2.2 | Statistical analyses
The basic population composition of captured wintering robins at Hutu Ranch was constructed after testing for the sex of the bird using molecular analysis. To determine potential morphological differences between females and males with the same-colored plumage, independent comparison tests were conducted based on the normality and homoscedasticity of the data for the eight morphological traits measured.
The optimal indexes to describe fat reserve on birds may vary among different species and populations (Labocha and Hayes, 2012). Therefore, it is necessary to test the significance of different body condition indices on body fat reserve. In this study, we dissected the carcass of robins (n = 9) which had unexpectedly died and weighed the fat reserve in the furcular and abdominal region. The sum of the fat reserve in the furcular and abdominal region was recorded as fat mass (±0.001 g). Furthermore, the proportion of fat mass in the whole body (fat mass / body mass) was recorded as fat percentage (%). Fat mass and fat percentage were used as response variables to explore the relationship between fat reserve and other body condition indices using ordinary least squares (OLS) regressions models. The significance of indices was mainly confirmed using the R2and p -value.
To study the dynamic body condition patterns in winter, we conducted an OLS regression analysis of suitable body condition indices with ordinal days. I Then, a comparison test was performed to compare the body condition indices among different winter stages. The research period in this study was divided into five stages: pre-winter (2022.11.6 – 2022.11.28), cold wave (2022.11.29 – 2022.12.1), early winter (2022.12.2 – 2023.1.13), midwinter (2023.1.14 – 2023.1.28) and late winter (2023.1.29 -2023.3.1), each stage based on the temperature change trend in that period. However, no more robins were captured or recorded after 2023.3.1. It is possible that these robins had already started their spring migration northwards in early March.
For further analysis of predictor variables, fat score was chosen as the response variable based on its significant regressive relationship with ordinal days and its significant variance among winter stages. The effects of the structural size of birds and daily capture time had been discussed in some earlier studies, as the structurally larger birds are usually heavier and can hold larger amounts of fat (Labocha and Hayes, 2012). Small passerine birds usually acquire more fat reserve throughout the day in winter (Colorado Z. and Rodewald, 2017). Therefore, we tested the effect of structural size (the pca1 result of eight morphological traits) and daily capture time on selected body condition indices and determined that only daily capture time had a significant positive effect on fat score (β = 1.2844, p = 0.007). The residuals of the fat score that extracted daily capture time by OLS regression were then used as dependent variables for further analysis.
The predictor variables were divided into two groups: external and internal. External factors consisted of environmental conditions, including local temperature and humidity, snowfall events, and invertebrate biomass. Internal factors consisted of the sex and capture status of the birds. Detailed descriptions of the factors mentioned above are available in Table S2 of the supplementary material. The effect of those predictor variables on the fat score residual was analyzed using a multiple linear regression model for the day of capture, the three-day average before capture, and the seven-day average before capture. Predictor variables were scaled using the scaling method in the R ‘arm’ package. All potential predictor variables were included in the initial version of the full model. A VIF test was then conducted to check for multicollinearity, and predictors were sequentially eliminated until all the VIF values of predictor variables were below 10. Then the full model final version was determined. All possible submodels were constructed from the full model, and an Akaike information criterion value (AIC) was then calculated for each submodel. Using the model averaging method, models having a ΔAIC < 2 were retained and averaged, and finally an optimal model was constructed. All statistical analyses were conducted in R version 4.2.1. The function ‘dredge’ and ‘model.avg’ in the R package MuMIn was used for model selection.