Skip to content

Specification Tests

Specification tests answer the most fundamental question in econometric modeling: is the model correctly specified? A misspecified model produces biased coefficients and misleading inference, regardless of how sophisticated the standard errors are.

What Specification Tests Check

Specification tests evaluate three aspects of model adequacy:

  1. Model type: Should you use Fixed Effects or Random Effects? (Hausman, Mundlak)
  2. Functional form: Is the linear specification appropriate? (RESET)
  3. Parameter stability: Do coefficients change over time? (Chow)
  4. Model comparison: Which of two competing specifications fits better? (J-Test, Cox, Encompassing)

Quick Comparison

Test Hâ‚€ Tests For Requires
Hausman RE is consistent FE vs RE choice FE and RE results
Mundlak RE is consistent Correlated random effects RE results only
RESET Correct functional form Omitted nonlinearities Any model results
Chow No structural break Parameter stability Break point
J-Test Model 1 is correct Non-nested model comparison Two model results
Cox / Encompassing Model 1 encompasses Model 2 Relative model adequacy Two model results

Follow this sequence when specifying a panel model:

Step 1: Hausman / Mundlak
    Choose between FE and RE
        |
Step 2: RESET
    Check functional form of chosen model
        |
Step 3: Chow (if applicable)
    Test for structural breaks
        |
Step 4: J-Test / Cox (if applicable)
    Compare alternative specifications

Start with Hausman

The Hausman test (or its Mundlak alternative) should be your first specification test. It determines the fundamental modeling approach -- whether unobserved heterogeneity is correlated with regressors.

Running All Specification Tests

Use the ValidationSuite to run applicable specification tests automatically:

from panelbox.validation import ValidationSuite

suite = ValidationSuite(results)
spec_results = suite.run_specification_tests(alpha=0.05)

for test_name, result in spec_results.items():
    print(f"{test_name}: p={result.pvalue:.4f} -- {result.conclusion}")

Hausman Requires Both Models

The ValidationSuite cannot run the Hausman test automatically because it requires both FE and RE results. Run it separately: 

from panelbox.validation.specification.hausman import HausmanTest
hausman = HausmanTest(fe_results, re_results)
result = hausman.run()

Manual Specification Testing

For full control over test parameters:

from panelbox.validation.specification.hausman import HausmanTest
from panelbox.validation.specification.mundlak import MundlakTest
from panelbox.validation.specification.reset import RESETTest
from panelbox.validation.specification.chow import ChowTest

# 1. FE vs RE
hausman = HausmanTest(fe_results, re_results, alpha=0.05)
print(hausman.summary())

# 2. Alternative: Mundlak test (RE results only)
mundlak = MundlakTest(re_results)
mundlak_result = mundlak.run(alpha=0.05)
print(mundlak_result.summary())

# 3. Functional form
reset = RESETTest(chosen_results)
reset_result = reset.run(alpha=0.05, powers=[2, 3])
print(reset_result.summary())

# 4. Structural break
chow = ChowTest(chosen_results)
chow_result = chow.run(alpha=0.05, break_point=2008)
print(chow_result.summary())

Non-Nested Model Comparison

When comparing models with different regressors:

from panelbox.diagnostics.specification.davidson_mackinnon import j_test
from panelbox.diagnostics.specification.encompassing import cox_test

# J-Test: does one model's predictions improve the other?
jtest_result = j_test(
    result1, result2,
    direction="both",
    model1_name="Baseline",
    model2_name="Extended"
)
print(jtest_result.interpretation())

# Cox test: likelihood-based comparison
cox_result = cox_test(
    result1, result2,
    model1_name="Baseline",
    model2_name="Extended"
)
print(cox_result.interpretation())

Nested vs. Non-Nested Models

Understanding whether models are nested determines which test to apply:

Relationship Definition Test
Nested Model 1 is a special case of Model 2 F-test, LR test, Wald test
Non-nested Neither model is a special case of the other J-Test, Cox test
FE vs RE Different assumptions about individual effects Hausman, Mundlak

Examples

Nested: \(y = \beta_1 x_1\) vs \(y = \beta_1 x_1 + \beta_2 x_2\) (set \(\beta_2 = 0\) to get Model 1)

Non-nested: \(y = \beta_1 x_1 + \beta_2 x_2\) vs \(y = \gamma_1 z_1 + \gamma_2 z_2\) (different regressors entirely)

Common Result Interface

All specification tests return objects with these attributes:

result.test_name            # str   -- Name of the test
result.statistic            # float -- Test statistic
result.pvalue               # float -- P-value
result.df                   # int or tuple -- Degrees of freedom
result.reject_null          # bool  -- Whether to reject at alpha
result.conclusion           # str   -- Interpretation text
result.metadata             # dict  -- Test-specific details

What to Do When Tests Reject

Test Rejects Meaning Corrective Action
Hausman RE inconsistent Use Fixed Effects
Mundlak Group means significant Use Fixed Effects
RESET Functional form wrong Add polynomials, interactions, or log transforms
Chow Parameters unstable Split sample or add interaction with time dummy
J-Test (both reject) Neither model adequate Develop new specification

Software Comparison

Test PanelBox Stata R
Hausman HausmanTest(fe, re) hausman fe re plm::phtest(fe, re)
Mundlak MundlakTest(re).run() Manual (add group means) plm::phtest(, method="aux")
RESET RESETTest(res).run() estat ovtest lmtest::resettest()
Chow ChowTest(res).run(break_point=T) Manual strucchange::sctest()
J-Test j_test(r1, r2) Manual lmtest::jtest()
Cox cox_test(r1, r2) Not built-in lmtest::encomptest()
LR likelihood_ratio_test(r1, r2) lrtest lmtest::lrtest()

See Also

References

  • Davidson, R., & MacKinnon, J. G. (1981). Several tests for model specification in the presence of alternative hypotheses. Econometrica, 49(3), 781-793.
  • Hausman, J. A. (1978). Specification tests in econometrics. Econometrica, 46(6), 1251-1271.
  • Mundlak, Y. (1978). On the pooling of time series and cross section data. Econometrica, 46(1), 69-85.
  • Ramsey, J. B. (1969). Tests for specification errors in classical linear least squares regression analysis. Journal of the Royal Statistical Society, Series B, 31(2), 350-371.
  • Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.). MIT Press.