Newey-West Standard Errors

Discussion in 'Economics' started by annaland, Aug 26, 2007.

  1. Hi,

    I’m looking for the appropriate criterion to determine lag length for the Newey-West standard errors. Can anyone recommend a source?

    Thanks in advance.

    I'm using Greene's Econometrics text but am quite dissatisfied with his notation, which is why I'm confused about lag determination and not sure if AIC/SIC applies to Newey-West.

  2. Thanks! Unfortunately, I'm not using Panel data (I'm neither testing nor adjusting for fixed or random effects); I'm using GMM on Time Series...

    I think it’s either the number of observations to the 1/4th power or it's 1/4 * (number of observations raised to the 1/3rd power), but definitely not AIC/SIC as I previously stated

    ... not quite sure...
  3. Now I'm lost by what you said. What is the 1/4 power? Never heard of it.

    For a moment I was puzzled by your AIC/SIC comment but realized that you meant Akaike and Bayesian Information Criteria (AIC/BIC). Why don't you think that they would work? Did you try DIC (Deviance Information Criterion) which is easier to calculate?
  4. Newey and West discuss some alternatives in:

    Newey & West (1994). Automatic Lag Selection in Covariance Matrix Estimation. Review of Economic Studies, v61, n4 (October 1994): 631-53.

    see also:

  5. The 1/4th power comes from Greene's Econometric Analysis (5th Edition) book, where he talks about Newey-West and I'm pretty sure he's referring to lag length. But I have a hard time staying with his notation and therefore get confused with the derivations. I don’t have the time (or desire) to go through the chapter and re-derive his estimators.

    The only reason I suspect AIC/BIC won’t work is because Greene doesn’t mention it and therefore, I think there may be a more robust criterion. I’m not too familiar with DIC, but I can look it into it. Thanks.
  6. sjfan


    I hate to be a jerk and give you citations, but it's been a very long time since I looked over Newey-West and I don't want to give you the wrong information. These two, I think, are the original papers (which means, of course, they will completely suck - but at least you can scan it quickly for 1/4 - or just read the abstract like everyone else):

    Newey WK & West KD (1987), A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix. Econometrica, 55, 703–708.
  7. Thanks for the replies. The number of lags is dependent on the process (moving average or autoregressive) and the order of autocorrelation that is significant. Generally, T^1/4 is the rule of thumb and experimentation is a good way to decide at which level lags can be ignored - where T is the # of observations. Newey and West extend White’s (1980) approach and provide great detail, which I am not interested in. For my purposes, I will be using the rule of thumb. Thanks again for the replies and interest.