Dynamic Binary Panel

Quick Reference

Class: DynamicBinaryPanel Import: from panelbox.models.discrete.dynamic import DynamicBinaryPanel Stata equivalent: xtprobit y L.y x1, re with Wooldridge initial conditions R equivalent: Custom implementation (no standard package)

Overview

Dynamic binary panel models address a fundamental question in applied economics: does past behavior causally affect current behavior? When we observe persistence in binary outcomes -- individuals who were employed last period tend to be employed this period, firms that exported last year tend to export this year -- this persistence may arise from two distinct sources:

1. True state dependence: past outcomes \(y_{i,t-1}\) have a genuine causal effect on current outcomes \(y_{it}\) (e.g., work experience builds human capital, making future employment more likely)
2. Spurious state dependence: unobserved individual heterogeneity \(\alpha_i\) creates persistence without any causal effect of past behavior (e.g., inherently motivated individuals are always more likely to be employed)

The dynamic binary panel model disentangles these two channels:

\[P(y_{it} = 1 \mid y_{i,t-1}, X_{it}, \alpha_i) = F(\gamma \, y_{i,t-1} + X_{it}'\beta + \alpha_i)\]

where \(\gamma\) captures true state dependence and \(\alpha_i\) captures unobserved heterogeneity. A significant \(\gamma > 0\) provides evidence of genuine state dependence beyond what heterogeneity alone can explain.

Quick Example

import numpy as np
from panelbox.models.discrete.dynamic import DynamicBinaryPanel

model = DynamicBinaryPanel(
    endog=y, exog=X,
    entity=entity, time=time,
    initial_conditions="wooldridge",
    effects="random"
)
result = model.fit()

print(f"State dependence (gamma): {result.gamma:.4f}")
print(f"RE std deviation:         {result.sigma_u:.4f}")
print(result.summary())

When to Use

• Persistence in binary outcomes: employment status, export participation, technology adoption, brand loyalty, poverty traps
• Disentangling causes of persistence: is past behavior truly causal, or is it driven by permanent traits?
• Policy evaluation: if \(\gamma \approx 0\), a temporary intervention has no lasting effect; if \(\gamma > 0\), temporary subsidies can shift long-run behavior
• Dynamic structural models: when the lagged dependent variable is part of the economic model

Key Assumptions

• Correct initial conditions specification: The initial observation \(y_{i0}\) is endogenous to \(\alpha_i\). Ignoring this leads to upward-biased estimates of \(\gamma\).
• Random effects: \(\alpha_i\) is modeled parametrically (normal distribution). If \(\alpha_i\) is correlated with \(X_{it}\), estimates are biased.
• Strict exogeneity of \(X\): Regressors must not be affected by past values of \(y\).
• Binary outcomes only: This model is designed for \(y \in \{0, 1\}\).

The Initial Conditions Problem

The key econometric challenge in dynamic binary panels is the initial conditions problem. Since the model includes \(\alpha_i\) as an unobserved effect and \(y_{i,t-1}\) as a regressor, the initial observation \(y_{i0}\) is correlated with \(\alpha_i\). Simply conditioning on \(y_{i0}\) and treating it as exogenous leads to biased estimates.

PanelBox offers three approaches:

Wooldridge (2005) -- Recommended

Models the distribution of \(\alpha_i\) conditional on the initial observation and time-averages of covariates:

\[\alpha_i = \delta_0 + \delta_1 \, y_{i0} + \delta_2' \bar{X}_i + a_i, \quad a_i \sim N(0, \sigma^2_a)\]

This is the most practical approach and is widely used in applied work.

model = DynamicBinaryPanel(
    endog=y, exog=X, entity=entity, time=time,
    initial_conditions="wooldridge",
    effects="random"
)
result = model.fit()

# Structural parameters
print(f"gamma (lag effect):    {result.gamma:.4f}")
print(f"delta_y0 (initial):    {result.delta_y0:.4f}")
print(f"delta_xbar (X means):  {result.delta_xbar}")

Heckman (1981)

Models the joint distribution of \((y_{i0}, \alpha_i)\) by specifying a separate reduced-form equation for the initial period:

\[P(y_{i0} = 1 \mid Z_i, \alpha_i) = F(Z_i'\pi + \alpha_i)\]

model = DynamicBinaryPanel(
    endog=y, exog=X, entity=entity, time=time,
    initial_conditions="heckman",
    effects="random"
)
result = model.fit()

Simple (Exogenous)

Treats \(y_{i0}\) as exogenous -- drops the first period for each entity and uses \(y_{i1}\) as the initial condition. This is biased when \(\alpha_i\) matters but can serve as a quick baseline.

model = DynamicBinaryPanel(
    endog=y, exog=X, entity=entity, time=time,
    initial_conditions="simple",
    effects="random"
)
result = model.fit()

Which Approach?

The Wooldridge approach is recommended for most applications. It is computationally simpler than Heckman and produces similar estimates in most cases. Use "simple" only as a diagnostic baseline -- if \(\gamma\) changes substantially between "simple" and "wooldridge", the initial conditions matter.

Detailed Guide

Data Preparation

The model requires balanced or unbalanced panel data. Entity and time identifiers are used to construct the lagged dependent variable internally.

import numpy as np
import pandas as pd

# Simulated panel data
n_entities = 500
n_periods = 10
N = n_entities * n_periods

entity = np.repeat(range(n_entities), n_periods)
time = np.tile(range(n_periods), n_entities)
x1 = np.random.normal(0, 1, N)
x2 = np.random.normal(0, 1, N)
X = np.column_stack([x1, x2])

# Binary outcome with state dependence
# (in practice, y comes from your data)
y = np.random.binomial(1, 0.5, N)

Estimation

from panelbox.models.discrete.dynamic import DynamicBinaryPanel

model = DynamicBinaryPanel(
    endog=y,
    exog=X,
    entity=entity,
    time=time,
    initial_conditions="wooldridge",
    effects="random"
)
result = model.fit()
print(result.summary())

Interpreting Results

The key parameters are:

Parameter	Attribute	Interpretation
\(\gamma\)	result.gamma	State dependence: effect of \(y_{i,t-1}\) on \(P(y_{it}=1)\)
\(\beta\)	result.beta	Covariate effects
\(\sigma_u\)	result.sigma_u	Heterogeneity: standard deviation of \(\alpha_i\)
\(\delta_{y0}\)	result.delta_y0	Initial value coefficient (Wooldridge)
\(\delta_{\bar{x}}\)	result.delta_xbar	Time-average coefficients (Wooldridge)

Interpreting \(\gamma\):

• \(\gamma > 0\) and significant: true state dependence -- past behavior causally affects current behavior
• \(\gamma \approx 0\): persistence is entirely due to unobserved heterogeneity
• Large \(\sigma_u\) with small \(\gamma\): most persistence comes from permanent individual traits

Predictions

# Predicted probabilities
probs = result.predict()

# Marginal effects at the mean
me = result.marginal_effects()

Configuration Options

Parameter	Type	Default	Description
endog	ndarray	required	Binary dependent variable (0/1)
exog	ndarray	required	Exogenous covariates
entity	ndarray	required	Entity identifiers
time	ndarray	required	Time identifiers
initial_conditions	str	"wooldridge"	"wooldridge", "heckman", or "simple"
effects	str	"random"	"random" or "pooled"

fit() parameters:

Parameter	Type	Default	Description
start_params	ndarray	None	Starting values (auto-computed if None)

Result Attributes

Attribute	Type	Description
params	ndarray	Full parameter vector
beta	ndarray	Covariate coefficients
gamma	float	Lag coefficient (state dependence)
sigma_u	float	Random effects standard deviation
delta_y0	float	Initial value coefficient (Wooldridge only)
delta_xbar	ndarray	Time-average coefficients (Wooldridge only)
llf	float	Log-likelihood
converged	bool	Convergence flag
n_iter	int	Number of iterations

Diagnostics

Testing for State Dependence

The primary diagnostic is whether \(\gamma\) is significantly different from zero:

result = model.fit()

# Check state dependence
print(f"gamma = {result.gamma:.4f}")
print(result.summary())  # Includes z-statistics and p-values

Sensitivity to Initial Conditions

Compare estimates across different initial conditions specifications:

from panelbox.models.discrete.dynamic import DynamicBinaryPanel

# Wooldridge (preferred)
m1 = DynamicBinaryPanel(endog=y, exog=X, entity=entity, time=time,
                          initial_conditions="wooldridge", effects="random")
r1 = m1.fit()

# Heckman
m2 = DynamicBinaryPanel(endog=y, exog=X, entity=entity, time=time,
                          initial_conditions="heckman", effects="random")
r2 = m2.fit()

# Simple (biased baseline)
m3 = DynamicBinaryPanel(endog=y, exog=X, entity=entity, time=time,
                          initial_conditions="simple", effects="random")
r3 = m3.fit()

print(f"Wooldridge gamma: {r1.gamma:.4f}")
print(f"Heckman gamma:    {r2.gamma:.4f}")
print(f"Simple gamma:     {r3.gamma:.4f} (upward biased)")

Expected Bias Pattern

The "simple" approach typically overestimates \(\gamma\) because it attributes some of the heterogeneity effect to state dependence. If "simple" and "wooldridge" produce similar \(\gamma\), the initial conditions problem is not severe in your data.

Comparing with Static Models

from panelbox.models.discrete.binary import RandomEffectsProbit

# Static RE Probit (no lag)
static = RandomEffectsProbit("y ~ x1 + x2", data, "id", "year")
static_res = static.fit()

# Dynamic (with lag)
dynamic = DynamicBinaryPanel(endog=y, exog=X, entity=entity, time=time,
                              initial_conditions="wooldridge", effects="random")
dynamic_res = dynamic.fit()

print(f"Static sigma_alpha:  {static.sigma_alpha:.4f}")
print(f"Dynamic sigma_u:     {dynamic_res.sigma_u:.4f}")
print(f"Dynamic gamma:       {dynamic_res.gamma:.4f}")
# sigma_u should be smaller than sigma_alpha if state dependence is real

Common Applications

Application	Outcome	State Dependence Interpretation
Labor economics	Employment status	Job experience makes future employment more likely
International trade	Export participation	Sunk costs of entering export markets
Technology adoption	Use of technology	Learning-by-doing, switching costs
Marketing	Brand loyalty	Habit formation, satisfaction feedback
Poverty	Poverty status	Poverty traps, asset depletion

Tutorials

Tutorial	Description	Link
Dynamic Discrete Choice	State dependence analysis with initial conditions

See Also

• Binary Choice Models -- Static Logit and Probit models
• Ordered Choice Models -- Extension to ordinal outcomes
• Marginal Effects -- Computing and interpreting effects
• Dynamic GMM -- Arellano-Bond for continuous dynamic panels

References

• Wooldridge, J. M. (2005). "Simple Solutions to the Initial Conditions Problem in Dynamic, Nonlinear Panel Data Models with Unobserved Heterogeneity." Journal of Applied Econometrics, 20(1), 39-54.
• Heckman, J. J. (1981). "The Incidental Parameters Problem and the Problem of Initial Conditions in Estimating a Discrete Time-Discrete Data Stochastic Process." In Structural Analysis of Discrete Data, ed. C. Manski and D. McFadden. MIT Press.
• Arulampalam, W. and Stewart, M. B. (2009). "Simplified Implementation of the Heckman Estimator of the Dynamic Probit Model and a Comparison with Alternative Estimators." Oxford Bulletin of Economics and Statistics, 71(5), 659-681.
• Stewart, M. B. (2007). "The Interrelated Dynamics of Unemployment and Low-Wage Employment." Journal of Applied Econometrics, 22(3), 511-531.
• Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data. 2^nd ed. MIT Press. Chapter 15.