-
Internal inconsistency on the first-stage estimator: the Abstract/Introduction state that a dynamic System GMM approach is used to capture endogenous feedback/persistence in investment (Sec. 1), but the Methods/Results present only a static fixed-effects regression of $\ln(s_{i,t})$ on $\ln(y_{i,t})$ (Eq. (3); Secs. 2.2, 3.1), with no lagged dependent variable, no GMM instrument strategy, and no GMM diagnostics. This undermines a central methodological claim (handling endogeneity/dynamics in investment) and makes it hard to interpret what the framework is actually validating.
Recommendation: Make the narrative and implementation consistent. Either: (i) implement the dynamic System GMM version of Eq. (3) explicitly (e.g., include $\ln(s_{i,t-1})$), fully document instrument sets/lag structure in Sec. 2.2, and report standard diagnostics (AR(1)/AR(2), Hansen/Sargan) in Sec. 3.1; or (ii) remove/modify System GMM claims in the Abstract/Sec. 1 and justify why FE is the intended estimator in this synthetic validation (including a brief discussion of simultaneity and why it does/does not matter under the DGP). Ideally, include FE vs. GMM as a robustness comparison if endogeneity is part of the motivation.
-
Synthetic DGP is under-specified, preventing meaningful validation against “known truth.” Sec. 2.1 does not provide the explicit equations and parameter values (production function, $\alpha$, $\delta$, technology/population growth, shock processes, cross-country heterogeneity), nor does it clearly state how trade openness and government expenditure are generated and whether they affect steady states (levels) and/or convergence speed (transition). Sec. 3 does not compare estimated parameters ($\gamma_1$, $\delta$, $\beta_1$, $\beta_2$) to their true DGP values, which is the key advantage of synthetic data.
Recommendation: Expand Sec. 2.1 (or add a dedicated Appendix) with full DGP documentation: write down the Cobb–Douglas production function and laws of motion, list parameter values and shock processes (including persistence/cross-country heterogeneity), and specify exactly how policy variables are generated and where they enter the true model (investment/steady state vs. speed). Then, in Sec. 3, add a validation subsection reporting true vs. estimated values (with bias and uncertainty) for $\gamma_1$, $\delta$, $\beta_1$, $\beta_2$. This should be presented as the main methodological deliverable, not a side note.
-
Lack of Monte Carlo evidence: results appear to be based on a single synthetic draw, so statistical significance (e.g., the $10\%$ trade openness interaction in Sec. 3.2) is hard to interpret as method performance rather than sampling luck. In a validation paper, single-sample $p$-values are much less informative than bias/RMSE/coverage across repeated simulations.
Recommendation: Add a Monte Carlo section: repeat the full two-stage procedure many times under the same DGP and report distributions for $\gamma_1$, $\delta$, $\beta_1$, $\beta_2$ (bias, RMSE, coverage of nominal $95\%$ CIs, and rejection rates under null moderation). Include scenarios where true moderation is zero vs. nonzero for each policy variable (trade openness/government expenditure) to quantify Type I/II error and power. Reframe the empirical claims in Sec. 4 in terms of estimator performance rather than one-off significance.
-
IV identification for the moderated transition model is not described, making $\beta_2$ difficult to trust. Sec. 2.3 labels Eq. (5) as an IV regression to handle endogeneity of $D_{i,t}\times\text{Policy}_{i,t}$, and Table 1 reports IV estimates, but the paper does not specify (a) which regressors are treated as endogenous (interaction only vs. also $D$ and Policy levels), (b) what instruments are used and how constructed from the DGP, (c) first-stage results and weak-IV diagnostics, or (d) overidentification tests. Also, Table 1 reports items (e.g., “Adj. R-squared”) whose meaning depends on the exact IV routine used.
Recommendation: In Sec. 2.3, explicitly list endogenous regressors and instruments for $D_{i,t}$, $\text{Policy}_{i,t}$, and $D_{i,t}\times\text{Policy}_{i,t}$ (e.g., lags, exogenous shifters from the DGP, or constructed instruments like $D\times Z$ where $Z$ instruments Policy). State whether estimation is 2SLS or IV-GMM and how SEs are computed. In Sec. 3.2/Table 1, report first-stage coefficients, partial $R^2$, Kleibergen–Paap (or analogous) $F$-statistics, and overidentification tests where applicable; adjust reported fit statistics to those appropriate for the estimator (or explain them). If the DGP implies exogeneity for some variables, say so and justify a non-IV baseline alongside the IV specification.
-
Generated-regressor and two-stage inference problem: $D_{i,t}$ is constructed using $\hat{s}_{i,t}$ from the first stage (Secs. 2.2–2.3; Eqs. (3)–(4)), making the second-stage regressor a function of estimated parameters and potentially shared shocks. Standard IV/OLS SEs in the second stage may understate uncertainty, and mechanical dependence of $D_{i,t}$ on $y_{i,t}$ (and fixed effects) can blur interpretation of “speed” versus “level” channels.
Recommendation: Address two-stage inference explicitly. At minimum, bootstrap the entire pipeline (first-stage estimation $\rightarrow$ $K^*$ construction $\rightarrow$ $D$ computation $\rightarrow$ second-stage IV) and report bootstrap SEs/CIs in Sec. 3.2. Also add a robustness check constructing $D_{i,t}$ using the true $s_{i,t}$ from the DGP (or using observed $s_{i,t}$ rather than $\hat{s}_{i,t}$) to quantify how first-stage estimation error affects $\beta_1$ and $\beta_2$. If feasible, discuss (or implement) a joint estimation approach as an alternative.
-
Ambiguity/possible inconsistency in the steady-state construction under Cobb–Douglas. Sec. 2.3 defines $K^*_{i,t}$ via $s_{i,t}Y_{i,t} = \delta K^*_{i,t}$, apparently using observed $Y_{i,t}$. But in a Cobb–Douglas model $Y$ depends on $K$, so a steady state typically requires evaluating $Y$ at $K^*$ (a fixed-point condition). Without a derivation, it is unclear whether $K^*_{i,t}$ is a structural steady state, an ‘implied’ accounting steady state, or a shortcut inconsistent with the stated DGP.
Recommendation: Provide a clear derivation in Sec. 2.3 (or Appendix): either (i) explicitly define $K^*_{i,t}$ as an “implied steady state” holding $Y$ fixed at observed $Y_{i,t}$ and justify why this is the target estimand, or (ii) derive the structural steady state under Cobb–Douglas (solve $s\cdot Y(K^*) = (\delta+g+n)K^*$ with the appropriate growth terms) and implement that version. Clarify whether you use $s_{i,t}$ or $\hat{s}_{i,t}$, and whether technology/labor are treated as fixed or evolving in the steady-state expression.
-
Timing/indexing is inconsistent across equations and affects interpretation of convergence speed. Eq. (1) uses $K_{i,t+1}$; Eq. (2) defines $\Delta\ln(K_{i,t+1})$ using period-$t$ objects; Eq. (5) uses $\Delta\ln(K_{i,t})$ with unclear dating of $D_{i,t}$ and $\text{Policy}_{i,t}$. The text interprets $\beta_1$ as “closing $\sim35\%$ of the gap per period” without clearly stating the time unit (annual $1990$–$2019$) or mapping to standard convergence metrics (e.g., half-life).
Recommendation: Standardize timing across Secs. 2.1–2.3: define $\Delta\ln(K_{i,t}) \equiv \ln(K_{i,t})-\ln(K_{i,t-1})$ (or the forward version) and date $D$ and $\text{Policy}$ consistently (typically $t-1$ on the RHS for a $t$ growth rate). State explicitly that the panel is annual and interpret $\beta_1$ accordingly; optionally report implied half-life ($\ln(2)/\beta$) for the baseline and at selected policy percentiles. Ensure Table 1 and Fig. 1 match the chosen timing.
-
Separation of “level” vs “speed” effects is asserted but not demonstrated. If policy variables affect investment/savings (level channel) in the DGP (or in real data), omitting policy from the first-stage investment function (Eq. (3)) can shift policy variation into the second stage and confound moderation ($\beta_2$) with misspecified steady states. Conversely, if the maintained restriction is that policy affects only speed, the paper should state and verify this against the DGP.
Recommendation: Make the maintained restriction explicit: do trade openness and/or government expenditure affect investment/steady states in the DGP, or only convergence speed? Then demonstrate the separation empirically: (i) include policy variables (and possibly their interactions) in the first-stage investment regression as a robustness check, (ii) recompute $K^*$ and $D$, and (iii) show how $\beta_2$ changes. In the Monte Carlo, add scenarios where policy affects levels only, speed only, both, or neither, to show when the two-stage approach correctly attributes channels.
-
External relevance is not yet established: the paper remains largely inside the synthetic environment, with limited guidance for real-data implementation (Sec. 4). Practical obstacles—capital stock measurement, depreciation estimation, policy endogeneity, structural breaks, shorter/noisier panels—are not discussed in a way that would help applied researchers adopt the framework.
Recommendation: Strengthen Sec. 4 with a concrete implementation roadmap for real data: data sources (e.g., Penn World Table, WDI), how to construct $K$ and $\delta$ (perpetual inventory vs. provided series), how to handle measurement error and policy endogeneity, and what instrument strategies might be plausible outside a synthetic DGP. If feasible, add a short illustrative empirical application (even a minimal replication on a standard dataset) to demonstrate feasibility; otherwise clearly scope the paper as a simulation-validation study and articulate what remains before applied use.