Human capital and knowledge

NOTE: The present notebook is coded in R. It relies heavily on the tidyverse ecosystem of packages. We load the tidyverse below as a prerequisite for the rest of the notebook - along with a few other libraries.

\(\rightarrow\) Don’t forget that code flows sequentially. A random chunk may not work if the previous have have not been executed.

library(tidyverse)     # Package for data wrangling
library(readxl)        # Package to import MS Excel files
library(latex2exp)     # Package for LaTeX expressions
library(quantmod)      # Package for stock data extraction
library(highcharter)   # Package for reactive plots
library(ggrepel)       # Better labels on plots 
library(WDI)           # To access World Bank data
library(broom)         # Neat regression output

Side note

The function below will be used to fill in missing data points, on a country-by-country basis (via grouping).
We allow copy-pasting the most recent value up to a maximum of six years.
This assumes some form of stability/persistence in the imputed variable.

impute <- function(v){ # Imputation function
    n <- 6
    for(j in 1:n){
        ind <- which(is.na(v))
        if(length(ind) > 0){if(ind[1]==1) {ind <- ind[-1]}}
        v[ind] <- v[ind-1]
    }
    return(v)
} 

The content of the notebook is heavily inspired from the book Advanced Macro-economics - An Easy Guide.

Context & motivation

The models we have covered so far do not allow for strictly positive growth in equilibrium. There are several ways in which they can be changed in order to try to foster endogenous growth. One might be that we can include some human capital into the model and can expect that better health and education can improve productivity.

But first, let’s see if the Human Capital Index from the World Bank can explain cross-sectional variations in growth across countries.

library(WDI)
wb_raw <- WDI(indicator = c("gdp_per_capita" = "NY.GDP.PCAP.KD",
                            "population" = "SP.POP.TOTL",
                            "gdp" = "NY.GDP.MKTP.CD",
                            "human_capital" = "HD.HCI.OVRL",
                            "gdp_growth" = "NY.GDP.MKTP.KD.ZG"), 
              #country = c('FR','US'), # here you can specify some countries
              extra = TRUE,
              start = 1960, 
              end = 2025) |>
  mutate(across(everything(), as.vector)) |>
  select(-status, -lending, -iso2c, -iso3c) |> 
  filter(region != "Aggregates", income != "Aggregates") |>
  arrange(country, year) |>
  group_by(country) |>
  mutate(across(everything(), impute)) |>
  mutate(HC_change = human_capital/lag(human_capital) - 1) |>
  ungroup()

Checking if imputation has worked.

wb_raw |> 
  select(country, year, population, gdp, human_capital) |>
  tail(7)

country	year	population	gdp	human_capital
Zimbabwe	2019	15271368	25715657177	0.4612416
Zimbabwe	2020	15271368	25715657177	0.4668931
Zimbabwe	2020	15526888	26868564055	0.4668931
Zimbabwe	2021	15797210	27240507842	0.4668931
Zimbabwe	2022	16069056	32789657378	0.4668931
Zimbabwe	2023	16340822	35231369343	0.4668931
Zimbabwe	2024	16634373	44187704410	0.4668931

Population and GDP have been updated in the data, but human capital has not…

In the cross-section

Now, let’s compute the average growth rate and GDP per capita of countries and plot them against human capital.
NOTE: the data is relatively sparse and the last point is in 2020. But still…

wb_raw |>
  filter(is.finite(human_capital), ) |>
  group_by(country, region) |>
  summarise(avg_growth = mean(gdp_growth, na.rm = T),
            avg_wealth = mean(gdp_per_capita/1000, na.rm = T),
            avg_human_capital = mean(human_capital, na.rm = T)) |>
  ggplot(aes(x = avg_human_capital, y = avg_wealth)) + geom_point() +
  geom_smooth(se = F) + scale_y_log10() + 
  theme_classic() + ylab("Average GDP per capita ($K, log-scale)") + xlab("Average Human Capital Index (HCI)") +
  geom_text_repel(aes(label = country)) 

The relationship is this case is pretty clear.

wb_raw |>
  filter(is.finite(human_capital)) |>
  group_by(country) |>
  summarise(avg_growth = mean(gdp_growth, na.rm = T),
            avg_wealth = mean(gdp_per_capita, na.rm = T),
            avg_human_capital = mean(human_capital, na.rm = T)) |>
  filter(avg_growth < 10) |>
  ggplot(aes(x = avg_human_capital, y = avg_growth)) + geom_point() +
  theme_classic() + ylab("Average GDP growth rate") + xlab("Average Human Capital Index (HCI)") +
  geom_text_repel(aes(label = country)) + geom_smooth(se = F, method = "lm")

Things are much less clear for economic growth…

Time-series

Let us now turn to single-country time-related links.
We estimate the following model: \[\Delta GDP_{t+1} = a + b X_t+ e_{t+1},\] where \(X_t\) is the variable of interest and \(\Delta GDP\) is in fact growth of GDP per capita.
This should measure the correlation (not causality!) between a high human capital at time \(t\) and high growth at time \(t+1\).

lm(gdp_growth ~ lag(human_capital), data = wb_raw |> filter(country == "United States")) |> broom::tidy()

term	estimate	std.error	statistic	p.value
(Intercept)	-0.2933042	9.653291	-0.0303839	0.9760349
lag(human_capital)	3.8652608	13.722084	0.2816818	0.7808202

The link (as measured through the \(t\)-statistics), is weak.

We could also investigate the link between growth and change in human capital. Indeed, there is a problem in the above model: the independent variable (human capital) is persistent. Given the fact that the regression is predictive, this causes autocorrelation in residuals and thus perturbs inference (the standard errors are under-estimated).

One “simple” solution to this is to instead consider variations in human capital (when it improves or deteriorates and the effect this has on future economic output).

lm(gdp_growth ~ lag(HC_change), data = wb_raw |> filter(country == "United States")) |> broom::tidy()

term	estimate	std.error	statistic	p.value
(Intercept)	2.448426	0.3023969	8.0967321	0.0000001
lag(HC_change)	10.271897	21.6312544	0.4748637	0.6400285

The results is not really more compelling.

Let us nonetheless turn to some theory!

The model

Production

In this section, the change comes from the production function:

\[Y = K^\alpha H^\beta (AL)^\gamma,\]

where \(H\) stands for human capital and returns to scale are constant, increasing or decreasing if \(\alpha+\beta+\gamma\) is equal to one, strictly above, or strictly below. The productivity (or technological) factor \(A\) is associated to labor in this case. Dividing by \(L\) under CRS, we get the per-capita version:

\[y = A^\gamma k^\alpha h^\beta , \tag{1}\]

Evolution through saving & depreciation

As in the simpler models (e.g., à la Solow), we assume savings rates for physical and human capital \(s_k\) and \(s_h\); as well as depreciation rates, too: \(\delta_k\) and \(\delta_h\).

Side note

Whether human capital depreciates should be open to debate…

Then, time-variations are given by

\[\dot{K} = s_kY-\delta_k K \quad \Leftrightarrow \quad \frac{\dot{K}}{L}=s_ky-\delta_kk\] \[\dot{H} = s_hY-\delta_h H \quad \Leftrightarrow \quad \frac{\dot{H}}{L}=s_hy-\delta_hh. \] Moreover, by definition, \(\dot{k}=\dot{K}/L-nk\) and \(\dot{h}=\dot{H}/L-nh\), where \(n\) is the constant growth rate of the population \((L)\). Plugging this into the above expressions, we get \[\dot{k}=s_k A^\gamma k^\alpha h^\beta-(\delta_k+n)k, \quad \text{and} \quad \dot{h}=s_h A^\gamma k^\alpha h^\beta-(\delta_h+n)h,\] and the corresponding relative growth rates are: \[\gamma_k=\frac{\dot{k}}{k}=s_k A^\gamma k^{\alpha-1} h^\beta-(\delta_k+n)\] \[\gamma_h=\frac{\dot{h}}{h}=s_h A^\gamma k^{\alpha} h^{\beta-1}-(\delta_h+n).\]

Steady states

The balanced growth path requires that \(\gamma_k\) and \(\gamma_h\) be constant. Since \(\delta_k+n\) and \(\delta_h+n\) are already constant, we need that \(s_k A^\gamma k^{\alpha-1} h^\beta\) and \(s_h A^\gamma k^{\alpha} h^{\beta-1}\) be constant. To simplify these quantities, we take their log (they, too, have to be constant), and we differentiate with respect to time (log(\(s_k\)), log(\(A^\gamma\)), and log(\(s_h\)) are constant so they vanish): \[\frac{\partial}{\partial t} \left((\alpha-1) \log(k)+\beta \log(h) \right)=(\alpha-1) \frac{\dot{k}}{k}+\beta \frac{\dot{h}}{h}=(\alpha-1)\gamma_k+\beta \gamma_h=0\] \[\frac{\partial}{\partial t} \left(\alpha \log(k)+(\beta-1) \log(h) \right)=\alpha \frac{\dot{k}}{k}+(\beta-1) \frac{\dot{h}}{h}=\alpha\gamma_k+(\beta-1) \gamma_h=0.\] Both l.h.s. expressions have to be equal to zero, i.e, from the second equation, \(\gamma_h=\frac{\alpha}{1-\beta}\gamma_k\) and substituting this into the first equation, \[\left((\alpha-1)+\frac{\beta \alpha}{1-\beta}\right) \gamma_k=\frac{(\alpha-1)(1-\beta)+\alpha \beta}{1-\beta}\gamma_k=\frac{\alpha+\beta-1}{1-\beta} \gamma_k=0.\] If we assume constant returns to scale, \(\alpha+\beta < 1\) (because else \(\gamma=0\) and in this case, labor and productivity are out of the game, which is not what we want!). And obviously, \(\beta > 0\) for the same reason (human capital), hence this imposes \(\gamma_k=0\).

Important

Again, this means zero growth in equilibrium!

We are not really more advanced - but this was to be expected because the model is too Solow-like (the basic mechanisms are the same).

Another model

This section follows Barro and Sala-ì-Martin, Section 5.1.
The production is of Cobb-Douglas type:

\[Y=AK^aH^{1-a}, \quad a \in (0,1),\] with \(H\) being human capital. The resource constraint is: \[Y=C+I_K+I_H,\] where \(I_K\) and \(I_H\) are the investments in capital and in the workforce. Importantly, consumption remains the only control variable, while investments are state variables. Moreover, \[\dot{K}=I_K-\delta_KK, \quad \dot{H}=I_H-\delta_HH.\]

The population seeks to maximize utility across generations: \(\int_0^\infty e^{-\rho t}u(C_t)dt\).
Given the 3 constraints listed above, the Hamiltonian is \[J=e^{-\rho t}u(C)+\nu(I_K-\delta_KK)+\mu(I_H-\delta_HH)+\omega(AK^aH^{1-a}-C-I_K-I_H).\]

The model then assumes a CRRA utility function, written as \(u(x)=x^{1-\theta}/(1-\theta)\). We have

\[\frac{\partial J}{\partial C}=e^{-\rho t}C^{-\theta} -\omega, \quad \frac{\partial J}{\partial I_K}=\nu -\omega, \quad \frac{\partial J}{\partial I_H}=\mu -\omega \] and setting all these expressions to zero, we get \[\omega = \nu = \mu = e^{-\rho t}C^{-\theta}.\] Hence, differentiating with respect to t, we obtain (equating with \(\partial J/\partial K\) and \(\partial J/\partial H\) because of the FOC for state variables),

\[\dot{\nu}=-\rho e^{-\rho t}C^{-\theta}-\theta e^{-\rho t}C^{-\theta-1} \dot{C}=\frac{\partial J}{\partial K}=\nu \delta_K-\omega a A K^{a-1}H^{1-a}\]

\[\dot{\mu}=-\rho e^{-\rho t}C^{-\theta}-\theta e^{-\rho t}C^{-\theta-1} \dot{C}=\frac{\partial J}{\partial H}=-\mu \delta_H-\omega (1-a) A K^{a}H^{-a}\]

And since \(\omega = \nu=\mu\), it must hold that \[\frac{\dot{\nu}}{\nu}=\delta_K-aAK^{a-1}H^{1-a}, \quad \frac{\dot{\mu}}{\mu}=-\delta_H+(1-a)AK^{a}H^{-a}.\] At the same time, from the expressions of \(\dot{\nu}\) and \(\dot{\mu}\), we derive \[\frac{\dot{\nu}}{\nu}=-\rho -\theta \frac{\dot{C}}{C}, \quad \frac{\dot{\mu}}{\mu}=-\rho -\theta \frac{\dot{C}}{C}.\]

Hence,

\[\frac{\dot{C}}{C}=\theta^{-1}(aA(K/H)^{a-1}-\delta_K-\rho)=\theta^{-1}((a-1)A(K/H)^{a}-\delta_H-\rho) \tag{2}\]

If we set \(\delta_K=\delta_H\) (big assumption!), it must hold that \(aA(K/H)^{a-1}=(a-1)A(K/H)^{a}\), i.e., \[K/H=a/(1-a). \tag{3}\] We can then plug this in the above equation to get:

\[\frac{\dot{C}}{C}=\theta^{-1}(Aa^a(1-a)^{1-a}-\delta-\rho)\]

If \(Y=AK^aH^{1-a}= AK (K/H)^{a-1}\) and Equation 3 holds, we have \[Y=AK\left(\frac{a}{1-a} \right)^{a-1}.\] \(\rightarrow\) The AK model! More on that soon.

If \(\delta_H \neq \delta_K\),

\[aA(K/H)^{a-1}-\delta_K=(a-1)A(K/H)^{a}-\delta_H,\] and \(K/H\) does not have a simple expression (when end up with an algebraic equation).

Two imperfect avenues towards growth

Returns to scale

One reason why growth is not possible in the extended Solow model is the assumption of constant returns to scale. Indeed, if we allow \(\alpha+\beta=1\) (and \(\gamma >0\), so that total returns to scale are increasing), then \(\gamma_k \neq 0\). Ok, but that’s an easy one.

And, in practice, at the aggregate level, returns seem to be close to, or slightly smaller than, one (but with sector heterogeneity), see for instance:

• Returns to Scale in U.S. Production: Estimates and Implications
  
• On the estimation of returns to scale, technical progress and monopolistic markups
  
• On estimating optimal α-returns to scale

Exogenous growth

If \(\gamma_k= s_k y/k-(\delta_k+n)\) and \(\gamma_h=s_h y/h-(\delta_h+n)\) are constant, then \(k/y\) and \(h/y\) are constant and all variables grow at the same pace. Let’s take log-output:

\[\log(y)=\gamma \log(A) + \alpha \log(k) +\beta \log(h), \]

hence, differentiating with respect to time (recalling a Cobb-Douglas function \(y=A^\gamma k^\alpha h^\beta\)), \[\frac{\dot{y}}{y} = \gamma \frac{\dot{A}}{A}+\alpha \frac{\dot{k}}{k} + \beta \frac{\dot{h}}{h}\]

Since \(x = \frac{\dot{y}}{y} =\frac{\dot{k}}{k} = \frac{\dot{h}}{h}\), if we write \(g\) for the relative growth rate of \(A\) we get \[x = \gamma g + \alpha x + \beta x,\] and assuming constant returns to scale \(1-\alpha-\beta=\gamma\), we get \(x=g\), i.e., all growth rates are equal to \(g\), the rate of increase of the productivity factor \(A\). But this is not endogenous growth…

The AK model

Assumptions and BGP

This is a special model because it assumes

\[y=Ak,\] hence its name.

As in the Ramsey model, if we assume \(n=0\), capital evolves as: \[\dot{k}_t=y_t-c_t=Ak_t-c_t. \tag{4}\]

We then set the utility function: \[u(c_t)=\left(\frac{\sigma}{\sigma-1} \right)c_t^{(\sigma-1)/\sigma},\] which implies \(u'(c_t)=c_t^{-1/\sigma}\) and \(u''(c_t)=-\sigma^{-1}c_t^{-1-1/\sigma}\) so that indeed \(-\frac{u'(c)}{u''(c)c}=\sigma\).

As before, the representative agent seeks to maximize discounted utility over an infinite horizon: \(\int_0^\infty u(c_t)e^{-\rho t}dt\).

The corresponding Hamiltonian is \[H=\left( \frac{\sigma}{\sigma - 1} \right)c_t^{(\sigma-1)/\sigma}+\lambda_t (A k_t - c_t),\] leading to

\[\frac{\partial H}{\partial c_t}=c_t^{-1/\sigma}-\lambda_t=0\] and, for the co-state variable, to \[\dot{\lambda}_t=-\frac{\partial H}{\partial k_t}+\rho \lambda_t=(\rho-A)\lambda_t.\]

The first FOC implies \(\lambda_t=c_t^{-1/\sigma}\), hence \(\dot{\lambda}_t=-\sigma^{-1}c_t^{-1-1/\sigma} \dot{c}_t\). If we divide this by the FOC, we get

\[\frac{\dot{\lambda}_t}{\lambda_t}=-\sigma^{-1} \frac{\dot{c}_t}{c_t}\]

From the second FOC, this implies \[\frac{\dot{c}_t}{c_t}=\sigma(A-\rho). \tag{5}\]

For the BGP, the variables need to grow at constant speed. This is already the case for consumption, but what about capital? From Equation 4, we have \[\frac{\dot{k}_t}{k_t}=A-c_t/k_t, \tag{6}\] hence we need the same growth rate for \(c_t\) and \(k_t\). In the model, the BGP is such that all variables, including output \(y_t\) grow at the same pace, \(\sigma(A-\rho)\). In addition, from Equation 6, we have \[\sigma(A-\rho)=A-c_t/k_t \Leftrightarrow k_t = c_t (\sigma(A-\rho)-A)\]

Hence, consumption is exactly proportional to capital (and output).

The transversality condition

In the model, the TVC \(e^{-\rho t}k_t \lambda_t \rightarrow 0\) is \(e^{-\rho t}k_t c^{-\sigma^{-1}_t} \rightarrow 0\). The differential Equation 5 has simple solution \(c_t=c_0e^{(\sigma(A-\rho))t}\). But remember that this is also true for \(k\): \(k_t=k_0 e^{(\sigma(A-\rho))t}\), hence the condition becomes \(k_0c_0^{-\sigma^{-1}}e^{(\sigma(A-\rho)-A)t}\rightarrow 0\), which requires \(\sigma(A-\rho)<A\).

Simulations

TO DO !! (in class?)

Knowledge & learning by doing

To generate growth, one way may be to rely on some form productivity. From a human capital perspective, the idea here is to assume that experience gets workers better at production. Below, we follow the survey Learning by doing, Section 5. See also On the mechanics of economic development by Lucas, wherein technology is put forward as driver of human capital (good transition with the next session!).

\[Y=AK^aL^{1-a},\] with \(A\) being knowledge. Labor increases as usual exogenously at rate \(n\), while \(K\) satisfies \(\dot{K}=sY\), i.e., via savings and without depreciation. Knowledge is linked to cumulative investment (e.g., R&D) via \(A=bK^\lambda\).

We simplify the production function to

\[y=Ak^a \tag{7}\]

and laws of motion are (same as in the Solow model below):

\[\frac{\dot{k}}{k}=sAk^{a-1}-n \tag{8}\]

and \[\frac{\dot{A}}{A}=\lambda n+\lambda \frac{\dot{k}}{k}. \tag{9}\]

Indeed, \(\dot{A}=b\lambda K^{\lambda-1}\dot{K}=\lambda A \frac{\dot{K}}{K}=\lambda A(\dot{k}/k+n)\) where the final equality comes from \(\dot{K}/K=sY/K=sA(L/K)^{1-a}=sAk^{a-1}=\dot{k}/k+n\) (from Equation 8).

This implies

\[\frac{\dot{y}}{y}=\frac{\dot{A}k^a+Aa k^{a-1}\dot{k}}{Ak^a}=\frac{\dot{A}}{A}+a\frac{\dot{k}}{k}=\lambda n+(\lambda +a)(sAk^{a-1}-n).\] In short, we have \(\frac{\dot{y}}{y}=\lambda n+ (\lambda+a)\frac{\dot{k}}{k}\): output growth is an affine function of capital growth. If we assume \(\frac{\dot{k}}{k}=0\), then growth comes from learning only if \(n>0\). We can find a situation where all growth rates are positive. It corresponds to \(\frac{\dot{y}}{y}=\frac{\dot{k}}{k}\) and \(\frac{\dot{A}}{A}=(1-a)\frac{\dot{k}}{k}\). This yields \[\frac{\dot{k}}{k}=\frac{\dot{y}}{y}=\frac{\lambda n}{1-a-\lambda}\] Note that \(n=0\) requires \(a+\lambda=1\) (constant returns to scale) to obtain nonzero growth. If population grows, then it must hold that \(a+\lambda<1\) (DRS).

NOTE: a simple alternative model can be found in Economic Growth, by Barro and Sala-i-Martin, Section 4.3.

Note

In the models above, production depends on human capital or knowledge.
But the reverse causal relationship my also hold (output \(\rightarrow\) human capital)…