Did a Dangerous Business Climate Strike the Sunshine State’s Economy?

• David Denslow, Ph.D.

[Warning: Not user or reader friendly. Requires some knowledge of statistical methods.]

Cloud-to-ground lightning hits a random square mile in Florida more often than in any other state, 25 times a year, nearly three times the national average.  The sunshine state is also the lightning state, something Floridians think about only slightly more often than they worry about alligator attacks. Four Danish economists, however, just published a paper in Harvard’s Review of Economics and Statistics (Thomas Barnebeck Andersen, Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya, “Lightning, IT Diffusion, and Economic Growth across U.S. States,” November 2012, pp. 903-924) implying that above-average lightning density caused Florida to lose, accumulated over the decade 1997 to 2007, over three percentage points of productivity growth. Growth of output per worker, they report, was held back by lightning-caused disruption of information technology. Computers and networks are highly sensitive to brief power surges, which restrained the spread of computers and the Internet during a decade in which information technology was a major source of rising output.

The Danish economists demonstrate the role of lightning in two steps:

1. Controlling for various standard variables that might be correlated with growth, they show that from 1997 to 2007 productivity growth was slower in states with greater lightning density.
2. They show that the states with more lightning were slower to adopt information technology. Further, aside from its effect on the adoption of IT, lightning had no further effect on growth.

Thus they conclude that lightning slowed productivity growth through a single, plausible channel: deterring the use of IT at a time when it this new general technology was boosting productivity.

My purpose in this note is to weigh the evidence that lightning slowed economic growth meaningfully. If they are correct, then they have provided a new exogenous variable useful for identifying the spread of IT and its effects on growth, an extra tool for studying the determinants of growth across cities, states, and countries. They have even extended their results to the effect of the internet on political corruption. (“Does the Internet Reduce Corruption? Evidence from U.S. States and across Countries,” The World Bank Economic Review Advance Access published May 27, 2011.) There is concern that the internet, by reducing funding for investigative journalism, will reduce the exposure of corruption. The Danish economists argue, in contrast, that the web makes it easier to expose corruption. A difficulty with studying that empirically is the risk of circular causation. Corrupt politicians are likely to restrict use of the internet. Lightning, however, may provide an exogenous determinant of the spread of IT that allows them to identify its effect on corruption. They conclude that the web does indeed expose corruption.

Thus there is a lot at stake. Whether they are correct that lightning reduces growth (and increases corruption) is important to regional and urban economists who study growth, to those who analyze the spread of general purpose technologies such as IT, and to political scientists and others interested in political corruption. A similar controversy long ago was whether the railroad was crucial to the development of the United States. Robert Fogel and Stanley Engerman argued, against the consensus, that it was not, that canals were a good substitute. That debate continued for years, enriching our understanding of American economic history. Though the Dane’s lightning hypothesis is unlikely to stir up the same degree of controversy, it may contribute to our understanding of regional growth at the end of the 20th century.

Andersen and his co-authors conclude from regressions that states with frequent cloud-to-ground lightning strikes per square kilometer as a result grew more slowly after 1997.Briefly, Andersen and his co-authors conclude from regressions that states with frequent cloud-to-ground lightning strikes per square kilometer as a result grew more slowly after 1997. Actually they argue that occurred from 1992 on, but their evidence for the years before 1997 is weak, so I will discuss only 1997 on. Their reasoning is straightforward. Computers are highly sensitive to power surges and disruptions, of which lightning is a major cause. They cite a statement in The Economist, March 22,2001, that “Every year, millions of dollars are lost to seemingly insignificant power faults that cause assembly lines to freeze, computers to crash and networks to collapse ….”  They also cite the owner of a computer service business in Tampa and technical manuals about the importance of power disruptions.

My intuition, confirmed by conversations with a few owners of IT startups and one specialized investor, is that the frequency of lightning strikes is irrelevant to firms’ location and investment decisions.My intuition, confirmed by conversations with a few owners of IT startups and one specialized investor, is that the frequency of lightning strikes is irrelevant to firms’ location and investment decisions. The billionaire investor said that in his 25 years of funding IT startups across the country, lightning had never been mentioned.  Why might that be? First, a large share of the cost of dealing with power disruptions is fixed. You need as much backup and surge protection for disruptions twice a year as for ten times a year. Second, the main source of productivity growth from IT comes from major changes in organization, in firm structure. The direct cost of the IT equipment is only a fraction of the total cost of structural change. Improvements in retail efficiency, for example, arose from large changes in the structure of the industry, led by a firm headquartered in Bentonville, Arkansas, a state with twice the national density of lightning strikes.  Third, spending on IT hardware is a relatively small share of overall investment. Manufacturing firms in 2003 devoted less than 6% of their machinery investment to information technology equipment. Third, only a fraction of power disruptions are caused by lightning. Overall, the demand for computers should be inelastic with respect to the frequency of lightning strikes.

I return to the observation in The Economist.  What does “millions of dollars” mean? In 2001, U.S. GDP was just over ten trillion dollars. Suppose “millions of dollars” means $100 million. That would be one dollar out of each $10,000 of GDP, not enough to cause measurable differences in productivity growth across states. As an order of magnitude, in contrast, their econometric results imply that if the U.S. had no lightning, its GDP in 2007 would have been higher by about $200 billion. That’s two thousand times $100 million, and perhaps five thousand times the damage caused by lightning. Even if “millions of dollars” means $500 million, it is hard to believe that would measurably reduce growth in the lightning states.

And yet the Danish economists provide strong statistical evidence that lightning matters. First they demonstrate a negative relation across states between overall productivity growth and lightning from 1997 through 2007. Second, they show that the statistical significance of this negative relation survives the inclusion, one at a time, of reasonable control variables. Third, as a falsification test, they show there was no relation between lightning and growth before 1991, when the Internet was introduced. Fourth, they show that across states in 2003, lightning was inversely correlated with household computer ownership, with household internet access, and with the share of manufacturing machinery investment in IT equipment. Controlling for this last, there was no further correlation between lightning and productivity growth. 

Here’s an example of the sort of evidence they provide:

In this regression, the dependent variable “dlprod” is the percentage change in productivity, measured as state GDP per employee, over six five-year intervals beginning in 1977 and one four-year interval, 2007 to 20ll. The independent variables are dummies d1982 through 2007 for six of the intervals, with the first one suppressed to allow a constant term; the log productivity level (lprod1977 through lprod2007) at the beginning of year interval, to allow for convergence, and the log frequency of lightning flashes per square mile interacted with period dummies (lflash1977 through lflash2007) to allow for varying effects of lightning intensity over time. (The lightning frequency itself is stationary. There is no time trend and little variation from year to year.)

The notable results are: (1) significant negative effects of lightning in the periods 1997-2002, 2002-2007, and 2007-2011, and (2) no lightning effects in the periods before 1997, the approximate year IT started to boost productivity.  I have modified their regression in three ways: (1) I add the period 2007 to 2011, to take advantage of recent data; (2) I do not restrain the coefficients on the convergence variables to be the same for all periods (clearly they are not); and (3) I use a slightly more recent lightning frequency measure, since what is easily available on the web has changed. The fact that I not only confirm the statistically significant negative impact of lightning in the periods beginning in 1997 and 2002 but also find a significant negative impact for 2007-2011 adds to the robustness of their results.

Regarding their other growth regressions, one strategy I could follow would be to search for economic variables that, when added to the regression, make lightning insignificant for the period 1992-2007, essentially telling a different story about those years. I have in fact done so, from my familiarity with the economic history of those years, but that’s just one plausible story against another. For 1997-2007, however, I have not found a plausible economic alternative that makes lightning statistically insignificant. How then can there be room for doubt? The first reason is that lightning is highly correlated geographically across states, striking often where there is high humidity coupled with hot summers. As noted, Florida is the most lightning-intensive state, with a thunder corridor where Gulf and Atlantic breezes meet in hot summers. The rest of the Southeast is next, with hot summers kept humid by air from the Gulf, which more and more mixes with moisture from the Atlantic as the prevailing winds from the west blow the warm air toward the Atlantic. The most lightning-free of the continental states is Washington, farther north and bounded by the cool Pacific. In fact lightning intensity declines almost monotonically as you go northwest from Tampa to Seattle.

To illustrate this I have constructed a variable to measure the “distance” from Florida in the direction of Washington state. To do so, I have reshaped the continental U.S. into a square, and measured the share of the distance traveled along the corresponding diagonal from SE to NW, as you go from state to state. I have constrained states along the east coast to lie directly north of Florida, even though that coast goes NE. Below is the plot of the log of lightning frequency against that distance (the R-squared for the fracpoly fit is 0.77):

The graph displays a fractional polynomial fit of lightning to distance. If you include distance in the growth-lightning regressions, lightning becomes insignificant, usually with absolute t-statistics less than one. Here’s an example:

This result also holds with the use of the predicted value of lightning based on distance and with the use of various periods and covariates. This raises the possibility that some set of economic conditions correlated with being north or west or northwest slightly boosts productivity growth since 1997. Perhaps it is the quality of a given educational attainment, or being closer to rising Asia, or attracting talented people to scenery and cool summers. The challenge is to find out what. In the IT age, the closer you are to Seattle, the faster you grow. Or maybe it’s just the free-wheeling West.

Another concern is that Andersen and his co-authors, as part of their falsification tests, include 11 climate-related variables other than lightning in regressions for three ten-year periods, starting with 1977-87. Of the 33 possibilities, seven are significant at the 10% level or better. That includes six out of eleven for 1977-1987, when growth varied positively with precipitation and tornado intensity, and negatively with hail size, wind speed, elevation, and cooling degree days. Were these true climate effects, or simply correlates of hard times for the agricultural plains, mining areas, and the rust belt? When lightning is included as a control, cooling degree days has a positive impact in 1997-2007. Hot and dry boosts economic growth?  These results highlight the possibility of accidental correlations, perhaps related to the influence of Kőppen-Geiger climate zones on historical developments and their legacies or geographies.

Of lesser significance, but still a concern, is that lightning starts becoming significant in the year in which the state GDP accounts are switched from the SIC to the NAICS classification. In 1997, the Bureau of Economic Analysis used both methods, allowing a comparison. If you regress the NAIC-based productivity estimate on the SIC-based estimate, the result is:

The log of lightning frequency is significantly positively correlated with the difference between the NAIC and the SIC estimates. The coefficient is slightly larger than the subsequent reduced productivity effect over the next five years. This is another illustration of a correlation between lightning and economic structure that is hard to interpret. The correlation may relate to subtle difference in economic structure in the lightning zone that came into being before the IT revolution, but that subsequently slowed IT-related growth in a way unrelated to lightning. If such a relation exists, the challenge is to find it. Associated with the switch from SIC to NAICS, incidentally, was a noticeable increase in sigma divergence across states.

Another reason for doubting the significance of lightning is Florida.  In spite of the state’s averaging three times the national density of lightning strikes (25 per square mile versus nine), in Florida 61% of households had personal computers in 2003 and 56% used the internet at home, just shy of the nation’s 62% and 58%.  Florida would be expected to fall below the nation because of its high concentration of senior residents. In 2003, the use of computers by age of household head was

Also, in spite of Florida’s being the lightning-density high outlier, in 2003 the state’s manufacturing firms devoted 5.7% of their machinery investment to information technology, compared to the 5.0% national average. Of course the Dane’s should counter that they included Florida in their regressions and thus the evidence from the remaining states is all the stronger. But somehow I feel uneasy that the state with the most lightning by far is a manufacturing exception and nearly a household exception, with its slight shortfall easily explained by demography.

As to the lightning-dense remainder of the Southeast, low income and high poverty strike me as plausible explanations. The relation between computer use and household income from the 2003 CPS survey is

The poverty rate fits well in cross-state regressions, with the expected negative sign, but the inclusion of that variable does not knock out lightning density. One suspects that there may be non-linearities, if a contagion model of the diffusion of an innovation is at work here. But this needs to be tested with individual data. State numbers are too coarse.

Clearly I have not shown that the Danes are wrong in claiming to have uncovered an inverse relation between lightning density and post-1997 information technology use and productivity growth. All I have done is explain why I am not yet convinced. How could this hypothesis be resolved, or if not resolved at least advanced? Here are some possibilities: (1) Use MSA instead of state lightning and productivity data to increase the degrees of freedom and explore alternative hypotheses. Place MSA productivity growth in the context of a Rosen-Roback model. (2) Use individual data from the several CPS surveys to analyze household computer ownership and Internet use. (3) Use detailed sector Census of Manufacturing IT investment data to explore whether states invest differently in IT because of differences in the sectoral composition of their manufacturing. (4) Look at computer use within other countries to test the lightning/IT correlation. A casual glance at Chinese data, for example, suggests that computers came first to the provinces with the greatest lightning density.