Cobb–Douglas Production Function

Charles Cobb, a mathematician, and Paul Douglas, an economist who later became a US senator, published this in the American Economic Review in 1928. Douglas had noticed that labour's share of US manufacturing output was remarkably steady over time, and asked Cobb for a functional form that would produce that. The answer was a power function in capital and labour, and the exponents they fitted to US manufacturing data from 1899 to 1922 were a quarter for capital and three quarters for labour, with a coefficient of 1.01.

The form has three properties that explain its survival. It is easy to fit, because taking logarithms turns it into a straight line — which is exactly what the exponent brains on this page do, and it is how the 1928 estimates were obtained. The exponents have an interpretation: under competitive factor markets and constant returns, $\alpha$ is capital's share of output and $\beta$ is labour's, which is why they are called factor shares. And the elasticity of substitution between capital and labour is exactly one at every point, which is convenient and is also the form's biggest limitation.

A is a residual, and the word means what it says. Total factor productivity is not measured; it is what remains once capital and labour have been credited with whatever the exponents say they contributed. Technology lands in it. So do management quality, institutions, infrastructure, the weather, capacity utilization, mismeasurement of the capital stock, and every specification error in the model. Moses Abramovitz called the residual "a measure of our ignorance", and the phrase has stuck because it is accurate. A change in A is a description of something unexplained, not an explanation.

A also has no units of its own — it absorbs whatever units K, L and Y were measured in. Refit the same data with capital in millions instead of thousands and A changes. It does not transfer between data sets, and a value of A quoted without its measurement units is not a number.

$\alpha + \beta = 1$ is an assumption, not a finding. Cobb and Douglas imposed constant returns to scale on their fit; most published estimates since have imposed it too. When it is imposed, the fact that the estimated shares match observed factor income shares is not independent evidence for anything — it is the constraint showing up in the output. Freeing the constraint lets the data speak, and estimated returns to scale then wander around 1 in ways that depend heavily on the sample and the deflators.

The 1928 fit itself does not transfer. It described US manufacturing over a specific twenty-four years, in a period whose capital stock, labour force and technology bear little resemblance to a modern service economy. Applying an $\alpha$ of 0.25 to a software firm, where capital is nearly all intangible and the marginal cost is nearly zero, is a category error rather than an approximation.

Finally, the substitution property. An elasticity of substitution of exactly one means a 1% fall in the wage relative to the cost of capital shifts the input mix by exactly 1%, everywhere, always. Real substitution possibilities are not like that — some processes admit almost none, others a great deal — and the CES production function exists precisely to relax this. When you see a Cobb–Douglas result that hinges on how easily one input replaces another, it is worth remembering that the answer was largely built into the functional form before any data arrived.