According to Quantum threshold theorem, error rate can be arbitrarily low. But when I use IBM Quantum Experience to measure a simple $|+\rangle$, it gives result of $|1\rangle$ with probability of 52.148%.
I'm not sure what your specific question is - the IBM Q processor does not implement an error correction scheme by default, so the theorem doesn't apply.
Furthermore, the statistic provided in the Wikipedia page suggests this chip would be incapable of sufficiently depressing the error rate:
At a 0.1% probability of a depolarizing error, the surface code would require approximately 1,000-10,000 physical qubits per logical data qubit
IBM's chips have 10-20 qubits, and your own figures indicate an error rate of >2%, so you'd either need significantly reduced error rates or significantly increased qubit volume.