2.2. Climate data

We acquired climate data from two sources: (1) the PRISM Climate Group (2013), which we used to extract both historical climate data and current climate data, and (2) the Moscow Forestry Sciences Laboratory (MFSL: Rehfeldt, Crookston, Warwell, & Evans, 2006), which we used to obtain current climate data and future climate scenarios. To develop historical climate surfaces, we created 30‐year (1910–1940) climate normals based on monthly interpolated conditions of precipitation, maximum temperature, minimum temperature, and mean annual temperature using data from the PRISM Climate Group (2013). From these four variables, we created five additional climate variables: December precipitation, July precipitation, precipitation seasonality, January minimum temperature, and July maximum temperature. We define precipitation seasonality as the degree of variability in monthly precipitation throughout the year and calculated this variable by calculating the coefficient of variation of precipitation across each year using the “raster” package (Hijmans, van Etten, & Cheng, 2015) as implemented in R (R Core Development Team, 2015). To directly compare between historical and current climate projections, we developed current climate surfaces using these same nine variables: we created 30‐year (1970–2000) climate normals based on monthly interpolated conditions of precipitation, maximum temperature, minimum temperature, and mean temperature using data from the PRISM Climate Group (2013). We used those four variables to calculate December precipitation, July precipitation, precipitation seasonality, January minimum temperature, and July maximum temperature (Table 1). We also tested 16 climate variables from the MFSL (Rehfeldt et al., 2006) for multicollinearity using the “rfUtilities” package (Evans & Murphy, 2014) as implemented in R (R Core Development Team, 2015). Using a multicollinearity threshold of 10⁻⁷, we removed six climate variables from our dataset. We included 10 nonredundant variables from the MFSL for current and future models (Table 2; Evans & Murphy, 2014; Evans, Murphy, Holden, & Cushman, 2011; Murphy, Evans, & Storfer, 2010). We included these particular variables (Tables 1 and 2) because they have been shown to be important in structuring the distribution of bird species (Peterson et al., 2007). Though we only modeled breeding distributions, we included variables related to winter conditions as winter climate influences forest dynamics of western North America where sapsuckers breed (Mantgem et al., 2008; Rehfeldt, Ferguson, & Crookston, 2009). Winter climate conditions can also influence migration time (Gienapp, Leimu, & Merilä, 2007; Hüppop & Hüppop, 2003; Stervander, Lindström, Jonzén, & Andersson, 2005) and the timing of breeding by affecting the period of peak insect abundance (Saino et al., 2011; Thomas, Blondel, Perret, Lambrechts, & Speakman, 2001), as well as affecting ecosystem productivity by altering year‐round water availability (Illán et al., 2014). We also included winter climate in our models because some populations of Red‐breasted Sapsuckers included in our study area are sedentary, and thus directly experience winter conditions (Walters et al., 2014b).

Climate variables from the PRISM Climate Group (2013) used in Random Forests models for both current and historical classifications. Number represents variable importance in each model as assessed using the Gini impurity index, which ranks variables based on the decrease of the accuracy of an RF model when a variable is excluded, such that important variables lead to a large decrease in model performance when excluded (Breiman, 2001; Liaw & Wiener, 2002). Dashes mean a variable was not included in the final model

Climate variables from the MFSL (Rehfeldt et al., 2006) used in Random Forests models for both Current classifications. Number represents variable importance in each model as assessed using the Gini impurity index, which ranks variables based on the decrease of the accuracy of an RF model when a variable is excluded, such that important variables lead to a large decrease in model performance when excluded (Breiman, 2001; Liaw & Wiener, 2002). Dashes mean a variable was not included in the final model

We chose three future climate scenarios from a general circulation model (GCM) based on the Canadian Center for Climate Modeling and Analysis (CGMC3). Three emission scenarios, representing low (B1), medium (A1B), and high (A2), were chosen for the year 2090 (IPCC, 2000).