2.1 Monetary policy and growth

The previous literature approaches the impact of monetary policy on growth from two fundamentally distinct perspectives. One strand, the literature on monetary neutrality, examines whether expansionary and contractionary monetary policy have persistent effects on output. The other strand focuses on institutional determinants of growth, analyzing how central bank features such as inflation targeting, central bank independence, and related governance arrangements shape economic performance. Both strands report mixed empirical findings.

Both King and Watson (Reference King and Watson1997) and Bernanke and Mihov (Reference Bernanke and Mihov1998) find that US postwar data is consistent with the neutrality – but not the superneutrality—of money. Yet Mankiw (Reference Mankiw2001) interprets the very findings of Bernanke and Mihov (Reference Bernanke and Mihov1998) as evidence of non-neutrality. More recently, Jordà et al. (Reference Jordà, Singh and Taylor2024) find that expansionary policy is neutral, whereas contractions leave permanent scars that reduce output. Serletis and Xu (Reference Serletis and Xu2024) emphasize that, for most countries, the data does not even permit meaningful tests of neutrality, but they confirm neutrality where such tests are feasible. Serletis and Koustas (Reference Serletis and Koustas1998) test the neutrality and superneutrality of money for ten developed countries and find neutrality in all cases and superneutrality in all but Italy. Italy stands out as the only country in their sample experiencing a persistent shock to money growth. This link between persistent money-growth shocks (and thus persistent changes to inflation) and the failure of superneutrality connects their findings to the broader literature on the growth effects of inflation (Bick, Reference Bick2010; López-Villavicencio and Mignon, Reference López-Villavicencio and Mignon2011), and, by extension, to the literature on how central bank institutions influence growth through their role in maintaining price stability.

Much of this work focuses on inflation targeting and central bank independence. For inflation targeting, some studies find that it can support long-run growth by anchoring expectations and reducing macroeconomic volatility (Mishkin, Reference Mishkin2007; Svensson, Reference Svensson2010), whereas others report little or no effect (Ball and Sheridan, Reference Ball and Sheridan2004; Brito and Bystedt, Reference Brito and Bystedt2010; Junankar and Wong, Reference Junankar and Wong2020). The evidence on central bank independence is similarly mixed. While some research argues that greater independence strengthens credibility and fosters growth (Akinci et al. Reference Akinci, Akinci and Yilmaz2015; Garriga, Reference Garriga2016), other studies find no systematic growth benefits (Neyapti, Reference Neyapti2001; de Mendonça, Reference de Mendonça2006; Anastasiou, Reference Anastasiou2009). A related, largely theoretical literature addresses the more fundamental question of which monetary policy framework—understood broadly as the rule or target variable guiding the conduct of policy—delivers the highest economic growth (see, e.g., Lai and Chin, Reference Lai and Chin2013; Chu and Ji, Reference Chu and Ji2016; Huang et al. Reference Huang, Yang and Cheng2017; Chu et al. Reference Chu, Ning and Zhu2019; Gil and Iglésias, Reference Gil and Iglésias2020).

By examining how the operational features of an IT framework shape growth, our paper bridges the two strands of literature on policy actions and institutional or rule-based frameworks. Concretely, we ask how a hawkish versus a dovish stance of the central bank within an IT framework affects growth. Empirical work on this aspect is still scarce. The paper closest to our question is Gil et al. (Reference Gil, Iglésias and Guimaraes2023), who provide theoretical support for the idea that strong policy responses to inflation—which are consistent with the Taylor principle—anchor expectations and create conditions that contribute to growth.

2.2 Time-varying monetary policy reaction functions

Our identification strategy exploits time variation in the Taylor rule. This allows us to estimate the effect of changes in Taylor-rule parameter—particularly the long-run inflation coefficient—on growth, while accounting for time-invariant country heterogeneity through fixed effects. Introducing time variation into estimated versions of the original backward-looking Taylor (Reference Taylor1993) specification is conceptually straightforward (see, e.g., Judd et al. Reference Judd and Rudebusch1998; Assenmacher-Wesche, Reference Assenmacher-Wesche2006; Kuzin, Reference Kuzin2006). However, since the seminal work of Clarida et al. (Reference Clarida, Galí and Gertler1998), it has become widely accepted that central banks typically respond to expected inflation and output rather than to current conditions. Future realizations—often used as proxies for expectations under the rational-expectations assumption—are endogenous to monetary policy by construction, since policy is explicitly designed to influence those future outcomes. Orphanides (Reference Orphanides2001) proposes using real-time forecasts, such as Greenbook or SPF projections, to address this endogeneity concern, since these forecasts are conditioned only on information available at the time and are therefore exogenous to contemporaneous policy shocks. Yet such forecasts are often unavailable, meaning that estimating a forward-looking policy rule requires an instrumental-variables approach that complicates time-varying-parameter estimation. Wesche (Reference Wesche2003) proxies inflation expectations using a VAR-based forecast in her Markov-switching framework, and, in a similar spirit, Trecroci and Vassalli (Reference Trecroci and Vassalli2010) use trend inflation—derived from a simple filtering procedure—as a measure of long-run expectations in his time-varying-parameter model. Both approaches amount to ad hoc modifications of a standard two-stage IV procedure to accommodate evolving monetary-policy parameters. A two-step procedure for estimating bias-corrected, forward-looking, time-varying Taylor rules was introduced by Kim and Nelson (Reference Kim and Nelson2006) and has been applied in various empirical settings, including the Brazilian case examined by Carvalho and Muinhos (Reference Carvalho and Muinhos2023). An alternative approach is the GMM framework proposed by Partouche (Reference Partouche2007), which Anderl and Caporale (Reference Anderl and Caporale2024) apply in their ten-country study. A much larger literature examines time variation in monetary policy using VAR-based approaches (see, e.g., Boivin and Giannoni, Reference Boivin and Giannoni2006; Cogley and Sargent, Reference Cogley and Sargent2005; Primiceri, Reference Primiceri2005; Sims and Zha, Reference Sims and Zha2006). However, these approaches do not permit identification of the policy parameters, rendering them unsuitable for addressing our research question.

In this paper, we employ the framework of Kim and Nelson (Reference Kim and Nelson2006) to model gradual shifts in the reaction function of 37 inflation-targeting countries.

3.1 Panel growth regression model

The model. We augment a standard panel growth regression by including a time-varying measure of policy hawkishness as our main explanatory variable.

Following the usual convention in the growth literature, we construct non-overlapping five-year periods and estimate our model on this quinquennial panel (see, e.g., Barro et al. Reference Barro2003; Bruns and Ioannidis, Reference Bruns and Ioannidis2020):

(1) \begin{equation} \Delta y_{i,t}= \delta _0 + \delta _1 y_{i,t - 1} + \delta _2 \text{Hawk}_{i,t-1} + \phi X_{i,t} + u_i + v_{t} + \varepsilon _{i,t}, \end{equation}

Here, $y_{i,t-1}$ is the log of real GDP per capita at the onset of the period, and $\text{Hawk}_{i,t-1}$ is an indicator for whether country $i$ exhibited a hawkish policy stance at that time. The vector $X_{i,t}$ contains standard growth determinants, including population growth, investment, government consumption (as a share of GDP), education, and trade openness, all measured at the end of period $t-1$ , corresponding to the beginning of the period $t$ . Country and time fixed effects are captured by $u_i$ and $v_t$ , and $\varepsilon _{i,t}$ is the idiosyncratic error term.

We define hawkishness as a strong response to inflation. In the public debate, this is often reduced to a preference for higher interest rates. This association arises because hawkishness attracts most attention when the trade-off between price stability and growth is most acute—that is, when combating inflation risks undermining economic expansion. However, hawkishness—in the sense of fighting inflation aggressively—is not necessarily associated with higher interest rates. If anything, low equilibrium inflation implies a low equilibrium nominal interest rate, and price stability may further reduce the real interest rate by compressing the average risk premium. Indeed, Figure A.1 shows no discernible correlation between real interest rates and hawkishness at moderate levels—if anything, it is extreme dovishness that tends to coincide with elevated rates. At the same time, expansionary policy offers only transitory support through the Phillips curve—once expectations adjust, the effect dissipates. Adherence to the Taylor principle, by contrast, may deliver permanent price stability. Since high and, more importantly, volatile inflation reduces allocative efficiency, hawkishness in our sense may well be beneficial for long-run growth.

Estimation. The specification in Equation (1) corresponds to an autoregressive process in the log level of GDP and can be rewritten equivalently as:

(2) \begin{equation} y_{i,t}= \delta _0 + (\delta _1+1) y_{i,t - 1} + \delta _2 \text{Hawk}_{i,t-1} + \phi X_{i,t} + u_i + v_{t} + \varepsilon _{i,t}, \end{equation}

Dynamic panel models such as Equation (2) require estimators that address the endogeneity introduced by the inclusion of the lagged dependent variable. A widely used approach in the growth literature is the system-GMM estimator of Arellano and Bover (Reference Arellano and Bover1995) and Blundell and Bond (Reference Blundell and Bond1998), which leverages additional moment conditions to enhance efficiency, especially when the variable of interest exhibits high persistence.

However, the validity of the assumptions required for system GMM when applied to growth regressions has been questioned, particularly the stationarity requirement for the level equation (Moral-Benito, Reference Moral-Benito2013, Reference Moral-Benito2016). Including country and time fixed effects can absorb much of the non-stationarity, as the stochastic trend in output is largely driven by global shocks to total factor productivity. Indeed, the autoregressive coefficient in Equation (2) is well below unity for almost all estimators we consider, indicating that stationarity is not a major concern. Yet to ensure robustness to the Moral-Benito critique, we rely on several estimation strategies rather than a single method.

First, we consider a standard fixed-effects estimator, which merely captures unobserved heterogeneity through fixed effects and leaves other identification issues unaddressed. Second, we use system-GMM to address potential endogeneity. Third, we employ the difference-GMM estimator of Arellano and Bond (Reference Arellano and Bond1991) as a robustness check against possible violations of the system-GMM moment assumptions. Finally, we use the quasi–maximum-likelihood estimator proposed by Kripfganz (Reference Kripfganz2016), which is asymptotically equivalent to GMM but offers improved finite-sample properties, particularly in the presence of endogenous regressors.

All results are reported in terms of Equation (2), in other words, the tables report $\hat {\delta }_1+1$ .

Sample and data. Our sample begins in 2000 and ends in 2020, the close of the most recent five-year block. We include all IT countries listed by Central Bank News (2022) for which sufficiently long time series are available. Most of the data for the growth model come from the World Bank’s World Development Indicators, with the exception of education, for which we use the latest vintage of the dataset originally introduced by Barro and Lee (Reference Barro and Lee2013). Details on variable definitions, transformations, and data sources are provided in Table A.1 and a list of IT countries used in this study in Table A.5 in the Appendix.

3.2 Identification of hawkishness

Our indicators of hawkishness and dovishness are based on a forward-looking specification of the Taylor rule with interest-rate smoothing. The central bank’s desired interest rate $r^{*}_{t}$ at time $t$ is given by:

(3) \begin{equation} r^{*}_{t} = b^{*}_{0,t} + b_{1,t}(E_{t}(\pi _{t,K}) - \pi ^{*}_{t}) + b_{2,t} E_{t}(y_{t,K}), \end{equation}

However, in practice the policy rate is typically only gradually adjusted toward the rule-implied rate $r^*$ , to avoid financial market volatility, signal policy consistency, and preserve room to avoid hitting the zero lower bound Rudebusch (Reference Rudebusch2002). This yields the policy reaction function,

(4) \begin{equation} r_{t} = (1 - a_{t}) r^{*}_{t} + a_{t} r_{t-1} + \mu _{t}. \end{equation}

Combining Equation (3) with Equation (4) implies that $b_{1,t}$ determines the long-run response of the policy rate to inflation, while the short-run response is $(1 - a_t)b_{1,t}$ . Throughout the remainder of the paper, the (policy) response to inflation or inflation coefficient refers to the long-run response $b_{1,t}$ .

Rather than imposing an exogenously determined threshold, we classify a country as hawkish at time $t$ if its long-run response to inflation, $b_{1,t}$ , exceeds the sample median.

Our approach follows Kim and Nelson (Reference Kim and Nelson2006). Their two-step estimation strategy addresses two problems inherent to time-varying forward-looking models.

First, they address endogeneity by rewriting the Taylor rule in terms of future realizations—rather than expectations—and adding a bias-correction term.

When substituting Equation (3) into (4) and replacing expectations with future realizations, we obtain:

(5) \begin{equation} r_{t} = (1 - a_{t})(b_{0,t} + b_{1,t}\pi _{t,K} + b_{2,t} y_{t,K}) + a_{t} r_{t-1} + e_{t}, \end{equation}

where

(6) \begin{equation} e_{t} = -(1 - a_{t})[ b_{1,t}(\pi _{t,K} - E_{t}(\pi _{t,K})) + b_{2,t}(y_{t,K} - E_{t}(y_{t,K})) ] + \mu _{t}. \end{equation}

The error term $e_t$ is correlated with the regressors $\pi _{t,K}$ and $y_{t,K}$ by construction, because both enter Expression (6). Recognizing that endogeneity arises from the forecast errors in inflation and the output gap, Kim and Nelson (Reference Kim and Nelson2006) recommend augmenting Equation (5) with orthogonalized prediction errors for $\pi _{t,K}$ and $y_{t,K}$ generated by an auxiliary model.

This yields the following expression:

(7) \begin{equation} r_t = (1 - a_t)(b_{0,t} + b_{1,t}\pi _{t,K} + b_{2,t}y_{t,K}) + a_t r_{t-1} + \rho _1\sigma _{e,t} v^*_{1t} + \rho _2\sigma _{e,t} v^*_{2t} + \omega _t, \end{equation}

where $v^*_{1t}$ and $v^*_{2t}$ are the orthogonalized forecast errors and $\omega _t$ is the now independent error term.

Second, although the interest-rate-smoothing parameter generally falls between $0$ and $1$ in constant-parameter specifications, TVP models can misread increases in the policy rate as explosive dynamics, yielding estimates of $a_t$ above one. In this case, the long-run policy coefficients in Equation (5) are no longer well identified

By defining $a_t$ as

(8) \begin{equation} a_t = \frac {1}{1 + \exp ({-}b_{3,t})} \end{equation}

and modeling $b_{3,t}$ rather than $a_t$ itself as random-walk process addresses this issue.

However, the resulting model is no longer linear. To obtain a tractable representation, Kim and Nelson (Reference Kim and Nelson2006) apply a first-order Taylor expansion to approximate the model linearly. The resulting linear specification, including the bias-correction terms described above, can then be estimated as a TVP model using a standard Kalman filter. In addition to the time-varying Taylor-rule parameters, their model also allows for heteroskedasticity modeled as stochastic volatility.

All Taylor-rules are estimated using quarterly data for the CPI and real GDP, where available, and unemployment otherwise. The data is sourced from the IMF and the Federal Reserve Economic Data FRED, hosted by the Fed St. Louis, and is seasonally adjusted using the X-13 ARIMA-SEATS. The output gap is computed using the Hodrick-Prescott filter. Where possible we use the policy rate, which is otherwise substituted by the lending rate. Where data availability allows, the samples used to estimate Taylor rules start in 1990.