Environmental Correlates of Lyme Disease Emergence in Southwest Virginia, 2005–2014

Abstract

Lyme disease is the most common tick-borne disease in North America. Though human infection is mostly transmitted in a limited geography, the range has expanded in recent years. One notable area of recent expansion is in the mountainous region of southwestern Virginia. The ecological factors that facilitate or constrain the range of human Lyme disease in this region remain uncertain. To evaluate this further, we obtained ecological data, including remotely sensed data on forest structure and vegetation, weather data, and elevation. These data were aggregated within the census block groups of a 9,153 km² area around the cities of Blacksburg and Roanoke, VA, an area with heterogeneous Lyme disease transmission. In this geographic area, 755 individuals were reported to have Lyme disease in the 10 yr from 2006 to 2015, and these cases were aggregated by block group. A zero-inflated negative binomial model was used to evaluate which environmental variables influenced the abundance of Lyme disease cases. Higher elevation and higher vegetation density had the greatest effect size on the abundance of Lyme disease. Measures of forest edge, forest integrity, temperature, and humidity were not associated with Lyme disease cases. Future southward expansion of Lyme disease into the southeastern states may be most likely in ecologically similar mountainous areas.

Lyme disease is a common bacterial infection caused by Borrelia burgdorferi sensu lato (Spirochaetales: Spirochaetaceae) and transmitted by Ixodes (Acari: Ixodidae) species ticks. Tens of thousands of cases are reported annually in the United States, but some estimates put the true disease burden in the hundreds of thousands. The endemic range of Lyme disease is geographically heterogeneous, with more than 85% of all reported cases concentrated in the northeastern coastal region from Virginia to northern New England. A growing body of both entomologic and public health data has revealed southward expansion of this endemic range through Virginia over the past two decades. This expansion has been particularly pronounced along the Appalachian Mountain range, with a growing focus of cases around the cities of Blacksburg, Lynchburg, and Roanoke, all located in southwestern Virginia (Li et al. 2014, Lantos et al. 2015).

To date, the factors driving or limiting this expansion, as well as the plausible endemic range and magnitude of human Lyme disease infections in this region, remain uncertain. Different host-seeking behavior and host preference of southern tick subpopulations, lower survivorship and densities of ticks in warmer climates, and different peak tick activity periods, all may contribute to a lower probability that residents living in the southeastern United States will be exposed to infectious ticks (Oliver 1996, Ginsberg et al. 2017). A leading hypothesis is that northern and southern subpopulations of I. scapularis (Say) (Ixodida: Ixodidae) have different host-seeking behavior as nymphs, the life stage that that leads to most human disease (Falco et al. 1999, Brinkerhoff et al. 2010, Arsnoe et al. 2019). Specifically, the tendency to host-seek higher in the vegetation by the northern subpopulation is both more likely to be exposed to mammals that serve as reservoirs for B. burgdorferi and more likely to feed on human hosts. Conversely, southern I. scapularis ticks, which tend to remain below the leaf litter, are more likely to be exposed to lizards that are a poor reservoir for B. burgdorferi, and ticks in this region appear to be less inclined to feed on human hosts (Felz et al. 1996, Diuk-Wasser et al. 2012, Stromdahl and Hickling 2012, Arsnoe et al. 2019). Ixodes scapularis populations with northern feeding behaviors may be expanding southwards, but it remains to be determined to what degree they account for the geographic expansion of human Lyme disease cases, or whether northern populations will flourish in more southern latitudes (Brinkerhoff et al. 2014, Lantos et al. 2015).

Our goal was to determine which environmental variables were most associated with Lyme disease in southwestern Virginia, an area where Lyme disease has recently emerged. We have chosen to study an area encompassing Blacksburg and Roanoke because there appears to be geographic heterogeneity in the risk of Lyme disease. This feature allows us to evaluate the predictors of local Lyme disease risk in a small geographic area, thus limiting the influence of variables (such as the association of latitude with climate) that vary over a large spatial scale. We included 10 yr of Lyme disease case data aggregated at the census block group level, a small census unit that typically has less than 3,000 people. Environmental data analyzed includes satellite imagery measures of vegetation and forest structure, temperature and humidity data, and elevation.

Methods

We performed a cross-sectional retrospective analysis using public health data. This study was approved by the Duke University Health System Institutional Review Board and by the Virginia Department of Health.

Study Area

ArcGIS 10.5.1 (ESRI, Redlands, CA) was used for spatial operations requiring geographic information systems (GIS). To select our study area, we used the ArcGIS buffer tool to compute 40 km radius circles around the cities of Blacksburg and Roanoke, VA. This distance was chosen based on the distance between the two cities (approximately 40 km). We then used the ArcGIS minimum bounding geometry tool to create a single ellipsoid shape encompassing both circles. There were 347 census block groups in Virginia that intersected this ellipsoid shape (Fig. 1). Two block groups with a listed population of 0 persons were excluded from analysis, leaving 345 total. The total study area was 9,153 km². This area was comprised of 345 block groups with a median area of 4.0 km² (mean 26.5 km², interquartile range 1.3–32.8 km²).

Fig. 1.

Study area in Virginia, United States.

Data Sources

Case data were all probable and confirmed cases of Lyme disease identified by the Virginia Department of Health from 2005 through 2014 (inclusive). These cases were classified according to the contemporaneous surveillance definition published by the Centers for Disease Control and Prevention. Digital elevation models (30-m) were obtained from the United States Geological Survey data interface, The National Map (www.nationalmap.gov). The 2011 National Land Cover Data (Homer et al. 2015) was used to represent forest by combining the three forest categories (deciduous, evergreen, and mixed). Blacklegged ticks are associated with forests that provide suitable habitat for ticks as well as their hosts (Spielman et al. 1984, Ginsberg and Zhioua 1996, Guerra et al. 2002, Ginsberg et al. 2004). The extent of forest fragmentation has also been hypothesized to affect the abundance of infected nymphs and therefore disease risk, but the relationship may depend on the spatial scale (Allan et al. 2003, Brownstein et al. 2005, Diuk-Wasser et al. 2006, Diuk-Wasser et al. 2010). Forest fragmentation metrics were calculated with the Landscape Fragmentation Tool 2 (Vogt et al. 2007) that identified edges (100 m from edge to interior forest), forest core areas of different sizes (beyond the 100-m edge), patches (forested areas lacking core area), and perforated areas (non-forest areas in the middle of forest core areas). The Normalized Difference Vegetation Index (NDVI) was calculated from a cloud-free, atmospherically corrected Landsat 8 image from May 2014 (30-m resolution, values scaled by 1,000). Vapor pressure deficit (VPD, in Mbars) and minimum temperature were calculated from the PRISM dataset (4 km) by averaging over the years, 2012–2014. VPD was averaged over the driest warmest months (July, August, September), while T_min was averaged over the fall months (September, October, November) when larvae may be particularly susceptible to cooling (Yuval and Spielman 1990), ArcGIS 10.5.1 (ESRI, Redlands, CA) was used to join data sources with census block group data, obtained from the United States census. Total cases of Lyme disease (both probable and confirmed) for the 10-yr study period were counted within the block groups. Zonal statistical tools in ArcGIS were used to collect average values for environmental variables within each block group. Continuous variables were centered on 0 by subtracting the mean, then scaled by dividing by two SDs (Gelman 2008).

Analytical Methods

We evaluated spatial heterogeneity of our dependent variables using Bayesian generalized additive models. These models were implemented using the brms package in the statistical programming language R (www.r-project.org) (Bürkner 2017) which facilitates the construction of Bayesian regression models that are then transferred to the program Stan (mc-stan.org) for sampling of the posterior distribution (Carpenter et al. 2017).

Observations, which were the number of Lyme disease cases per census block group, contained a large number of zero values. We assumed that a zero count could occur for two reasons: either case occurred but were never recognized or reported, or they were true zero counts. After comparing several modeling approaches for zero-inflated count data, we found that a zero-inflated negative binomial model was the most suitable for our data. Zero-inflated models are based on a mixture of a zero-generating process (i.e., a degenerate distribution that produces zero values) and a count distribution. In other words, some zeroes may be due to missing data (true cases that went unobserved/unrecorded) and some may be due to true zero counts. In this case, we considered a negative binomial distribution to model count values. The underlying population of each block group was accounted for using a population offset term in the negative binomial component and using population as a linear predictor in the logistic component. In the latter case, this was because we assumed a 0 count (whether because of an unrecognized and unreported case of Lyme disease or because of truly 0 cases) would be more likely among block groups with lower populations.

Regression models normally assume independence of observations. In a spatial model, however, in which spatial units constitute the observations themselves, it is typical that there is autocorrelation among observations that are geographically closer to one another. We used an intrinsic conditional autocorrelation structure in our models to incorporate the spatial relationships of observations.

Because reported Lyme disease cases are investigated and classified by county health departments, we hypothesized that reporting biases may be clustered by county due to unmeasured variability in surveillance and reporting efforts. To account for this, we added a random intercept for county identity within the negative binomial component of our model.

For priors in our model, we applied a normal distribution with mean 0 and SD 1 to the intercept and other regression coefficients. These were relatively non-informative, regularizing priors, that did not appear to bias the results of our models. For the county level random intercept, a student T distribution with 3 degrees of freedom, location parameter 0, and scale parameter 1 was chosen. The model was run for two chains with 32,000 sampling iterations (16,000 after warmup), and the chains were combined for inference (see Supplementary material [online only]).

Results

A total of 755 cases of probable or confirmed Lyme disease were identified in the area during the 10-yr study period (Table 1). The range of cases per census block group was 0 to 52. Zero cases were recorded in 175 block groups (51%), and 129 block groups had four or fewer cases (38%). One block group with 52 cases was an outlier; the next highest count was 26. Eighteen census block groups had at least 10 cases of Lyme disease (5%) reported over the study period.

Table 1.

Summary of environmental variables among the 345 census block groups in the study area

	Mean	Median	SD	IQ Range
Mean elevation (Meters)	460.5	391.8	155.3	322.9–614.1
NDVI (units * 1000)	630.2	757.6	299.5	689.5–805.1
Minimum temperature Sept.–Nov. (°C)	6.78	7.12	1.02	5.61–7.73
Average vapor density, July–Sept. (Mbar)	17.15	17.59	1.89	15.78–18.75
Forest core area < 250 acres (%)	6	1	9	0–8
Forest patch area (%)	30	7	39	1–58
Forest edge area (%)	24	19	24	0–46

Next, we examined the influence of our environmental covariates (Table 2). Higher elevation was associated with more cases of Lyme disease; every 310-meter increase in elevation would be expected to increase the number of cases 1.86-fold per block group (95% credible interval 0.71–4.91). Forest edge area was also correlated with an increased number of cases, with a 2 SD increase in the percent of the forest considered edge associated with a 1.35-fold increase in Lyme disease cases (95% credible interval 0.88 to 2.10).

Table 2.

Effect of environmental variables on the distribution of Lyme disease cases per block group

Count model
	RR	95% CI	P ≠ 1
Mean elevation	1.86	0.71–4.91	0.90
NDVI	1.14	0.69–1.91	0.69
Forest patch area	1.01	0.67–1.52	0.52
Forest edge area	1.35	0.88–2.10	0.91
Forest core area	0.97	0.73–1.29	0.60
Minimum temp (Sept.–Nov.)	0.70	0.28–1.78	0.77
Vapor pressure deficit	0.71	0.23–1.78	0.77
Logistic regression model
	OR	95% CI	P ≠ 1
Mean elevation	0.43	0.07–2.41	0.83
NDVI	0.52	0.12–2.72	0.81
Forest patch area	0.72	0.17–2.85	0.67
Forest edge area	0.79	0.15–3.69	0.60
Forest core area	0.91	0.18–3.30	0.52
Minimum temp (Sept.–Nov.)	1.40	0.28–7.20	0.66
Vapor pressure deficit	0.81	0.18–3.79	0.62

The model was divided into a count (zero-inflated negative binomial) component and a binary (logistic) component. RR = Relative Risk, OR = Odds Ratio, 95% CI = 95% Credible Intervals, P ≠ 1 = probability that the effect size was not equal to 1, reflecting that probability that the environmental variable has a positive or negative influence on the outcome variable.

The variable most strongly associated with the presence or absence of any Lyme disease cases was NDVI. A 2 SD increase in NDVI was accompanied by an odds ratio (OR) of 0.56 of a block group having no Lyme disease cases (95% credible interval 0.12 to 2.72 to +1.4). In other words, greater vegetation density was associated with the presence of Lyme disease cases and less vegetation associated with their absence. No other environmental variables, including temperature, humidity, and forest fragmentation, were associated with either the presence or the abundance of Lyme disease cases.

Lyme disease cases were distributed heterogeneously, with the greatest case counts observed in the higher elevation environs of Blacksburg (Fig. 2). By contrast, Lyme disease cases were less abundant in the lower elevations around Roanoke. The adjusted model predicted a similar case distribution, albeit the spatial trends from the model appeared smooth by comparison to the native case counts.

Fig. 2.

Reported and predicted cases of Lyme disease per 1,000 people in the general population. Reported cases were most numerous south of the city of Blacksburg, corresponding to higher elevation terrain. Cases were less abundant in the lower-lying region to the northeast, including the city of Roanoke, and there were few cases in the northern part of the study area near the West Virginia border. The predicted distribution of cases resulting from our statistical model was smoother than the raw case counts. Adjustment for environmental variables, however, did not markedly change this distribution, suggesting that there remain unmeasured variables influencing the distribution of Lyme disease in this area.

Discussion

Higher elevation, greater vegetation density, and greater forest edge area were all associated with Lyme disease cases in the mountainous regions of southern Virginia. Temperature and humidity were not correlated with Lyme disease cases. We did not find an association between Lyme disease cases forest fragmentation. Forested mountain habitat in adjacent regions of North Carolina and Tennessee may be a suitable landscape for the continued southward expansion of Lyme disease transmission.

Our findings were similar to a previous examination of environmental and demographic correlates of Lyme disease cases between 2006 and 2010 in Virginia.(Seukep et al. 2015) In this study, the investigators found a positive correlation between percent herbaceous vegetation per tract and incidence and a negative correlation between percent developed land and incidence. Edges between forest and developed land were not associated with Lyme disease incidence, and there was an inconsistent relationship between herbaceous-forest edges. Herbaceous-developed edge areas were negatively associated with Lyme disease incidence. Measures of forest fragmentation were not found to be significant in this study.

Although there was a positive association between higher elevation and Lyme disease cases in this study, prior studies on the vector predicted a negative relationship between higher elevation and the presence of I. scapularis nymphs (Brownstein et al. 2005, Diuk-Wasser et al. 2010) and B. burgdorferi-infected nymphs (Diuk-Wasser et al. 2012). This discrepancy may reflect simply the rate of invasion dynamics of northern populations of I. scapularis and/or it may reflect an improvement in habitat suitability over time. Specifically, climate change may have made higher elevations more suitable for survivorship and development of I. scapularis ticks and their hosts, as has been seen in Europe with expansion of I. ricinus into higher elevations (Gern et al. 2008).

There is a sizeable body of literature in which remotely sensed data has been used to evaluate landscape features associated with Lyme disease. The earliest of these studies used Landsat imagery to evaluate two communities in Westchester, NY and found that Lyme disease risk was positively associated with both moisture and vegetation density (Dister et al. 1997). Studies since then have variably but not uniformly associated Lyme disease risk with forest fragmentation (Brownstein et al. 2005, Estrada-Peña 2009, Diuk-Wasser et al. 2012, Tran and Waller 2013, Larsen et al. 2014, Simon et al. 2014, McClure and Diuk-Wasser 2018). The majority of this literature, however, comes from the northeastern United States and may not apply to southern Virginia or other southeastern states.

Notwithstanding some local heterogeneity in Lyme disease risk, the northeastern region has been highly endemic for decades. Not only is Lyme disease less abundant in the southeast, but the rate of change in Lyme disease risk is considerably greater, where risk has markedly increased over a mere decade. A robust signal of thousands of cases, as is the case in the northeast, may strengthen statistical associations with ecological variables.

In addition to differences in case abundance and case stability, the environmental factors that either permit or constrain Lyme disease transmission are likely to differ between the southeastern and the northeastern states. Entomologic factors, such as host-seeking behavior of immature ticks, appear to be an important factor differentiating northern from southern tick populations (Nadolny et al. 2011, Kelly et al. 2014, Arsnoe et al. 2015, Ogden et al. 2018, Arsnoe et al. 2019). This, in turn, may be driven by tick population genetics, abiotic conditions such as temperature that affect survivorship, as well as host availability (Apperson et al. 1993, Ginsberg et al. 2014, Sakamoto et al. 2014, Van Zee et al. 2015, Ginsberg et al. 2017). Forested regions and forest edges at higher elevations appear to have the strongest association with human Lyme disease risk, and it is entirely plausible that ecologically similar areas of neighboring North Carolina and Tennessee will become endemic in the coming years. Already an increase in blacklegged ticks in eastern Tennessee, located on the other side of the Appalachian Mountains, has been observed over the last decade (Hickling et al. 2018).

This study has several limitations. First, we have aggregated 10 yr of case data, which prevents us from making temporal inferences about how Lyme disease emerged within this time period. With 755 total cases distributed over 10 yr and 345 block groups, a more granular temporal analysis may not have been fruitful. It is also possible, though unlikely that measures of forest structure and vegetation density could change during this interval. Worth considering in future analysis are the unmeasured effects of real estate development and residential mobility in the years surrounding the housing market crisis beginning in 2008, which falls in the middle of our study period.

A second limitation is that Lyme disease cases reported to public health departments represent only a minority of total cases, perhaps as little as 10% of total cases. At the same time, case definitions, even for confirmed Lyme disease, leave room for positive misclassification. Both these factors are likely occurring in southern Virginia, an area where Lyme disease will be less commonly considered or diagnosed than in more endemic regions. Moreover, the effort and funding available for Lyme disease surveillance efforts may vary between county-level public health departments. The comparative abundance of Lyme disease in Floyd County, VA in this study may be due to increased reporting capacity.

Third, census block group of residence was used to identify the locations of cases in this study, but this may not represent where all individuals acquired Lyme disease. For instance, individuals may reside in low-risk areas but acquire Lyme disease in nearby areas where they travel for recreation or work. We aggregated cases of Lyme disease by census block group, as this allowed us to consider the underlying population distribution. While convenient, census units seldom correspond to natural features. In environmental studies, census may create a biased sampling of environmental variables in that their area, shape, and number of neighbors may be spatially heterogeneous.

Finally, there is considerable ecologic heterogeneity in this region, including a diversity of soil types and vegetation. Elevation and vegetation indices, for instance, may be less predictive of Lyme disease transmission risk in areas south of Virginia or to the west of the Appalachian Mountains.

Despite these limitations, our study has several strengths. By focusing on a relatively small geographic area in which there is heterogeneous intensity of Lyme disease; this allows us to minimize the effect of factors that influence Lyme disease transmission over larger spatial scales. Examples of such factors, which may be relevant in a statewide or multi-state study, would include significant climate or weather differences, the effect of major natural or manmade barriers, demographic and economic factors that influence exposure to tick habitat, and the relative abundance of white-tailed deer. Block groups are generally smaller than zip code tabulation areas. They are, by definition, smaller than census tracts and counties, providing greater spatial resolution than these other commonly used units of aggregation for case-population research.

In summary, we have identified elevation, forest edge, and vegetation density as variables associated with the expansion of Lyme disease in Virginia. Suitable habitat for expansion can likely be found in adjacent areas of North Carolina and Tennessee, warranting increased active and syndromic surveillance in neighboring regions. Further entomologic and ecologic research are needed to determine whether these factors or human behavior drive the heterogeneity in Lyme disease risk in our small study area. Continued field research and human case surveillance will both be necessary to monitor the further expansion of Lyme disease into the southeastern states.