1. Introduction

Without data, you’re just another person with an opinion.

(William Edwards Deming)

The National Innovation System (NIS) focuses on a broad range of variables, activities, institutions, and their interactions that can foster economic growth and development in countries [1]. However, this literature underrepresents the global South. One of the major problems for this lack of reasonable representation stems from the lack of data for low- and middle-income countries (LMICs). By resulting in the exclusion of LMICs in empirical analyses, missing data lead to either positively or negatively biased results that manifest themselves in over and underestimated effect sizes.

Despite the general limitations, several studies have recently investigated NIS and its relationship with growth and development in some developing economies [2–5]. Other studies, using capacities as a way to operationalize NIS, have employed available data for diverse samples of countries to estimate the quantitative impact of financial, technological, and social capacities of countries on their economic growth and development process [6–11].

Inspired by the studies on capacities and economic development, Khan [12] has recently rigorously operationalized a thorough list of capacities that capture innovation, knowledge absorption, and learning processes in LMICs and further included those capacities in a formal framework of National Absorptive Capacity System (NACS). A firm-level concept of “absorptive capacity,” as advanced by Cohen and Levinthal [13], particularly motivates the NACS framework. As a modified version of NIS, NACS considers an LMIC an “economic learning” entity that absorbs, creates, and deploys knowledge, learning, and skills subject to the strength of its local capacities [14].

To study NACS and its evolution in LMICs and to further examine the impact of the framework capacities on economic development in LMICs, complete panel data (country-year observations) on variables that measure capacities are required. Unfortunately, such variables are not wholly available across LMICs eligible for the World Bank’s International Development Association (IDA) support. The IDA eligibility depends mainly on a country’s relative poverty, defined as the Gross National Income (GNI) per capita below an established threshold updated annually (1,185 US dollars in the fiscal year 2021). IDA also supports some countries, including several small island economies, that are above the operational cutoff but lack the creditworthiness needed to borrow from the International Bank for Reconstruction and Development (IBRD). Since IDA eligibility is based on GNI per capita, countries graduate and reinter (More information on IDAs can be accessed here: https://ida.worldbank.org/en/about/borrowing-countries). Such IDAs are the foci of this study (I have data on 82 countries—74 among them are still eligible for IDA resources and 8 countries recently graduated, together I call them LMICs). Since the LMICs are data-impoverished, there is a dire need to fix the problem of missing data for those LMICs, presumably prime candidates for development, learning, and innovation. Therefore, in this article, I build a complete and recent dataset on variables constituting capacities within LMICs, using established statistical and machine learning techniques.

Data incompleteness, commonly called the missing data problem, severely hampers empirical research. Various research fields have extensively investigated missing data dynamics, consequences, and possible remedies [15–20]. However, the innovation system and absorptive capacity literature have yet to thoroughly investigate missing data’s nuances, processes, and implications. One significant repercussion of missing data is that the current empirical literature on NIS and economic growth suffers from an imbalance. The literature either focuses on many countries within a limited period [7] or analyzes a few economies for an extended time [21, 22]. The former strand of literature can only provide a limited study of the evolution within NIS and NACS, whereas the latter prevents analyses in many LMICs. Hence neither is ideal; while the former is static, the latter is not representative of the LMICs.

This article systematically compiles, estimates, and imputes an incomplete dataset to alleviate the missing data problem in LMICs eligible for IDA support. It employs multiple imputation (MI) approach that efficiently and consistently estimates missing data and generates a panel dataset for 82 LMICs between 2005 and 2019. MI uses state-of-the-art statistical methods to address the missing data problem [18, 23]. By treating missing variables as outcomes and complete variables as predictors, MI statistical methods either impute all incomplete variables in a single computation step (multivariate regression model) or impute one variable at a time in a series (univariate regression models). Many research fields in physical and biological sciences have embraced such techniques [24–27]. This work explicitly employs univariate regression modeling, a variable-by-variable (sequential or chained) predictive mean matching (PMM) technique [28]. As an MI conditional modeling approach, PMM imputes missingness dependent on observed data in continuous, panel variables that do not have to be normally distributed [28–30]. This technique returns meaningful imputations that respect the data distribution of the original incomplete dataset (observed dataset).

Castellacci and Natera [31] conducted a similar data compilation study (CANA hereon). The researchers estimate a CANA dataset for 134 countries between 1980 and 2008 using an MI algorithm developed by Honaker and King [32]. The proposed MSK dataset is similar to CANA dataset as both are panel datasets estimated using novel MI techniques. Similarly, both datasets have a roughly identical structural build of NACS and NIS. For instance, they contend that such systems are measured by dimensions (CANA) and capacities (MSK), which, in turn, are captured by many variables interacting in multiple ways.

Although this article builds on CANA, it is different in several ways. First, as opposed to the CANA dataset, the MSK dataset estimated here focuses on relatively more data-deficient and economically poor IDA-eligible countries.

Secondly, though the MSK dataset employs some CANA dataset variables, it has an entirely different functional and operational conception of the capacities and the variables used to operationalize those capacities. Particularly, Public Policy and Social Capacity are operationalized very differently. In the MSK dataset, the Public Policy Capacity includes variables about public sector management and institutions, economic management, structural policies, the strength of legal rights, and statistical capacity scores of countries, whereas the Social Capacity includes variables on policies for social inclusion, human resource rating, social protection rating, equity of public resource use, poverty headcount ratio, and social contributions. On the other hand, in the CANA dataset, the analogous dimensions are the Political-institutional dimension (which comprises freedom of press and speech, human rights, women’s rights, and political rights, among other factors) and the Social capital dimension (which includes the importance of friends, family, marriage, trust, happiness, and Gini Index).

Additionally, the MSK dataset includes an extended set of other relevant variables to measure capacities. The MSK dataset consists of 47 variables for all economies in the dataset. In contrast, CANA consists of 34 variables for all economies and another seven variables for a restricted set of countries within the dataset.

Fourth, the timeframe for this study is truncated to fifteen years, not only because it is a decent period for panel analysis but also because of pragmatic concerns regarding data availability, particularly on public and social policy capacity variables. The World Bank Group’s country offices started collecting these variables in the IDA-eligible countries from 2005 onwards [33].

The last vital distinction worth considering is that the CANA dataset is estimated using Honaker and King’s [32] Expectation-Maximization algorithm. The MSK, on the other hand, is estimated using the Multiple Imputation by Chained Equations Predictive Mean Matching (MICE PMM) algorithm. Although the EM algorithm is efficient and undoubtedly suitable for panel data, it forces a normal distribution on the imputed data regardless of the distribution structure (skewed, unimodal, bimodal) in the observed data [34]. In contrast, the MICE PMM algorithm preserves the distribution pattern of observed data in the imputed values [35], and it has been used for panel data imputation [36]. Besides preserving the distribution pattern in the imputed values, the MICE PMM is best suited for this study because the data structure is heteroskedastic (Variances of the variables in data mostly differ; for instance, variance for days to enforce a contract is 80 times larger than the variance for days to start a business) and associations among variables are nonlinear as can be seen in scatterplots.

In short, this article contributes to the literature by constructing a complete dataset and establishing its relevance for panel analyses of NACS and economic growth, among other analyses, in LMICs. A standard MICE PMM algorithm is employed to construct this dataset. The panel dataset, hence obtained, is complete with no missing values. It consists of 47 variables grouped into six vital capacities for each country: technological capacity, financial capacity, human capital capacity, infrastructural capacity, public policy capacity, and social capacity. The incomplete (original or observed) dataset, which contains many missing values, is constructed from reputable data sources such as the World Bank, International Monetary Fund (IMF), International Labor Organization (ILO), United Nations COMTRADE, and United Nations Educational, Scientific and Cultural Organization (UNESCO), among others (see S2 Table). The MSK dataset is estimated from this observed dataset, which provides information on 82 LMICs between 2005 and 2019 (total observations are 1,230 country-year observations). A four-way quality check establishes this dataset’s reliability and usefulness for researchers interested in panel analyses of absorptive capacity and innovation system, economic development, economic policy, and convergence analysis within LMICs.

The rest of the paper is shaped as follows. Section 2 gives a brief literature landscape, the association between NIS and NACS, and discusses the missing data and its implications on methodologies. Section 3 further discusses the importance of handling missing data, strategies to address missingness, and underlying missing data mechanisms. Section 4 elaborates on Multiple Imputation and MICE PMM technique. Section 5 discusses the MSK dataset and the steps taken to develop this dataset. Section 6 carries out a brief descriptive analysis of the MSK dataset, and Section 7 conducts a quality check of the estimated dataset. Lastly, Section 8 concludes by summarizing the results and implications of this work. The Supporting Information includes graphs and tables, conveying more information on how the database is constructed and other dataset characteristics.

2. From NIS to NACS: Comparative analyses of national systems and growth, and development and the problem of missing data in LMICs

The concept of NIS emerged in the 1990s [37–39]. It considers systems, activities, institutions, and interactions as the driving force behind economic growth and development [1, 40]. The strength of these factors explains cross-country differences in growth, development, and innovation. Around the time NIS emerged, Cohen and Levinthal developed the idea of “absorptive capacity” to explain how learning is consolidated in a firm and how it impacts its growth [13]. In the early 2000s, researchers extended the firm-level concept to a national level [41, 42]. They developed a theoretical framework for aggregating national absorptive capacities upwards from a firm level. Other empirical studies also applied the idea nationally [8]. These works used different capacities emerging in NIS literature (such as technological and social capacities) as proxies for national absorptive capacity. In this essence, NACS is essentially an offshoot of NIS.

Earlier, foundational theoretical and empirical work on NIS focused mainly on prosperous economies [37, 43]. Later, NIS literature theoretically included developing countries, as they considered developing countries “national economic learning” entities and “imitation” centers [4, 44, 45]. National-level capacities literature examining the impact of capacities on economic development also included some developing economies in their analyses [8]. However, because of the lack of data in LMICs, such studies had to compromise operationalizing the complex and multifaceted capacities proposed in NIS and NACS. Similarly, the lack of data on many vital variables perhaps trimmed the list of essential capacities in their analyses.

Another critical challenge that missing data poses is limiting the application of study methodologies in many LMICs. In general, quantitative studies of capacities and development use mainly two different methodologies: panel regression analyses and composite indicator analyses.

Panel regression analyses examine the empirical relationship between a few capacity variables and comparative national differences in GDP per capita growth across countries [46, 47]. While powerful as they consider the dynamic nature of capacities, such panel studies ignore or drop off many LMICs because longitudinal data for many variables are missing in these countries. As a result, the coefficients of interest obtained through panel analyses do not provide information about the economically poor economies. Using econometric terminology, the estimates from such studies exhibit an upward or downward bias by overestimating or underestimating the effect of capacities on economic growth.

On the other hand, composite indicator analyses establish a country’s comparative standing against other countries by building aggregate or composite indicators that denote different dimensions of technological and social capabilities [7, 48]. Compared to panel analyses, the composite analyses consider many countries, including some LMICs. However, since most LMICs have limited data, such studies are usually static (one-year studies), ignoring how NACS evolved. Also, not all LMICs have data on all the variables of interest available for one particular year. Therefore, even composite analyses cannot possibly include all LMICs.

Generally, data availability restricts the number of countries and periods used in the analyses. Both methodologies are challenging for developing countries, particularly LMICs eligible for IDA resources, which are the foci of this study. This article contributes to alleviating the problems stemming from missingness by constructing a new complete panel dataset. A statistical technique called MICE PMM is employed to estimate the missing values in the original incomplete data sources [23]. Out of many imputation suites, this article considers MICE PMM because they are powerful, efficient, consistent, convenient, and reliable. The following section elaborates on why it is essential to adequately handle missing data and what strategies could be used to deal with missing data.

3. Properly handling missing data- why it is crucial, mechanisms underlying missing data, and strategies to handle missing data

It is essential to carefully consider the missing data problem to obtain accurate estimates of the parameters of interest in any analysis. Missing data pose many dilemmas in data analysis. The chief dilemma is that if a researcher uses original data by excluding subjects with missing data from the study, the researcher will not use all the existing information in the data, most likely causing over- or underestimated parameters (aka ‘biased parameters’). To treat bias in parameters due to the exclusion of subjects in the analysis, a researcher can impute the missing data. During the imputation process, however, the researcher should take utmost care in preserving variability found in existing data and incorporating uncertainty underlying any missing data. Therefore, employing proper and standard imputation methodologies is imperative to estimate a reliable dataset.

Provided that the imputation technique is sound, one may get reliable imputations. The first step in getting the imputation technique right essentially means being mindful of the missing data pattern and what might have caused it. The literature considers three potential mechanisms underlying missing data [49].

4. The multiple imputation method and predictive mean matching

Rubin [56] first introduced the multiple imputation methodology as an efficient statistical methodology to estimate missing values in a dataset. Several other researchers also explain this technique [56, 57]. Over the years, this methodology has evolved into various methods, catering to missingness in diverse data models. MI overcomes many of the problems associated with deletion and other single imputation techniques [58, 59]. In addition, the methodology returns efficient and accurate estimates and preserves variability, which is otherwise lost using other single imputation techniques (such as mean or cold deck imputation).

MI is valid under MAR (Missing at Random) assumption [59]. Therefore, MI estimates missing values using available, observed data [60].

Since there is uncertainty about missing data values, the estimation process is repeated m times (this step refers to the imputation stage). From the imputation stage, m complete datasets are generated. In the next stage (analysis stage), econometric analyses of interest are separately performed on m datasets. Finally, all these multiple results are combined (pooled) to obtain a final value of the coefficient of interest, for instance, regression coefficients (pooling stage). In short, a standard MI process produces multiple imputed datasets, applies a statistical analysis model to each dataset, and then integrates all analysis results to create an overall result (see Fig 1 below).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Shows a standard multiple imputation process.

In the first step (imputation stage), missing data at hand, shown in white dots, are imputed (all in blue now showing imputation happened) to create m imputed datasets. Following imputation, each imputed dataset is separately analyzed using standard methods (such as OLS regression). Lastly, the analysis results are combined using Rubin’s rules [22].

https://doi.org/10.1371/journal.pone.0274402.g001

Suppose the imputation model at the imputation stage is specified correctly and the data exhibit a normal distribution. In that case, MI yields consistent parameter estimation and confidence intervals that incorporate uncertainty because of the missing data [29]. To clarify, the correct specification of an imputation model entails the inclusion of variables considered to predict missingness and variables associated with the variable being imputed, and the outcome variable of the analysis model [29, 61].

One of the common parametric approaches for MI execution is Multivariate Normal distribution MI (MVN). This approach assumes all imputed variables to follow a joint multivariate normal distribution. Conversely, MI by Chained Equations (MICE) is a semiparametric approach that does not take a joint MVN distribution but considers a different distribution for each imputed variable [62]. Unlike MVN, MICE employs a sequential (variable-by-variable) approach while incorporating functional relationships among variables and data characteristics such as ranges. Within MICE, one can either use Linear Regression or Predictive Mean Matching (PMM) for continuous variables. This article carries out the PMM technique to impute missing values. PMM relaxes most of the assumptions of parametric MI techniques [30]. Hence, it is handy for imputing quantitative variables that are not normally distributed [63]. In the PMM, the missing value for an observation (considered as a ‘recipient’) is imputed by the observed value from another observation (called a ‘donor’) with a similar predicted mean outcome as follows [30, 64]:

In the imputation stage, for every missing value, the PMM algorithm structures a small set of donors (typically 5 or 10) from all complete cases that have predicted values closest to the predicted value for the missing value. Next, one donor is randomly drawn from the neighborhood pool. The observed value of such a donor is assigned to the missing value. This procedure is conducted m times, which generates m datasets. After the imputation stage, the analysis and pooling stages follow the same pattern as any standard MI. Like any MI, in the analysis stage m times analyses are conducted, and in the pooling stage, these results are combined to get a single estimate.

A more step-by-step computational process within the imputation stage of PMM is explained below:

Suppose there is a variable (X) that has missing values and another set of variables (Vs) to be used to impute X, the software (STATA or R) carries out the following computations in the imputation stage:

• Firstly, it estimates a linear regression of X on Vs for complete observations (those with no missing values). This step produces a set of coefficients a.
• Secondly, it randomly draws from the “posterior predictive distribution” of a (the posterior predictive distribution is the distribution of possible unobserved values conditional on the observed values [65]). This step generates a new set of coefficients a*. (this step ensures variability in the imputed values produced later on).
• Thirdly, the software uses coefficients a* to generate predicted values for X for all observations.
• Fourthly, for each observation with a missing value of X, the software identifies a set of observations with observed X (called donors or neighbors) whose predicted values are roughly close or similar to the predicted value for the observation with missing data.
• Lastly, from the neighborhood pool identified, it randomly chooses one donor and designates its observed value to fill in for the missing value.

For each completed dataset, steps 2 through 5 are conducted. The key idea is constructing the right donor pool from where observations with missing data will be matched with observations with available data [66]. Researchers have answered how many donors or neighbors should be in the donor pool [29, 66]. They assert that the size of the pool depends on sample size. In general, for most situations, these studies suggest k = 10 or k = 5. The default in the Stata MI command is k = 1.

In short, PMM is simple to perform and a versatile method. It relaxes the normality distribution assumption, which is not always observed in continuous data. Since PMM imputations are based on observed neighborhood values, they are much more realistic. Unlike other techniques such as EM or MVN, PMM does not produce imputations outside the observed values; thus, they overcome the problems with meaningless imputations. Compared to other suites such as Normal Linear Regression imputation, PMM is also less susceptible to model specification and can handle many variables irrespective of their distributions [36]. While imputing from the neighboring donor candidates, it incorporates nonlinearities (nonlinear associations among variables) and returns the same distribution for missing data present in the observed data [36].

5. MSK panel dataset

Here I am presenting the main features of the MSK dataset. The dataset has been compiled and estimated after applying the MICE predictive mean matching technique described in the previous section. The complete dataset consists of information for many pertinent variables and for all LMICs eligible for IDA support over time (panel data). Specifically, the dataset contains complete data for 47 variables for 82 countries between 2005 and 2019 (1,230 country-year observations).

This new complete dataset offers ample statistical content to conduct longitudinal comparative country analyses of national absorptive capacity systems (NACS) within LMICs. Among other valuable insights, such analyses illustrate the relative standing of LMICs. Similarly, the dataset’s time-series feature enlightens how LMICs’ NACS evolved in the last one and a half decades. Immediate use of the dataset would entail estimating the relationship between the variables within the dataset (capacities constituting NACS) and the LMICs’ economic development. Such an exercise will offer crucial lessons on economic growth and development to leading and lagging LMICs. Similarly, another use will involve clustering LMICs into different groups based on capacities scores.

Since NACS are multifaceted, any analysis of NACS would involve a large number of possibly relevant variables interacting in many ways. Therefore, the MSK dataset embraces a multidimensional operationalization of NACS. In this dataset, the NACS constitutes six capacities drawn from the literature. In addition, various incoming flows from abroad (learning, knowledge, skills, and technology) also may influence the NACS. Fig 2 represents these capacities of NACS while alluding to the incoming flows. The six capacities are: 1) Technological capacity, 2) Financial capacity, 3) Human capacity, 4) Infrastructural capacity, 5) Public Policy capacity, and 6) Social capacity. The discussion of all these capacities (and incoming flows) and how encompassing they are compared to other narrow definitions of capacities is beyond this article’s scope (please see [12] for this discussion). However, the central hypothesized idea behind this dataset’s construction is that LMICs that are severely lacking in data need to appreciate that these capacities and their dynamic interaction drive economic development and science, technology, and innovation (STI) in those economies. For this purpose, development economists and STI policymakers need access to panel statistical data (country-year observations) on these capacities, which would help them conduct empirical analyses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Shows National Absorptive Capacity System (NACS) and its capacities.

These six capacities constitute NACS. Incoming flows mediate capacities within NACS. Figure Source: Khan [12].

https://doi.org/10.1371/journal.pone.0274402.g002

Literature on NIS helped identify 64 variables, likely constituting one of these capacities in NACS. After performing imputation analysis, the list of variables was reduced. Resultantly, the MSK dataset consists of 47 variables, as shown in Table 1. As a matter of good practice, the table also compares descriptive statistics (mean, standard deviation, minimum, maximum, and observation count) of the variables in the new (complete) dataset with descriptive statistics for corresponding variables in the observed (incomplete) dataset. The last column of the table reports the share of missing data present in the original dataset. As can be seen, the missingness is very high for some variables; missingness ranges from 0.89% to about 87%. A quick look at the table shows that descriptive statistics of the two data (complete and incomplete) do not differ much. This is one of the many ways to show that the complete dataset is sufficiently reliable (this will be elaborated on in the forthcoming section).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics of new MSK dataset vs. incomplete observed dataset (for more details on the variables, please consult S2 Table).

https://doi.org/10.1371/journal.pone.0274402.t001

The dataset was constructed in five main steps (also illustrated in S1 Fig).

6. Descriptive analysis of the MSK dataset

To empirically illustrate the usefulness of the MSK dataset and how it can be used to study absorptive capacity systems across countries, I have conducted a detailed analysis in another article [12]. A brief descriptive analysis of the MSK dataset is conducted here. This analysis offers insights into the trends in capacities constituting NACS in LMICs and how they evolve over time. Three brief analyses are conducted: distribution (kernel density) of select few variables of interest within each capacity at the start, middle, and the end of the study period (i.e., 2005, 2010, and 2019); time trends (2005–2019) of the variables of interest for select countries (six countries, one from each region in our countries of study); and comparative ranking of countries based on composite capacity indices.

7. Quality check of the estimated MSK dataset

A quality check is conducted to determine the usefulness and vitality of this dataset.

As mentioned in section 5, I collected 64 variables to measure countries’ capacities to construct the database. After carrying out imputations and evaluation, I shortlisted 47 variables to be included in the dataset for an entire range of 1,230 country-year observations (15 years for 82 countries). The remaining 17 variables were rejected either because the system could not impute them (three variables) or the results produced (14 variables) were not of good quality.

In order to assess the imputation procedure and the reliability of the variables included in the MSK dataset, this article conducts a four-way quality check: first descriptive statistics of the two datasets (complete and observed) are conducted; secondly, distributions of completed and observed datasets are observed; thirdly, correlation tables of the observed and complete variables are compared; and fourthly, trends within imputations and convergence pattern are observed.

8. Conclusion and implications

Comparative country analyses on absorptive capacity and economic development in LMICs lack because of the lack of complete data availability. To address this problem, this article employed Rubin’s Multiple Imputation to impute missing values in variables. Specifically, it used Multiple Imputation by Chained Equations with a Predictive Mean Matching approach to estimate the MSK panel dataset. The dataset consisted of six country-related capacities. A total of 47 continuous variables measured these capacities. This dataset was estimated from an observed dataset containing many missing values. The complete dataset contained 82 countries for the period 2005–2019, for 1,230 country-year observations.

The MSK dataset provides a rich panel (across countries and over time) of statistical content that can be used in several ways. For instance, this dataset can be used to estimate the impact of absorptive capacities on economic growth in LMICs. Similarly, the capacities can be aggregated for different LMICs to find the relative standing of one economy vis-à-vis other economies. Further, such an exercise can be used to investigate the factors of development within leading and lagging LMICs. Finding leading and lagging economies within LMICs at the same level of development offer lessons to lagging economies on how they can catch up. Here, I demonstrated how a simple descriptive analysis of capacities within the complete dataset could be used to gain insights into the dynamic evolution of such capacities in different countries.

On the methodological front, MICE PMM for estimating dataset for the comparative analyses of capacities and economic growth in LMICs is powerful compared to other solutions such as mean imputation or deletion. MICE PMM is powerful because it retains variability in data as the imputed value is randomly taken from the suitable donor pool. Moreover, PMM is a good technique because it reduces bias by keeping information on all variables: variables for which partial data is available are imputed rather than deleted. Similarly, the technique preserves representation (by keeping all economies even if they have partial data rather than dropping them of analysis), returns accurate or realistic data (imputed data is taken from neighboring data pool), and captures dynamic evolution for all economies (which is compromised by using other imputation techniques).

However, MI returns multiple datasets, which indicates the uncertainty underlying missing data values. Thus, no matter how rigorous MI is, no imputation can claim with 100 percent certainty the accuracy of imputed values. Therefore, the dataset generated through MI must be carefully used for any analysis. The results of such analysis must make a disclaimer about the process through which the dataset was obtained. The reliability or quality check must be performed on the newly generated dataset, just as conducted for the MSK dataset. The MSK dataset generated here passed the quality check as the observed and complete dataset exhibited almost similar distributions, descriptive statistics, and correlation coefficients, and the process through which the dataset was imputed returned a healthy convergence among iterations.

As the MI-generated dataset is reliable, such a dataset can be valuable for hypothesis generation in LMICs suffering from poor data environments. Results from analyses based on original datasets for countries (and LMICs) with reasonably complete datasets can be compared with those based on imputed datasets. This may give some insights into the relative vitality of the completed dataset alongside interesting findings on what drives economic development in various countries. Moreover, future research may estimate datasets generated assuming MNAR pattern and then compare the datasets from both MAR and MNAR analyses to further investigate the strength of the MSK dataset.

Supporting information

Acknowledgments

This article was presented at the 17th Globelics International Conference, Heredia, Costa Rica, November 2021, the 35th Annual General Meeting and Conference, Pakistan Society of Development Economics, Pakistan Institute of Development Economics in Peshawar, Pakistan, November 2021, Shanghai University of International Business and Economics, Institute of Artificial Intelligence and Change Management, in Shanghai, China (online), December 2021, and the Schar School of Policy and Government, George Mason University, in Virginia, US, June 2022. I am highly grateful to the conference participants, the anonymous referees, and the editor of this journal for their valuable comments and suggestions. I also want to thank David M. Hart, James L. Olds, Maurice D. Kugler, and Lucas Núñez for their feedback on this article.