My fixed effect model and methodology I'm doing my master thesis on FDI effect on Chinese wage inequality. I am new to quantitative econometrics so I have no idea if my wage equation is correct.
$$W_{it} = β X_{it} + λ_t + η_i + ε_{it}$$
Where the wage paid by firm $i$ in year $t$ is denoted as $W_{it}$. $X_{it}$ contains a set of control variables (including include total sales, total exportations, total labor compensations, firms’ fixed asset, firms’ R&D expenditures, employee’s turn over rate and a dummy variable of foreign ownership). A time effect, $λ_t$, controls for time varying elements that affect all establishments in a given year. An individual effect, $η_i$, captures time invariant element that differ across establishments. An error term, $ε_{it}$. All variables are measured in logarithm units.
Is that correct? If yes, what should I do next? I don't know if I should use pooled OLS or GMM....I have already reshaped and treated my data. I really appreciate your help. Thank you in advance.
 A: Two years of econometrics classes is a start. What you need to check in any case, is whether the various series that you will use are 2nd-order ("weakly") stationary or not - you cannot avoid that. If they are not, you cannot overlook their non-stationarity. Hopefully if they are non-stationary, they will be integrated of order 1, and then you should modify your regression specification accordingly to obtain a stationary setup.
Assuming then that you have stationary data in your hands:  
1) The first simplifying assumption should be that error terms are stochastically independent cross-sectionally, and in the time dimension also (and are identically distributed). Also, that regressors are independent from the error terms (in some cases independence is more than needed -a way to learn and also show effort is to study and realize how strong your assumptions on stochastic relations need be in your particular model).  
2) You have an individual effect, $\eta_i$. Model it as a "fixed effect" and not as a "random effect" to avoid dealing with (and choose among) the different sets of assumptions regarding its relation with the error terms. This is one of the "simplifying assumptions" I was talking about in my comment: as Hsiao(2003, 2nd ed.) "Analysis of Panel Data" p.41 writes, 

when $T$(time dimension) is finite and $N$ is large, whether to treat
  the effects as fixed or random is not an easy question to answer. It
  can make a surprising amount of difference in the estimates of the
  parameters.

..and then provides an illustrative example from a research paper by Hausman (1978). But you won't be able to avoid a "random-effects-like" approach, because:
3) You have a common time-varying component, $\lambda_t$. It is difficult to argue that it is not random. Assume then that it is random and stationary, it has a zero-mean, and it is not autocorrelated. And that it is also independent from the error term. This lambda links the cross-sections stochastically and makes the variance-covariance matrix of the stacked model not a scalar matrix.   
This implies that Generalized Least Squares on the stacked model is a prospect. If this is what you eventually choose, you are looking at an initial estimation step to obtain consistent estimates of the two variances (of $\lambda_t$ and of $\varepsilon_{it}$), and then apply  FeasibleGLS. As for GMM, I understand it is currently in vogue since Hansen received the Nobel prize, but it needs careful steps in order to be sure that you understand what you are doing with it -it can be a complicated business.
