大非平衡面板数据的异方差与自相关检验技术咨询

First off, let's break down whether your current approach makes sense for your large unbalanced panel, then refine the code, and walk through how to interpret the results.

1. Is Your Current Test Approach Reasonable? #

Let's unpack each part:

Heteroskedasticity Check with xtgls #

Using xtgls with panels(heteroskedastic) and igls is a valid way to correct for panel-level heteroskedasticity (it gives ML-consistent estimates). But a heads-up: this is more of a correction method than a formal test for heteroskedasticity. If your goal is to first confirm whether heteroskedasticity exists, you should add explicit tests before jumping to GLS.

Autocorrelation Check with xtserial #

xtserial is a standard, reliable test for first-order autocorrelation in fixed-effects panel models, and it works perfectly well with unbalanced panels. That part is solid—just note that it assumes no cross-sectional dependence (if your data has that, you’ll need extra checks, which I’ll mention later).

2. Code Optimization Suggestions #

Let’s tweak your code to be more robust, efficient, and aligned with best practices:

Step 1: Add Explicit Heteroskedasticity Tests #

Before diving into GLS, run formal tests to confirm heteroskedasticity. For panel data, two common options are:

• Modified Wald test (for group-wise heteroskedasticity, paired with fixed effects):

  xtreg adjusted_volume ibn.rounded_time i.TRD_EVENT_DT, fe
  estat hettest
  

  Note: I dropped i.id here because the fe option already absorbs individual fixed effects—including it is redundant.
• Breusch-Pagan test (if you prefer a random-effects baseline):

  xtreg adjusted_volume ibn.rounded_time i.id i.TRD_EVENT_DT, re
  estat hettest
  

Step 2: Streamline xtserial Installation #

Your current install steps are redundant. Use this one-liner to grab the latest version of xtserial:

ssc install xtserial, replace

This is the standard way to install user-written commands from Stata’s official module repository, and replace ensures you overwrite any outdated versions.

Step 3: A Complete, Optimized Workflow #

Here’s a cohesive pipeline that starts with baseline models, runs tests, and applies corrections as needed:

* First, make sure your data is properly panel-structured
xtset id TRD_EVENT_DT

* Baseline Fixed Effects Model (absorbs individual fixed effects automatically)
xtreg adjusted_volume ibn.rounded_time i.TRD_EVENT_DT, fe
estimates store fe_baseline

* Test for heteroskedasticity
estat hettest

* Test for first-order autocorrelation
ssc install xtserial, replace
xtserial adjusted_volume ibn.rounded_time i.TRD_EVENT_DT

* If heteroskedasticity is confirmed: run iterated GLS with heteroskedastic panels
xtgls adjusted_volume ibn.rounded_time i.id i.TRD_EVENT_DT, igls panels(heteroskedastic)
estimates store hetero_corrected

* If both heteroskedasticity AND autocorrelation exist: correct for both with AR(1) panels
xtgls adjusted_volume ibn.rounded_time i.id i.TRD_EVENT_DT, igls panels(correlated)
estimates store hetero_ar1_corrected

Step 4: Handle Unbalanced Panels #

Good news: both xtserial and xtgls automatically handle unbalanced data, as long as your xtset is correctly defined (panel ID + time variable). No extra hoops needed here.

3. How to Interpret the Results #

Heteroskedasticity Test Results #

• Modified Wald/Breusch-Pagan tests: If the p-value is below your significance level (e.g., 0.05), reject the null hypothesis of constant variance across panels—this means heteroskedasticity is present and biasing your baseline standard errors.
• xtgls output: Compare the standard errors from your corrected GLS model to the baseline FE/RE model. If they’re noticeably larger or smaller, that’s a clear sign heteroskedasticity was distorting your original results.

Autocorrelation Test Results (xtserial) #

• The test outputs a Lagrange Multiplier (LM) statistic and p-value. A small p-value (<0.05) means there’s significant first-order autocorrelation in your idiosyncratic errors.
• If autocorrelation exists, using xtgls with panels(correlated) will correct for both it and heteroskedasticity (it assumes AR(1) correlation within each panel).

Extra Consideration: Cross-Sectional Dependence #

If your data has cross-sectional correlation (e.g., firms in the same industry moving together), xtserial and xtgls won’t handle this. Test for it with:

xtcsd, pesaran

If significant, use Driscoll-Kraay standard errors (xtscc) instead of GLS—this method is robust to heteroskedasticity, autocorrelation, and cross-sectional dependence.

内容的提问来源于stack exchange，提问作者Amin Karimi