about exitflag of lsqnonlin

When I used the lsqnonlin, I found the data could affect the exitflag when the optimization stoped. When my data are perfect, the ef=4. But when the data contain noise, the ef=3. I tried to improve the accuracy of TolX and TolCon from 10^(-6) to 10^(-10).Then the exitflag changed form 4 to -3. How could I make the optimization has the same convergence criteria ef=3.