Over the past few decades, there has been an explosion in the amount of publicly available sequencing data. This opens new opportunities for combining data sets to achieve unprecedented sample sizes, spatial coverage or temporal replication in population genomic studies. However, a common concern is that nonbiological differences between data sets may generate patterns of variation in the data that can confound real biological patterns, a problem known as batch effects. In this paper, we compare two batches of low-coverage whole genome sequencing (lcWGS) data generated from the same populations of Atlantic cod (Gadus morhua). First, we show that with a "batch-effect-naive" bioinformatic pipeline, batch effects systematically biased our genetic diversity estimates, population structure inference and selection scans. We then demonstrate that these batch effects resulted from multiple technical differences between our data sets, including the sequencing chemistry (four-channel vs. two-channel), sequencing run, read type (single-end vs. paired-end), read length (125 vs. 150bp), DNA degradation level (degraded vs. well preserved) and sequencing depth (0.8× vs. 0.3× on average). Lastly, we illustrate that a set of simple bioinformatic strategies (such as different read trimming and single nucleotide polymorphism filtering) can be used to detect batch effects in our data and substantially mitigate their impact. We conclude that combining data sets remains a powerful approach as long as batch effects are explicitly accounted for. We focus on lcWGS data in this paper, which may be particularly vulnerable to certain causes of batch effects, but many of our conclusions also apply to other sequencing strategies.