We analyze the changes in the mean and variance components of a quantitative trait caused by changes in allele frequencies, concentrating on the effects of genetic drift. We use a general representation of epistasis and dominance that allows an arbitrary relation between genotype and phenotype for any number of diallelic loci. We assume initial and final Hardy-Weinberg and linkage equilibrium in our analyses of drift-induced changes. Random drift generates transient linkage disequilibria that cause correlations between allele frequency fluctuations at different loci. However, we show that these have negligible effects, at least for interactions among small numbers of loci. Our analyses are based on diffusion approximations that summarize the effects of drift in terms of F, the inbreeding coefficient, interpreted as the expected proportional decrease in heterozygosity at each locus. For haploids, the variance of the trait mean after a population bottleneck is var(Δz̄) = Σⁿ_k=1 F^kV_A(k), where n is the number of loci contributing to the trait variance, V_A(1) = V_A is the additive genetic variance, and V_A(k) is the kth-order additive epistatic variance. The expected additive genetic variance after the bottleneck, denoted ⟨V^*_A⟩, is closely related to var(Δz̄); ⟨V^*_A⟩ = (1 − F) Σⁿ_k=1 kF^k−1V_A(k). Thus, epistasis inflates the expected additive variance above V_A(1 − F), the expectation under additivity. For haploids (and diploids without dominance), the expected value of every variance component is inflated by the existence of higher order interactions (e.g., third-order epistasis inflates ⟨V^*_AA⟩). This is not true in general with diploidy, because dominance alone can reduce ⟨V^*_A⟩ below V_A(1 − F) (e.g., when dominant alleles are rare). Without dominance, diploidy produces simple expressions: var(Δz̄) = Σⁿ_k=1 (2F)^kV_A(k) and ⟨V^*_A⟩ = (1 − F) Σⁿ_k=1 k(2F)^k−1V_A(k). With dominance (and even without epistasis), var(Δz̄) and ⟨V^*_A⟩ no longer depend solely on the variance components in the base population. For small F, the expected additive variance simplifies to ⟨V^*_A⟩ ≃ (1 − F)V_A 4FV_AA 2FV_D 2FC_AD, where C_AD is a sum of two terms describing covariances between additive effects and dominance and additive × dominance interactions. Whether population bottlenecks lead to expected increases in additive variance depends primarily on the ratio of nonadditive to additive genetic variance in the base population, but dominance precludes simple predictions based solely on variance components. We illustrate these results using a model in which genotypic values are drawn at random