Inflation targeting in low- income countries: Does IT work?

Previous research on inflation targeting (IT) has focused on high-income countries (HICs) and emerging market economies (EMEs). Only recently has enough data accumulated for the performance of IT in low-income countries (LICs) to be assessed. We show that IT has not so far been effective in reducing inflation in LICs, unlike in EMEs. Weak institutions, a typical feature in LICs, help explain this result, particularly under floating exchange rate regimes. Our interpretation is that poor institutions, leaving fiscal policy unconstrained, impair central banks’ ability to conduct monetary policy in a way consistent with IT.


Introduction
Inflation targeting (IT) was first adopted in 1990 by New Zealand, followed by a number of other high-income countries (HICs) and emerging market economies (EMEs). An IT central bank, which has price stability as its overriding objective, publicly announces a mediumterm numerical target for inflation and commits to it using inflation expectations as an intermediate target. Existing empirical studies suggest that IT has significantly reduced inflation in EMEs, but has made little difference in HICs. Only in the twenty-first century have low-income countries (LICs) begun to adopt IT. To our knowledge this is the first study to analyze the performance of IT in LICs as a separate group. 1 This study adds to the literature by showing that IT effects in LICs are significantly different from in EMEs, and by providing possible explanations behind these heterogeneous effects within non-HICs.
Specifically, using an updated dataset covering up to 182 countries for the 1980-2016 period, we show that IT is not effective in reducing inflation in LICs, unlike in EMEs. With IT being less effective in HICs than in EMEs, the relation between the effectiveness of IT and income levels is non-monotonic. To understand why IT effects are different between LICs and EMEs, we examine the role of institutional quality. In particular, acknowledging that institutions in LICs often fail to make governments accountable to the public, we test if government accountability plays a role in the effect of IT on inflation rates. We find that, within a pooled sample of LICs and EMEs, accountability is negatively associated with the effectiveness of IT. This is the case particularly when exchange rate flexibility is taken into account as a prerequisite for successful IT framework. Our interpretation is that low government accountability tends to be associated with fiscal dominance, i.e., the subordination of monetary policy to fiscal requirements, and impairs central banks' ability to conduct monetary policy in a way consistent with IT.
The rest of the paper is organized as follows. Section 2 provides a background for this study. Section 3 motivates why it is interesting to study IT effects in LICs, and presents 1 How exactly we define LICs is clarified below. testable hypotheses. Section 4 explains the empirical methodology to address the hypotheses.
Section 5 describes the data. Section 6 presents the results. Section 7 offers discussion and concluding remarks.

Background
The theoretical foundation of inflation targeting is rooted in the literature on commitment and discretion in monetary policy propounded by Kydland and Prescott (1977) and Barro and Gordon (1983). The theoretical literature has proposed a number of ways to deal with the inflation bias, which can be classified broadly into: reputational approaches (Barro and Gordon (1983)), delegation to a conservative central banker (Rogoff (1985)), the optimal contracts approach (Walsh (1995)) and inflation targeting (Svensson (1997)). In line with Svensson (1997), the empirical literature has tested whether inflation is significantly lower under IT than without it. This section highlights that little is known about whether IT works in LICs. Table 1 lists the countries with IT experiences, together with their income classes and the adoption dates. We ensure that our income classification reflects income levels over the sample period. Specifically, we classify countries through the following three steps. First, for each of the years when PPP-adjusted GDP per capita are available , countries are sorted into four groups: the highest 25th percentile, 25th-50th, 50th-75th and 75th-100th. 2 Second, counting the number of times each country appears in those four groups over the period, we classify countries that appear in the top 25th percentile most frequently as high income countries; likewise countries appearing in the 25th-50th (50th-75th, 75th-100th) most frequently as upper-middle (lower-middle, low) income countries. Last, we reclassify the four groups into three, by combining the bottom two groups, resulting in HICs, EMEs,and LICs. 3 This yields 11 (14,14) IT adopting countries in LICs (EMEs, HICs). For information, the 2 PPP-adjusted GDP per capita is from World Bank's World Development Indicator (WDI). 3 We take this measure, to ensure that we have sufficient number of IT adopters in LICs.
table also shows the income classification used by the World Bank in 2016, which is based on income levels in 2015 alone.
The last two columns in the table give the year of IT adoption for each country. Following the literature, we consider two alternative dates: strict and loose adoption dates. The difference between these years is that the former corresponds to the time when countries simply announce inflation targets without strong commitment, possibly using other nominal anchors at the same time. The latter, on the other hand, is the year when a strong commitment is made to achieve the target. Those years largely follow Samarina et al. (2014), except that for countries not included in their study, the dates are taken from other sources including respective central bank websites. For some countries such as Israel, Colombia, Chile, Peru and Ghana, the difference between the alternative years is substantial (more than 5 years).
The main message of Table 1 is that IT is a recent phenomenon particularly in LICs.
For example, according to strict IT adoption years, 9 out of 11 LICs adopted IT after 2005 (inclusive), and 5 adopted IT after 2009. Thus, samples used in many of the previous works omit IT-adopting LICs. To illustrate, Table 2 lists several empirical studies on IT, divided into three categories according to their country coverage: both advanced countries (roughly our HICs) and non-advanced countries (EMEs and LICs); only advanced countries; only nonadvanced countries. The recurrent finding is that IT helps reduce inflation in non-advanced economies, but not in advanced economies. Importantly, however, because the time periods covered by many of those studies end in the mid-2000s, little is known about the effects of IT in LICs as a separate group. This paper aims to fill in this gap. 4 4 To note, a few works include LICs in their sample. First, Samarina et al. (2014), using data till 2011, cover a few IT adopters from LICs. However, they highlight the difference in IT effects between advanced economies and others, and do not consider the possible heterogeneity in the effects within non-advanced economies. Next, Gemayel et al. (2011) highlight IT in LICs, defined as countries eligible for the Poverty Reduction and Growth Trust (PGRT), which include Albania, Armenia, and Ghana (we also categorize these countries as LICs). However, due to the fact their data covers only till 2008, they use IT-adopting EMEs as a proxy for IT-adopting LICs, while acknowledging this approximation as one of caveats of their analysis (page 17 of their paper). Last, Bleaney and Francisco (2016) use a more updated dataset till 2013, but their focus is Sub-Saharan African countries, thus missing out a number of IT adopting LICs.  Notes: A negative effect on inflation means that IT implementation significantly reduces the inflation level. Did (PSM) stands for differences-in-differences (propensity score matching).

Hypotheses
Why is it interesting to examine the effect of IT on inflation in LIC as a separate group?
Indeed, if IT performances in LICs and EMEs are alike, such a study may not be necessary, because as seen in Table 2 the previous studies already suggest that IT helps reduce the level of inflation in non-advanced countries. However, in what follows, we indicate that LICs and EMEs are not necessarily alike, and argue why IT performances might be different between them. We then clarify hypotheses which we test in the following sections.
First, the relevant fact is that the quality of institutions is generally lower in LICs than EMEs. 5 Figure 1 compares the quality of institutions between LICs and EMEs, alongside HICs, using the measures of "executive constraints" and "democracy/autocracy" (both from Polity IV, Marshall et al. (2013)). Each bar represents the cross-country average of a countrylevel average of the corresponding variable over the sample period  for each income group. As elaborated below, these variables essentially reflect the degree to which a government is constrained and made accountable to the general public. With the larger value corresponding to stronger institutions, the message is that a government in LICs is generally less accountable to the public than in EMEs. Democracy/autocracy Notes: Executive constraints (Democracy/autocracy) ranges between 1 and 7 (−10 and 10). Each bar represents an average of country−level averages across LICs, EMEs, and HICs over the sample period. Larger values correspond to stronger institutions. Source: Authors' calculations Next, why may the degree of government accountability be relevant for IT performance?
We argue that this is because an unaccountable government may be associated with fiscal dominance, defined as the subordination of monetary policy to fiscal requirements. While it is admittedly difficult to measure the degree of dominance, one proxy would be the extent to which legal restrictions limit a central bank's lending to the government, as quantified that affect interactions between a government and the general public. For example, institutions such as competitive elections and free media can help a government to be more accountable to the public.
by Cukierman et al. (1992) for the 1980-89 period and subsequently updated by Crowe and Meade (2007)  rate. Therefore, even if IT has a potential to help reduce inflation (as shown by the previous studies for EMEs), fiscal dominance under an unaccountable government may hinder the potential from being fulfilled.
6 Legal restrictions that limit a central bank (CB)'s lending to government is one of the four aspects of a central bank's independence measured by Cukierman et al. (1992) and Crowe and Meade (2007). Other three aspects of independence are 1) whether CB's management is protected from political pressure by secure tenure and independent appointment, 2) whether the government can participate in or overturn the CB's policy decisions, 3) whether the legal mandate of the CB sets a clear objective for monetary policy. 7 A value of lending restrictions in 1980-89 period (one value per country) is from Cukierman et al. (1992), and a value in 2003 is from Crowe and Meade (2007). Only for a limited number of countries, two observations (1980-90, and 2003)   Democracy/autocracy Notes: Added variable plots based on OLS estimations, with standard errors clustered by country.
Lending restriction reflects the degree of restrictions that limit central bank's lending to governmnet. The larger value of lending restriction corresponds to less of a fiscal dominance problem.

Source: Authors' calculations
As a caveat, however, while we argue that different qualities of institutions within non-HICs may yield different IT effects on inflation, it is important to acknowledge that there are other factors which affect the IT performance. In particular, as highlighted by Masson et al. (1997), the type of exchange rate scheme should also have a critical bearing as a prerequisite for a successful IT framework. This is because when countries use the pegging of nominal exchange rates as an alternative anchor, monetary policy is already significantly constrained, so that the additional effect of IT on inflation expectation may be small. Thus, under fixed exchange rates, whether or not institutions eliminate fiscal dominance may be of second-order relevance. 9 To summarize, our argument is that in low-income countries, where government accountability is generally low, the problem of fiscal dominance is present, which in turn reduces 9 While considering the strict IT adoption dates often precludes the case where a central bank pursues exchange rates as a nominal anchor, we confirm that exchange rates are often categorized as soft peg even after the strict adoption dates. Therefore, it is still important to pay an attention to exchange rate flexibility.
prospect for a successful IT performance, particularly under a floating exchange rate regime.
Based on this argument, this paper tests the following two hypotheses.
1: IT is significantly less effective in reducing level of inflation in LICs than in EMEs.
With IT also being less effective in HICs than in EMEs (as the previous literature shows), the relation between the IT effect and income levels is non-monotonic.

2:
The different IT effects between LICs and EMEs are explained by differences in the degree of government accountability across the country groups. The role of accountability is particularly evident under floating exchange rate regimes.

Empirical Methodology
The standard regression specification tests for an IT effect by adding to an inflation regression a dummy variable that is equal to one when an IT regime is in place, and zero otherwise. The reference regression model for inflation in country i in year t is as follows: The lagged inflation term, π i,t−1 is expected to be always positive and significant, reflecting the persistence of inflation shocks. IT i,t , a dummy variable, takes the value of one if an IT regime is adopted in country i in year t. z i,j,t represent a vector of control variables, including exchange rate regime dummies (for a hard peg and for a float, so the omitted category is a soft peg); a dummy for a parity change (usually a devaluation) in a pegged regime in the current or previous year; and a dummy for a currency crisis in the current or the previous year. The latter two variables reflect the fact that devaluations and currency crises tend to be associated with spikes in the inflation rate. µ i and ν t are the country fixed effect and time dummies, capturing unobserved time-invariant country characteristics and global variations in inflation, respectively. Last, importantly, the right-hand side also contains a country-specific linear time trend, trend i,t . This is to address the possibility that initially high-inflation countries converge to the mean irrespective of implemented policies, including IT. This so-called "regression-to-the-mean" is consistent with the observation that even amongst non-IT countries, there are significant differences in time trend of inflation over the sample period.
To investigate how the effects of IT may differ across different income levels (Hypothesis 1), we consider two alternative specifications. The first is: where LIC i is a time-invariant dummy variable, which takes the value of one if country i is LIC (as defined above) and zero otherwise. EM E i and HIC i are also dummies defined likewise. Our main interest is to compare coefficients on the interactions between income group and IT (i.e., β L , β E and β H ). The second equation is: where y i,t is the log of real GDP per capita (in US dollar) in country i in year t. The idea is to make use of the time-variation of income levels to estimate how they interact with the IT effect. To allow for possible non-monotonicity between income levels and the IT effect, we add the interaction between squared income and the IT dummy as well. The coefficients of our interest are the ones on interaction terms, i.e., ζ and ψ. Both Eqs.2 and 3 include a country-specific linear trend to control for regression-to-the-mean.
Next, we examine the role of institutions as a factor which differentiates IT effects between LICs and EMEs (Hypothesis 2). The reference equation is: (4) Account i,t is an institution variable which measures the degree to which governments are accountable to the public in country i in period t. As indicated above, we use "executive constraints" and "democracy/autocracy" as a proxy. Although these variables vary over our sample period, particularly within the sample of LICs and EMEs, they generally do not show frequent year-to-year variations. Thus, we estimate Eq.4 without country fixed effects as well to make use of cross-country variations in government accountability. Further, to take account of exchange rate flexibility as a prerequisite for successful IT performance, we interact IT i,t , Account i,t and F loat i,t , which takes the value of one when exchange rate is floating and zero when a soft peg is adopted or there is no legal tender of their own. 10 The resulting equation is: The three-way interactions allow us to examine the role of institutions in the IT effects on inflation conditional on an exchange rate regime. We estimate Eq.5 without country time fixed effects as well.
Having clarified the regression equations, it is important to realize that the estimation of the above dynamic panel data models using ordinary least squares (OLS) produces biased 10 z i,j,t in Eq.5 do not include exchange rate regime dummies. tends to be weak. Another possible option is to use a differences-in-differences approach (DiD, cf. Table 2) to address the causality of IT on inflation. 12 However, this method is not free from weaknesses either. In particular, a non-negligible arbitrariness is bound to arise when defining the dividing line for non-IT-targeters used as a control group. Therefore, on balance, we prefer the panel regression approach for simplicity and greater robustness.  Table 3 shows that average inflation rates in LICs (EMEs, HICs) are 11.18, 13.31, 3.98%, respectively.
Annual real GDP per capita (in US dollars), used to estimate Eq.3, is from WDI. The average figure is highest in HICs (34,405 dollars) and lowest in LICs (1,583 dollars). We use two proxies for institutional quality to measure the degree to which governments are accountable. The underlying assumption for the choice of proxies is that governments are more (less) accountable when they are more (less) constrained. The first proxy, "executive constraints", is from Polity IV, measuring the degree of institutionalized constraints on the decision-making powers of chief executives. 15 Second, we use "democracy/autocracy", also from Polity IV, which measures not only the degree of institutionalized constraints (as in "executive constraints") but also other democratic elements such as the extent to which citizens' political participation is guaranteed. 16 The participation of the citizens in the governance process should prompt governments to be more accountable for their policy actions.
For both variables, the larger value corresponds to the higher government accountability.
13 See above for how countries are classified by income levels and how IT adoption dates are defined. 14 In our dataset, correlation of inflation data (log difference of CPI) between WDI and WEO is 99 percent. 15 This variable is often used in the literature on institutions and development, including Acemoglu et al. (2001). In Polity IV, the variable name is "XCONST". 16 The variable name in Polity IV is "POLITY2". The average of both proxies is highest (lowest) in HICs (LICs). While one may argue that corruption measures are also relevant proxies for government accountability, our view is that institutional features such as constraints on politicians and citizens' political participation are the ones that are more relevant in relation to fiscal dominance, rather than corruption as an outcome of such features. 17 Turning to control variables, exchange rate regime data and information on parity changes are dummy variables based on Bleaney and Tian (2017).

Results
This section tests the two main hypotheses presented above. We first test if IT is significantly less effective in LICs than EMEs (Hypothesis 1), and then test if the difference in government accountability between the two country groups is a possible explaining factor for the different IT effects (Hypothesis 2).
is institutions that restrict government's rent seeking (e.g., competitive elections), rather than the level of corruption. 18 An alternative is Reinhart and Rogoff (2004), which tends to under-record floats, as discussed in Bleaney and Tian (2017). 19 Specifically, the authors define that this takes 1 when the EMPI is in the upper quartile of their dataset (spanning 1980-2012).
6.1 IT effects across different income levels 6.1.1 Using time-invariant income dummies Table 4 shows estimation results of Eq.1 for an unconditional effect of IT on inflation, and also results of Eq.2 for conditional effects upon income levels. The conditional effects are estimated using time-invariant country group dummies (LIC i , EM E i and HIC i ). Acknowledging the difficulty of defining IT adoption dates, we estimate equations using both strict and loose adoption dates. Also, given that using extra control variables (z i,j,t ) restricts the sample size greatly, results are shown with and without them. As noted, to take account of regressionto-the-mean, we include a country-specific linear trend as well as time dummies.
The first two columns estimate the equations without the controls using the strict IT adoption dates. Column (1) shows the unconditional IT effects, based on all the observations regardless of country's income levels. The coefficient on the IT dummy of −0.04 is insignificant, implying that the adoption of IT is not associated with a change in inflation rates when using the entire observations. However, Column (2), estimating the IT effects conditional on income levels, shows that for EMEs, the adoption of IT is significantly associated with lower inflation by 4.78 percentage points. Meanwhile, the coefficients for IT*LIC and IT*HIC are positive (2.49 and 2.27), albeit insignificant for the former. Notice that the coefficient is significantly more negative (i.e., IT is more effective) in EMEs than in LICs and HICs. This is based on the observation that p-values from testing the equality of interaction coefficients between IT*LIC and IT*EME (see LIC EME in the table) and IT*HIC and IT*EME (see HIC EME) are 0.035 and 0.0040. The lagged inflation variable is highly significant, showing that inflation is persistent.
Column (3) and (4) add extra control variables, still using strict adoption dates. They confirm the heterogeneous effects of IT across income levels: only for EMEs, the IT dummy is negatively associated with inflation rates, and the coefficient is significantly more negative than in LICs or HICs. Turning to controls, a floating exchange rate is significantly associated Notes: Fixed-effect estimations. Constant, time dummies and country-specific linear trends are not shown for brevity. LIC EME (HIC EME) gives p-value from testing the equality of coefficients on IT between LIC and EME (HIC and EME). Inflation rate is calculated as a log difference of CPI. Countries with the average inflation of over 50 percent over the sample period are not included. t-statistics are in parentheses. Clustered standard errors are used to adjust for correlation of error terms within countries. *** p < 0.01, ** p < 0.05, * p < 0.1.
with higher inflation than the omitted category of a soft peg with no parity change, and the coefficient on a hard peg is negative, though insignificant. A currency crisis in the current or the previous year is always associated with significantly higher inflation, as is a current (but not lagged) parity change in a pegged regime. Columns (5) to (8) present results using loose IT adoption dates. Unlike Columns (2) and (4), coefficients on IT*HIC in Columns (6) and (8) are negative, albeit insignificant, and the difference between coefficients on IT*EME and IT*HIC is marginally insignificant in Column (8) (p-value is 0.106). Still, heterogeneous IT effects across income levels are observed with or without extra controls. Overall, in LICs IT is significantly less effective than in EMEs, and more broadly, the relation between the effectiveness of IT and income levels appears to be non-monotonic.  (3) and (4) indicate that the results are robust to the use of loose IT adoption dates. Therefore, together with the results from the analysis using time-invariant income dummies (Table 4), we argue that Hypothesis 1 has been supported. We showed above that IT is significantly less effective in reducing inflation in LICs than in EMEs. We now test the hypothesis that the degree of government accountability is a possible driving force behind this result (Hypothesis 2). We first consider the role of government accountability in the IT effect on inflation rates without taking account of the possible relevance of exchange rate regimes as a prerequisite for successful IT performance. Specifically, using the sub-sample of LICs and EMEs, Table 6 estimates the association between IT and inflation rates conditional on government accountability (Eq.4), proxied by "executive constraints" and "democracy/autocracy". As mentioned, because these institutional variables do not show frequent time variations, we show results without country fixed effects as well, which exploit cross-country variations of accountability. For brevity, the table only shows results based on strict IT adoption dates (Results using loose adoption dates are in Table 10 in Appendix A).

Interaction with per capita GDP
Columns (1) to (4) are using executive constraints as a proxy for government accountability. The former (latter) two columns are without (with) extra controls, and Columns (1) and (3) include fixed effects, while Column (2) and (4) do not. In all these four columns, coefficients on the interaction between IT dummy and executive constraints, which reflects the role of accountability in the marginal effect of IT, are significantly negative. A rise in executive constraints (which ranges between 1 and 7) by the value of one corresponds to a fall in the marginal effect of IT by 1.15 to 3.03 percentage points. In Columns (5) to (8), democracy/autocracy is used as an accountability proxy. Again, the signs of interaction coefficients are all negative, although the coefficient is significant only in Column (5). Notice, however, that the coefficients being insignificant do not necessary indicate that the role of accountability in the effect of IT is not robust, because the relevance of exchange rate regimes is not taken into account yet. To illustrate the implication of the negative interaction coefficients, Figure 3 plots marginal effects of IT together with 90 percent confidence interval for different levels of executive constraints. Sub-figures (a) to (d) correspond to Columns (1) to (4) of Table 6. They show that apart from sub-figure (a), IT is associated with significantly negative marginal effect of IT Executive constraints EMEs Notes: Executive constraints are used as a proxy for government accountability. Executive constraints range from 1 to 7. The higher the value is, the more government is constrained.
Source: Authors' calculations when the proxy takes the value of 7. Now, notice from Figure 4 (the histograms of executive constraints for LICs and EMEs) that about 45 percent of observations from EMEs take the values of 7, whereas only about 10 percent of observations from LICs do. 21 Therefore, even when the relevance of exchange rates is not considered, there is some indication that government accountability works as a driving factor behind the different IT effects across income levels. A similar observation can be made for the case when democracy/autocracy is used as a proxy for government accountability (see Figure 6 and Figure 7 in Appendix A). 22

Relevance of exchange rate regimes
Next, to take account of the possible relevance of exchange rate regimes as a pre-requisite for successful IT performance, Table 7 estimates Eq.5 which entails three-way interactions among IT dummy, accountability proxy, and floating dummy. Using the notation of Eq.5, the marginal effect of IT on inflation is given as: Thus, the effects under different exchange rate regimes are: In Eq.7, λ + υ represents the role of accountability in the marginal effect of IT on inflation rates under a floating regime, and λ indicates the role of accountability under a fixed regime. Table 7 has the same structure as Table 6, considering the alternative accountability proxies and regression equations with and without extra control variables. The key message of the table is simple: as shown in the rows on the size and p-value of λ+υ, it is significantly negative in most of the cases considered (except for Column (8), where the p-value is 0.104), whereas λ is insignificant for all the cases. This means that particularly under floating exchange rates, an increase in government accountability is associated with significantly more effective IT in terms of reducing inflation rates. For example, λ + υ in Columns (1) to (4) indicates that under floating rates a rise in executive constraints by one is associated with a fall in the marginal effect by 2.25 to 3.82 percentage points. Under fixed exchange rates, however, government accountability plays virtually no role. Table 11 in Appendix A shows that the results are robust to the use of loose IT adoption dates.
To complete the analysis, Figure   Notes: Based on the sub-sample of LICs and EMEs. Strict IT adopt dates are used. Constant, time dummies and country-specific linear trends are not shown for brevity. Executive constraints (democracy/autocracy) ranges from 1 to 7 (-10 to 10). λ + υ reflects how the marginal effect of IT on inflation changes as government accountability rises under floating exchange rates; λ reflects how the effect changes under fixed exchange rates. Inflation rate is calculated as a log difference of CPI. Countries with the average inflation of over 50 percent over the sample period are not included. t-statistics are in parentheses. Clustered standard errors are used to adjust for correlation of error terms within countries. *** p < 0.01, ** p < 0.05, * p < 0.1.

Alternative explanation: the role of initial inflation
Having shown that government accountability helps explain why IT may not be effective in LICs unlike in EMEs, we here examine an alternative possible explanation. That is, one may argue that IT has reduced inflation more in EMEs than in LICs (and HICs), simply because the pre-IT inflation rate in EMEs was higher than other countries. Indeed, Table 8 shows that the initial inflation, calculated as a 5-year average before the adoption of IT, is particularly higher in EMEs on average (18.19%, 13.69% without Brazil) than in LICs and HICs (7.32% and 4.44%), whereas the 5-year average after IT adoption in EMEs (5.53%, 5.31% without Brazil) is rather close to the corresponding figure in LICs and HICs (5.31% and 2.17%).
Notice that the fact that we always include a country-specific linear trend (to control for regression-to-the-mean) does take account of the effects of initial inflation to some degree.
Still, however, there is an explicit way to address this issue (though not entirely satisfactory, as explained below), which is simply to interact the IT dummy with initial inflation rates.
This is feasible despite the fact that initial inflation rates themselves, being time-invariant, are absorbed into country fixed effects. This is because for IT adopters, the interaction between the IT dummy and initial inflation shows time variations. What is unsatisfactory with this approach, however, is that for non-IT adopters, initial inflation (inflation before IT adoption) does not exist by definition. 23 Nonetheless, since the IT dummy is always zero for these countries, the level of initial inflation would not matter at least for an estimation purpose. 24 With this caveat, Table 9 estimates Eq.5 with the additional interaction term, "IT*Initial Infl", where "Initial Infl" is the 5-year average inflation prior to the IT adoption. The table has the same structure as Table 7, except that only the regressions with country fixed effects are shown. This is because as emphasized, the initial inflation is not defined for countries that have never adopted IT. The results show that the coefficient of the new interaction variable is always negative, as expected, and significant (at the 1 percent level). However, the institutional variables also retain similar coefficients and significance levels to those in Table 7 under floating exchange rates (see the odd-numbered columns of Table 7 with fixed effects). 25 This suggests that, although the initial-inflation effect is significant and thus there is some truth in the alternative explanation, our institutional story is robust to its inclusion. Notes: Initial inflation is the 5-year average of inflation rates just before the adoption of IT. "Change" is obtained as the average of inflation just before IT adoption minus the (5-year) average just after the adoption.
IT adoption year is based on the strict definition (cf. Table 1). When inflation data is not available for 5 years after IT adoption (e.g., Japan), the average is calculated using as many observations as available.

Discussion and Concluding Remarks
The standard result in previous research is that inflation targeting has made little difference to the inflation rate in the advanced countries, but has significantly reduced inflation in non-advanced countries (as indicated by Table 2). Because LICs have been slower to adopt inflation targeting than EMEs (Table 1), the samples of non-advanced countries used in previous research have contained very few LICs. Now that more time has passed, it is possible to consider the effectiveness of IT in LICs separately from EMEs. Our basic result is that IT has been far less effective in LICs than in EMEs. By using panel regression methods rather than differences-in-differences or propensity score matching, our results are able to control for unobserved country characteristics (through country fixed effects), for unexplained fluctuations in inflation that affect all countries equally (through time fixed effects), and for variation in the speed of disinflation in different countries (through country-specific time trends).
We gave a story as to why this should be the case. Specifically, we examined the role of institutions which affect the degree of government accountability in the effectiveness of IT in a sample of LICs and EMEs. Measures of institutional quality based on political arrangements are more structural and less subjective, and also less likely to be endogenous to outcomes, than those based on survey data such as perceptions of corruption. The results indicate that IT was more effective with stronger institutions. Various authors (e.g. Masson et al. (1997); Thornton (2016)) have pointed out that in lower-income countries pegging the exchange rate can also be an effective nominal anchor. If that is the case, the benefits from IT should be greater when the exchange rate is floating than when it is pegged. Allowing for this, we found that the institutional effect is particularly marked under floating rates, and not significant when the exchange rate is pegged. This is still true even when we control for the significant effect of the pre-IT inflation rate, which has tended to be particularly high in EMEs, on the reduction in inflation achieved under IT. Overall, given that institutions are generally weaker in LICs than in EMEs (Figure 1), we believe that government accountability does help us grasp why IT may be less effective in LICs.
We argued that the reason why government accountability matters in the IT effect is that fiscal dominance under an unaccountable government creates inflationary pressures of a fiscal origin, and impairs the ability of central banks to align private sector's inflation expectation to their target rate. However, we admit that this is merely a conjecture, where a formal model of institutional quality and inflation would be desirable. For example, Acemoglu et al.  Notes: Based on the sub-sample of LICs and EMEs. Loose IT adoption dates are used. Constant, time dummies and country-specific linear trends are not shown for brevity. Executive constraints (democracy/autocracy) ranges from 1 to 7 (-10 to 10). Inflation rate is calculated as a log difference of CPI. Countries with the average inflation of over 50 percent over the sample period are not included. t-statistics are in parentheses. Clustered standard errors are used to adjust for correlation of error terms within countries. *** p < 0.01, ** p < 0.05, * p < 0.1. Notes: A marginal effect with 90% confidence interval is shown. Democracy/autocracy initially ranges from −10 to 10. The higher the value is, the more government is constrained. It is rescaled to 0 to 20 to be compatible with "Margins" Stata command.

A Supplementary results
Source: Authors' calculations Democracy/autocracy EMEs Notes: Democracy/autocracy initially ranges from −10 to 10. The higher the value is, the more government is constrained. It is rescaled to 0 to 20 to be compatible with "Margins" Stata command.
Source: Authors' calculations Notes: Based on the sub-sample of LICs and EMEs. Loose IT adopt dates are used. Constant, time dummies and country-specific linear trends are not shown for brevity. Executive constraints (democracy/autocracy) ranges from 1 to 7 (-10 to 10). λ + υ reflects how the marginal effect of IT on inflation changes as government accountability rises under floating exchange rates; λ reflects how the effect changes under fixed exchange rates. Inflation rate is calculated as a log difference of CPI. Countries with the average inflation of over 50 percent over the sample period are not included. t-statistics are in parentheses. Clustered standard errors are used to adjust for correlation of error terms within countries. *** p < 0.01, ** p < 0.05, * p < 0.1. A marginal effect with 90% confidence interval is shown. Democracy/autocracy initially ranges from −10 to 10. It is rescaled to 0 to 20 to be compatible with "Margins" Stata command.