brycewang-stanford/Auto-Empirical-Research-Skills / skills/32-dylantmoore-stata-skill/plugins/stata/skills/stata
stata
VerifiedComprehensive Stata reference for writing correct .do files, data management, econometrics, causal inference, graphics, Mata programming, and 20 community packages (reghdfe, estout, did, rdrobust, etc.). Covers syntax, options, gotchas, and idiomatic patterns. Use this skill whenever the user asks you to write, debug, or explain Stata code.
packages/xtabond2.md
View on GitHub# xtabond2: Dynamic Panel GMM Estimation
## Overview
`xtabond2` implements Arellano-Bond (1991) difference GMM and Arellano-Bover/Blundell-Bond (1995, 1998) system GMM for dynamic panel data. Designed for:
- Small T (time periods), large N (panel units)
- A lagged dependent variable as a regressor
- Independent variables that are not strictly exogenous
- Fixed individual effects correlated with regressors
---
## Installation
```stata
ssc install xtabond2, replace
ssc install ranktest * Required dependency
* Verify installation
webuse abdata, clear
xtabond2 n L.n w k, gmm(L.n) iv(w k) robust
```
---
## The Dynamic Panel Problem
Standard FE estimation with a lagged dependent variable produces **Nickell bias**: the within-transformed lag correlates with the transformed error. Bias is O(1/T), severe when T is small.
**Key insight**: The true parameter on the lagged DV lies between FE (downward biased) and OLS (upward biased). GMM should fall in this range.
```stata
* Demonstrate bias bounds
xtset id t
xtreg y L.y, fe * Biased downward
reg y L.y * Biased upward
xtabond2 y L.y, gmm(L.y, lag(2 3)) robust * Should fall between
```
---
## Difference GMM (Arellano-Bond)
First-differences to remove fixed effects, then instruments with lagged levels:
- y_i,t-2 is uncorrelated with delta-epsilon_it (if epsilon has no serial correlation)
```stata
* Difference GMM
xtabond2 n L.n w k, ///
gmm(L.n, lag(2 .)) ///
iv(w k, equation(diff)) ///
robust
```
**Best for**: Low/moderate persistence (AR parameter < 0.8), small T (3-10 periods).
**Limitations**: Weak instruments when series are highly persistent; loses one time period to differencing.
---
## System GMM (Blundell-Bond)
Combines difference and level equations to address weak instruments in difference GMM:
1. **Difference equation**: instruments = lagged levels
2. **Level equation**: instruments = lagged differences
**Additional assumption**: E[delta-y_i,t-1 * mu_i] = 0 (changes uncorrelated with fixed effect).
```stata
* System GMM (default when no equation(diff) restriction)
xtabond2 n L.n w k, ///
gmm(L.n, lag(1 .) collapse) ///
iv(w k) ///
robust twostep
```
| Feature | Difference GMM | System GMM |
|---------|---------------|------------|
| **Equations** | First differences only | Differences + levels |
| **Instruments for levels** | None | Lagged differences |
| **Efficiency** | Lower for persistent series | Higher |
| **Additional assumption** | None | E[delta-y * mu] = 0 |
| **Number of instruments** | Fewer | More (needs collapse) |
The **Difference-in-Hansen** test (reported automatically for system GMM) evaluates the additional level-equation instruments.
---
## Basic Syntax and Usage
```stata
xtabond2 depvar [indepvars] [if] [in] [weight], ///
gmm(varlist [, lag(start end) collapse orthogonal equation()]) ///
[iv(varlist [, equation(diff|level|both)])] ///
[onestep | twostep] [robust] [small] [orthogonal]
```
### Variable Treatment
**Strictly exogenous** (use `iv()`):
```stata
xtabond2 y L.y x1 x2, gmm(L.y) iv(x1 x2) robust
```
**Predetermined** -- may respond to past shocks but not current (use `gmm()` with `lag(1 .)`):
```stata
xtabond2 y L.y x, gmm(L.y x, lag(1 .)) robust
```
**Endogenous** -- may correlate with current shocks (use `gmm()` with `lag(2 .)`):
```stata
xtabond2 y L.y x, gmm(L.y x, lag(2 .)) robust
```
### Multiple Variables with Different Treatments
```stata
xtabond2 y L.y x1 x2 x3, ///
gmm(L.y x3, lag(2 .)) /// /* Endogenous */
gmm(x2, lag(1 .)) /// /* Predetermined */
iv(x1) /// /* Strictly exogenous */
robust twostep
```
---
## Instrument Creation
### The lag() Option
```stata
gmm(L.y, lag(2 .)) * Lags 2 through all available
gmm(L.y, lag(2 3)) * Only lags 2 and 3
gmm(L.y, lag(1 4)) * Lags 1 through 4
```
Without collapse, instrument count grows quadratically with T (one column per variable per time period per lag).
### The equation() Suboption
```stata
gmm(L.y, lag(2 .) equation(diff)) * First-differences equation only
gmm(L.y, lag(2 .) equation(level)) * Levels equation only
gmm(L.y, lag(2 .) equation(both)) * Both (system GMM default)
iv(x, equation(diff)) * Standard IV for diff equation only
iv(x) * Default is equation(both)
```
### Mixed Instrument Example
```stata
xtabond2 n L.n L(0/1).w L(0/1).k yr1980-yr1984, ///
gmm(L.n, lag(2 4)) /// /* Lags 2-4 of employment */
gmm(L.(w k), lag(1 3)) /// /* Lags 1-3 of wages & capital */
iv(yr1980-yr1984) /// /* Year dummies strictly exogenous */
robust twostep
```
---
## Lag Limits and the Collapse Option
### The Instrument Proliferation Problem
Without collapse: # instruments = T(T-1)/2 per GMM variable. With T=10 and 3 GMM variables, that is 135 instruments.
**Problems**: Overfitting, weak Hansen test, finite sample bias toward OLS.
**Rule of thumb: Keep instruments <= number of groups.**
```stata
* Check after estimation
display e(j) // Number of instruments
display e(N_g) // Number of groups
assert e(j) <= e(N_g)
```
### Limiting Lag Depth
```stata
gmm(L.y, lag(2 3)) * Only lags 2 and 3 (instead of lag(2 .))
```
### The collapse Option
Creates one instrument per variable per lag instead of one per variable per time period per lag.
```stata
gmm(varlist, lag(start end) collapse)
```
```stata
* Without collapse (many instruments, likely > 100)
xtabond2 n L.n w k, ///
gmm(L.n, lag(2 .)) gmm(L.(w k), lag(1 .)) ///
iv(yr1980-yr1984) robust
* With collapse (likely 20-40 instruments)
xtabond2 n L.n w k, ///
gmm(L.n, lag(2 .) collapse) gmm(L.(w k), lag(1 .) collapse) ///
iv(yr1980-yr1984) robust twostep
```
**Use collapse when**: instruments approach/exceed groups, Hansen p > 0.90.
**Optimal strategy**: Always use collapse for system GMM, limit lags to 2-3, check j < N_g.
---
## Orthogonal Deviations
First differences lose TWO observations per gap in unbalanced panels. Orthogonal deviations (subtract future mean) lose only ONE, preserving more data.
```stata
xtabond2 depvar indepvars, ///
gmm(...) iv(...) ///
orthogonal ///
robust
```
**Standard practice** -- use orthogonal with system GMM:
```stata
xtabond2 y L.y x, ///
gmm(L.y, lag(1 .) collapse) ///
iv(x) ///
robust twostep orthogonal small
```
---
## One-Step vs Two-Step Estimation
**One-step** (default): Fixed weighting matrix. Faster, more robust in finite samples.
**Two-step**: Optimal weighting matrix from first-step residuals. Asymptotically efficient but SEs are downward-biased in small samples -- requires Windmeijer correction via `small`.
**Best practice -- use two-step with Windmeijer correction**:
```stata
xtabond2 y L.y x, ///
gmm(L.y, lag(2 3) collapse) ///
iv(x) ///
robust twostep small
```
Typical pattern: coefficients similar across one-step/two-step; uncorrected two-step SEs are smaller (biased); `small` correction brings them to comparable levels.
---
## Standard Errors
**Always use `robust twostep small`**:
| Type | Syntax | Notes |
|------|--------|-------|
| Conventional | (no options) | Not recommended; assumes homoskedasticity |
| Robust | `robust` | Sandwich estimator; standard in applied work |
| Windmeijer-corrected | `robust twostep small` | Essential for two-step; typically inflates SE by 10-40% |
| Cluster-robust | `robust cluster(group_var)` | Rarely needed; panel structure already handles within-unit correlation |
---
## Diagnostic Tests
### 1. Arellano-Bond AR Tests
Tests serial correlation in **differenced** residuals.
- **AR(1)**: Should be significant (p < 0.05) -- differencing induces MA(1) by construction
- **AR(2)**: Should NOT be significant (p > 0.05) -- validates no serial correlation in level errors
```
Arellano-Bond test for AR(1) in first differences: z = -2.49 Pr > z = 0.013 (expected)
Arellano-Bond test for AR(2) in first differences: z = -0.28 Pr > z = 0.777 (good)
```
**If AR(2) rejects**:
```stata
* Add more lags of DV, start instruments deeper
xtabond2 y L(1/2).y x, gmm(L.y, lag(3 .)) iv(x) robust
```
### 2. Hansen Test (Overidentifying Restrictions)
Tests instrument validity. H0: instruments are exogenous.
- **p > 0.10**: Cannot reject -- instruments appear valid
- **p < 0.10**: Instruments may be invalid
- **p > 0.90**: Test weakened by too many instruments -- reduce them
**Ideal range: 0.10 < p < 0.25**
**Hansen vs Sargan**: Sargan is not robust to heteroskedasticity but not weakened by many instruments. Hansen is robust but weakened. Focus on Hansen when using `robust`.
### 3. Difference-in-Hansen Test
Tests subsets of instruments. Reported automatically for system GMM level instruments.
```stata
* Manual difference-in-Hansen
quietly xtabond2 n L.n w k, gmm(L.(n w)) iv(k) robust twostep small
scalar H_full = e(hansen)
scalar df_full = e(hansendf)
quietly xtabond2 n L.n w k, gmm(L.n) iv(w k) robust twostep small
scalar diff_H = H_full - e(hansen)
scalar diff_df = df_full - e(hansendf)
display "P-value: " chi2tail(diff_df, diff_H)
```
### 4. Coefficient Bounds Check
GMM estimate on lagged DV should fall between FE (lower) and OLS (upper):
```stata
quietly reg n L.n w k
scalar beta_ols = _b[L.n]
quietly xtreg n L.n w k, fe
scalar beta_fe = _b[L.n]
quietly xtabond2 n L.n w k, gmm(L.n, lag(2 3)) iv(w k) robust twostep small
assert beta_fe < _b[L.n] & _b[L.n] < beta_ols
```
### Complete Diagnostic Checklist
```stata
xtabond2 n L.n w k, gmm(L.n, lag(2 3) collapse) iv(w k) robust twostep small
display "AR(2) p > 0.05: " cond(e(ar2p) > 0.05, "PASS", "FAIL")
display "Hansen p 0.10-0.90: " cond(e(hansenp) > 0.10 & e(hansenp) < 0.90, "PASS", "FAIL")
display "Instruments < Groups: " cond(e(j) < e(N_g), "PASS", "FAIL")
```
---
## Complete Dynamic Panel Workflow
### Steps 1-3: Data Prep and Baselines
```stata
webuse abdata, clear
xtset id year
xtsum n w k
* Pooled OLS (upward biased)
reg n L.n w k
estimates store ols
scalar beta_ols = _b[L.n]
* Fixed effects (downward biased)
xtreg n L.n w k, fe vce(cluster id)
estimates store fe
scalar beta_fe = _b[L.n]
```
### Steps 4-5: GMM Estimation
```stata
* Difference GMM
xtabond2 n L.n w k, ///
gmm(L.n, lag(2 3)) iv(w k, equation(diff)) ///
robust twostep small
estimates store diff_gmm
* System GMM (preferred)
xtabond2 n L.n w k yr1980-yr1984, ///
gmm(L.n, lag(1 2) collapse) iv(w k yr1980-yr1984) ///
robust twostep orthogonal small
estimates store sys_gmm
* Test time effects
test yr1980 yr1981 yr1982 yr1983 yr1984
```
### Step 6: Robustness
```stata
* Sensitivity to lag structure
forvalues maxlag = 2/4 {
quietly xtabond2 n L.n w k, ///
gmm(L.n, lag(2 `maxlag') collapse) iv(w k) ///
robust twostep small
display "Max lag = `maxlag': coef=" %6.4f _b[L.n] ///
" AR2p=" %6.4f e(ar2p) " Hansenp=" %6.4f e(hansenp) " j=" e(j)
}
* Treat wages as predetermined
xtabond2 n L.n w k, ///
gmm(L.n, lag(2 3) collapse) gmm(w, lag(1 2) collapse) ///
iv(k) robust twostep small
```
### Step 7: Postestimation
```stata
* Long-run effects
nlcom (lr_wage: _b[w] / (1 - _b[L.n])) ///
(lr_capital: _b[k] / (1 - _b[L.n]))
* Export results
esttab ols fe diff_gmm sys_gmm using "results.tex", replace ///
b(%9.4f) se(%9.4f) star(* 0.10 ** 0.05 *** 0.01) ///
stats(N N_g ar2p hansenp j, ///
labels("Observations" "Firms" "AR(2) p-value" "Hansen p-value" "Instruments")) ///
mtitles("OLS" "FE" "Diff-GMM" "Sys-GMM")
```
---
## Best Practices
### Standard Specification (Use as Default)
```stata
xtabond2 depvar L.depvar indepvars, ///
gmm(L.depvar, lag(1 2) collapse) ///
iv(exog_vars) ///
robust twostep orthogonal small
```
Why: System GMM (efficient), limited lags with collapse (avoids proliferation), orthogonal deviations (handles gaps), Windmeijer correction (valid SEs).
### Instrument Selection
- Keep instruments < groups (`assert e(j) < e(N_g)`)
- Use collapse + limit lag depth to 2-3
- Only GMM-instrument truly endogenous variables; use `iv()` for exogenous ones
### Common Mistakes
**DON'T**: Use `lag(2 .)` without collapse; ignore AR(2) test; omit `robust`; use two-step without `small`; trust Hansen p > 0.90.
**DO**: Always `robust twostep small`; keep instruments < groups; use `collapse`; limit lag depth; report all diagnostics.
### Sample Size Requirements
- N (groups): At least 50-100
- T (time periods): At least 3-4
- Very large T (> 20): FE may be better; N ~ T: not suitable for dynamic panel GMM
### When NOT to Use xtabond2
- No lagged dependent variable -> use `xtreg, fe`
- Large T relative to N (T > 20, N < 100) -> use FE with cluster-robust SEs
- Very persistent series (AR near 1) -> consider `xtlsdvc` or cointegration methods
### Reporting Requirements
Report: coefficients with robust SEs, N and groups, AR(1)/AR(2) p-values, Hansen p-value, number of instruments, specification details (one/two-step, lags, collapse).
---
## Troubleshooting
### AR(2) Test Rejects
```stata
* Add more lags of DV, start instruments deeper
xtabond2 y L(1/2).y x, gmm(L.y, lag(3 4)) iv(x) robust
```
### Hansen Test Rejects (p < 0.10)
```stata
* Reduce instruments; move suspect exogenous vars to gmm()
xtabond2 y L.y x, gmm(L.y, lag(2 3) collapse) iv(x) robust
xtabond2 y L.y x z, gmm(L.(y x)) iv(z) robust * Treat x as endogenous
```
### Hansen Test Too High (p > 0.90)
```stata
* Reduce instruments dramatically
xtabond2 y L.y x, gmm(L.y, lag(2 2) collapse) iv(x) robust twostep small
* Target j < N_g/2
```
### GMM Outside OLS-FE Bounds
- GMM < FE (weak instruments): Switch to system GMM
- GMM > OLS (invalid instruments): Try difference GMM only with `equation(diff)`
### "Could not calculate robust VCE"
```stata
* Check for collinearity; drop units with zero within-variation
bysort id: egen sd_y = sd(y)
drop if sd_y == 0
```
### Very Large Standard Errors
Use system GMM, reduce instruments, check first-stage relevance:
```stata
reg L.y L2.y L3.y x
test L2.y L3.y * F > 10 suggests reasonable instruments
```
### Insufficient Observations
```stata
* Reduce lags; use orthogonal deviations; drop short panels
xtabond2 y L.y x, gmm(L.y, lag(2 2)) iv(x) robust orthogonal
bysort id: gen T = _N
keep if T >= 4
```
---
## Quick Reference
```stata
* Difference GMM
xtabond2 y L.y x, gmm(L.y, lag(2 3)) iv(x, equation(diff)) robust
* System GMM (recommended)
xtabond2 y L.y x, ///
gmm(L.y, lag(1 2) collapse) iv(x) ///
robust twostep orthogonal small
* Diagnostics
display "AR(2): " e(ar2p)
display "Hansen: " e(hansenp)
display "Instruments: " e(j) "/" e(N_g)
```
### Related Commands
- `xtabond`: Official Stata Arellano-Bond (less flexible)
- `xtdpdsys`: Official Stata system GMM (less flexible)
- `xtdpd`: General dynamic panel data estimator
### Key References
- Roodman (2009). "How to Do xtabond2." *Stata Journal*, 9(1), 86-136.
- Windmeijer (2005). Finite sample correction for two-step GMM. *JoE*, 126(1), 25-51.
```stata
help xtabond2
```
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |