Vector network analyzer accuracy depends almost entirely on calibration. The three methods I see most often:

SOLT (Short-Open-Load-Thru)
- Uses known standards characterized by a calibration kit definition (models for open capacitance, short inductance, load impedance, thru delay).
- Works in coax with commercial kits; fast and familiar.
- Accuracy relies on how well the standards are known. The load is usually the limiting standard at high frequencies.
- Good for coaxial devices and general lab work.

TRL family (Thru-Reflect-Line, plus variants like LRL and LRM)
- The line standard's characteristic impedance defines the reference impedance; the reflect standard does not need to be precisely known, only identical on both ports.
- Ideal for non-coaxial media: on-board fixtures, waveguide and on-wafer.
- Line length limits bandwidth: a single line is usable roughly where its phase differs from the thru by about 20 to 160 degrees. Multiple lines extend coverage.
- Low-frequency coverage needs long lines or a different method (LRM uses a match instead).

Electronic calibration (ECal)
- A module with switched, factory-characterized impedance states.
- Fast, repeatable and reduces operator error; good for production and for frequent recalibration.
- Accuracy is tied to the module's characterization and its connector condition.

Practical advice
1. Calibrate at the plane you want to measure. For fixtures, either TRL on the fixture or calibrate in coax and de-embed the fixture with its S-parameters.
2. Let the analyzer and cables warm up; temperature drift is real.
3. Use phase-stable cables and do not move them after calibration.
4. Torque connectors to specification with a proper torque wrench.
5. Verify the calibration with a known device (an airline, a verification kit, or simply a mismatched load you measured earlier).
6. Choose IF bandwidth and averaging consistently; narrow IF bandwidth reduces trace noise at the cost of sweep time.