- Authors and contributors: Katia Aristova, Simon Guo, John Lee, Adrianne Leung
A data analysis (inferential) project for MDS DSCI522 (Data Science Workflows) from Group 26
Bumble bee population are experiencing rapid decline in North America. In the last 20 years, certain bumble bee species abundance decreased by 96% across the United States (Cameron et al., 2011). This phenomenon is not limited to North America, European bee colonies also experience collapses and sudden bee deaths (Bacandritsos et al., 2010; Scheper et al., 2014). Researchers proposed a variety of associations and causes ranging from pesticide use (Henry et al., 2012), to mite infestation (Brettel & Martin, 2017), as all as habitat loss and climate change. The status of bee population health, diversity and abundance remains an important topic.
As part of Ontario's Pollinator Action Plan, the Ministry of Environment, Conservation and Parks conducted Bumble Bee Diversity and Abundance Survey. The dataset for this project comes from this survey which can be found on the Open Data Ontario website:
link. The raw csv dataset can be accessed via this link: dataset file.
Surveyors collected bumble bee counts at 67 location in southwestern Ontario, according to the dataset (please note, that the website says 46 but the dataset contains 67 locations) between 2015 and 2017. The methodology is absent from the website. Each site contains multiple observations for the same day. Sites were encoded as agricultural, natural or semi-natural remnant.
The features include, species and bee type (e.g. borealis species worker), county, year, date, plant information (sparse, many missing values).
Fig. 1 displays locations of each site. Southern Ontario is a highly populated sprawling metropolitan area with agriculture surrounding the urban and suburban areas. Natural sites are limited to conservation areas, parks and other protected areas. There are three different types of bee sites, namely:
- Agricultural
- Natural
- Semi-natural remnant
In this project, we are aiming to answer the following question: Do the bumble bee counts vary depending on the type of site?
Null hypothesis: The median bee count values are the same at agricultural, natural and semi-natural remnant sites in Southern Ontario.
Alternative hypothesis: The median bee count is not the same at agricultural, natural and semi-natural remnant sites in Southern Ontario.
Answering this question is important to see how the site type would impact the bee counts occurring in agricultural, natural and semi-natural sites. This can help direct planning in favour of pollination and bee health in agriculture, natural and semi-natural ecosystems.
Result of the analysis will be reported as a table of the hypothesis test statistics and p-value in the final report.
Exploratory data analysis (EDA)
Each row of the data set represents the number of bee counts of each species at the sampled bee site at the specified sample date. We are interested in the total bee count regardless of species, time, year, or bee type (worker, queen, etc.). Data wrangling was necessary to summarise the data set by grouping the sample date by year, bee site type and site id. Each row of the simplified dataset represents the number of bee counts per site.
Prior to analysis, we performed EDA to assess the type of distribution and density of bee counts at each type of site, as well as to assess median and mean of bee counts at each site type. The distribution of bee counts of types of site are represented with histogram plot:
Figure 1: The plot above displays the right skew in the bee count. The majority of samples have a count of zero. This is not surprising considering that bees are likely to congregate in certain areas and be absent from others.
and the density is displayed with violin plots combined with indication of median and mean values of bee counts on the same plot:
Figure 2: The plot shows that visibly there are difference in the mean and median values between sites. However, since they are counts, they cannot be simply compared with averages. The plot also underlines how the data is skewed.
The full EDA report can be found here.
As observed in EDA, the the data is right skewed. To answer the inferential question posted above, we plan to do a hypothesis test for independence of a difference in mean bee counts by site type using Poisson regression analysis. Since the response variable, count of bees, is discreet, Poisson is a more suitable regression test. In contrast with Linear regression, the coefficients in Poisson regression are multiplicative. The coefficient estimates need to be converted by putting e to the power of the estimate. The resulting number is the multiplicative difference in average bee counts (i.e. site type Natural has x times more bees on average than Agricultural).
Note: an even better model to analyze the dataset would be the zero-inflation model. However, the authors of this project lack knowledge of this model.
One vs all
Example:
- Agricultural mean vs natural mean
- Agricultural mean vs semi-natural remnant mean
- Natural mean vs semi-natural remnant mean
Assumptions
The Poisson regression analysis assumes that all observations are independent. As mentioned earlier, this assumption is likely satisfied as multiple observations were taken on the same day at each site using random sampling technique. More assumptions related to the use of regression are described in the full report.
Limitations
The dataset is small. The exact collection methodology is unknown. Since the dataset was collected by the Ministry of Environment in Ontario, we assume that the data was sampled randomly and independently. The researchers controlled for seasonality by only sampling between May and August.
The final report can be found here.
The results show that the p-values for the Agricultural and Natural pair at 0.064. Therefore we fail to reject the null hypothesis that there is a difference in the average counts for this combination. However, agricultural vs semi-natural remnant and natural vs semi-natural remnant are 0.000 and 0.005 respectively, indicating significance at the alpha level of 0.05. Therefore, we reject the null hypithesis.
The average count of bees is 1.222 times higher for the semi-natural remnant sites, and 1.132 times higher than in the natural sites.
These indicate that the semi-natural sites offer some attractive qualities to the bees. In the future, it would be important to examine these in more detail and attempt linear regression or machine learning models to see which site features are the most critical.
Side note To see results of the zero-inflation model please see the jupyter notebook section 5.3 here.
- To replicate the report and analysis: clone the GitHub repo. Run the following command from the root folder to create the report:
docker run --rm -v /$(pwd)://home//rstudio//bee_count johnwslee/bee_count make -C //home//rstudio//bee_count all
- To reset the repo to its original clean state, run the following:
docker run --rm -v /$(pwd)://home//rstudio//bee_count johnwslee/bee_count make -C //home//rstudio//bee_count clean
- To replicate the report and analysis: clone the GitHub repo and install dependencies listed below. Run the following command from the root folder to create the report:
make all
- To reset the repo to its original clean state, run the following:
make clean
- Python 3.9.5 and Python packages:
- docopt==0.6.2
- requests==2.25.1
- pandas==1.3.2
- R version 4.1.1 and R packages:
- knitr==1.33
- docopt==0.7.1
- tidyverse==1.3.1
- ggplot2==3.3.5
- boot==1.3.28
- testthat==3.1.0
- broom==0.7.9
Dependency diagram (created by makefile2graph):
The materials used in this project are licensed under the Open Government Licence - Ontario. If re-using/re-mixing please provide attribution and link to this webpage.
Bacandritsos, N., Granato, A., Budge, G., Papanastasiou, I., Roinioti, E., Caldon, M., ... & Mutinelli, F. (2010). Sudden deaths and colony population decline in Greek honey bee colonies. Journal of invertebrate pathology, 105(3), 335-340.
Brettell, L. E., & Martin, S. J. (2017). Oldest Varroa tolerant honey bee population provides insight into the origins of the global decline of honey bees. Scientific reports, 7(1), 1-7.
Cameron, S. A., Lozier, J. D., Strange, J. P., Koch, J. B., Cordes, N., Solter, L. F., & Griswold, T. L. (2011). Patterns of widespread decline in North American bumble bees. Proceedings of the National Academy of Sciences, 108(2), 662-667.
Henry, M., Beguin, M., Requier, F., Rollin, O., Odoux, J. F., Aupinel, P., ... & Decourtye, A. (2012). A common pesticide decreases foraging success and survival in honey bees. Science, 336(6079), 348-350.
R Core Team. 2021. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. https://www.R-project.org/.
Roback, Paul, and Julie Legler. 2021. Beyond Multiple Linear Regression: Applied Generalized Linear Models and Multilevel Models in r. CRC Press.
Scheper, J., Reemer, M., van Kats, R., Ozinga, W. A., van der Linden, G. T., Schaminée, J. H., ... & Kleijn, D. (2014). Museum specimens reveal loss of pollen host plants as key factor driving wild bee decline in The Netherlands. Proceedings of the National Academy of Sciences, 111(49), 17552-17557.
Wickham, Hadley, Mara Averick, Jennifer Bryan, Winston Chang, Lucy D’Agostino McGowan, Romain François, Garrett Grolemund, et al. 2019. “Welcome to the tidyverse.” Journal of Open Source Software 4 (43): 1686. https://doi.org/10.21105/joss.01686.
Xie, Yihui. 2014. “Knitr: A Comprehensive Tool for Reproducible Research in R.” In Implementing Reproducible Computational Research, edited by Victoria Stodden, Friedrich Leisch, and Roger D. Peng. Chapman; Hall/CRC. http://www.crcpress.com/product/isbn/9781466561595.