inst/shiny/docs.md

In psolymos/moosecounter: Adaptive Moose Surveys

General workflow

The over all workflow is to set Settings, load your Data, model Total moose counts and then model moose Composition.

Within the two modelling steps are further steps: Explore the variables, create Models, check model Residuals (in Total), create Prediction Intervals and then Explore or Summarize these prediction intervals.

Next, we will go over these steps in detail.

Settings

You can adjust the settings for the analyses by modifying options:

• response (total vs. cows),
• model optimization methods (Nelder-Mead, etc.)
• bootstrap iterations
• alpha level
• weighting scale
• sightability
• random seed*

* particularly useful for troubleshooting and reproducibility

At the bottom of the settings page is a section, Current Settings, which you can peruse to see all settings in effect, including those not currently available for modification.

Note that for reproducibility and troubleshooting, settings are included in the data downloads (under Total > Explore Predictions and Composition > Summary)

Data

To start an analysis, you’ll first want to load your data. We include an example data set if you would like to explore but do not have your own data yet. Click on “Browse” to choose a data set.

Explore your data using Interactive Table and Data Structure tabs to ensure the data you’ve loaded looks as it should.

Note that predictors/explanatory variables are inferred and defined as variables which are not Response variables or Metadata variables. See the Data Structure tab so see how your variables have been categorized.

Potential response variables are: BULL_SMALL, BULL_LARGE, LONE_COW, COW_1C, COW_2C, LONE_CALF, UNKNOWN_AG, MOOSE_TOTA, COW_TOTA

Potential metadata variables are: SURVEY_NAM, YT_REGION, SURVEY_YEA, SURVEY_ID, S_SET_ID, SU_ID, ID, SUS_, SUS_ID, S_YEAR_ID, S_TYPE, S_SEASON, PROJECT_ID, CENSUS_ID, In1Out0, SU_STRATUM, ALL_STRATA, IDLATDEG, IDLATMIN, IDLONDEG, IDLONMIN, CENTRLAT, CENTRLON, REGION, MMU_ID, GMU, GMU2, USE_SCALE, SRC_SCALE, SRC_NOTES, AREA_KM, PERIM_KM, AREA_MI, PERIM_MI, FEATURE_ID, SUBSET_NAM, SUBSET_ID, Sampled, srv, Kluane_ID

Omit variables with too few surveyed levels.

In order for a predictor to be useful, it’s levels must exist in cells that were surveyed as well as in cells that were not. If they don’t have this variability they can cause problems when modelling. While you can simply ignore these variables in the models, it can be simpler to omit them here. Any variables that may be problematic will be automatically selected. You may choose to not omit them if you wish, by removing them from this selection.

Convert integer to categorical

When R loads data scored as numbers, it assumes they are numeric. However, depending how you’ve created your data, integer values may actually be categorical. Click on the “Data Structure” tab to see what the current format of your data is. To ensure your models are appropriate, add any variable which should be categorical but are instead listed as integer (int) to this field. They should then switch to categorical (Factor) in the Data Structure display.

Filter data

If you wish to model only a subset of your data, choose the appropriate grouping levels in this section. Then use the Interactive Table and Data Strucutre tabs to double check your selection.

Modelling moose totals

Exploring predictors

Before modelling you should explore your potential predictors either one-by-one (Univariate exploration) or in tandem (Multivariate exploration).

In a univariate exploration, choose the variable of interest and the distribution you want to model. The figures below reflect Total Moose Counts as a function of the variable chosen by density of samples, spatial sampling, and the total moose response.

In a multivariate exploration, choose at least two variables of interest and the alpha level you want to use to define a split. The figure below shows the conditional inference tree (non-parametric regression tree) defining relationships among the variables.

Build models

In the Models window you can build different models to explore. You can assign variables to be modelled as count or zero or both. You can choose the distribution family to model (P = Poisson, NB = Negative Binomial, ZIP = Zero Inflation Poisson, and ZINB = Zero Inflation Negative Binomial), and whether the variables should be weighted or not (weighting moderates potentially influential observations). By default, models are named alphabetically, but you can give each model a distinct ID if you choose.

Once you have added models, they will appear in the model table and in the AIC Model Comparison table. You can delete models by clicking on the “X” next to them.

Any model problems (convergence, etc.) will be displayed in the Error messages section.

Exploring Residuals

The Residuals window will show you the AIC Model Comparison table along with plots showing both the spatial (left) and general (right) distributions of residuals for the selected model.

You can compare model fits and, if necessary, remove models by clicking on the X next to the model name in the Models window.

Prediction Intervals

After you’ve settled on the models you want to use, go to the Prediction Intervals window to calculate the PIs.

Choosing Models

First choose the model(s) you want to use. If you choose more than one, you have the option of using the best among those models, or averaging over the models. Once you’re satisfied, click the “Calculate PI” button. Calculating the prediction intervals can take some time, watch the progress bar in the lower right hand corner to see how it goes.

Note that if you change anything used to calculate Prediction Intervals (Settings, Data, Models), the button will turn Yellow and prompt you to re-run the calculation.

Outputs

The PI output has several different parts. First there is the overall “Summary” of estimated Total moose counts and density. You are reminded of the settings and alerted to potential issues under “Issues and Options”. In the lower half of the window you have the model outputs. These are divided in to sub tabs.

The first tab is Diagnostic Plots. As in the Residuals window, we have plots showing spatial (left) and general (centre) distributions of residuals. We also have a third plot (right) showing spatial accuracy. The model used is noted in the title of these plots. If you chose to average over models, it will have “Avg” as the title.

The second tab is Moose Predictions. This tab shows density plots of the prediction intervals over all (left) and for specific cells in the data (right). Use the new input in the upper right box called “Cell to plot for predictions” to choose a cell to plot.

If the right-hand, cell-level plot is empty, with text such as “Observed Count = X”, this means that the cell was actually surveyed, so there is a single observation and no prediction for that cell.

The third tab is Bootstrap Results. This shows an interactive table of bootstrap results per SU_ID.

Note that these data are all available to download as sheets in an Excel file in the next window, Explore Predictions.

Explore Predictions

In this window, you can subset the data you want to explore by specifying the column defining the group and then the groups you want to keep. By default all data is displayed (this only applies to the display, the download contains all data).

For the map, you can pick different calculated values to display. You can also adjust the number of colour bins to ensure the best differentiation.

For example, the starting value shows you “observed values” which are the actual values surveyed. “Cell.pred” shows you the predicted moose counts based on your models, which may help identify un-surveyed locations with potential high numbers of moose.

Click on the “Plot” tab to explore the predictions results as a function of different explanatory variables, separated by surveyed and non-surveyed results. The grey ribbon represents the prediction intervals.

You can download the run info, settings, data, and bootstrap runs as an Excel file by clicking “Download full results as Excel file”.

Modelling moose composition

Explore

As with the Total count models, you should first explore your potential predictors. Use the drop down menu to choose a predictor to explore. The figures show moose composition as a function of the total moose count.

Models

Next, build your models by defining the model ID (defaults to letters of the alphabet) and the predictors you want to include. As with Totals, you can define several models and compare them with the AIC table.

You can delete models by clicking on the X next to the Model ID. Error messages, if any, will appear in the lower left corner.

Prediction Intervals

Choosing Models

To calculate Prediction Intervals for Composition, you will need models for Total counts as well as models for composition. Total count models can be averaged, therefore you can select more than one. However, you can only select one composition model. As in Total > Prediction Interval, if you select more than one Total model you can decide whether you want to use the best model or average over the models.

Once you are satisfied with your selection, click “Calculate PI”. Again this may take some time, see the progress bar in the lower right corner.

Note that if you change anything used to calculate Prediction Intervals (Settings, Data, Total Models, Composition Models), the button will turn Yellow and prompt you to re-run the calculation.

Outputs

Potential issues and options are defined in the lower left corner, while the Summary and Bootstrap Results are presented as tabs on the right side of the window.

The Summary table shows the percentile for the prediction intervals of total moose counts as well as by composition.

The Bootstrap Results show an interactive table of bootstrap results per SU_ID. (This table may take time to load).

Summary

Finally we have the Composition Summary window. Here you can see the prediction intervals for total moose counts and composition by SU_ID.

You can subset the data you want to explore by specifying the column defining the group and then the groups you want to keep. By default all data is displayed (this only applies to the display, the download contains all data).

You can download the run info, settings, data, and bootstrap runs as an Excel file by clicking “Download full results as Excel file”.

psolymos/moosecounter documentation built on Feb. 25, 2024, 4:43 p.m.