A couple of notes:
1. As has been mentioned, not all BIA devices are created equal. Multifrequency BIA devices (such as those produced by Inbody) use multiple frequencies of current to estimate body composition, and do tend to produce values that are more similar to those produced by DEXA. There are a fair number of studies that validate multifrequency BIA against DEXA, with multifrequency BIA (on average, across studies) underestimating BIA against DEXA by 2-3%.
2. Multifrequency BIA body composition estimates tend to be highly reproducible (they show excellent test-retest reliability) as long as several assumptions are met. As has been mentioned, it is important not to eat or exercise for 2-3 hours before the test. Moreover, a test participant must be standing for at least 10 minutes (preferably 20) before the test to allow fluid to settle in the lower extremity in a reproducible fashion (body fat estimates can increase up to 3% over a 20 minute period from sitting to standing). Finally, arm position when using an Inbody matters (I have experience with the 770 model) - participants can "gain" up to 3% body fat when moving the arm from a fully adducted position to a position of 90 degrees of abduction. This final point is generally not reported in the validation literature, which I think is a real limitation based on our own lab observations and pilot data. We use a position of 30 degrees of arm abduction, which is similar to the only reported value I could find the published literature. For the best (most precise) results, I encourage individuals be assessed at the same time of day under the same conditions during each test.
3. It is important to remember that DEXA, like BIA or underwater weighing, arrives at a body fat "value" through an estimation equation. DEXA does not measure body fat, and I don't believe studies been conducted to validate DEXA using actual human cadaver data. DEXA has been validated against animal cadavers, and I encourage the reading of these studies if one has an interest in the validity of DEXA values.
An older, although informative, analysis of body composition methodology was published by Ellis in 2000 and is freely accessible via the Journal of Applied Physiology link at PubMed (
https://www.ncbi.nlm.nih.gov/pubmed/10747204). A newer, updated review may include much of the information on multifrequency BIA, as much of the validation information specific to this methodology has been published in the past 5 years.