Skip to main content

Cookies on Macro by Mark.We use essential cookies for sign-in, locale, and security. With your permission, we also use optional cookies for preferences, for analytics, and for advertising measurement with the Meta (Facebook) Pixel. Read our cookie policy.

Macro by MarkMacro by Mark
  • Home
Loading…
Loading…
← Models
MacroConceptsPolicyModels

Smooth transition regression Model

Regression with gradual regime weights, allowing coefficients to transition smoothly as a threshold variable changes.

OverviewGraphMathReferences

Understand the mathematics

Follow the same equations and assumptions one section at a time or in a continuous reading.

Hover or focus a symbol defined in this model to find its exact matches. Select it to keep the highlight; Escape clears the selection.

Model Specification and Transition Functions

The STR model augments a linear regression with a nonlinear component governed by a transition function. We present both the logistic (LSTAR) and exponential (ESTAR) forms and show how the linear model and the threshold model are nested as special cases.

The transition function is the single source of nonlinearity. Everything else - regressors, innovations, estimation - follows standard regression mechanics.

yt=ϕ′xt+θ′xt⋅G(st;γ,c)+ut,ut∼iid(0,σ2)y_t = \phi^\prime x_t + \theta^\prime x_t \cdot G(s_t; \gamma, c) + u_t, \quad u_t \sim \text{iid}(0, \sigma^2)yt​=ϕ′xt​+θ′xt​⋅G(st​;γ,c)+ut​,ut​∼iid(0,σ2)
G(st;γ,c)G(s_t; \gamma, c)G(st​;γ,c)σ2\sigma^2σ2

The effective coefficient at time ttt is ϕ+θG(st)\phi + \theta G(s_t)ϕ+θG(st​). When G=0G = 0G=0, only ϕ\phiϕ matters; when G=1G = 1G=1, the coefficient is ϕ+θ\phi + \thetaϕ+θ.

GL(st;γ,c)=11+exp⁡(−γ(st−c)),γ>0G^L(s_t; \gamma, c) = \frac{1}{1 + \exp\bigl(-\gamma(s_t - c)\bigr)}, \quad \gamma > 0GL(st​;γ,c)=1+exp(−γ(st​−c))1​,γ>0
sts_tst​γ\gammaγccc

The logistic form: GGG approaches 0 as st→−∞s_t \to -\inftyst​→−∞ and 1 as st→+∞s_t \to +\inftyst​→+∞. Asymmetric: behavior differs above and below ccc.

GE(st;γ,c)=1−exp⁡(−γ(st−c)2),γ>0G^E(s_t; \gamma, c) = 1 - \exp\bigl(-\gamma(s_t - c)^2\bigr), \quad \gamma > 0GE(st​;γ,c)=1−exp(−γ(st​−c)2),γ>0
sts_tst​γ\gammaγccc

The exponential form: behavior at the center (near c) differs from behavior in the tails (far from c in either direction). Symmetric nonlinearity.

lim⁡γ→0GL(st;γ,c)=0.5⇒yt=(ϕ+0.5θ)′xt+ut\lim_{\gamma \to 0} G^L(s_t; \gamma, c) = 0.5 \quad \Rightarrow \quad y_t = (\phi + 0.5\theta)^\prime x_t + u_tγ→0lim​GL(st​;γ,c)=0.5⇒yt​=(ϕ+0.5θ)′xt​+ut​
sts_tst​γ\gammaγccc

As γ\gammaγ approaches zero, the transition function flattens to a constant, and the STR degenerates to a linear model with modified coefficients.

lim⁡γ→∞GL(st;γ,c)=1(st>c)⇒Threshold model\lim_{\gamma \to \infty} G^L(s_t; \gamma, c) = \mathbb{1}(s_t > c) \quad \Rightarrow \quad \text{Threshold model}γ→∞lim​GL(st​;γ,c)=1(st​>c)⇒Threshold model
sts_tst​γ\gammaγccc

As γ\gammaγ approaches ∞\infty∞, the logistic function becomes a step function at ccc, and the LSTAR model becomes a Threshold Regression.

Scale dependence of γ

The magnitude of γ\gammaγ depends on the scale of sts_tst​. Dividing γ\gammaγ by the sample standard deviation of sts_tst​ (or standardizing sts_tst​ before estimation) makes γ\gammaγ comparable across variables and datasets. This is standard practice but not always reported.

Transition checks

Regime switch
yt=F1(zt;θ1)1{qt≤c}+F2(zt;θ2)1{qt>c}y_t=F_1(z_t;\theta_1)\mathbf{1}\{q_t\le c\}+F_2(z_t;\theta_2)\mathbf{1}\{q_t>c\}yt​=F1​(zt​;θ1​)1{qt​≤c}+F2​(zt​;θ2​)1{qt​>c}
ccc

The indicator changes the active equation at c. Continuity at the boundary is an extra restriction, not a default property of the model.

Read the cutoff as a structural boundary. The law on one side need not match the law on the other.

Do not connect the two regimes with a smooth transition unless the model estimates one.

Variable timing

Switching variable

Rt=1{qt>c}R_t=\mathbf{1}\{q_t>c\}Rt​=1{qt​>c}
ccc

The state qtq_tqt​ selects the active equation. The curve can kink at ccc even if each side is smooth.

Observable

yt=FRt(zt;θRt)y_t=F_{R_t}(z_t;\theta_{R_t})yt​=FRt​​(zt​;θRt​​)

The reported series follows the equation for the active regime, not a single global law.

Keep going

Read next

Continue with a related concept, model, or example.

Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
Help
Privacy PolicyTerms of UseEthics and Compliance

© 2026 Mark Jayson Nation LLC