areas, which despite having higher burdens of risk for opioid overdose and bloodborne disease transmission, lack the resources to address these risks. Public health jurisdictions had been expressing interest and the need for an assessment for several years, and collaborations and conversations that laid the groundwork for the assessment had begun more than two years ago.

#### Rhode Island Vulnerability Assessment Project

Rhode Island’s 2019 Opioid JVA project employed a more sophisticated statistical model approach featuring machine learning techniques. The model was proposed by Brown University partners as the best approach and ultimately included more than 300 variables which explored geographic units as far down as census tracts and ZIP codes.

##### Composite Index Scores Read More

The first main approach was the Composite Index Score (CIS) that uses methods similar to those developed for the Social Vulnerability Index (SVI) by the Geospatial Research, Analysis, and Services Program (GRASP). The SVI assesses the vulnerability of communities to disasters, such as earthquakes, hurricanes, and floods by ranking each census tract on fifteen factors and grouping them into four related themes. Each census tract receives a separate ranking for each of the four themes, as well as an overall ranking. The mechanics behind the CIS are similar. For this assessment, however, indicators relate to vulnerability, opioid overdose and/or to non-sterile IDU and are selected by subject matters experts. The rankings from the CIS then represent an indirect method of assessing risk for counties. An illustrative example is provided using a subset of the indicators identified in the National Vulnerability Assessment [1]: opioid prescriptions, drug arrests, drug overdose deaths, and per capita income. Similar to the SVI, if multiple variables on the same topic area are included, consider separate ranks for each topic or weighting the overall rank to account for the number of variables by topic area.

CIS uses the rankings of indicators within each jurisdiction to assess which counties (or other subdivisions) are most vulnerable to opioid overdoses and/or infections from non-sterile IDU. This method is less statistical than regression approaches which have more robust modeling and predictive power, but is very adaptable and works well in applied situations.

##### Composite Index Score Example:

For this example, we used the following category/indicator pairs, based on publicly available data:

- Socioeconomics: Per Capita Income (2013-2017 American Community Survey)
- Crime: Drug Arrest Rate (Uniform Crime Reporting Program, 2016)
- Prescriptions: Per Capita Opioid Prescriptions (IQVIA Xponent, 2017)
- Mortality: Drug Overdose Rate (CDC Wonder, 2013-2017)

To handle missing data and zeros, we made the following assumptions. We assumed that zero values were a legitimate reflection of zero events and assigned the nationwide median for missing (N/A) values, as a simple form of imputation. Assumptions regarding zero values and imputation methods could be adjusted based on local context and comfort with other imputation methods.

We ranked each indicator among the 3,412 counties and county equivalents, with a “higher is worse” categorization (see Table 2). We then took the average of the four ranks to calculate the overall CIS rank for each county. Analysis was performed using widely available spreadsheet software (Microsoft Excel).

Geography | Drug Arrests (Rate, Rank) |
Opiod Prescriptions (Rate, Rank) |
Drug Overdose Deaths (Rate, Rank) |
Per Capita Income (Value, Rank) |
Average Rank |
---|---|---|---|---|---|

Bell County, Kentucky | 113 (3063) | 228 (3137) | 60 (3126) | 14754 (3104) | 3107.5 |

Clay County, Tennessee | 87 (2920) | 191 (3125) | 49 (3108) | 17667 (3002) | 3038.75 |

Grundy County, Tennessee | 83 (2891) | 141 (3046) | 35 (3009) | 15824 (3080) | 3006.5 |

Logan County, West Virginia | 109 (3046) | 170 (3109) | 70 (3137) | 21074 (2522) | 2953.25 |

Scott County, Tennessee | 100 (3006) | 154 (3075) | 29 (2841) | 18748 (2902) | 2956 |

Figure 4 presents a map of the overall CIS rank from this illustrative example by county using Microsoft Power BI. Specific counties can then be further explored to understand the underlying factors contributing to their overall CIS rank as well as the different preventive services and interventions available in the community.

When interpreting the overall CIS rank, it may be helpful to compare it with an appropriate outcome metric for context. Jurisdictions who used the CIS approach often included or compared their CIS ranks with diagnosis data for HCV or HIV or administrative claims data for endocarditis, HCV or HIV.

When doing this type of contextual comparison, we offer the following considerations:

- Areas with high outcome metric and high overall CIS rank: these represent areas where additional resources may be needed. Health departments should look at available resources within these areas and see what steps are being taken, and could be taken, to improve prevention programs in those communities.
- Areas with low outcome metric and high CIS rank: these areas may have unmeasured protective factors such as syringe services programs, Naloxone availability, or programs by local public health officials. Conversely, such areas may reflect a lack of testing for the outcome metric, such as lack of capacity to test for HIV/HCV. Health departments should consider if available testing and prevention services are meeting local needs.
- Areas with high outcome metric and low CIS rank: these areas may reflect unmeasured exogenous factors, such as drug traffic patterns or unique population characteristics. When working with one state, a county identified with high HCV incidence was found to be a county that contained a major interstate highway and located near a large urban area in a neighboring state.

##### Missouri Vulnerability Assessment Project Read More

#### Missouri Vulnerability Assessment Project

Missouri’s Opioid 2019 JVA project focused on the development and communication of an approach similar to the Social Vulnerability Index (SVI) by aggregating multiple factors related to opioid-related disease risk, and relating this risk across a geography including 114 counties (115 when including St. Louis). The primary intent of Missouri’s selected approach was to manage time and to offer simplicity in communication with the public. Overall, the results were a success and the methodology they employed seemed to be more transparent and easier to share.

The team applied a rank sum methodology by individually ranking each individual outcome and community factor such that a rank of 1 always indicates a better outcome or more resources. These ranks were then summed, weighting the individual outcome ranks more heavily than the community factor ranks, to get an overall rank for each county. Stakeholders involved in the process were helpful in suggesting that some of the social and economic variables were possibly influencing the model with greater weight than needed. These suggestions moved the JVA team to experiment with adjustments to increase the profile of individual outcomes as compared to the more general social and economic factors. Ten variables were ultimately included in the assessment model, five of these variables captured outcomes such as HCV infection, overdose, bloodborne infections, and five were community factors.

Data for their SVI approach were sourced from several different agencies. Overdose data came from the Bureau of Vital Statistics, emergency department data came from the Bureau of Health, mental health departments shared drug-related arrests, and the American Community Survey (ACS) was sourced for community factors.

The method and results were discussed and shared across six stakeholder meetings. In these meetings variable weighting and variable changes were discussed and determined. For example, concerns about the potential for correlation and common rank results arose for some community factors like poverty rate and uninsured. The locations of these meetings corresponded with HIV care regions, where approximately 50 invitations were sent out per meeting. A wide variety of attendees contributed to the process including mental health staff, police, coroners, non-profit staff, and even school representatives.

A powerful visual tool, maps were developed and shared at stakeholder meetings to help communicate the draft results of the assessment and ranking outcomes. Maps were provided with final outcomes as well as individual variables such as some of the community factors. ESRI software including ArcGIS [5] was used to publish maps for key variables. Certain variables were chosen when presenting data to avoid small number issues for some variables.

For outreach, Missouri’s team placed community-level calls with vulnerable counties and used data to apply for action grants for HIV bloodborne work. Five local agencies were reviewed and selected for funding. The assessment work was also used for individual project evaluation in the state and, if appropriate staffing is made available, results will be used again next year for evaluation of certain project successes.

Missouri finished their assessment and has made their full report available on their website along with related resources. Additionally, for each county that scored as more vulnerable in at least one of their assessments, presentations were made via live webinars to provide targeted assistance. Dissemination efforts also included posters and outreach calls. Additional focus has been provided to rural counties which often fail to get as much attention, but where health outcomes can be critical. For example, Missouri recently had a significant HAV outbreak in a rural county area.

Missouri’s team found that overall, the work took longer to complete than originally anticipated, and the team had to set limits to the project scope with only one year of funding available. It was also challenging for their team not to have funding available to hire more staff. Despite challenges, Missouri found that conducting stakeholder meetings in the middle of the process timeline was an effective and useful means of gathering information that was then used to draft content for the final report. For these meetings, facilitators were hired to collect written content and to capture stakeholder ideas from parties less willing to speak publicly. Missouri gathered written comments and obtained helpful input from stakeholders in this way.

Overall, success for Missouri included working closely with stakeholders, and ultimately having work that was well received and appreciated by other agencies and the public. The work increased team credibility, with both internal agency groups and multiple work groups outside the agency. It was also a success for Missouri’s team to get final draft approval, and additional funding for the model.

- Statistical Modeling for Count Data
Statistical modeling was the second main approach used for the assessments. Statistical modeling can be used to identify the specific factors associated with the outcome of interest. Using this approach analysts can calculate the estimated rate of the outcome and better understand the factors relevant in the community. Count (or rate) outcome data were principally modeled using Poisson or Negative Binomial regression. To determine which candidate variables to include in a multivariable model, two steps were generally taken. First, each independent variable was bivariably fit, and outcome associations with p-values less than 0.1 (i.e., marginal significance) were retained for multivariable model consideration. Second, the remaining independent variables were checked for multicollinearity. Independent variable pairs that strongly correlated together were flagged (generally >0.7), and the independent variable from the pair with the stronger outcome association was retained, with the other being dropped. Thus, independent variables that had at least a marginal bivariable association with the outcome and low collinearity with other characteristics were considered for multivariable modeling. Additional characteristics that may epidemiologically confound or bias outcome associations were further included, when appropriate. Multivariable regression modeling used either forward, backward, or stepwise selection, and the final models were selected after evaluation of fit statistics, adjustment for confounding, and error residuals. Predicted outcome rates were generated at the unit of analysis level (e.g., county/ZIP code) and ranked from highest rates to lowest rates. A task flow diagram for regression modeling best practices is provided.

- Regression Model Options for Count Data
When assessing count or rate data as an outcome, Poisson or Negative Binomial are the preferred regression methods. Both Poisson and Negative Binomial regression methods appropriately weigh both the count of events and exposure, either in time or population. For example, in 2017, the crude death rates in Ada County and Clark County, Idaho were 7.1 and 8.0 per 1,000 population, respectively. Ada County had 3,229 deaths with a population of 456,849; Clark County had seven deaths with a population of 873. Poisson and Negative Binomial regression both assign a weight to Ada County hundreds of times larger than that of Clark County (proportional to the population size difference); whereas an ordinary least squares regression would inappropriately weigh the counties equally, despite the massive population difference.

In a Poisson model, the analyst assumes the outcome variable to follow a Poisson distribution, and that the natural-log of the expected value (i.e., the mean) of the outcome can be estimated by a linear combination of the independent variables. A critical assumption of the Poisson distribution is that the raw mean and variance of the outcome are approximately equal. In practice, this is rarely true, with variances either far exceeding the mean (overdispersion, most common) or being below the mean (underdispersion, less common). Use of Poisson regression in the presence of overdispersion or underdispersion can lead to biased standard errors and incorrect conclusions regarding statistical significance.

Negative Binomial regression is a generalization of Poisson regression that uses a mixture of Poisson and Gamma distributions to model the outcome, loosening the assumption of an equal raw mean and variance. In practice, Negative Binomial regression is often appropriate over Poisson regression, but the analyst should not simply default to Negative Binomial regression for all analysis cases. When the outcome data truly follow a Poisson distribution, Poisson regression will provide the least biased standard errors and p-values, compared to Negative Binomial regression. A chi-square goodness of fit (GOF) test for overdispersion exists and should be considered when equality of mean and variance is in question. Both this question and others regarding statistical modeling are addressed in the FAQ document. For SAS [6] users, code and a discussion on GOF testing and regression modeling are shown in the previous webinar presentation.

- Outcome
The outcome variable for count-based regression models should always be a raw count, with the natural-log of the population included as an offset characteristic. It is not advisable to transform the count outcome or to include the population as a numeric independent variable. Likewise, as seen in the Idaho example, using rates for the outcome gives improper weight. In count-based regression, offset variables determine exposure, and are modeled with the outcome variable, not the independent variables.

- Independent Variables
When determining independent variables for an analysis, jurisdictions should consider a variety of data sources a priori. See the Data Collection and Exploration sections for common sources. Common independent variables used in the assessments were: drug overdose deaths, prescription opioid sales, mental health services, insurance coverage, urgent care access, vehicle access, access to interstate, buprenorphine prescribing potential, education, income, poverty, race/ethnicity, unemployment, population density, urban/rural status, and drug-related arrest data. Additional characteristics germane to specific states and/or geographic regions should be further considered and included.

- Transformations
For count-based regression analysis, it is not recommended to transform the outcome data. The continuous independent variable data, however, can be transformed to meet model assumptions and assist with interpretability. Log10 and natural-log transformation were most prominently used in the assessments. Log10 transformations are useful if an independent variable is on a scale where order of magnitude differences are important, for example, per capita income. Natural-log transformation are useful for data that are right-skewed, for example, miles to nearest healthcare facility. Other common transformations include: arcsine (for values between 0 and 1), square root (for count data), and inverse hyperbolic sine (for right skewed data that include 0 and/or negative values).

- Variable Selection

Given a moderate to large set of potential independent variables, an analyst must choose a limited number for regression modeling, contingent on sample size and multicollinearity (i.e., high correlation among two or more independent variables). The process most frequently used for assessments involved three steps: 1) bivariable regression modeling, 2) correlation of independent variables, and 3) selection procedures for the final multivariable model. In Step 1, each independent variable is separately associated with the count outcome using either Poisson or Negative Binomial regression, as appropriate. Variables that are statistically significant at p<0.1 and/or have epidemiological significance are retained. Remaining variables are then correlated with each other using Pearson or Spearman correlations (Step 2). Independent variable pairs that correlate >0.7 are flagged for multicollinearity, and the variable from the pair with stronger outcome association is maintained. Finally, in Step 3, the remaining independent variables are considered for multivariable inclusion and modeled using either stepwise, forward, or backward selection. For advanced models (e.g., those modeling longitudinal data), stepwise procedures are often not available, and the analyst must choose between either forward or backward selection. As previously inferred, the analysis should not be guided by statistical significance alone, but also informed by industry standards and subject matter expertise, while building the final multivariable models. For analysts modeling correlated (or longitudinal) data, issues with model convergence sometimes arise. In these circumstances, please see the Convergence Questions resource that provides tips for such situations.

In addition to correlating variables for checks on multicollinearity, a more sophisticated approach employs methods from principal components and factor analysis. Principal components are a type of dimension (data) reduction that clusters correlated variables into groups (i.e., components). The first step is generating a covariance matrix between variables, then calculating the components and eigenvalues (entities that represent the variance explained by each component). Sums of these eigenvalues are visualized using a Scree plot. The user is tasked with picking the number of components that explain the most variance, without overfitting, a technique called the “elbow method”. In Figure 5, a 24-item (variable) psychological measure has most of the variance explained by the first three or four components (at the “elbow” of the curve). As such, a researcher would consider three or four principal components as candidates to include in a follow-up factor analysis.

A factor analysis rotates principal components in geometric space to create orthogonal (i.e., uncorrelated) constructs on which variables load. The advantage of factor analysis is that it separates variables into independent groupings, with higher absolute loadings representing greater associations with the factor constructs. Consider the results below from a state involved in the assessment (Table 3). Based on this table, one of YPLL (Years of Potential Life Lost), MME90 (90 Milligrams Morphine Equivalent per day), or opioid prescribing could be selected to represent Factor 1 (top three absolute value (+/-) loadings); one of rural, poor health, or HIV could be selected to represent Factor 2; and one of percent without high school diploma, teen birth rate, or percent vacant housing could be selected to represent Factor 3. A user may choose more than one variable to represent each Factor (e.g., both YPPL and MME90 for Factor 1), but would need to again check for correlation between these characteristics.

Factor 1 Factor 2 Factor 3 Gini Index 12 53 8 HIV Rate 2 57 14 MME90 79 9 0 Opioid Prescribing 82 -23 4 Percent Unemployed 4 53 56 Percent Vacant Housing 40 -4 57 Percent Without High School Diploma 11 38 60 Poor Health 48 66 5 Rural 25 -67 6 Syphilis Rate -26 4 48 Teen Birth Rate 2 1 57 YPLL 85 35 2

Once candidate variables have been identified, appropriate statistical software (e.g., SAS, R, STATA [7]) can be used to create the final multivariable model using previously described procedures. Coefficients for Poisson and Negative Binomial regression are measured on a logarithmic odds scale, and as such, are interpreted as the difference in logs of expected counts per unit change in the variable, given the other independent variables in the model are held constant.

Table 4 is an example of how to present results from a regression analysis. In addition to descriptive statistics (mean, median, interquartile range), it has the risk ratio and p-values for the bivariable and multivariable models, as well as the regression coefficients and standard errors from the multivariable model.

Table 4. Measures of central tendency, regression coefficients, standard error, and significance level for variables used in vulnerability assessment.

Bivariable Multivariable Variable Mean Median First Quartile Third Quartile Risk

RatioP-Value Regression

CoefficientSE Risk Ratio P-Value Drug overdoses for all drugs,

2017 to 2018, per 100,000 population414.61 414.13 359.10 468.93 1.09 0.004 Log of MME for all prescribed drugs 17.66 17.81 10.24 24.39 1.70 <0.001 0.49 0.04 1.63 <0.001 Percent of population that is non-Hispanic white 65.40 70.19 54.29 93.72 1.49 0.001 0.39 0.15 1.48 0.014 Rate of sexually transmitted infections,

per 100,000 population717.23 555.07 430.29 803.30 1.12 <0.001 0.09 0.03 1.09 0.005

For ease of interpretation, and in addition to the multivariable regression coefficients, many states chose to report ranked vulnerability scores. Vulnerability scores are simply the predicted rates for study outcomes, calculated from multivariable Poisson/Negative Binomial regression models. To get these rates, a user calculates predicted values directly from their multivariable regression equation. These raw, predicted values are in the form of log-counts per person. These values can be exponentiated to get predicted counts per person (i.e., person-rates). This rate is not population-specific, and can be comparable across other sampling units (e.g., counties). Many times, it was reasonable to multiply this person-rate by a constant to get an epidemiologic rate (e.g., rate per 100,000). These epidemiologic rates were used as the final vulnerability scores and ranked for reporting.

- Spatial Epidemiology Methods
The third approach used for the assessments included various spatial epidemiology methods. The most common form of spatial epidemiology used as part of opioid-related vulnerability was the choropleth map, which maps a variable using color intensity. ColorBrewer is a useful resource for identifying appropriate color schemes including sequential (for increasing rates), diverging (for comparing above and below a baseline), and qualitative (using no meaningful sort order for outcomes). Figure 6 illustrates per capita income from the American Community Survey using a sequential yellow-orange-red color scheme, mapped using R and Leaflet and uploaded to RPubs, a public website for visualizations created in R.

While most jurisdictions used county-level data, some jurisdictions focused on smaller geographic units, such as ZIP code-level data (ZIP code tabulation area) and formatted it for mapping purposes such that low population ZIP codes could be merged with adjacent ZIP codes to facilitate more stable rate calculations.

Additionally, jurisdictions introduced point layers including physical address-level data consisting of: buprenorphine-waivered physicians, syringe service programs, Naloxone providers, drug detox, outpatient and inpatient services, and other community-based resources. This type of approach added community health needs assessment-type content to the overall vulnerability assessment. Drive-time analyses were then conducted to estimate the proportion of a jurisdiction with access to the specific service type. [8,9]

A final risk and resource map could then be produced showing the intersection and availability of relevant community resource locations with the highest-risk communities for each indicator as well as overall vulnerability groups/ranks.

Combining GIS, spatial epidemiologic, and statistical modeling analyses allows for a more comprehensive review of the spatial pattern of risk, exploration of correlation, and the potential alignment of community health assets to help highlight the communities that face the highest risk and that may require additional resources.

##### Oregon Vulnerability Assessment Project Read More

#### Oregon Vulnerability Assessment Project

Oregon’s County-Level JVA provided a thorough example of spatial epidemiology using a variety of choropleth map approaches, as introduced above. Oregon used multiple methodological approaches beginning with a high number of key indicator variables that were eventually reduced in a backward stepwise regression model until only significantly associated variables were included. Their final model included four variables: high intensity drug trafficking area, premature deaths, risky opioid prescribing, and lack of transportation or vehicle availability. Using this final model, a vulnerability score was generated. Oregon proceeded to use ESRI software to generate a map showing the overall vulnerability scores using Jenks natural breaks to categorize into groups described as: lowest, low, mid-range, high, highest priority for intervention as shown in Figure 7.

Oregon also produced a similar stylized county-level map for each variable in the final model. Breaks in additional maps were chosen based on the statistical properties of each variable mapped, and additional geospatial layers were added for relevance. For example, Oregon layered interstate and highway data on top of the county map showcasing high intensity drug trafficking areas. Layering here helped demonstrate the reliance on interstate travel for drug movement.

Oregon faced limitations in the analysis due to the small number of counties in the state, which reduced their statistical power. Geographically, large sized rural counties made it difficult to discern risk arising in specific communities. Oregon was able to present the choropleth maps along with contextual information gathered from other projects, to provide further detail of potential risk to stakeholders interested in these areas of the State.

##### Illinois Vulnerability Assessment Project Read More

#### llinois Vulnerability Assessment Project

Illinois undertook their JVA to inform work at the state level to conduct prevention activities, such as working with local health departments to develop jurisdictional response plans. The State leveraged national (e.g., Census Bureau) and local data sources (e.g., I-NEDSS; vital records) to derive 30 independent variables to test for association with the outcome of interest, HCV infections in individuals less than age 40. Data at the ZIP code-level were obtained, and Illinois employed Negative Binomial regression methods after observing overdispersion in the outcome. For this analysis, Illinois used PROC GENMOD in SAS v.9.4 for regression modeling. Specifically, analysts ran bivariable regression models for each independent variable (approximately 30 models) with counts of HCV infections among individuals less than age 40 as the outcome, and population as the offset. Prior to regression modeling, Illinois combined data over two years (2017-2018), and all sampling units were treated as independent. Bivariable regression model results eliminated nine variables from consideration for the multivariable model due to statistically insignificant associations with the outcome at the 0.05 level. The remaining 21 variables were added to a multivariable model and backward selection was used to derive the final model containing eleven significant variables. Model assumptions and fits were checked and predicted person-rates were calculated for each ZIP code. These person-rates were used as the final vulnerability scores and ranked for reporting and mapping.

Illinois mapped both rates and counts of combined HCV incidence and drug overdose for 2017-2018 along with the final vulnerability scores, based on model-estimated rates. When working with local partners, Illinois found that ZIP code-level data gave far more detail than county-level maps, allowing local officials to focus on specific neighborhoods of high risk and high outcome. Partners found outcome rates (for relative comparisons), raw counts (for prioritizing resources) and vulnerability scores all to be useful. They further defined high risk as the bottom (i.e. most vulnerable) ranked 10% of counties and plotted that in a separate map.

Illinois used a number of software packages for mapping. Initial mapping was done in both R and GeoDa [10]. Illinois also had an academic collaborator who was able to produce maps using Carto, an online package that produces interactive maps [11].

Despite the large number of ZIP codes in Cook County (2019 population of over five million people), there was only a single ZIP code that was in the top 10% for the vulnerability scores. Therefore, Illinois did not highlight Cook County in the statewide maps; they plan to produce ZIP code-level maps specifically for Cook County and Chicago partners during their regional meetings. Overall, the state found that ZIP code data worked well and provided an additional level of detailed information compared to just using the county level.

The state has used the assessment findings to inform work at the state level and to undertake prevention activities. They are working with two local health departments to develop jurisdictional response plans and had invited all local health departments to review their findings and provide feedback. Local health departments have said that the assessment has been a great tool for them to affirm areas they already knew are vulnerable and have begun conversations about how to turn the assessment findings into a tool for policy-making.

#### A Comparison of the Approaches: Composite Index Score, Regression Modeling, and Spatial Epidemiology

Each of the three main approaches used in the JVAs has pros and cons. The CIS approach is very adaptable, works well in applied situations, and may be easier to communicate to broader audiences. However, it is less statistical, relies on existing subject matter expertise to identify the appropriate indicators, and may orient ‘vulnerability’ in the wrong direction if true associations differ from the expected direction based on subject matter expertise. Regression modeling utilizes dependent variables or outcomes, and independent variables or risk factors, similar to those in the CIS; however, regression has the advantage of employing quantitative methods, such as maximum likelihood, to determine the strengths and directions of associations between independent variables and outcomes. Moreover, validated statistical models can predict outcomes, whereas, CIS is limited to county rankings, based on vulnerability inputs. Yet, regression modeling requires more technical skills, can be challenging to communicate to general audiences, and may also inadvertently exclude important indicators. Spatial epidemiology is less focused on the relationship between variables and more on the spatial relationships between variables. Spatial epidemiology has the advantage of incorporating techniques to account for human geography, but is technically advanced and has limited application without using CIS or statistical modeling to inform the spatial epidemiology approach. The choice of approach ultimately depends on both the familiarity of the analyst with the data and method options, as well as the statistical literacy of the audience; an analyst must not only be familiar with the technique used, but also be able to explain the technique and results to others.

#### Identifying Prevention Gaps

Jurisdictions used three key steps to identify prevention gaps using the results of the assessments. The first step was to identify key interventions needed to reduce opioid overdoses and injection-related bloodborne infections. The second step was to determine where these key interventions are currently available and accessible, and the third step was to analyze gaps between where interventions are needed and where they are available and accessible.

Many jurisdictions were able to identify prevention gaps using their assessment findings. Often, these states already had some services in place and sought to use the results as a way to identify the gaps within their states where services might not be adequately meeting the populations’ needs. The identified gaps could be related to rural and urban divides within the state, such as in frontier areas, or among particular population groups, such as tribal populations. States cited important aspects of the process to identify prevention gaps such as meeting at the local level with community planning groups to talk about resources. Several states credited their assessment results with enabling them to better have conversations about gaps and needs because they were able to move from citing anecdotal evidence to sharing data that better demonstrated their populations’ true vulnerabilities. Others were aware of the existing vulnerabilities but the assessment provided them a new voice, informed by data, that allowed them to make compelling cases for the need for expanded and new services. One state spoke about the impact the years of potential life lost (YPLL) measure had in understanding the human losses related to opioid misuse compared to just using traditional mortality rates.

##### Nebraska Vulnerability Assessment Project Read More

#### Nebraska Vulnerability Assessment Project

Nebraska’s Department of Health and Human Services (DHHS) sought to conduct a vulnerability assessment to both identify the high burden jurisdictions in the state, and to take a holistic approach to the entire state population. Nebraska has both rural and frontier areas, which despite having higher burdens of risk for opioid overdose and bloodborne disease transmission, lack the resources to address these risks. Public health jurisdictions had been expressing interest and the need for an assessment for several years, and collaborations and conversations that laid the groundwork for the assessment had begun more than two years ago.

Initial efforts looked at resources and gaps, especially for the five high burden jurisdictions. One striking gap noted was the correlation between drug overdoses and income levels. Another was the lack of treatment providers in less populated areas, with certain areas having only one provider that would require a two to three-hour drive for those seeking treatment. A third gap noted was that while individuals who overdose might get emergency treatment, many do not go on to treat the underlying addiction that led to the overdose. The grant funding allowed the state to continue and expand upon what they were already doing to address the public health risks.

Nebraska leveraged existing stakeholder coalitions to determine how to best use the assessment findings to address the identified prevention gaps for the state. To address the service provider gap, the locations of testing sites for HCV and HIV along with treatment providers and recovery resources will be listed by county and local health departments on the DHHS website, as well as the website for the local health department association. To address the income and risk gap, existing resources such as behavioral health programs were identified to assist with payment for treatment for very low-income populations. And by working closely with law enforcement, emergency medical services, the state patrol, and other public safety organizations, treatment information and resources can better reach the populations experiencing overdoses.

While Nebraska is still working through the best ways to address newly identified populations at risk for opioid overdose and bloodborne disease transmission, they are continuing their ongoing interventions to match existing resources with identified gaps.

#### Collaboration, Communication, and Dissemination

Consistent collaboration and communication with stakeholders was key in ensuring jurisdictions’ success in all stages of the JVA project. From identifying and prioritizing indicators to implementing the plan to address intervention gaps, stakeholder engagement allowed jurisdictions to achieve buy-in for the assessment and build the partnerships necessary to address gaps in services. Jurisdictions were encouraged to take an asset-based approach [13] to engaging stakeholders and establishing partnerships. The asset-based approach is focused on identifying community assets that can be mobilized for improvement. This approach fosters a sense of independence, pride, and possibilities among stakeholders, and creating a positive sense of engagement is important in avoiding stigmatization of jurisdictions at high risk. An asset mapping tool is available to assist with the process of identifying and developing collaborative relationships with individuals, community organizations, and institutions. The technical assistance team also provided a webinar on how to disseminate the findings of the assessments, and how to tailor messages to different audiences for engagement. A communication and dissemination process model tool to assist with target audience profiling is also available.

##### Rhode Island Vulnerability Assessment Project Read More

#### Rhode Island Vulnerability Assessment Project

Rhode Island’s 2019 Opioid JVA project employed a more sophisticated statistical model approach featuring machine learning techniques. The model was proposed by Brown University partners as the best approach and ultimately included more than 300 variables which explored geographic units as far down as census tracts and ZIP codes. Rhode Island’s administrative geography is limited to five U.S. counties, so to look closer at population characteristics it was important and helpful to look at census tracts and even ZIP codes in some cases. The granularity of census tract data helped the state health department understand and implement community-level approaches with greater detail and accuracy.

Input and output data from the model aligned well with overdose surveillance programs in the State which provided overdose deaths that were used as a proxy for injection drug use, HCV, and HIV incidence rates. Other critical variables included single member households, complete kitchen in housing units, divorce, housing unit type, race, less than $10,000 in annual income, and binary variables for non-employed households and households with no vehicles.

Rhode Island used ESRI to help visualize geospatially the model results across census tracts. Using the scores from the machine learning model, tracts were ultimately ranked and categorized into groups from lower to higher vulnerability. Choropleth shades of blue and red spectrums were used and worked well with audiences. Some risk factor data provided high enough definition to identify individual neighborhoods with heightened risk. In addition to ESRI ArcGIS, Adobe Illustrator was used to inset the maps and publish content online as well as layer content onto Google Maps for street-level review.

For Rhode Island, the JVA work has been a turning point in terms of advancing the State’s HIV program, overdose prevention, Health and Human Services programs, and for the Governor’s Overdose Prevention and Intervention Task Force. Overall, the work has generated much interest by other agencies and the general public. Tremendous stakeholder activity has come as a result. Maps were developed for and presented to community leaders in Providence and in more rural areas of the state. Additionally, Rhode Island’s team offered extended 3-hour trainings for technical support in developing planning efforts and investment scenarios, and collaboration building.

Rhode Island found that success has come through refinement of language. Language development has been key for the making the most of their opportunity. At the final stakeholder meetings fifty to seventy individuals or representatives were in attendance, and the team provided a 35-page report for general orientation to the method and results. A lot of time was spent working on language prepared in the map layout in this report. Findings will also eventually be shared at community meetings, and at the Governor’s Task Force meeting which includes more than 100 attendees. This meeting was originally scheduled for March 2020, but was postponed due to COVID-19.

Another experience that has been unique to Rhode Island is the detail that has encouraged action at the neighborhood level. Partnering with Brown University and other subcontractors has helped Rhode Island make use of detailed variables and geography. And working closely with CDC has expanded their analytical capacity. A Request for Proposal (RFP) announcement was released soliciting proposals for the “Syringe Services Access and Drug User Health” program which includes syringe services, naloxone distribution, safe injection equipment and referrals for overdose treatment where appropriate, particularly in vulnerable areas. RIDOH required applicants to use the results of the map to ensure that they were reaching these vulnerable areas. Based on the results of this RFP syringe service activities have expanded throughout the state. RIDOH has continued to have ongoing discussions with key stakeholders including the City of Providence to identify future opportunities for the collaboration. Overall, it is felt that the assessment has boosted the understanding of the essential services provided through this program, leveraging additional sources of funding to address vulnerabilities and facilitating action-oriented successes. Rhode Island’s The VILLAGE Prevention Plan and a one-page summary can be accessed here.