Abstract Background The objective of the present study was to investigate how variation in the faecal microbial composition is associated with variation in average daily gain (ADG), backfat thickness (BFT), daily feed intake (DFI), feed conversion ratio (FCR), and residual feed intake (RFI), using data from two experimental pig lines that were divergent for feed efficiency. Estimates of microbiability were obtained by a Bayesian approach using animal mixed models. Microbiome-wide association analyses (MWAS) were conducted by single-operational taxonomic units (OTU) regression and by back-solving solutions of best linear unbiased prediction using a microbiome covariance matrix. In addition, accuracy of microbiome predictions of phenotypes using the microbiome covariance matrix was evaluated. Results Estimates of heritability ranged from 0.31 ± 0.13 for FCR to 0.51 ± 0.10 for BFT. Estimates of microbiability were lower than those of heritability for all traits and were 0.11 ± 0.09 for RFI, 0.20 ± 0.11 for FCR, 0.04 ± 0.03 for DFI, 0.03 ± 0.03 for ADG, and 0.02 ± 0.03 for BFT. Bivariate analyses showed a high microbial correlation of 0.70 ± 0.34 between RFI and FCR. The two approaches used for MWAS showed similar results. Overall, eight OTU with significant or suggestive effects on the five traits were identified. They belonged to the genera and families that are mainly involved in producing short-chain fatty acids and digestive enzymes. Prediction accuracy of phenotypes using a full model including the genetic and microbiota components ranged from 0.60 ± 0.19 to 0.78 ± 0.05. Similar accuracies of predictions of the microbial component were observed using models that did or did not include an additive animal effect, suggesting no interaction with the genetic effect. Conclusions Our results showed substantial associations of the faecal microbiome with feed efficiency related traits but negligible effects with growth traits. Microbiome data incorporated as a covariance matrix can be used to predict phenotypes of animals that do not (yet) have phenotypic information. Connecting breeding environment between training sets and predicted populations could be necessary to obtain reliable microbiome predictions.