Extended Data Table 1 Specification of generalized linear models for analysis of paternal plastid transmission and plastid inclusion
From: Control of plastid inheritance by environmental and genetic factors
- Models 1–5 are the models analyzed in the manuscript. Models within a model family (Model Fam.) that are designated by a lowercase m are intermediates in the modelling process (that is, have parameters that were eliminated or added during the AICc minimization process). All models have ‘log’ as link function.
- †Dispersion parameter θ estimated from the data. The standard error of the estimate is given in parentheses.
- ‡ Includes the intercept, parameters for the effects of variable, and interactions. In negative binomial models, it also includes the dispersion parameter θ. The extra parameters (+) or the missing parameters (-) that the respective model includes compared to the selected model are indicated in parentheses. T, Temperature; G, Genotype; E, Experiment (in the case of Model 2, E refers to the experiments performed at different stages of development, whereas in the case of Model 3, E refers to the three independent experiments conducted to determine the rate of paternal plastid transmission).
- §LRT compares two models by checking whether λLR (difference in residual deviances) surpasses the critical value χ20.05 with Δdf degrees of freedom, where Δdf is equivalent to the difference in the number of parameters between the two models. For all LRTs: α = 0.05; ns, P > 0.05; **, P < 0.01; ***, P < 0.001. Null hypotheses are formulated depending on the comparisons. LRTs are single-tailed by definition, and no corrections for multiple comparisons are performed.
- || LRT between a candidate model and the saturated model built from the same data (also called deviance test) is a general measure of goodness of fit of the candidate model. The saturated model has as many parameters as observations, and it fits maximally to the data by definition (residual deviance is zero). Thus, λLR is equal to the residual deviance of the candidate model. The null hypothesis is that the candidate model fits the data as well as the saturated model. Non-rejection of the null hypothesis is interpreted as evidence that the tested model fits well the data.
- ¶ LRT between a negative binomial model and its nested Poisson is a test for data overdispersion (when the conditional variance is higher than the conditional mean). The null hypothesis of the test is that the Poisson and negative binomial models fit to the data equally well. A rejected null hypothesis indicates the negative binomial model fits better to the data.
- $ Model Family 3 includes the calculation of the interaction between genotype and temperature across all experiments, as well as the interactions involving Experiment 2. By contrast, triple interaction terms of the form GxTxE could not be calculated, as well as the GxE and TxE interactions involving Experiment 3. The independent experiments of paternal plastid transmission did not always contain experimental groups of all levels of genotype and temperature, thus leaving those interactions undefined for the dataset.
