A Rising Tide Lifts All Boats? IT growth in the US over the last 30 years

Size: px

Start display at page:

Download "A Rising Tide Lifts All Boats? IT growth in the US over the last 30 years"

Rosanna Warren
5 years ago
Views:

1 A Rising Tide Lifts All Boats? IT growth in the US over the last 30 years Nicholas Bloom (Stanford) and Nicola Pierri (Stanford)1 March 25 th ) Executive Summary Using a new survey of IT usage from 1987 to 2014 we establish some stylized facts about the adoption of Information Technology. 1. This has been a gradual process with a steady increase of our key measure of PCs per employee since 1987, rather than a sudden burst from the mid-1990s to the mid-2000s as implied by the timing of the IT productivity miracle. 2. This increase in IT intensity has occurred across all US firms rather than being concentrated in certain high-tech firms. In particular, the 25 th, 50 th and 75 th percentiles of firm-level IT intensity have increase at about the same rate across firms over this period 3. This increase in IT intensity has also been reasonably balanced geographically. While California and the East Coast has a higher level of IT intensity, all areas of the country have seen similar trends in the growth of IT intensity 4. The two major drivers of IT intensity appear to be education in particular the share of employees in a geography with a college degree and industry. As such, the continued increase in the education levels of US workers will naturally lead to an upward drift in IT intensity over and above that generated by technological advance. 1 Contact details: 579 Serra Mall, Stanford CA 94305, nbloom@stanford.edu. We are grateful for financial support from the Toulouse Network for Information Technology.

2 2) Our Information Technology Data The paper combines a variety of datasets, most obviously the Aberdeen Ci Technology Database. The Ci Technology Database is a product of the marketing company Aberdeen group who collects information on hardware and software installed in thousands of establishments worldwide. Data are constructed from interviews with plants' IT-decision makers. The same data source (formerly known as Harte Hanks) has been extensively used in economics literature,2 and it is considered to be relatively highly reliable because of disciplined enforced by Aberdeen's customers' directing marketing and sales efforts based on CiTDB. We access several waves of the survey of plants in USA: all the years between 1986 and 1996, plus 1999, 2003, , 2008, 2009, 2010, 2012 and Figure 2 shows this widening coverage of the survey over time. One break point is evident: between 1995 and 1996 the number of counties where at least one plant was surveyed increases from around 2 thousands to more than 3 thousands.3 Hence, in our analysis we are careful to check results are robust to a balanced and unbalanced panel of plants, plus including controls for survey coverage. The set of variables included in the database has been evolving over more than three decades of activity. Information about the number of personal computers installed and on number of employees working in the establishment are available for all years. Therefore, we choose the number of PCs per employee as our main proxy for plant IT adoption. Each wave collects also information about the adoption of latest or trending technologies, at least for a subset of the interviewed plants. For instance, later editions document the diffusion of Cloud-related solutions. Furthermore, estimates of plants budget for IT spending are available from 2010 on. Consequently, we compute three measures of counties IT intensity: 1. Average of PCs per employee of all the plants present in the county.4 This measure proxies for stock of IT capital installed. 2. Average IT budget per employee. This measure is available for most counties in 2010, 2012 and While it cannot be used for long-term comparisons, it includes both software and hardware IT investments and it weights each new purchase with its market price. It aims to capture the flow, rather than the stock, of IT capital. 2 See, for instance Bresnahan et al (2002), Beaudry et al (2010), Forman, Goldfarb, and Greenstein (2012), and Bloom et al (2012, 2016) 3Moreover, between 2009 and 2010 the number of surveyed establishment goes from around 50 thousands to more than 300 thousands. 4 We winsorize the top 2% of establishments to mitigate problems related to measurement errors.

3 3. Share of plants adopting any Software-as-a-Service solution. This measure captures establishments adoption of Cloud Computing technologies, and it proxies for counties closeness to the technology frontier. Furthermore, for each county we compute the share of establishments in Aberdeen Ci Technology Database belonging to each SIC 2 digits industry. We supplement the analysis with county-level educational attainment and other information from Decennial Census and American Community Survey. We use the Rural-Urban continuum classification of counties provided by provided by US department of Agricultural Economic Research (1993, 2003 and 2013). We collect data on share of individuals holding a scientific or engineering bachelor degree in 2009 for almost 8 hundreds counties. A set of descriptive statistics is included in Table 1. 3) The roll-out of IT across the US Figure 1 shows the roll out of computers across the US, showing a steady increase since the late 1980s. Interestingly, despite the period being associate with the IT productivity miracle and a period of rapid roll-out of IT, we do not see any particular acceleration in the adoption rates of computers. Indeed one striking fact from this graph is the steady roll-out of computerization across the US - both on average but also at the lower and upper tails in terms of the 25th and 75th percentiles of computer use. Hence, the last 25 years has seen a gradual adoption of computerization - at least as measured by PCs per employee in US firms - across the broad range of firms in the US. As a second measure of this we look at the histogram of PCs per employee for our first year of good data (1987), our last year (2014) and two years in-between (1996 and 2003). In Figure 3 we see a steady movement to the right of the computer intensity over time - again this supports the view that computerization has not been driven in the US particularly by a tail of hi-tech firms, but instead has been gradual and broad-based over the last 20 years. Another measure of this is the spread of computerization geographically as shown by Figure 4. We can see in our first map in 1987 that computerization was heavily concentrated in the West Coast (Northern California and Washington State) and around New York and Boston. As time progresses again we see a gradual and steady increase in computerization spreading across the US, with the map strongly dark green by around 2010 as pretty much all areas of the country

4 have experienced heavy computer adoption. So much like as we saw across firms the geographical spread of adoption exists with some low and high adoption regions, but this spread has not increased over time as all areas have experienced gradual increases in IT intensity. 4) The drivers of IT Adoption What does explain heterogeneity in adoption patterns across US counties? The main drivers in our analysis are county industrial compositions and education. To perform such analysis we run county-level IT adoption measures (PCs per employee, IT budget per employee and share of plants adopting any SaaS solution) on a set of county characteristics: share of people above 25 with bachelor degree or higher, natural logarithm of median household income, Rural-Urban classification in 2013 (RUCC) and natural logarithm of counties population. Furthermore, we include the share of graduates holding a scientific or engineering degree in 2009 as a proxy for STEM intensity of the educated labor force5. Finally, for each county we compute the share of plants belonging to each SIC2 industry to create a set of controls for sectoral composition of local economy. We only include counties with at least 15 surveyed plants and weight each observation for number of establishments in the county. Results are presented in Table 2 which shows that, first, the share of a county with a bachelor degree (or higher) is strongly and robustly predictive of computer adoption. This simple measure of education of the workforce explains alone almost a quarter of the variation in the county stock of IT capital, as shown by the R square of column (1). Hence, areas with skilled employees have seen persistently stronger adoption rates of computers - the results are for 2014 but we find very similar patterns looking at any other year. In column (1) we see a coefficient of on the share of employees with a college degree which means going from 10% to 50% in a county - which is equivalent to moving from the bottom 5 to the top 2 percentile, would be associated with an increase of computers per person of 0.07, which is about 10% of the sample average (see the descriptives statistics in Table 1). In column (2) we include a measure of STEM intensity of the highly educated workforce and we see the coefficient on the main educational variable (only) slightly decreases. In columns (3) and (4) we add a set of controls for other income and population factors and indeed find the educational variable becomes even more important. Finally in column (5) we add an industry control and we see a big increase in 5 Unfortunately, we obtained educational attainment data divided by area of study only for 792 US counties. Therefore we impute the median value of STEM intensity to all others. Results are robust to exclusion of these imputed counties.

5 the R square of the regression fit, which jumps from 28% to almost 70%, indicating how important industry variation is for predicting computer intensity. We perform the same analysis using the two alternative measures of IT intensity. For each plant in the survey, we compute establishment level IT spending per employee. Then,6 we take the logarithms of county-level averages as measure of local IT investment. Results in Table 3 indicate not only the stock of current computers but also the flow of new expenditure is strongly related to regional skills levels. An increase of one standard deviation of the share of individuals with a bachelor degree or higher is associated with an increase in the mean IT budget per person of around 1.5%. Notice that, while education seem to explain a smaller share of the variance of the flow of IT investment than of the stock,7 the inclusion of other county-level controls magnifies the skill-technology relation. Furthermore, we investigate the relation between skills availability and the adoption of frontier IT. For each county, we compute the share of establishments deploying any Software as a Service solution in 2014 and use it as a measure of Cloud Computing. Table 4 reveals that education is strong predictor of the presence of Cloud solutions. In fact, 38% of the variance of the dependent variable, can be explained by the share of individuals with bachelor or higher. Going from 10% to 50% in a county increases the share of establishment using Cloud Computing by 10%, which is equivalent to go from the bottom 1% to the top 10% of county adoption. Furthermore, STEM intensity is another important factor, being positively associated with SaaS in any regressions. In Table 5 we regress the growth of PCs per employee8 between 1996 and 2014 (rather than the levels) against the growth of education and our other key drivers, and again see a computerization-skills relationship. Interestingly, there is also a heavy degree of mean reversion in that counties with historically high levels of computer use have lower rates of growth (as indicated by the strong negative on the lagged level of PCs per employee). The level of IT in 1996 adoption explains 96% of the subsequent growth rate. Therefore, a substantial geographical convergence between low and high IT counties has taken place during the last 20 years. The pattern of the spread of PCs seem to be more of a tide lifting all boats rather than contributing to geographical inequality. 6 After winsorization of top 2% of plants. 7 The R square in column (1) of Table 3 is only 4.3%. 8 We compute variables growth rates according to the formula of Davis, Haltiwanger, and Schuh (1996): 2

6 Bibliography 1. Beaudry, Doms, Lewis Should the Personal Computer Be Considered a Technological Revolution? Evidence from U.S. Metropolitan Areas, Journal of Political Economy 2. Bloom, Sadun & Van Reenen Americans do I.T. Better: US Multinationals and the Productivity Miracle, American Economic Review 3. Bresnahan, Brynjolfsson and Hitt, 2002, Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence, The Quarterly Journal of Economics 4. Davis, Haltiwanger and Schuh, 1996, Job Creation and Destruction, Cambridge: MIT Press 5. Forman, Goldfarb and Greenstein The Internet and Local Wages: A Puzzle, The American Economic Review

7 Table 1: Summary Statistics at level PCs per employee IT budget per employee Share of Plants adopting any SaaS Share of Population with Bachelor or Higher Share Graduates with Science or Engineering Degree Median Family Income Population Whole Panel Mean Median Standard Deviation N 48,743 9,370 3,036 48, ,701 48,688 Year 1987 Mean Median Standard Deviation N 1, ,841. 1,841 1,841 Year 2014 Mean Median Standard Deviation N 3,136 3,136 3,036 3,136. 3,136 3,136 Notes: Sources are Aberdeen Ci Technology Database, American Community Survey and US Decennial Census. IT budget is in US dollars, Median Income is in thousands of US dollars, Population is in thousands of individuals. Columns are blank when the survey contains no information on that variable in that year.

8 Table 2: IT adoption in 2014 and some county characteristics - PCs per employee VARIABLES (1) (2) PCs per employee PCs per employee (3) PCs per employee (4) PCs per employee (5) PCs per employee Education 0.181*** 0.160*** 0.184*** 0.198*** *** (0.017) (0.023) (0.025) (0.020) (0.011) ST EM (0.052) (0.049) (0.043) *** (0.023) Median Family Income 0.005*** (0.002) (0.001) Population *** ** (0.003) (0.002) RUC C SIC2 Controls Observations 2,995 2,995 2,995 2,994 2,994 R-squared Notes: the dependent variable is the county-average number of PCs per employee for all plants surveyed by Aberdeen Ci Technology Database. Education is the share of individuals above 25 years old holding a bachelor degree or higher. STEM is the share of individuals holding science or engineering Bachelor degree (or higher) over total number of individuals with a Bachelor degree (or higher); it refers to Median Family Income is the logarithm of the median income of families living in a county as reported by ACS and Decennial Census. Population is the logarithm of the number of individuals living in a county. RUCC is a set of fixed effects for counties Rural-Urban continuum classification by US department of Agricultural Economic Research in SIC2 Controls are the shares of plants (in Ci Technology Database) belonging to each SIC2 industry. Standard errors in parentheses are clustered at State level. Observations are weighted by number of plants *** p<0.01, ** p<0.05, * p<0.

9 Table 3: IT adoption in 2014 and some county characteristics - Logs of IT budget VARIABLES (1) IT budget per employee (2) IT budget per employee (3) IT budget per employee (4) IT budget per employee (5) IT budget per employee 0.313*** (0.0253) 0.327*** (0.0652) *** Education 0.180** 0.192* 0.421*** 0.456*** (0.0827) (0.102) (0.0988) (0.0881) ST EM (0.101) (0.0671) (0.0638) Median Family Income *** ( ) ( ) Population *** *** (0.0104) ( ) RU CC SIC2 Controls Observations 2,995 2,995 2,995 2,994 2,994 R-squared Notes: the dependent variable is the natural logarithm of the county-average IT budget per employee for all plants surveyed by Aberdeen Ci Technology Database. Education is the share of individuals above 25 years old holding a bachelor degree or higher. STEM is the share of individuals holding science or engineering Bachelor degree (or higher) over total number of individuals with a Bachelor degree (or higher); it refers to Median Family Income is the logarithm of the median income of families living in a county as reported by ACS and Decennial Census. Population is the logarithm of the number of individuals living in a county. RUCC is a set of fixed effects for counties Rural-Urban continuum classification by US department of Agricultural Economic Research in SIC2 Controls are the shares of plants (in Ci Technology Database) belonging to each SIC2 industry. Standard errors in parentheses are clustered at State level. Observations are weighted by number of plants. *** p<0.01, ** p<0.05, * p<0.1

10 Table 4: IT adoption in 2014 and some county characteristics - Cloud Computing (2) Share of Plants using SaaS in the (3) Share of Plants using SaaS in the (4) Share of Plants using SaaS in the 0.211*** 0.197*** 0.206*** (0.029) (0.034) (0.0337) ST EM 0.148** 0.131** 0.114* (0.063) (0.061) (0.0592) Med. Family Income 0.006*** (0.002) Population (0.004) RU CC VARIABLES (1) Share of Plants using SaaS in the 0.253*** Education (0.032) (5) Share of Plants using SaaS in the 0.179*** (0.021) 0.147*** (0.050) 0.002* SIC2 Controls (0.001) (0.003) Observations 2,960 2,960 2,960 2,959 2,959 R-squared Notes: the dependent variable is the share of plants surveyed by Aberdeen Ci Technology Database adopting any Software-as-a-Service solution. Education is the share of individuals above 25 years old holding a bachelor degree or higher. STEM is the share of individuals holding science or engineering Bachelor degree (or higher) over total number of individuals with a Bachelor degree (or higher); it refers to Median Family Income is the logarithm of the median income of families living in a county as reported by ACS and Decennial Census. Population is the logarithm of the number of individuals living in a county. RUCC is a set of fixed effects for counties Rural-Urban continuum classification by US department of Agricultural Economic Research in SIC2 Controls are the shares of plants (in Ci Technology Database) belonging to each SIC2 industry. Standard errors in parentheses are clustered at State level. Observations are weighted by number of plants. *** p<0.01, ** p<0.05, * p<0.1

11 Table 5: Changes in -Level IT adoption and evolution of Characteristics VARIABLES (1) Growth Rate of PCs Per Employee (2) Growth Rate of PCs Per Employee (3) Growth Rate of PCs Per Employee (4) Growth Rate of PCs Per Employee Growth Rate of Education ** 0.131* PCs per employee (1994) Education (1996) Growth Med Fam Income Median Family Income (1996) Growth of Population Population (1996) *** (0.0369) (0.0878) *** (0.0457) *** (0.0849) (0.0816) *** (0.0407) *** (0.0818) (0.0758) *** (0.041) *** (0.0931) * ( ) 0.132* (0.0789) (0.0363) ** (0.0493) (5) Growth Rate of PCs Per Employee RUCC SIC2 Controls 0.123* (0.0684) *** (0.0369) *** (0.0994) ( ) (0.0766) (0.0345) (0.047) Observations R-squared Notes: growth rates are computed according to the formula proposed by Davis, Haltiwanger and Schuh (1996, Job Creation and Destruction MIT Press). The dependent variable is the growth rate of county-average number of PCs per employee for all plants surveyed by Aberdeen Ci Technology Database. Education is the share of individuals above 25 years old holding a bachelor degree or higher. Median Family Income is the logarithm of the median income of families living in a county, as reported by ACS and Decennial Census. Population is the logarithm of the number of individuals living in a county. RUCC is a set of fixed effects for counties Rural-Urban continuum classification by US department of Agricultural Economic Research in SIC2 Controls are the shares of plants (in Ci Technology Database) belonging to each SIC2 industry. Standard errors in parentheses are clustered at State level. Observations are weighted by number of plants *** p<0.01, ** p<0.05, * p<0.

12 Figure 1: PCs per employee Notes: mean and quantiles of PCs per employee are calculated on all US establishments surveyed by Aberdeen Ci Technology Database.

13 Figure 2: Counties covered by the Aberdeen data survey Notes: Number of US counties with at least one plant surveyed by Aberdeen Ci Technology Database

14 Figure 3: Distribution of PCs by year Notes: PCs per employee is winsorized at top 2% (establishment level) before taking average at county level. Only counties present in the dataset since 1989 are included. Source is Aberdeen Ci Technology Database.

Figure 4: PCs per employee across the US: 1990 to 2014 1987 1996 2006 2014 Notes: Maps show county average PCs per employee.

15 Figure 4: PCs per employee across the US: 1990 to Notes: Maps show county average PCs per employee. Values winzorized at top 2% of establishments before calculating county averages. Counties with missing values are assumed to be in the bottom 10% of IT adoption for illustrative purposes. Source is Aberdeen Ci Technology Database.

Wage Inequality and Establishment Heterogeneity

VIVES DISCUSSION PAPER N 64 JANUARY 2018 Wage Inequality and Establishment Heterogeneity In Kyung Kim Nazarbayev University Jozef Konings VIVES (KU Leuven); Nazarbayev University; and University of Ljubljana