My dependent variable will be drop developed in Han and Goetz (2015). It is reciprocal for resistance, which is calculated as the amount of impulse that a county experiences from a shock (the percentage of deviation of the actual employment from the expected employment during the Great Recession).

I want to find what factors affect drop. Then I run an OLS regression using drop as the dependent variables and following independent varaibles:

Income inequality: the Gini coefficient

Income distribution: poverty rate and the share of aggregate income held by households earning $200,000 or more

Control variables: Population growth rate from 2001-2005, % Black or African & American (2000), % Hispanic or Latino (2000)

Capital Stock variables:

Human Capital: % population with Bachelor’s degree or higher, age group (20-29, 30-49), female civilian labor force participation rate (2000)

Natural Capital: natural amenity scale (1999)

Social Capital: social capital index (2005)

Built Capital(2000): median housing value (2000)

Financial Capital (2000): share of dividends, interest and rent(2000)

Economic structure:

Employment share of 20 two-digit NAICS industries (manufacturing, construction, etc. other services (except public administration) omitted)

Significant variables and diagnostic tests results are shown in the following table.

Among the significant variables, I want to explore population growth rate from 2001 to 2005 because it has the larges coefficient in size among control variables.

Moreover, since the Konker BP statistics are significant, the results are non-stationary. Then I run the GWR for population growth rate 2001-2005.

GWR

The following maps are the distribution and GWR coefficients of population growth rate 01-05

For the distribution of population growth rate 01-05, blue regions show negative population growth while the rest counties are all growing in population. Orange and red counties are high in population growth rate.

For the GWR coefficients, negative relationship between population growth rate and drop exist in blue, grey and yellow regions – higher population growth, lower drop (employment loss)

Positive relationship exist in orange and red regions – higher population growth, higher drop (employment loss)

My explanation are :

In an expanding economy (regions high in population growth), there are more people marginally attached to the labor market. They are easily fired in the Great Recession.

In regions with negative population growth, there more deaths than births and population aging rises.

• Older people have higher accumulated savings per head than younger people, but spend less on consumer goods.
• Less available active labor
• An increase in population growth will decrease employment loss.

However, my local R-squared, range from 0-0.620 with the mean 0.0247. Only in Orange County and San Deigo, CA, local R-squared is higher than 0.5, 0.603 and 0.620 separately. This situation (low local R-squared) happened to all my other significant control and capital stock variables ()  There might be misspecification problems in my model, I will try adding new  explanatory variables.

Population growth rate

• Local R-squared, 0-0.620, mean 0.0247
• Only in 06059 Orange County and 06073 San Deigo, CA, local R-squared is higher than 0.5, 0.603 and 0.620 separately

Natural amenity scale

• Local R-squared, 0-0.519, mean 0.0216
• 12087 Monroe County, Florida 0.519

% Black or African American

• Local R-squared, 0-0.347, mean 0.0326
• 26095 Luce county, Michigan, 0.347

% Hispanic or Latino

• Local R-squared, 0-0.379, mean 0.0276
• 35029, Luna County, New Mexico, 0.379

% pop with Bachelor’s degree or higher

• Local R-squared, 0-0.391, mean 0.055
• 06073 San Deigo, CA, 0.391

Goal: Where does statistically significant spatial clusters of high values (hot spots) and low values (cold spots) of drop lie?

Variable: drop (employment loss)

• Monthly employment Jan 2003- Dec.2014
• Data availabl:2840 counties
• 

Tool : Hot spot analysis

Step:

• Add Mean center of population.shp (3141 counties), U.S. county.shp (3141 counties), drop value (2840 counties)
• Join the drop to the mean center, export as a new layer file.
• Project the data
• Run Hot Spot Analysis (Default for Conceptualization of Spatial Relationship, and Distance Method)
• Make a Map

1. Distribution of drop, using natural Break

2 Hot spot Analysis using 3141 counties.

• Low employment loss gathers in north part
• High employment loss lies in west and southeast part

Questions:

• Previous map is made of 3141 counties, however, there are only 2840 counties with drop value. Missing value will be noted as 0, will this affect the result?
• Hawaii and Alaska are included. Will exclusion of HI and AK affect the result?

Both: Yes

• Redraw the map
• Only keep the matched records – 2838 counties
• Exclude Hawaii and Alaska

3. Hot spot Analysis using 2838 counties

4. Hot spot Analysis using U.S. contiguous counties, Hawaii and Alaska removed.

Conclusion: Hot Spot analysis is sensitive to spatial outlier. It only identifies low value clusters or high value clusters. What if I want to find an outlier, for example low value unit among high value cluster?

I should use spatial autocorrelation tool- Anselin Local Moran’s I

Background:

In August 2015, the BLM released the National Seed Strategy which calls for the use of genetically appropriate seed for the restoration of damaged ecosystems. Bluebunch wheatgrass is a long-lived native perennial species that grows in most western states in the United States as well as British Columbia. This species has been shown to be genetically adapted to different climate patterns across its range. Adaptation is evident because bluebunch wheatgrass plants from distinct climates exhibit different phenotypic (physical) traits. Bluebunch wheatgrass seeds should therefore only be dispersed in areas where they are adapted to the local climate conditions.

In order to determine how far seed from this species can be moved across its range, species specific seed zones have been developed for the Great Basin ecoregion (St. Clair et al 2013). These seed zones were delineated using spatial analysis of climate patterns and observed plant traits from many populations of bluebunch wheatgrass in different climates in the Great Basin.

In order to test the efficacy of the seed zones for bluebunch wheatgrass common gardens have been established. Common gardens are a way of minimizing the effect of site history (i.e. grazing, fire, and management practices) on the growth habits of bluebunch wheatgrass and illuminate the climate-caused genetic adaptations. Wild seeds are collected from dispersed locations in the Great Basin, reared in a greenhouse to the young adult stage, and co-located into a common garden (see figure 1). Once the plants are placed within the common garden they can be monitored for phenotypic trait variability and these variations can be “mapped” back to their home climate.

Although the relationship between bluebunch wheatgrass phenotypes is fairly well understood, there is a knowledge gap surrounding bluebunch phenotypes and soils. I aim to use data gathered from existing common garden studies and soil maps to determine if relationships between soils and bluebunch phenotypes exist.

Study Design:

Two common gardens are situated within each of 8 seed zones resulting in 16 gardens total. At each common garden, 4-5 populations from each zone are represented. Twice per growing season, phenotypic trait data such as leaf length and width, crown width, and number of reproductive stalks are gathered for each individual plant at all common gardens. The resulting dataset contains population level means for each trait at each garden.

Generalized Common Garden Schematic

(figure 1)

Research Question:

Do phenotypic traits of bluebunch wheatgrass vary with soil order?

Tools and Approaches:

Soils Dataset:

• Soils data were gathered from the Geospatial Data Gateway and were downloaded separately for each state in the study region to reduce the file size.
• Soils raster layers for each state were loaded into ArcMap.
• Tabular data for soil order was joined to raster data using the “Join Field” tool in ArcMap.
• Soil order layers were symbolized using graduated colors by category.

Bluebunch Trait Dataset:

• Latitude and longitude decimal degrees (X Y coordinates) were added to each population in the common garden plant trait dataset.
• Blank cells were replaced with NA values.
• X Y locations for each row in the dataset were “jittered” to remove duplicate coordinates (see figure 2 & 3).
  • Using the rand() function in excel, two new columns with random positive decimal values were added.
  • The difference between the two random value columns was subtracted from the X and Y coordinates for each row in the dataset.
  • The resulting “jittered” X and Y coordinates were stored in new columns.
• The bluebunch trait dataset was loaded in to ArcMap and visualized using the jittered X and Y coordinates.
• Hot spot analysis was performed separately for each of four plant traits; leaf width, leaf length, crown width, and number reproductive stalks and visualized with soil order.

(figure 2) Un-jittered population X Y locations

(figure 3) Jittered population X Y locations

(figure 3) Jittered population locations

Results:

In the hot spot analysis of leaf width one hotspot and three cold spots were revealed. This indicates that a significant number of plants that were sourced from those areas had either wide or narrow leaves.

In the hot spot analysis of leaf length, one hotspot and two cold spots were revealed. This indicates that a significant number of plants that were sourced from those areas had either wide or narrow leaves. In this analysis, the hot spot was in a different location than previously indicating that wider leaves were not necessarily longer or vice versa. Contrastingly, the two cold spots in this analysis were in a similar location as with leaf width. This indicates that short narrow leaves tend to also co-occur.

In the analysis of crown width, the hotspot occurred in the same location as with the leaf width analysis. This leads me to believe that plants from this vicinity tend to be larger in general than plants from other areas.

In the analysis of reproductive stalks only one large hot spot was found. This indicates that the larger plants in this region also tend to produce significantly more flowering stalks per plant than in other areas.

Critique:

■  The jittering effect could be reduced to make it easier to distinguish the populations. It may also be possible that this type of analysis is not appropriate for this data set because the phenotypic trait data is for one population being grown at multiple common gardens and this creates a situation where the data values have the same X Y coordinate. Even though, the jittering effect was done at random, there is still a chance that the results we are seeing are a product of the data transformation itself.

■  90-95% CI might be too stringent for reasonable genetic inference. In many ecological contexts p-values of less than 95% are a regular occurrence. In this case, leaf width / length, crown width, and the number of reproductive stalks of each plant probably co-vary. It does not appear that this type of analysis is meant to deal with co-variance.

■  Soil traits are de-emphasized. Although soil order was used as the background it was difficult to see a pattern between soils and plant trait hot spots. This may be partially due to the large scale of the analysis or may also be a product of the jittering.

■  Only one phenotypic trait can be visualized at once. Since only one trait can be mapped in the hot spot analysis at once, there is no way to know if the observed similarity in hot spot locations is truly significant.

■  Hot spot analysis does not work with raster data. In the future, I would like to find a way to look at the clustering patterns within soils and it appears that this analysis does not recognize raster data.

■  A less regional, and more zoomed in analysis might improve the interpretation. Since the soils are highly variable in this study area it may be more productive to narrow the scope of inference to include smaller areas. Doing this may make the soil – plant trait interpretation easier.

References:

St. Clair Bradley John, Francis F. Kilkenny, Richard C. Johnson, Nancy L. Shaw, and George Weaver. 2013. “Genetic Variation in Adaptive Traits and Seed Transfer Zones for Pseudoroegneria Spicata (bluebunch Wheatgrass) in the Northwestern United States.” Evolutionary Applications 6 (6): 933–48. doi:10.1111/eva.12077.

For the ArcGIS hot spot analysis I used the tutorial available at https://www.arcgis.com/home/item.html?id=6626d5cc81a745f1b737028f7a519521. Note that the tutorial must be downloaded using the “open” button below the image on the web page. This will download a folder with pdf file of instructions and a dataset for their analysis.

I followed the tutorial, but substituted a dataset of fire ignitions logged by the Oregon Department of Forestry between the years 1960 and 2009 (Fig. 1). This dataset does not include some fires that occurred in the state during that time, for instance some on federal land and many on non-forested land.

Figure 1: Oregon Department of Forestry logged fire occurrences from 1963 to 2009.

The tutorial was straightforward and easy to follow. The only “problem” I had was with the spatial autocorrelation step. My data did not produce any peaks in the z-score graph, so I arbitrarily chose a distance towards the lower end of the graph (tutorial steps 10 f and g).

My results from the hot spot analysis step (Fig. 2) showed areas with hot spots. These tended to be near roads and cities in most cases, with a few instances at high elevation in northeastern Oregon.

Figure 2: Results from hot spot analysis.

Results from interpolating the hot spot results using the IDW tool (tutorial step 12) (Fig. 3) produced anomalous edge effects (Fig. 3). Note the areas of high values spreading from hot spot areas into areas with no data along the coast. A close up image of the Medford area (Fig. 4) shows edge effects spreading into areas with no data.

Figure 3: Hot spot analysis interpolated surface produced by the application of the Arc IDW tool. Note the edge effects, especially along the coast and northern state boarder.

Figure 4: Interpolated hot spot surface produced by the Arc IDW tool along with the original ODF ignitions dataset points. Note the edge effects in the southwest corner of the image spreading from the area of high fire occurrence into areas with no data.

As part of a study I did for a master’s thesis, I used the ODF data along with another dataset to do a logistic regression analysis of ignition occurrences in the Willamette watershed excluding the valley floor. Within this area, the hot spot analysis placed hot spots along some roads and population centers in several instances (Fig. 5A). These results are consistent with the results from my study, which showed that human-caused fires were more likely to occur along roads and near areas of high population, and that lightning-caused fires were likely to occur at high elevations (Fig. 5B). The lack of a match between the two methods at high elevation is due to data points that were used in the logistic regression but not in the hot spot analysis. However, one disadvantage of the hot spot analysis is that it is univariate, or strictly occurrence based. The logistic method allows for correlation and the application of that correlation to the prediction of future occurrence

Figure 5: Results from A) hot spot analysis and B) logistic regression analysis for fire occurrence in the hills and mountains of the Willamette watershed. Redder areas are “hotter” in the hot spot analysis and have a higher ignition probability in the logistic regression analysis.

In conclusion, hot spot analysis is a good technique for finding areas of higher occurrence, but has no predictive power beyond analyzing past occurrences. The ArcGIS IDW can produce an interpolated surface, but it has an edge effect issue that could lead to questionable results. Regression methods, for example logistic regression, may produce results more suited to analyzing correlation of independent variables with event occurrence.