Regional Innovation Disparities in Kazakhstan: Input–Output Misalignment in an Official Panel, 2014–2025

  1. Yevgeniya Tanassoglo
  2. Dilek Çetin
  3. Aigul Myltykbayeva

Reviewed by

Read the full article

Listed in

Log in to save this article

Abstract

How are innovation inputs and measured outputs distributed across the regions of a resource-dependent post-Soviet economy, and how closely do they correspond? We use a balanced official panel of Kazakhstan’s 17 regions for 2014–2025 (204 region–year observations) covering four indicators: R&D expenditure, innovation-active enterprises, the innovation activity rate and innovative product output. Granted patents for inventions are available by region for 2021–2025 only; this series is shorter than the panel, so it is used in the 2025 cross-sectional and descriptive analyses and does not enter the panel or machine learning models. On the panel, we estimate dispersion measures, panel regression with region-clustered standard errors and tree-based machine learning interpreted with SHapley Additive exPlanations; on the 2025 cross-section, we compute correlations, clustering and spatial association statistics. Inputs are highly concentrated: Almaty and Astana together accounted for 65.1% of national R&D expenditure in 2025, and cross-regional dispersion did not narrow over the period. In the 2025 cross-section (n = 17), R&D expenditure, innovation-active enterprises and patent grants are strongly inter-correlated (r = 0.96–0.97), but none of the three is significantly correlated with innovative product output. This input–output misalignment is a pattern whose mechanisms we do not test. The panel specification has modest explanatory power (R2 = 0.188), which bounds the weight that can be placed on any single coefficient. Within it the innovation activity rate is the only significant predictor of output (β = 0.177, p < 0.001), an association that survives exclusion of the three regions affected by the 2022 administrative reform; because that indicator and the output measure are compiled within the same system of enterprise observation, it is reported as a within-dataset association requiring validation against independently measured indicators. Under region-grouped cross-validation, the machine learning models show no out-of-region predictive ability and are reported solely as in-sample interpretation. The contribution is a documented diagnosis of input–output misalignment in an understudied regional system, with an explicit account of what such data can and cannot support, and a prospective statement of what institutional readiness for algorithmic decision support in regional innovation policy would require.

Article activity feed

  1. Version published to 10.3390/urbansci10100588
  2. PREreview

    This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/21796276.

    Summary

    The paper combines regional innovation indicators for Kazakhstan with PCA, TOPSIS, clustering, panel regression, tree models, SHAP, and spatial statistics. I found the contrast between where R&D money is concentrated and where innovative product output appears especially interesting. The manuscript is also candid about several data limitations. My main concern is that some analyses are described as independent confirmation even though they inherit the same geographic and validation problems. The six points below are the changes I think are needed before the central claims can be evaluated reliably.

    Major comments

    1. Section 2.1 does not describe a geographically stable panel. The paper says it preserves 17 pre-2022 regions, but the 2022-2025 observations for Karaganda, East Kazakhstan, and Almaty use their narrower post-split territories without adding Ulytau, Abai, and Zhetisu back to the parent regions. This means the unit called Karaganda, for example, changes during the series. Calling the resulting 204 observations a balanced panel is misleading even if no cells are missing. The authors should harmonize every year to one boundary system. At minimum, they should rerun the coefficient-of-variation, panel, clustering, and spatial results after excluding the affected units and after ending the panel in 2021. This is important because an administrative split can look like divergence even when the underlying activity has not changed.

    2. The five-fold cross-validation in Sections 2.5 and 2.7 appears to split 204 region-year rows at random. If so, training and validation folds contain both the same regions and adjacent years. The reported R-squared values of 0.384 and 0.314 are then not clean tests of performance on a new region or future year. Please repeat the analysis with leave-one-region-out or GroupKFold validation, and add a time-blocked test that trains on earlier years and predicts later years. Scaling and preprocessing should be fitted inside each training fold. A simple baseline, fold-level scores, and confidence intervals would make the gain from the tree models easier to judge.

    3. Section 3.6 overstates agreement between the random-effects model and the SHAP analysis. In Table 7, INP is the only significant predictor in the preferred model. NOEI is negative and not significant. In Table 8, however, NOEI has the largest mean absolute SHAP value. Those are not the same ordering, and raw regression coefficients are not comparable as importance measures when predictors use different scales. I suggest describing the panel model as conditional association and the tree models as predictive attribution, then discussing the disagreement rather than presenting it as a third confirmation of one hierarchy.

    4. The spatial results need enough detail to distinguish the statistic from its test. Table 6 is headed Moran's I (z), but entries such as 19.065 and 7.330 appear to be z statistics, not Moran's I values. Please report the actual I estimate, the expected value under the null, the permutation p-value, and the number of permutations. The row standardization and complete W1 and W2 matrices should also be provided. For the LISA results, p = 0.044 for Zhambyl and p = 0.035 for Atyrau are fragile when many regions, outcomes, and years are tested. The authors should apply or justify a multiple-testing procedure and show whether these clusters remain.

    5. Several central results come from a separate 2003-2024 GIS analysis, while the tables and figures also contain recomputations on the 2014-2025 panel and an author-constructed approximate queen-contiguity matrix. I had difficulty tracking which dataset supports each claim. A table listing every result, time window, boundary system, and weights matrix would help. More importantly, the derived data, Python scripts, fold assignments, adjacency matrices, and environment file should be deposited in a versioned public repository rather than supplied on request. The paper's reproducibility claim depends on those materials.

    6. The manuscript states that k = 2 gives the best silhouette score, then selects k = 3 because the three groups are more useful for policy. That may be a reasonable policy choice, but it is not purely an empirical validation of three natural clusters. Please show the silhouette and stability results across plausible k values, repeat the analysis without the dominant Almaty observation, and test sensitivity to indicator scaling and selection. The three policy tiers should be presented as one defensible operational grouping, not as uniquely identified by the data.

    Additional points

    The productivity-driven diffusion paradox is an appealing label, but the current evidence is associational. Sector mix, regional price levels, firm size, and the definition of VOIP could all produce part of the R&D-output disconnect. These alternatives deserve direct tests or a narrower claim. For the panel models, please report region-clustered or Driscoll-Kraay standard errors and clarify the treatment of year effects. The paper should also separate statements supported by the 2014-2025 analysis from statements imported from the 2003-2024 companion analysis.

    Recommendation

    I support a substantial revision. A boundary-harmonized panel, leakage-resistant validation, transparent spatial inference, and public replication materials would make the paper's strongest empirical result, the mismatch between innovation inputs and outputs, much more convincing.

    Competing interests

    The author declares that they have no competing interests.

    Use of Artificial Intelligence (AI)

    The author declares that they used generative AI to come up with new ideas for their review.

  3. Version published to 10.20944/preprints202607.1407.v1