Estimation of time-varying average treatment effects using panel data when unobserved fixed effects affect potential outcomes differently