Accounting for Innovations in Consumer Digital Services: IT still matters

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Accounting for Innovations in Consumer Digital Services: IT still matters David Byrne, Carol Corrado 2019-049 Please cite this paper as: Byrne, David and Carol Corrado (2019). “Accounting for Innovations in Consumer Digital Services: IT still matters,” Finance and Economics Discussion Series 2019-049. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2019.049. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Accounting for Innovations in Consumer Digital Services: IT still matters David Byrne∗ and Carol Corrado†‡ June 15, 2019 (Original Draft, March, 2017) Abstract This paper develops a framework for measuring digital services in the face of ongoing innovations in the delivery of content to consumers. We capture what Brynjolfsson and Saunders(2009)call“freegoods”asthecapitalservicesgeneratedbyconnectedconsumers’ stocks of IT digital goods; this service flow augments the existing measure of personal consumption in GDP. Its value is determined by the intensity with which households use their IT capital to consume content delivered over networks, and its volume depends on the quality of the IT capital. Consumers pay for delivery services, however, and the complementarity between device use and network use enables us to develop a quality-adjusted price measure for the access services already included in GDP. Our new estimates imply that accounting for innovations in consumer content delivery matters: The innovations boost consumer surplus by nearly $1,800 (2017 dollars) per connected user per year for the full period of this study (1987 to 2017) and contribute more than 1/2 percentage point to US real GDP growth during the last ten. All told, our more complete accounting of innovations is (conservatively) estimated to have moderated the post-2007 GDP growth slowdown by nearly .3 percentage points per year. Keywords: Consumer Digital Services; Information and Communication Technology; Consumer Durables; Consumer Surplus; Innovation; Digital Content Delivery; National Accounting, and Price Measurement ∗BoardofGovernorsoftheFederalReserveSystem,Washington,D.C.Theviewsexpressedinthispaperarethoseof the authors and do not necessarily reflect those of the Board of Governors or other members of its staff. †The Conference Board and Center for Business and Public Policy, McDonough School of Business, Georgetown University. Corresponding author (cac289@georgetown.edu) ‡ThispaperwaspreparedfortheNBER/CRIWConference,“MeasuringInnovationinthe21stCentury,”Georgetown University, Washington, D.C., March 10-11, 2017. We have benefited from presentations of this paper at the 5th IMF Statistical Forum in Washington, D.C. (November 2017), the ESCoE Measurement Conference in London (May 2018) andthe5thWorldKLEMSConferenceinCambridge,Mass(June2018). Wereceivednofinancialsupportforthispaper.

1 Introduction Capturing the impact of innovations in consumer content delivery in conventional well-being measures, e.g., GDP, presents significant challenges. It also seemingly requires a new approach because the manifestationoftheseinnovationsinconsumerwelfare(e.g.,timespentconsuminghighqualitycontent via networked IT devices) does not involve a market transaction at the time of consumption, which is where price collectors/estimators look to pick up new goods as they appear. Figure 1 shows that innovationsinconsumercontentdeliveryhavebeenveryrapidsincetheturnofthiscentury, suggesting their impacts may be missed in existing GDP; indeed, they are clustered in the mid-2000’s when the slow down in the trend GDP growth emerged. Is it possible that the substitution of uncounted, so-called free goods for purchased counterparts is a culprit in this much-discussed slowdown? This paper adapts a not-so-new approach—capitalization of consumer digital goods—to address this question, but the standard approach is augmented with an accounting for how IT devices and subscription network access services are used and consumed.1 To understand why a use-adjusted version of an “old” approach is both (a) needed and (b) up to the task of capturing 21st century innovations, consider first that it is consumer-owned devices with advanced processing technology— computers, powerful smartphones, smart TVs, and video game consoles—that enable the consumption of high quality content in many homes (and elsewhere), and these services currently are uncounted in national accounts (though their paid-for predecessors often were). Consider next that the spread of broadband since 2000 and rise of social media since 2004 suggests that the use of services that enable the delivery of content to consumer has risen dramatically (see figure 2). The rise in use of network servicesimpliesgreaterconsumptionvolume(foragivennumberofsubscriptions)becausesubscription costs do not fully depend on use rates. All told, we translate the problem of capturing the innovations shown in figure 1—including what Brynjolfsson and Saunders (2009) call “free goods”—into a quest for comprehensive measurement of (a) consumer services derived from IT device use and (b) consumer network service volumes in constant-quality terms. (a) involves an imputation to GDP for the missing services and (b) involves creating a new price index for the paid-for services. Because consumers’ IT capital use is inextricably tied to households’ utilization of public broadband, wireless, and cable networks (including their take up of over-the-top (OTT) media and personal 1The standard approach refers to the productivity literature that capitalizes consumer durables, originally due to Christensen and Jorgenson (1969, 1973); see also Jorgenson and Landefeld (2006). The U.S. national accounts do not capitalize consumer durables in headline GDP.

Figure 1: Timeline of Innovations in Consumer Content Delivery Source: Authors’adaptionandextensionofinformationinTotalAudienceReport,TheNeilsenCompany,December3,2014, availableathttp://www.nielsen.com/us/en/insights/reports/2014/the-total-audience-report.html. cloud services), its imputation must be linked to paid-for services. In other words, home services and paid-forservicesexhibitdemandcomplementarity2, andajointanalysisofthesetwotypesofconsumer digital services is required. This aspect of the approach to capitalization of consumer digital capital is novel with this paper. A related literature addresses the measurement of “free goods” using alternative methods and very different frameworks (Nakamura, Samuels, and Soloveichik, 2016; Nakamura, Soloveichik, and Samuels, 2018; Brynjolfsson, Collis, and Eggers, 2019; Brynjolfsson, Collis, Diewert, Eggers, and Fox, 2019); we compare our findings to these works later in this paper. Figure 2: Consumer Digital Capital Use (a) Broadband Use (b) Social Media Use (c) Mobile Device Use Source: PewCenterfortheInternet 2Thanks to Shane Greenstein for suggesting this interpretation. 2

The roadmap of this paper is as follows. Section 2 sets out our framework for thinking about how the standard framework for capitalizing consumer digital goods needs to be adjusted to take into account the dramatic increase in household digital asset use shown in figure 2. Then we review the relationship between device use rates and the volume of services that deliver content over networks; this forms the basis for the quality-adjusted price index for network access services developed in this paper, the details of which are covered in an appendix. Section 4 summarizes our empirical findings in terms of impacts on real GDP and consumer surplus. Section 5 concludes. Our new estimates imply that accounting for innovations in consumer content delivery matters: The innovations boost consumer surplus by nearly $1,700 (2017 dollars) per connected user per year for the full period of this study (1987 to 2017) and contribute more than 1/2 percentage point to US real GDP growth during the last ten (2007 to 2017). All told, our more complete accounting of innovations is (conservatively) estimated to have moderated the post-2007 US real GDP growth slowdown by nearly .3 percentage point per year (out of a 1-1/2 percentage points total slowdown). Only a portion, about .2 percentage point per year, contributes to business productivity for reasons explained below. 2 Framework: Demand Complementarity Digital device services and network access services work together to deliver consumer content. This section illustrates how their demand complementarity can be exploited to capture and account for quality change in consumer digital services. 2.1 Definitions Because consumer digital services reflect both households’ use of digital devices and households’ take up of network access services, the value of total consumer digital (T) services, PSTS T , is expressed as the sum of two components: (1) PSTS T = PS T H S T H +PS T B S T B . The components are nonmarket (or “home”) and market (or “paid-for”) services, respectively, where superscripts on the component digital services volume indexes (the S’s) denote location of the capital used to deliver each type service, i.e., business sector (B) or household sector (H). 3

Homeservices,PS T HS T H,aregeneratedviahouseholds’useofITgoodspurposedforaccessingdigital networks.3 Paid-for services, PS T BS T B, are derived from subscriptions to networks, e.g., payments for internet access, cellular access, etc. Where are the seemingly “free” services provided by Google, Facebook and other apps? Our answer is that they are embodied in both nonmarket and market services in this framework. The demand for consumer IT capital is a derived demand induced by the availability of search engines, social networks (and so forth) that push users to purchase higher quality equipment for, e.g., streaming YouTube and Netflix videos. The intensity of use of network access services is increased because the “free” services require that data—pictures, videos, search results— need to be delivered from the cloud for configuration and display by browsers and/or apps on the home device. It is tempting to associate the capture of “free goods” as solved by the imputation for home services that we propose in this paper, but the derived demand dynamic underscores it is equally important to use quality-adjusted price statistics for the purchased parts of content delivery systems, as improvements in quality are also seemingly “free.” Quality change is reflected in the price indexes of both components of (1). It stems from (a) the qualityoftheequipmentusedtoaccesscontentvianetworks(e.g.,thestoragecapacityofsmartphones, etc.), (b) the quality of network services (e.g., download and upload speeds of broadband service, channelvarietyinvideoservice, etc.), and(c)theuseintensityofthecombinedcontentdeliverysystem (i.e., the equipment plus the access service). After controlling for the quality of systems (equipment cum access services) at the time of their purchase, the change in system use intensity reflects changes in the system’s performance, i.e., change in the marginal product of its combined net capital stocks (just as ex post private capital income reflects changes in the return to capital). Not much of (b) and none of (c) is in existing GDP, and while (a) is included to a significant degree, we improve its capture in this paper. Network use intensity reflects how consumers use their IT devices and is revealed by the take up of paid-for network access services. Denoting network use intensity by λ, and letting N be the number of users on the network (i.e., consumer accounts, from the perspective of the service provider), then average network use intensity is defined as: SB (2) λ = T N 3IT goods used without network access produce uncounted services as well, such as personal computer used to work on local files. This use is outside the scope of our analysis. 4

where SB is the volume of paid-for access services consumed, per equation (1). λ and N are most T easily understood from a producer perspective, i.e., λ is an intensive per customer use margin and N is an extensive margin whose increases reflect customer growth, e.g., for broadband providers, the number of “customers” N is households with broadband subscriptions. For cellular service providers, N is individuals with cellular phone subscriptions.4 There are other, largely demographically-driven, dimensions of use, e.g., the number of users per householdandtheageofusers,asthisfeedsintohoursofuseperconnection. Notethatperequation(2) these distinctions in margins of use are implicit in λ to the extent they are not counted in N. 2.2 Home services Our starting point is the Christensen and Jorgenson (1969, 1973) framework, based on Jorgenson (1963), for imputing service flows from capitalized consumer durables. Letting KH denote the net T stock of digital goods held by consumers and PK T H the per period rental price for use of a unit of those stocks, then the value of their capital services PK T HK T H in the standard formulation would be given by: (3) PK T H K T H = (ρ+δ T H)PI T H K T H where ρ is an ex ante real household discount rate, δH is a depreciation rate for household IT stocks, T and PI T H is a quality-adjusted asset price index for new investments in those stocks. Nominal home services for consumer digital goods, the PS T HS T H term in equation (1), does not correspond to equation (3) because it is essentially a capacity flow; i.e., (3) does not reflect actual, or ex post, consumption.5 The trend in use rates illustrated in figure 2—and demand complementarity— suggeststhatincorporatingan“ITuse”dynamicisnecessarytocapturetheservicesdefinedbyPS T HS T H, i.e., there has been increased use of IT capital for consuming digital content over networks. The IT device use dynamic is specific to each device type, which implies we need to define a use rate ψ for each asset type a, e.g., for computers, for mobile phones, for TVs, etc. We thus have the a 4Although households have other modes of network service (e.g., cable, OTT) and all such services are considered in ourempiricalanalysis,forsimplicity,thediscussioninthissectionconsidersN asthenumberofsubscriptionstoasingle service, i.e., connected households. 5Private industry capital income is generally understood to include a utilization effect when the rate of return is calculated on an ex post basis as in Jorgenson and Griliches (1967). When consumer durables are capitalized, service flows are imputed using an ex ante return as in (3), and therefore a utilization effect is not be “automatically” present. See Hulten (2009) for a discussion. 5

following (cid:0)Dev a(cid:1) (4) ψ = a D∗ where Dev is the number of hours per day device type a is used to connect to networks, and D∗ is a the potential number of hours per day any device can be used. We can then define an “effective” stock of network access equipment and software, KeH, that T accounts for how the use of a given stock of network access equipment and software expands, in which case the value of nonmarket consumer digital services in equation (1) is given by: (5a) PS T H S T H = PK T H KeH T (5b) = ψ·PK T H K T H . where ψ reflects the appropriately weighted aggregate of the individual ψ ’s. A related issue is that a some consumer digital capital goods are not used for the consumption of content over networks, e.g., digital cameras, suggesting it is necessary to identify a relevant group of IT devices—call this network access equipment (NAE)—for generating the relevant capital services flows. The relevant IT products comprising NAE stocks will be identified in measurement; for now we proceed with the assumption that only NAE products are included in the capital measures subscripted by “T”. Consider next how to measure the implicit volume of services whose value is given by 5b. Log differentiation of equations (5b) and (3), holding ρ and δH constant, suggests that the growth of T • • • • • • nominal free services PS T H +S T H is equal to PK T H +KeH T . This in turn implies that PS T H = PI T H, and • • that growth of real services SH equals the growth of the effective stock KeH, or T T • • (6a) SH = KeH T T • • (6b) = KH +ψ . T 2.3 Paid-for services Digital access services are typically sold as subscriptions, where households pay a monthly fee for a “plan” in return for access to a range of services, e.g., broadband, smartphone, cable TV, subscription 6

video-on-demand. Each plan has a fixed set of characteristics, e.g., download speed, upload speed, number/availability of videos or video channels, etc., for the services involved. Plan heterogeneity by service type and service type characteristics is ignored (for now) for ease of exposition. Producers offer digital access service plans at prices PO T B. Offer prices are subscription contract prices set at the outset of the period, and the average price each customer pays is expressed as (7) P O T B = PO T BO T B N where PO T BO T B are producer revenues from consumer sales of N plans. Nominal consumer payments, PS T BS T B of equation (1) equals this producer sales revenue. We assume that producers’ capacity is constrained in the short run (the period of the contract) and, after accounting for the usual issues regarding peak load planning, that producers set offer prices based on a preferred rate of capacity utilization determined by anticipated average customer usage, λa. These assumptions imply that OB is a planned quantity of delivered services and not necessarily T equal to SB, the actual quantity of services consumed by users—unless of course actual usage λ is T perfectly anticipated, i.e., λa = λ. It follows that the offer price index PO T B does not necessarily equal the consumption price index PS T B of equation (1). Let u be an index of actual capacity utilization, whereu = 1denotesthesituationwhereλa = λ. Wethenhaveλ = λau, inwhichcasetherelationship between real services consumption and real services offered, and between consumption prices and offer prices is given by (8) SB = OBu. T T (9) PS T B = PO T B . u Equation (9) states that the consumption price index PS T B is a utilization-adjusted contract price. Equations (8) and (9) are not very helpful for conventional, timely price measurement (as in a monthly CPI) because producers’ preferred utilization rate u is not readily observed. However, substitution of (8) into (9) reveals that the consumption price may be alternatively written as: (10) PS T B = PO T BO T B . SB T 7

which suggests that consumption prices for access services may be obtained by dividing producer revenue by a relevant, consistently-defined volume measure, i.e., that ideally, SB ≡ VOL where VOL is T such a measure. What might that volume measure be? We know that total consumption increases along with the number of users and/or hours of use, but these are very coarse indicators that do not capture consumptionintensityorservicequality. Anidealmeasurewouldcaptureconsumers’useintermsofthe potential performance of communication networks and where utilized performance is a comprehensive measure capable of being consistently defined in the face of rapid technical change, e.g., Internet Protocol data traffic (IP) measured as optimally compressed megabytes/petabytes per year, i.e., that (11) ST ≡ VOL = IP . B Arangeofservicesaredeliveredovernetworks,anddataflows/IPtrafficmaynotalwaysbetherelevant indicator of quality, but for internet access services via computers of mobile phones IP traffic would appear to be a solid choice (e.g., see Abdirahman, Coyle, Heys, and Stewart, 2017). For video services, qualityisnotsosimple;cross-countrystudieshavefoundthatthequalitydimensionforvideoservicesis captured by a range of controls, including the number of channels (HD and standard), and availability of premium channels and 4K display resolution (Corrado and Ukhaneva, 2016, 2019; D´ıaz-Pin´es and Fanfalone, 2015). 2.4 Use intensity, λ With real services captured by a performance measure, the changes in network and device intensity of • use, λ, can be shown to reflect the difference between changes in the average price paid by users for a plan and the price index for access services, i.e., it reflects changes in the quality of services consumed. To see this, log differentiate (2): • • • (12) λ = SB −N . T 8

• After adding and subtracting the nominal change in paid services, PO T BO T B, and combining terms, we obtain: • • (13) λ • = (cid:0) PO T BO T B (cid:1) − (cid:0) PO T BO T B (cid:1) . N SB T Substitution of (7) and (10) for the first and second terms yields: • • (14) λ • = P O T B −PS T B . • In equation (13) the change in use intensity λ reflects the difference between the rate of change in a per user price and a unit volume price, or per equation (14), the difference between the rate of change in the price index for access services and rate of change in the average price per plan, i.e., quality change. Statistical agencies generate price indexes in terms of offer prices PO T B, not consumption prices PS T B. Consider now the relationship between λ • and the quality change in official price indexes for networkaccessservice(basedonofferprices),e.g.,qualitychangethatmightbecapturedusinghedonic techniques that account for improvements in speeds and other capabilities in subscription telecom • service plans.6 Noting first that the change in the offer price index, PO T B, also can be decomposed into the rate of change in quality of offered plans, ν•, and the rate of change in the average price per plan, • • • P O T B , i.e., PO T B = P O T B −ν•. Next, from log differentiation of (9), after subtracting the result from (14) and combining terms, the relationship between ν• and λ • is readily shown as (15) λ • = ν• +u• , which says that the quality change in real network access services consumption is equal to the quality change in offered plans (at offered prices) plus the unanticipated change in network service provider utilization. 2.5 Network utilization, u Considernowhowonemightmeasureu. Wedonotneedtomeasureubutithelpsinterpretandanalyze our new price measures, e.g., it can help understand how very small increases in ν might coexist with 6As done, e.g., at the BLS (see Williams, 2008). 9

notable declines in consumption prices for network access services. As suggested by equation (15), this situation occurs when increases in use intensity augment measured quality change in offer prices.7 Aspreviouslyindicated,privateindustrycapitalincomeisgenerallyunderstoodtoincludeautilization effect, and previous work has considered how to extract a measure of network capital utilization fromproductivitydataforinternetserviceproviders, orISPs(Corrado,2011;CorradoandJ¨ager,2014; seealsoCorradoandvanArk,2016). Thebasicideaintheseworksisthatwhenanexanteapproachis used to determine an industry’s return, a utilization factor can be calculated so as to exhaust observed capital income—provided that the industry’s aggregate net stock of capital is not particularly sensitive to composition differences in asset use, i.e., it acts more or less as a single capital good (Berndt and Fuss,1986;Hulten,1986). ThisisarguablythecasefornetworkservicesprovidersintheUnitedStates, whose capital stock is a physical network whose parts largely operate as a single good. Employing this assumption, Corrado (2011) found a substantial difference between the U.S. ISP industry’s ex post calculated nominal rate of return and the market interest rates typically used in ex ante productivity analysis; the difference was able to be interpreted as network utilization. The network services-providing industry’s ex post gross return is defined as (16) ΦISP = (rISP +δISP −πISP) where rISP is an ex post nominal net return determined residually (e.g., as in Jorgenson and Griliches, 1967), given depreciation δISP and revaluation of the industry’s capital stock πISP. Now define the industry’s ex ante gross return as (17) Φ ISP = (r+δISP −πISP) where r is an ex ante nominal rate of interest. Let uISP be the industry’s capital utilization rate. As shown in the appendix, this utilization rate is given by ΦISP (18) uISP = ISP Φ 7Our estimates of access prices may differ from official prices for another reason, namely that the change in offered services quality is mis-measured; but this, too, is difficult to judge without triangulation with measures of changes in producer utilization u and household use intensity λ. 10

which suggests that the underlying relationship between the ex post and ex ante net rate of return, i.e., r versus r, for an industry or sector is an indicator of its capital utilization.8 2.6 Summary Tosummarize,changesinthequantitiesandpricesofconsumerdigitalservicesassetoutinequation(1) are as follows: • • • (19a) SH = KH +ψ T T • • (19b) PS T H = PI T H • • (19c) SB = Vol T • • (19d) PS T B = P OB T −λ • . where λ and ψ were defined above, and PIH is a quality-adjusted asset price index for network access T equipment. 3 Measurement Thissectionsummarizeshowthepricesandquantitiesoftheprevioussectionaremeasuredandpresents some key results; details of measurement procedures are spelled out in the appendix. We begin with the new network access services price index, describing how this index may be built using alternative • volumemeasures. Wethenpresentresultsforλandforourcalculationsofutilizationfromthebusiness side, u•. A second subsection sets out how our consumer digital capital stocks, their connectivity use rates, and digital capital services are obtained. 3.1 Access prices, household use intensity, and network utilization To calculate a price index for each of the IT services provided by the business sector—cable, internet, mobile, and video streaming—we begin with nominal spending and divide by a measure of aggregate time spent using the service adjusted for quality. These individual price indexes are aggregated to create an overall access price index used to deflate nominal spending on access services and produce a measure of real spending. For exposition and analysis, we consider price indexes constructed using 8Inmodelsthatintroduceimperfectcompetitioninanotherwisestandardneoclassicalgrowthframework(e.g.,Rotemberg and Woodford, 1995), utilization is absorbed in a more general inefficiency wedge capturing, among other things, the ability of firms to maintain a price markup. 11

four alternative measures of quantity: the number of households subscribed to the service, the number of individual users, time spent on the service, and time spent adjusted for quality (our preferred measure for deflation). Thus four alternative indexes are calculated for each of the four services by dividing revenue by each of the alternative measures of quantity, yielding prices paid per household, SB SB SB SB per individual, per unit of time, and per unit of constant-quality time: P T, P T, P T, and P T. H I D Q (Note: D is the notation used for time, i.e., as in hours per Day). The calculation of the alternative price indexes for digital access services can be summarized as follows: Let (PO T BO T B) j be payments for service type j within total payments PO T BO T B. Price change for each type of price index covering J types of services is then (20) ∆lnP S T B = (cid:88) J w ∆ln (cid:16)PO T BO T B(cid:17) where v = H,I,D,Q v j VOL v j j=1 and w is a Divisia payments share for digital access service type j and VOL is its volume measure for j v,j price index concept v. Depending on the contract arrangement, the price observed by the consumer, PO T B, may be any of these four prices. For example, if a consumer pays a cable company a fixed amount to keep the household connected each month, PO T B equals P H S T B . If a consumer pays an internet provider a fixed amount to have unlimited access each month, PO T B equals P I S T B . If the consumer has a prepaid plan for a certain number of hours of talk time on a feature phone, PO T B equals P D S T B . And, if the consumer has a contract for smartphone use based on data traffic consumed, PO T B equals P Q S T B . We construct P H S T B , P I S T B , P D S T B , and P Q S T B for each service but we do not observe PO T B because we do not have information on the contract arrangements. The contract price PO T B is not needed for the analysis in this paper; note, however, contract prices are the basis for official price measures. In the appendix, a table documents the data sources for each access service price index and a table reports their time series.9 In terms of the framework set out in section 2, we have the following: (21) ∆lnPS T B =∆lnP S T B Q • SB SB (22) λ=−(∆lnP T −∆lnP T) Q H With the suite of indexes constructed along margins of use, changes in the quality-adjusted price index and in • ouruseintensity,λ,whichisbasedonhouseholds,canbedecomposedintocontributionsfromI,T,andQ—i.e., 9Pricesbyaccessservice,alongwithaggregatepricesconstructedusingthealternativemeasuresofvolume,areshown in this table. 12

into contributions from growth in individuals using the service per household, time spent on the service per individual user, and the quality of an hour of use of the service, respectively. Theaggregatequality-adjustedpriceindexforaccessservicecorrespondingtoequation(21)falls12.4percent • per year, on average, over the full period of this study. Household use intensity, λ, increases 13.9 percent at an annual rate. Results for the overall price index by sub-periods are plotted in panel (a) of figure 3 below; spending shares for its subcomponents by type are shown in panel (b) of the figure. As these figures show, the decline in the overall quality-adjusted access price index accelerates over time, first as internet service accounts forarisingshareofspending(1997to2007),thenassmartphoneaccessbecomesmoreimportant(2007to2017). Thesetrends inthe aggregatenetworkaccess priceindexreflectnotonly changesin spending sharesbymodeof access but also large differences in the contributions by access mode. Decompositions for each mode of access are shown in figure 4. Contributions to the overall volume price change by each intensity margin (i.e, volume measure) show little difference between changes in per individual user prices relative to per household prices; as a result, only the contribution of changes in the price per household shows through in figure 3. As may be seen in this figure, quality change contributes significantly to the overall decline in network access prices in most sub-periods, and consumers’increaseintimeconnectedprovidesasubstantial,additionalkickfrom1997to2012. Timeconnected is especially important in driving price change for spending on mobile network access and SVOD. Figure 5 reports year-by-year changes in our overall quality-adjusted network access services price index SB • (P T)andtheimpliedchangeinhouseholduseintensity(λ)perequation(22). Apriceindexfornetworkaccess Q services constructed using components of BEA’s PCE price index and our per household price index (i.e., the SB average price per household, P T) also are shown. Note first that our new access services price index (the gray H line) falls much faster than the implicit price index in existing GDP (the black line); the growth implications of this finding will be reviewed in the next section of this paper. Note second that changes in the BEA price index hovers about changes in our per household price from about 2000 on. If the BEA index accurately represents changesincontractprices,theresultimpliesthatthereisverylittlequalitychangeinmeasuredofferpricesfrom • from 2000 to 2017, i.e., ν has shown no change since 2000. How much of this implied zero rate of offered service quality change may be a measurement problem? It is • • tempting to compare ν with λ but the two are conceptually different in our framework. Using IP traffic as a • volumemeasureimpliesthatλwillreflecttrendsinconsumerusageoverandaboveimprovementstothequality of content delivery systems. And as previously noted, the price index based on hours per day of network use shows the contribution of household usage rates (“time”) to our overall price index, and as illustrated in panel (a) of figure 3, the contribution of time accounts for nearly one-half of the declines in access service prices from 1997 to 2012. The detail by type of service in figure 4 reveals that the contribution of time stems from different 13

Figure 3: Network Access Services Price Index (a) Contributions to Price Change by Volume Measure (b) Shares of Spending by Mode of Network Access modes of service during different sub-periods, with new modes tending to show increased hours of use at the expense of older ones. • To underscore this result and the fact that small increases in ν might coexist with notable declines in consumption prices for network access services, we calculate values for producer network utilization u. This result is shown in figure 6 and covers the period from 2000 to 2016. While this utilization measure bounces about year by year, it rises more than 4-1/2 percent per year, on balance. This pattern is interesting for several reasons—but before we offer our interpretation, note that the measure in figure 6 pertains to the entire telecommunications and broadcasting industry, i.e., it includes commercial and enterprise customers (as well as other services) and thus does not solely reflect the interaction between the demand and supply of consumer content delivery services as defined in this paper. That said, per equation (15), the rather sharp rise in u 14

Figure 4: Price Decompositions by Mode of Network Access (a) Cable (b) Mobile (c) Internet (d) SVOD • • ratifies our previously stated results that a significant fraction of the large divergence between ν and λ after 2000reflectedincreasesinuserates. Intermsofthemodelofsection2, λconsistentlyexceededλe, loweringthe effective price paid by each consumer (holding per plan quality constant).10 10 Recallthattheuthatwecalculatereflectsarelationshipbetweenexpostandexantereturns,andarisingureflects a growing margin between the ex post rate of return to capital in the telecommunications/broadcasting industry and the overall market rate (proxied by the AAA corporate bond rate). While the rise in the industry’s relative profitability suggests a strengthening of producer pricing power for network access services, on a per household basis, changes in households’ average prices decelerated after 2002 relative to earlier experience (see again figure 3). Put differently, the rise in the industry’s relative profitability reflected a demand shock. 15

Figure 5: Consumer Network Access Services Figure 6: Implied Network Utilization 3.2 Digital Net Stocks, Capital Services, and Asset Prices Table 1, column (1), lists the 14 product classes of durable goods considered to be consumer durable digital (or IT) goods. This list ranges from TVs, to computers and software, to cell phones.11 Consumer spending for most of these products may be developed from underlying detail in the U.S. national income and product accounts (NIPAs); indeed, the first 12 product classes shown in the table directly correspond to categories of digital goods reported in the annual personal consumption expenditures (PCE) bridge table.12 For the analysis inthispaper,estimatesoftheretailvalueofconsumercellphonepurchasesaredevelopedfromindustrysources; see the appendix for further details on how this series and the other telephone equipment series are estimated. 11Gameconsoles,whichhaveembodiedmassiveinnovationintheperiodofthisstudy,arenotincludedforlackofdata. 12BEA’s annual PCE bridge table begins in 1998 and does not extend through the most recent NIPA year. Nine categories of PCE spending on digital goods are reported on NIPA table 2.4.5U, however, and these data are used to develop the more detailed, bridge table-based series from 1970 to 1997 and for the year 2017. 16

Table 1: PCE durable digital goods Product class Depreciation Groupa NAE Groupb (1) (2) (3) 1. Televisions A Y 2. Cameras B N 3. Other photographic equipment A N 4. Other video equipment A N 5. Audio equipment A N 6. Recording media A N 7. Computers B Y 8. Data storage equipment B Y 9. Monitors B Y 10. Computer peripherals B Y 11. Miscellaneous office equipment A N 12. Software and accessories B Y 13. Cell phones B Y 14. Other telephone and communications equipment A N Notes: a. A = 9 year service life, B = 5 year service life. b. NAE = Network Access Equipment. In terms of service lives, the products are grouped into two categories, those with a 9 year service life (A) and those with a 5 year service life (B). These groupings are indicated in column (2) of the table, and are a (slight) simplification of the service life categories used by BEA in their fixed asset accounts.13 To compute net stocks we follow BEA and Hulten and Wykoff (1981a,b) and use a declining-balance rate of 1.65 for these goods, which implies geometric rates of depreciation of .1833 and .3300 for groups A and B, respectively. An end-of-year(EOY)netstockofeachproductclassaintable1iscalculatedusingtheperpetualinventorymethod with geometric depreciation, again following BEA (see page M-7 in U.S. Department of Commerce.Bureau of Economic Analysis, 2003): δH (23) KH =IH(1− a )+(1−δH)KH a,EOY a,t 2 a a,EOY−1 where IH is annual real investment for each asset class a in year t. a,t When considering the demand complementary of payments for digital access services with stocks of the devices shown in table 1, only the equipment used for cable TV, subscription video, internet or mobile network accessisrelevant. Column(3)ofthetableisanindicatorofwhethertheassetclassisincludedintheNAEstocks relevant to the analysis in this paper. To obtain “effective” NAE stocks, an estimate of ψ =N/N∗ per equation (4) is required. Our baseline estimates are based on N = number of households (or broadband connections), in 13 Compared BEA’s methods, the major simplification we make is to use geometric depreciation for computers. 17

Figure 7: Consumer Digital Capital Services, 1987 to 2017 which case N∗ in (4) is the fraction of each NAE component available for the consumption of digital content over the network. The measurement of ψ is detailed in the appendix. Thenominalvalueoftheflowofservicesfromeachconsumerdigitalassetiscalculatedviaequation(3)with the gross ex ante rental rate formed using the 10-year constant maturity government bond rate, the relevant depreciation rate as described above, and actual price change for the asset type.14 Summing over all asset types yields an estimate of consumer capital services based on total digital goods stocks, and summing over the asset types included in network access equipment yields the subcomponent for total NAE stocks. When we adjust services from NAE stocks for the extensive margin (as reflected in device use on networks) ψ, we obtain our estimate of PS T HS T H, services generated via households’ use of IT goods purposed for accessing digital networks. These three versions of consumer digital capital services are shown in figure 7, plotted relative to GDP adjusted to include them. Ourestimateofhome-generateddigitalservicesrelativetoGDP,thesolidblacklineinfigure7,risessteadily overthethirtyyearsshowninthefigure, reflectingboththeincreaseinrelativeimportanceofNAEstocksinall digitalstocks(theblackdottedlineversusthethingrayline)andtheincreasedportionofNAEstocksconnected tonetworks(thesolidversusthedashedlines)thatreflectsincreasesinψ. TheratioofPS T HS T H relativetoGDP 14In the implementation of (3), mid-period stocks computed from EOY stocks are used. 18

stood at 1.04 percent of GDP in 2017, up from .48 percent ten years earlier. This trajectory is roughly similar to estimates of free services prepared using a very different approach (the red dots in the figure).15 Figure 8: Network Equipment Price Change, 1988 to 2017 Source: Elaborationofpriceindexesdevelopedforthispaper. Therealinvestmentusedtodevelopnetstockestimatesviaequation(23)iscalculatedbydeflatingnominal spending on each product class using asset price indexes based on the sources documented in the appendix. These prices are research indexes largely adapted from prior work (Byrne and Corrado, 2015a,b; Byrne, 2015; Byrne and Corrado, 2017a,b). In new moves, we incorporate two quality-adjusted price indexes from Statistics Japan and exploit work by Copeland (2013) on consumer game software in combination with results from the Bureau of Labor Statistics producer price index for game software. Our price index for the 14 consumer digital goodslistedintable(1)falls11.7precentperyearfrom2007to2017,2.6percentagepointsfasterthanitsofficial counterpart (based on published PCE prices); see the appendix for further details. Theimplicitdeflatorforconsumerdigitalassetsdependsontheweightingofthecomponentsintheaggregate. Figure8showsannualpricefortotalNAEstocksversuseffectiveNAEstocks. Asmaybeseen, theweightingof the underlying components produces very similar results for effective NAE stocks versus a simple aggregate of those stocks. Our price index for home services, PS T H, is the Jorgensonian rental price index for effective NAE stocks (the red line in the figure), which is driven by the asset price and falls 11.2 percent per year over the full period of our study and 14 percent since 1997. 15Nakamura et al. (2018) estimate the costs of producing both professionally-created and user generated consumer content. The red dots in the figure are their estimates for the digital component of their professionally-created free content plus their estimate of the value of user-generated free content. 19

4 Results and Implications ThissectionreportsthenewdigitalconsumptionmeasuresanddiscussestheirimplicationsforrealGDPgrowth and increase in consumer surplus. 4.1 GDP Our results for GDP are summarized in the table below. These results are calculated under the conservative assumption that overall real GDP is unaffected by differences the PCE IT goods investment price indexes developed in this paper and official prices used in GDP because these goods are primarily imported (whether for “effective” investment or all IT goods spending); recall too that we are unable to include the rapid quality change in game consoles in our price indexes. Table 2: Changes in Consumer Digital Services, 1987 to 2017 1987 to 1987 to 1997 to 2007 to 2007 to 2012 to 2017 1997 2007 2017 2012 2017 Percent change, annual rate (1) (2) (3) (4) (5) (6) Capital services: 1. Nominal 11.9 10.9 13.6 11.2 10.5 11.9 2. Real 26.0 17.1 32.4 28.8 35.8 22.2 Access services: 3. Nominal 11.5 15.1 13.3 6.2 7.7 4.7 4. Real 26.2 18.1 29.5 31.2 33.7 28.8 Memos: Effective NAE investment: 5. Nominal 12.2 10.5 15.6 10.5 11.2 9.8 6. Real 27.7 19.4 37.3 27.0 33.1 21.2 • Contrib. to GDPa,b 7. Consumer digital services .32 .09 .30 .55 .55 .56 7a. Net of existing .24 .05 .17 .44 .38 .50 8. Capital services .08 .02 .06 .15 .14 .16 9. Access servicesb .24 .07 .24 .40 .40 .40 9a. Net of existing .16 .03 .11 .29 .23 .34 Notes: a. Percentage points. c. GDP contributions are calculated assuming that differences between PCE digital goodsinvestmentsandtheirpriceindexesandtheirofficialcounterpartshavenoimpactonexistingGDPbecausethey arelargelyimported. Thekeytakeawaysfromtable2are,first,asshownonline2,column(1),realservicesfromuseofconnected digital systems grow very strongly, averaging 26.0 percent per year for the full period of the study. Second, our new results for real access services (line 4) are also very strong; as shown in column (5), real growth averaged nearly34percentperyearduringtheGreatRecessionanditsimmediateaftermath(from2007to2012). Third, this paper’s approach to accounting for innovation in consumer digital services shows that it is possible to “see” 20

digitalizationinGDP. IfourmethodsweretobeincorporatedinthenationalaccountsoftheUnitedStates,the contribution of consumer digital services (both components) to real GDP growth would average .55 percentage points from 2007 to 2017 (line 7, column 4), and annual real GDP growth would be .44 percentage points per year higher (line 7a, column 4). These impacts are substantial. The impacts of our work on consumer IT prices on overall PCE price measures are summarized elsewhere Byrne and Corrado (2019); this includes the new services measures we have introduced in this paper as well as the durable goods prices reflected on lines 5 and 6 of the table and detailed more extensively in the appendix. With regard to changes in the trend rate of real GDP growth, the impact of using our framework for measuring consumer digital services boosts the rate of real GDP growth from 2007 to 2017 relative to ten years earlier (1997 to 2007) by .27 percentage point (line 7a, column 4 less column 3)—a notable acceleration. Both the GDP boundary expansion (adding imputed real digital capital services) and the adoption of a qualityadjusted consumption price index for network access services contribute to this acceleration, with about 60 percent stemming from the net contribution of the new access services price index (.16 percentage point). The latter contribution also boosts business productivity growth; as with services from owner-occupied housing, the imputation for self-generated digital capital services is not factored into conventional measures of productivity change. 4.2 Consumer Surplus The consumer surplus stemming from innovations in consumer content delivery can be calculated using an index number approach if the quality-adjusted price indexes used in the analysis fully capture the benefits of the changes in question. Assuming our price indexes are up to the task, we compute consumer surplus as the macroececonomic gain from the relevant continuing commodities following (Diewert and Fox, 2017) as: (24) .5 (cid:0) ∆ΠS T H ∆S T H(cid:1) +.5 (cid:0) ∆ΠS T B ∆S T B(cid:1) +.5 (cid:0) ∆ΠI T eH ∆I T eH(cid:1) where ∆ is a long difference, and the ∆Π’s are changes in the relative prices, i.e., (25) ΠS T H = PS T H , ΠS T B = PS T B and ΠI T eH = PI T eH , PPCE PPCE PPCE where PPCE is the overall price index for consumer spending. In the textbook exposition of consumer surplus, the price drop from the Hicksian reservation price to the transaction price of the new good or service is the welfare gain stemming from the innovation in question. To fully capture this gain, benefits of individual innovations are quantified using techniques that rely on estimates of demand elasticities or estimated parameters of utility functions in conjunction with transactions prices and revenue data, e.g., Petrin (2002) and Greenwood and Kopecky (2013). There are many individual innovations 21

underlyingourpriceindexes,however,andforthisreasonalone,eschewingaparametricapproachandusing(24) has many advantages. Using (24) views innovations in digital content delivery as serial Schumpterian change whereindividualinnovationsarelaunched, gainmarketshare, andthenlosemarketshareasanotherinnovation is introduced. We believe this process is well captured by our quality-adjusted price indexes even though they do not explicitly incorporate Hicksian reservation prices. Our comprehensive accounting of use intensity captures the benefits of ubiquitous connectivity/networks cum powerful equipment in our estimates. Many of the products that undergird these estimates experience constant change via incremental quality improvements andintroductionofnewformsandvarieties,andweareabletoincorporatethesechanges,e.g.,enhancementsto personal computing via new forms (tablets and cell phones) and improvements in performance (speed, storage) are incorporated in our estimates. 16 The results of computing (24) are presented in table 3. Changes from the beginning of our sample (1987, arguably also the beginning of the Internet) to the beginning of social media and mobile broadband (taken as 2004) are assessed, as are changes from this point to 2017, the last year of our estimates. As may be seen on row 1, the consumer surplus due to innovations in digital content delivery from 1987 to 2004 (18 years) was nearly $900 billion in 2017 dollars (column 1) and $4.5 trillion over the next 14 years (column 2). These are substantial amounts. On a per user basis, rows 5 through 8, the gain hovered about $25,000 over each period (in 2017 dollars). While these numbers seem very large (implying a per user gain in economic welfare of nearly $1,800 per year, on average, during the latter period), they are in the same neighborhood as estimates of consumer surplus obtained by Brynjolfsson et al. (2019) using massive online choice experiments. The sum of their median willingness-to-pay estimates for the items included in their surveys (search engines, email, maps, video, e-commerce, social media, messaging, and music) was $32,232 in 2017 (Brynjolfsson et al., 2019, table 7, sum of items in column 2). We compare our long difference estimates with the single-point-in-time survey results of Brynjolfsson et al. (2019)basedonaconjecturethatrespondentsintheirmassiveonlineexperimentsarethinkingaboutwhatthey would have to pay to “return” to life before social media, smart phones, and mobile broadband. Brynjolfsson et al. (2019) also report median willingness-to-pay estimates for a survey conducted in 2016, and these values sum to $26,150, expressed in 2017 dollars.17 Using (24) with a long difference from 2004 to 2016 (i.e., dropping 16NotefurtherthatevenforaninnovationassignificantastheiPhone,theimpactoftheomissionofHicksianreservation prices on a price index is very small because the revenue weight on the unobserved initial price drop is likely to be so tiny that ignoring this change has very little impact GDP or consumer surplus. As reported in Apple’s financials, total iPhone revenue in the quarter of introduction in 2007 was $8 million ($32 million at an annual rate). GDP was $14,452 billion and our digital services series was $278,334 million in that year. One half of the revenue gain from the iPhone in its introductory quarter at an annual rate was then .11×10−5 relative to GDP and .057×10−3 relative to our series for connected digital capital services. Consider now the following thought experiment: Assume the change from the reservation price to the actual price of the iPhone was a ginormous -1000 percent in the quarter of introduction. Then our price index would be off 5.7 percentage points in the initial quarter (1.4 percentage points for the year) but GDP would be essentially unaffected. 17The simple sum of their figures is $25,697. 22

Table 3: Consumer Surplus from Innovations in Content Delivery Systems 1987 to 2004 2004 to 2017 (1) (2) Surplus, in billions of 2017 dollars: 1. Digital goods and services, total 879 4,450 2. Capital investment 257 788 3. Capital services 303 1,409 4. Access services 319 2,254 Surplus, in thousands of $ per usera: 5. Digital goods and services, total 26,926 23,081 6. Capital investment 7,855 4,087 7. Capital services 9,292 7,036 8. Access services 9,779 11,689 Annual surplus per user: 9. Digital goods and services, total 1,548 1,775 10. Capital investment 462 314 11. Capital services 547 562 12. Access services 575 889 Note: Allfiguresarein2017dollars. a. The per user figure is obtained by dividing the results on rows 1 to 4 by the average numberofconnectedusersduringtheperiodindicated. the last year, and dividing by a slightly lower number for the average number of users) yields an estimate of the consumersurplusof$19,502peruser—againinthesameballparkand,webelieve,strengtheningourconjecture. 5 Conclusion The household is an important locus of the digital revolution and one of its most visible since smartphones and social media became widespread. Entertainment, communication, and work from home have been supercharged by advances in hardware, software, and communication. Hardware innovation has proceeded at an especially blisteringpaceasthemajorhouseholdplatforms—smartphones,tablets,televisions,gamingconsolesandallthe appsthatrunonthem—havebecomeextraordinarilypowerful(andcheap)andasdatacenterinnovation(i.e. the cloud) has charged ahead in the background. Faster communication speeds—both wireline and wireless—have beenessentialofcourse;forexample,nearlyone-thirdofallIPtrafficin2016wasaccountedforbyNetflixalone, a usage volume not possible one or two years earlier. The highly visible innovations in consumer content delivery raises the question of whether existing national accountsaremissingconsequentialgrowthinoutputandincomeassociatedwithcontentdeliveredtoconsumers viatheiruseofdigitalplatforms. ThechangingproductionborderfordigitalcontentdeliverysuggeststhatGDP (as well as other macroeconomic measures, such as PCE prices) need to account for the substitution away from market-based digital services consumption. While substitution between market activity and household activity 23

haslongbeenanissueinnationalaccounting,arguablyitdoesnottypicallysurfaceasafirst-ordermeasurement issue beyond the case of owner-occupied housing, which is addressed in the accounts.18 We believe the digitization of consumer content delivery presents a case akin to the selective treatment of housing in national accounts and that an imputation for omitted services for connected IT capital needs to be made to avoid imparting a bias to GDP. The case for imputation of owner-occupied housing is based on the size of the omitted services and the importance of accounting for it in international comparisions. The case we offer for accounting for the digitization of consumer content via an imputation for connected IT capital services is based on the relatively fast growth of the omitted services, i.e., the case rests on the fact that the omitted services provide an extra kick to real GDP due to their declining relative price (i.e., like business IT investment and service prices, as set out in Byrne and Corrado, 2017b). To restate the empirical relevance of the case, we estimate that consumer welfare due to growth in digital content consumption has been enhanced to the tune of $1,775 per connected user per year from 2004 to 2017 (2017 dollars). And when our demand complementarity framework is incorporated into existing GDP, we find that real consumer digital services contributes nearly .6 percentage points per year to U.S. economic growth from 2007 to 2017. 18Food produced and consumed on farms is another example of an imputation long included in national accounts to address substitution between home-produced and market activity, but it is not large in advanced economies. 24

References Abdirahman, M., D. Coyle, R. Heys, and W. Stewart (2017). A comparison of approaches to deflating telecoms services output. Technical Report Discussion Paper 2017-04, ESCoE and ONS. Aizcorbe, A., D. M. Byrne, and D. E. Sichel (2019). Getting smart about phones: New price indexes and the allocation of spending between devices and services plans in personal consumption expenditures. Working Paper No. 25645, National Bureau of Economic Research. Berndt, E. R. and M. A. Fuss (1986). Productivity measurement with adjustments for variations in capacity utilization and other forms of temporary equilibrium. Journal of Econometrics 33(1), 7–29. Brynjolfsson,E.,A.Collis,W.E.Diewert,F.Eggers,andK.J.Fox(2019). GDP-B:Accountingforthevalueof new and free goods in the digital economy. Working Paper No. 25695 (March), National Bureau of Economic Research. Brynjolfsson, E., A. Collis, and F. Eggers (2019). Using massive online choice experiments to measure changes in well-being. Proceedings of the National Academy of Sciences 116(15), 7250–7255. Brynjolfsson, E. and A. Saunders (2009). Wired for innovation: How information technology is reshaping the economy. MIT Press. Byrne, D. M. (2015). Prices for data storage equipment and the state of IT innovation. FEDS Notes (July 15), Federal Reserve Board, Washington, D.C. Byrne, D. M. and C. A. Corrado (2015a). Prices for communications equipment: Rewriting the record. Finance and Economics Discussion Series 2015-069 (September), Board of Governors of the Federal Reserve System, Washington, D.C. Byrne, D. M. and C. A. Corrado (2015b). Recent trends in communications equipment prices. FEDS Notes (September 29), Federal Reserve Board, Washington, D.C. Byrne, D. M. and C. A. Corrado (2017a). ICT asset prices: Marshalling evidence into new measures. Finance and Economics Discussion Series 2017-016 (February), Board of Governors of the Federal Reserve System, Washington. Byrne,D.M.andC.A.Corrado(2017b). ICTServicesandtheirPrices: WhatdotheytellusaboutProductivity and Technology? International Productivity Monitor (33), 150–181. Byrne, D.M. andC. A.Corrado (2019). Consumer price misstatement: IT prices inthe 21stcentury. Technical report, (forthcoming). Christensen, L. R. and D. W. Jorgenson (1969). The measurement of US real capital input, 1929–1967. Review of Income and Wealth 15(4), 293–320. Christensen, L. R. and D. W. Jorgenson (1973). Measuring economic performance in the private sector. In M. Moss (Ed.), The Measurement of Economic and Social Performance, pp. 233–351. NBER. Copeland, A. (2013). Seasonality, consumer heterogeneity and price indexes: the case of prepackaged software. Journal of Productivity Analysis 39, 47–59. Corrado, C. (2011). Communication capital, Metcalfe’s law, and U.S. productivity growth. Economics Program WorkingPaper11-01, TheConferenceBoard, Inc., NewYork. Availableathttp://papers.ssrn.com/sol3/ papers.cfm?abstract_id=2117784. Corrado,C.andK.J¨ager(2014). Communicationnetworks,ICT,andproductivitygrowthinEurope. Economics Program Working Paper 14-04, The Conference Board, Inc., New York. Corrado, C. and O. Ukhaneva (2016). Hedonic Prices for Fixed Broadband Services: Estimation across OECD Countries. Technical report, OECD/STI Working Paper. 25

Corrado,C.andO.Ukhaneva(2019). HedonicPriceMeasuresforFixedBroadbandServices: Estimationacross OECD Countries, Phase II. Technical report, OECD/DSTI/CDEP Report to CISP and MADE, revised. Corrado, C. A. and B. van Ark (2016). The Internet and productivity. In J. M. Bauer and M. Latzer (Eds.), Handbook on the Economics of the Internet, pp. 120–145. Northamption, Mass.: Edward Elgar Publishing, Inc. D´ıaz-Pin´es, A. and A. G. Fanfalone (2015). The role of triple- and quadruple-play bundles: Hedonic price analysis and industry performance in France, the United Kingdom, and the United States. Paper presented at 43rd research conference on commmunications, information and internet policy, George Mason University School of Law, Arlington, VA. Diewert, W. E. and K. Fox (2017). The digital economy, GDP and consumer welfare, paper presented at the CRIW workshop, NBER Summer Institute, Cambridge MA (July 17–18). Gordon, R. J. (1990). The Measurement of Durable Goods Prices. Chicago: University of Chicago Press. Greenwood, J. and K. A. Kopecky (2013). Measuring the welfare gain from personal computers. Economic Inquiry 51(1), 336–347. Hulten, C. (1986). Productivity change, capacity utilization, and the sources of efficiency growth. Journal of Econometrics 33, 31–50. Hulten, C. (2009). Growth accounting. Working paper, NBER Working Paper 15341 (September). Hulten, C. R. and F. C. Wykoff (1981a). The estimation of economic depreciation using vintage asset prices. Journal of Econometrics 15, 367–396. Hulten, C. R. and F. C. Wykoff (1981b). The measurement of economic depreciation. In C. R. Hulten (Ed.), Depreciation, Inflation & the Taxation of Income from Capital, pp. 81–125. The Urban Institute. Jorgenson, D.W.(1963). Capitaltheoryandinvestmentbehavior. American Economic Review 53(2), 247–259. Jorgenson, D. W. and Z. Griliches (1967). The explanation of productivity change. The Review of Economic Studies 34(3), 249–283. Jorgenson, D. W. and J. S. Landefeld (2006). Blueprint for expanded and integrated U.S. accounts: Review, assessment, and next steps. In D. W. Jorgenson, J. S. Landefeld, and W. D. Nordhaus (Eds.), A New Architecture for the U.S. National Accounts, Volume 66 of NBER Studies in Income and Wealth, pp. 13–112. Chicago: University of Chicago Press. Available at http://www.nber.org/chapters/c0133.pdf. Nakamura,L.,J.Samuels,andR.Soloveichik(2016). Valuing‘free’mediainGDP:Anexperminentalapproach. Technical report, Bureau of Economic Analysis. Nakamura,L.,R.Soloveichik,andJ.Samuels(2018).“Free”InternetContent: Web1.0,Web2.0,andthesources of economic growth. Technical Report Research Paper WP 18-17, Federal Reserve Bank of Philadelphia. Petrin, A. (2002). Quantifying the benefits of new products: The case of the minivan. Journal of Political Economy 110(4), 705–729. Rotemberg, J. J. and M. Woodford (1995). Dynamic general equilibrium models with imperfectly competitive product markets. In T. F. Cooley (Ed.), Frontiers of Business Cycle Research, pp. 243–293. Princeton University Press. U.S. Department of Commerce.Bureau of Economic Analysis (2003). Fixed Assets and Consumer Durables Goods in the United States, 1925–99. Washington, D.C.: U.S. Government Printing Office. Williams, B. (2008). A hedonic model for Internet access service in the Consumer Price Index. Monthly Labor Review (July), 33–48. 26

Appendixes A1 Network utilization This appendix provides a derivation of equation (18) in the main text, i.e., we set out how to extract a measure of network capital utilization from productivity data and documents the calculations reported in section 3.1. A1.1 Derivation Whatfollowsisbasedontheframeworksetoutforanalyzingcommunicationnetworksandnetworkexternalities in Corrado (2011), in which it is assumed there no markups due to imperfect competition or other inefficiency wedges; see also Corrado and J¨ager (2014) and Corrado and van Ark (2016). In sources-of-growth accounting, the contribution of private capital is expressed in terms of the services it provides. Let the value of the relevant private stocks be denoted as PIK where the price of each unit of capital PI is the investment price and the real stock K is a quantity obtained via the standard perpetual inventory model. Inourapplication,thevaluePIK representsthereplacementvalueofnetworkserviceprovidercapitalin terms of its capacity to deliver digital services (i.e., including in this application, the value of the “originals” for the content the provider can diseminate). The value PKK represents the service flow provided by that capital. The price PK is an unobserved rental equivalence price, but which is related to the investment price by the user cost formula, PK =PI(r+δ−π)T, where r is an after-tax ex post rate of return, δ the depreciation rate used inthe perpetual inventory calculation, π is capitalgains, and T is theHall-Jorgenson tax term. The rental equivalence price is simplified by defining the gross return Φ=(r+δ−π)T, so that when capital services PKK are equated with observed capital income via the residual calculation of an ex post after-tax rate of return r, we have (A1) observed capital income=PIK∗Φ When capital services are computed on the basis of an ex ante financial rate of return r, the value for capital income of network providers must be expressed differently. Defining the ex ante gross return Φ=(r+δ−π)T accordingly, network provider capital income is expressed as (A2) observed capital income=PIKuISP ∗Φ where uISP is network capital utilization and, via Berndt and Fuss (1986), capital utilization uISP (rather than r) exhausts capital income. Equating expressions (A1) and (A2) PIK∗Φ=PIKuISP ∗Φ and solving for uISP yields Φ (A3) uISP = . Φ This equation states that under the conditions set out in Berndt and Fuss (1986) the relationship between the ex post and ex ante gross rate of return for an industry or sector reflects its capital utilization. A1.2 Calculations Theimpliednetworkutilizationcalculatingaccordingtoequation(A3)whererinthedefinitionofΦiscalculated followingJorgensonandGriliches(1967)astheexpostreturnforthecombinedMotionPicture,SoundRecording, Telecommunications, and Broadcasting industries (NAICS 512,515,517) and where r in the definition of Φ is set to Moody’s AAA corporate bond rate. The ex post net return and the δ and π components of Φ and Φ were calculated by the authors for the combinedsectorusingdatafromBEA’sindustryaccounts(accessedOctober2018). Theresultsforuareshown in text figure 6. 27

A2 Access Service Prices and Consumption Tocalculateapriceindexforeachofthenetworkaccessservicesprovidedbythebusinesssector—cable,internet, mobile,andsubscriptionvideostreaming—webeginwithnominalspendinganddividebyameasureofaggregate time spent using the service adjusted for quality. These individual price indexes are aggregated to create an overall access price index used to deflate nominal spending on access services and produce a measure of consumption. For exposition and analysis, we also consider price indexes constructed using four alternative measures of quantity: the number of households subscribed to the service, the number of individual users, time spent on the service, and time spent adjusted for quality (our preferred measure for deflation). Thus four alternative indexes are calculated for each of the four services by dividing revenue by each of the alternative measures of quantity, yielding prices paid per household, per individual, per unit of time, and per unit of constant-quality time: P , H P , P , and P . I D Q Data sources and calculation methods for service prices are summarized in 9. A2.1 Nominal Spending For nominal spending, we use figures from the national accounts published by Bureau of Economic Analysis, table 2.4.5U, “Personal Consumption Expenditures by Type of Product.” In the cases of mobile access and video on demand, we developed additional detail as explained below. Cable Spending is taken from table line 215, “Cable, satellite, and other live television services.” We use “cable” as shorthand for spending in this category, which includes spending on the services of multi-channel video programming distributors (MVPDs) of all kinds, including in addition to cable television, programming provided via telecommunications service provider, direct broadcast satellite, home satellite dish, wireless cable, master antenna, and open video systems. Internet Spending is taken from table line 285, “Internet access.” Spending on internet services includes access via “dial-up” service and access via broadband whether obtained through a telecommunications service provider, a cable system, or a satellite system. We extrapolate a spending figure for 1987 using the growth rate of internet households. Mobile Spendingistakenfromtableline281,“Cellulartelephoneservices.” Mobileservicesspendingincludes access to broadband via smartphone as well as access to conventional features such as voice and text using smartphone or feature phone. We split nominal access spending between smartphone service and feature phone service,forwhichweconstructdistinctquantitymeasures,usingthenumberofsubscribersofeachtype(derived as explained below) and a judgmental assumption that price paid for a smartphone contract is four times the price paid for a feature phone contract. (At the time of writing, a casual review of prices on the Worldwide Web showed basic plans with no data were $10-15 per month and common smartphone plans were $40-60 per month.) Video Total video spending is taken from table line 220, “Video streaming and rental.”19 We focus on subscription video on demand (SVOD), which we use as an indicator for the broader category, due to data limitations.20 In particular, we construct estimates of revenue for the three most prominent SVOD providers– Netflix,AmazonPrime,andHulu—basedoncompanyfinancialreportsandpressreports. Netflixreportsrevenue per subscription beginning in 2012, which we extrapolate back to 2007 using the modest 2012-2013 growth rate. Revenue per subscription for Amazon and Hulu are assumed to be their standard charges ($7.99 per month for Huluand$79peryearforAmazonPrimethrough2013and$99peryearafterward). Thesefiguresaremultiplied by the number of households for each service estimated as described below. 19BEA also provides revenue for “Audio streaming and radio services (including satellite radio)”. We did not develop a price index for this category. 20In addition to SVOD, video streaming and rental as defined in the NIPAs encompasses one-off video on demand, such as sports events, and rental of DVDs, for which we do not have data. 28

A2.2 Households Cable Periodic reports from the FCC, “Status of Competition in Markets for the Delivery of Video Programming,” provide household subscription figures for 1990 to 2015, citing reports by consulting firm SNL Kagan. Earlier years were collected from Statistical Abstracts of the United States, which reports figures from Census of Housing. Figures for 2016 and 2017 were extrapolated using available reports from cable, telecom, and satellite service companies (Chartered, Comcast, AT&T, Verizon, DIRECTV, and DISH). Internet Periodic reports from the FCC, “Internet Access Services: Status,” provide household figures for broadband access for 1999-2016 and dial-up access for 2001-2009. Prior to 1999, we assume all access was via dial-up service. Dial-up service figures for years not covered by FCC reports were available from financial reports and press reports for America Online, Compuserve, Prodigy, Microsoft Network, AT&T Worldnet, and Genie. The company series were judgmentally extrapolated to the year of introduction for each service. Dial-up subscribers from 2010 onward were extrapolated using figures from America Online (AOL) through 2014 and the 2011-2014 rate of AOL subscription decline for 2015-2017. Mobile We do not have data on the number of households with cell phone service. We assume the share of households with service equals the share of individuals in the adult population with service. Video Netflix reports the number of paying members beginning in 2009, which we extrapolate back to 2007 using the 2009-2010 growth rate. Hulu and Amazon subscribers are collected from press reports, which typically cite estimates from eMarketer. Because eMarketer figures estimate the number of active users using an assumption of 2.5 users per subscribing household, we multiply these reported user figures by 0.4 to estimate the number of households, assuming one subscription per household. A2.3 Individuals Cable WescalecablehouseholdfiguresusingthenumberofresidentsatleasttwoyearsofageperTVhousehold reported by Nielsen for 1985, 1990, 1995, 2000, 2005, and 2010, interpolated and extrapolated. Internet In 1998, 2000, 2001, 2003, 2007, 2009-2013, 2015, and 2017, the Current Population Survey supplementalsurveyoncomputerandinternetuseprovidedestimatesoftheshareofpeoplelivinginahouseholdwith aninternetconnectionandtheshareofindividualsgoingonlineathome. Weusethisinformationtoconstructa timeseriesfortheshareofpeoplewhousetheinternetathomefor1998-2017foradultsandchildrenseparately. We extrapolate these shares back to 1987 using the growth rate for 1998-2009. These shares are applied to the average composition by age of U.S. households to derive the total number of home internet users by year. Mobile Number of cellphone users (smartphone and feature phone collectively) is taken from Consumer Telecommunications Industry Association (CTIA) estimates as reported in Statistical Abstracts of the United States for 1987-2004. Estimates for 2005-2017 are from population shares reported by the Pew Research Center (Pew) times the U.S. population. Pew also provides separate estimates for smartphone users, which are subtracted from total cellphone users to get (solely) feature phone users. Video For each SVOD service, the number of users is estimated by multiplying the number of households times the average household size reported by the U.S. Census for the year. That is, we assume all household members make use of the service. A2.4 Time Use Cable The Nielsen Corporation (Nielsen) provides time spent per day on live and time-shifted television by age group (2-11 years old, 12-17 years old, at least 18 years old) beginning in 1992, which we extrapolate to 1987 using the value for 1992. We weight these figures by the U.S. Census-reported share of the population at least2yearsoldforeachagegrouptogetanaveragehoursperdayforresidentsofhouseholdswithcableaccess which we multiply by total users to get total hours. 29

Internet Hoursspentusingtheinternetfor1992-2008wereusedfromStatisticalAbstractsoftheUnitedStates, variousyears,reportingestimatespublishedbyVeronis,Suhler,Stevenson(VSS).For2011-2017,Nielsenreports oftimespentaccessingtheinternetonacomputerwereused. Estimatesfor2009-2010wereinterpolatedandfor 1987-1991 a growth rate of 50 percent per year was assumed, yielding a trivial level for 1987 in order to match a report from VSS that hours were negligible prior to 1987. Mobile Our measure of time use for feature phones is talk time. Minutes of talking is calculated as a threeyear centered moving average of estimates taken from FCC reports, citing CTIA surveys for 1993-2014 and extrapolated. For smartphones, we use a three-year centered moving average of estimates from eMarketer available for 2011-2017 of average time spent per day with smartphones for U.S. adults, which we extrapolate back to 2005. Video To calculate hours spent on each SVOD service, we first estimate the data used in streaming using the share of internet traffic for each service reported by Sandvine, Inc. multiplied by the quantity of fixed internet traffic for the North American consumer market reported by Ciscos Visual Networking Index (VNI) reports. Sandvinereportsareavailableannuallyfrom2010to2014andfor2016; thesharesfor2014arelinearly interpolatedandtheshareforeachservicefor2017issetequaltoits2016value. Thenwedividebythenumber of bytes required to stream an hour of video to get the number of hours. The estimate of bytes per hour used is a weighted average of the number of bytes used for standard definition and high definition video streaming, where the share is estimated using VNI reports. In particular, VNI provides a high-definition share for SVOD of 0.59 for 2014. This estimate is extrapolated to 2010 using the growth reported in VNI for the high definition share of global managed IP video-on-demand traffic. The share is extrapolated further back to 2007 using a 5 percent growth rate, and forward to 2017 using the VNI forecast published in 2014, the last VNI vintage where Cisco provided data on the subject. A2.5 Quality-Adjusted Hours Cable To account for the increase in quality associated with the programming choices available to viewers, we scale hours by the average number of channels per cable system reported by the FCC. We use a natural log transformation, assuming for example that the additional quality obtained going from 100 to 200 channels equals the increase in quality obtained going from 10 to 20 channels. Internet Our indicator for quality of internet service is the VNI estimate of IP traffic for consumer fixed internet use for North America. We use North American traffic in the absence of information on the U.S. share, essentially assuming that the U.S. share of North American traffic is unchanged over time. Direct measures of theindicatorisavailablefor2005-2016,alongwithaforecastfor2017fromthelatestVNIreports,variousyears. We extrapolate back to 1994 using overall fixed internet traffic estimates for North America, and back to 1990 using global fixed internet traffic from VNI reports. For 1987-1989 we use the 1990-1993 growth rate. Mobile We assume the quality of talk time is unchanged over time, so no quality adjustment is necessary for feature phones. For smartphones, we use the volume (petabytes) of consumer mobile IP traffic per month for the North American market reported by VNI for 2005-2017, extrapolated to 2002 using the average growth rate for 2005-2008. Video Our quality-adjusted series is raw hours of viewing time scaled by a library quality indicator and multiplied by high-definition video share. Our indicator for the quality of SVOD service is the natural log of the size of the video library for each service measured in the number of equivalent feature films available for streaming. FCC reports in 2013 and 2016 provide data on the number of films and the number of TV seasons available on each service. Estimates from the press were found for 2010 and 2018. Netflix press releases provide data for 2007 and 2008. Missing years are interpolated. We reweight TV seasons using the judgmental assumption that two episodes of a television show are equivalent to one feature film and TV seasons have 15 episodes. The high-definition share adjustment employed to calculate hours of viewing time above is reverse to produce the quality-adjusted hours indicator, implying that the quality of high-definition viewing is 1.67 times thequalityofstandard-definitionviewing,correspondingtotheratioofdatatransmissionrequiredforeachtype, 5 megabits per second and 3 megabits per second, respectively. 30

A2.5.1 Price Indexes Table 13 shows the quality-adjusted price index for each access service and price indexes for each concept of quantity. Our aggregate quality-adjusted price index for access service, shown in the right-most column, falls 12.4 percent per year, on average, over the full period of this study. The price index decline accelerates over time,firstasinternetserviceaccountsforarisingshareofspending,inthe1997-2007period,thenasmobileand video on demand access become more important in the 2007-2017 period. (Decomposition of of growth rates in the final index into contributions from each margin is discussed in the paper.) 31

secirP ecivreS sseccA rof sdohteM dna secruoS :9 erugiF tenretnI dna elbaC )a( 32

secirP ecivreS sseccA rof sdohteM dna secruoS :9 erugiF enohptramS dna enohP erutaeF )b( 33

secirP ecivreS sseccA rof sdohteM dna secruoS :9 erugiF dnameD no oediV noitpircsbuS )c( 34

Figure 10: Price Indexes for Access Services 35

A3 Consumer IT Durable Prices and (Household) Investment DataSourcesandmethodsusedforconstructingnominalconsumerdurablespendingandpriceindexesusedfor deflation are summarized in 12. A3.1 Nominal Spending Nominal spending estimates were based on detailed personal consumption expenditures reported by BEA In particular, detailed annual-frequency estimates of spending by product type were allocated to the more detailed categories used in the paper based on the 2007 input-output tables. (The quinquennial “benchmark” inputoutput table from 2007 provides not only detailed product spending information but also commodity codes corresponding to the primary products of the industries of the North American Industry Classification System (NAICS).)Forexample,theannual-frequencyestimatesofPCEdetailedspendingincludeacategoryfor“video, audio,photographic,andinformationprocessingequipment”withfurtherdetailprovidedfor8commoditycodes, including “computer and electronic products.” The 2007 input-output table provides for the 6-digit industry of origin of the products within this category, allowing one to distinguish among personal computers, computer monitors, televisions, and so forth. In the case of cellular phones and digital cameras, outside sources were used. Although these categories can bederivedusingthemethoddescribed, theirshareofexpenditurehaschangedrapidlysince2007, renderingthe allocation process inaccurate. Expenditures on other products which share the relevant higher-level categories areoffsetproportionallytoaccomodatetherisingspendingoncellphonesandtheriseandsubsequentrapidfall in spending on digital cameras. Cellphones WeuseanestimateofcellularphonespendingintheU.S.consumermarketprovidedbyIDC,Inc., rather than estimates reported in the NIPA PCE detail tables for several reasons. Cellular phone equipment spending is not reported separately, appearing instead as part of a broader category, “telephone and related communication equipment”. And, as noted in Aizcorbe, Byrne, and Sichel (2019), this broader NIPA spending line does not account for the substantial portion of the relevant acquisition of consumer stocks of cell phones whichtakesplaceinconjunctionwiththepurchaseofcellularphoneservices. Incontrast, theestimatesforIDC impute a value for cell phones acquired as part of a service contract using the price a consumer would pay for the phone if acquired without a contract commitment. The IDC estimates thus provide a consistent estimate of the retail value of all phones acquired over time, which serves the purpose of measuring the household capital stock. As shown in 11, the IDC estimate of consumer cell phone expenditures is substantially higher than the NIPA estimate for the category containing cell phones. To corroborate the IDC estimate, we constructed an alternativeestimateusingU.S.salesatwholesaleprices,providedbyGartnerthrough2007,extrapolatedbycell phone importswhich dominate the U.S. marketreported by ITC, and inflated by 50 percent, a rough estimate of the retail margin in the cell phone market. This coarse indicator, shown by the gray line, is quite close to the IDC estimate. Digital Cameras Unit sales of digital cameras for the Americas market provided by the Camera and Digital Products Association are scaled by an average price series constructed by interpolating between estimates reportedinthepress(fallingfromroughly$4,000in1987toroughly$200in2007andremainingstablesincethen). AU.S.shareoftotalAmericasspendingisconstructedusingtherelevantlinefromthebenchmarkinput-output tables for 2007 for consumer spending on digital cameras, which yields a share of approximately 48 percent, which we assume is constant in our period of study. A3.2 Price Indexes Forequipmentprices,weuseeitherofficialestimatesorsubstitutesdrawnfromtheauthors’researchandinsome cases other national statistical agencies. Aggregate prices for three broad categories are shown in ??: audiovisual equipment (televisions, digital cameras, photographic equipment excluding digital cameras, other video equipment,audioequipment,andrecordingmedia),informationprocessingequipment(personalcomputers,data storage equipment, monitors, and peripherals), and communications equipment (cellular phones and telephone equipment excluding cellular phones). 36

Figure 11: Estimates of U.S. Consumer Cell Phone Spending Billions of dollars 80 70 Retail 60 Wholesale 50 40 Retail Alternative 30 PCE Category 20 10 0 1987 1992 1997 2002 2007 2012 2017 Source. IDCInc. (retail); authors' calculations (wholesale, retail alternative); Bureau of Economic Analysis (PCE category). Televisions We use the BEA PCE deflator for televisions, which corresponds to the BLS CPI for televisions. Digital cameras We use the CPI for cameras from Statistics Japan. Photographic equipment excluding digital cameras We use the BEA PCE deflator for photographic equipment, which corresponds to the BLS CPI for photographic equipment. Other video equipment We use the CPI for video cameras from Statistics Japan, available from 1990 forward, extrapolated backward using the Japanese CPI for cameras. Audio equipment We use the BEA PCE deflator for audio equipment, which corresponds to the BLS CPI for audio equipment. Recording media We use the BEA PCE deflator for recording media, which corresponds to the BLS CPI for video discs and other media. PersonalComputers WeusethepriceindexfromByrneandCorrado(2017a)forpersonalcomputersthrough 2014, extrapolated by the BEA PCE price for computers and peripherals augmented by the average difference between the growth rate of the BEA price index and the growth rate of the Byrne-Corrado price index for the 2009-2014 period. Data storage equipment We use the price index published by the Federal Reserve Board for computer storage equipment, which extends the price index developed in Byrne (2015). Monitors We use the BEA PCE deflator for televisions, which corresponds to the BLS CPI for televisions. 37

Computer peripherals We use the price index from Byrne and Corrado (2017a) for peripherals through 2014, extrapolated by the BEA PCE price for computers and peripherals augmented by the average difference between the growth rate of the BEA price index and the growth rate of the Byrne-Corrado price index for the 2009-2014 period. Other information processing equipment We use the BEA PCE deflator for calculators, typewriters, and other information processing equipment. Software and accessories WeusethepriceindexforprepackagedsoftwarefromByrneandCorrado(2017a) for non-game PCE software, extrapolated for 2015-2017 using the 5-year average growth rate. For gaming PCE software, we use the BLS producer price index for game software publishing, available for 1998-2009 and 2014- 2017, adjusted for the average difference between the PPI and Copeland (2013) over the 1998-2004 period. The 2010-2013 period is interpolated using the average growth rate in our index for the 2005-2009 period. For the 1987-1997 period, we use the BEA PCE price index for computer software and accessories. Cell phones We use the Byrne and Corrado (2015a) price index for cell phones for the 1987-2010 period and the Aizcorbe et al. (2019) index for 2010-2017. Telephone equipment excluding cellular phones We use the Byrne and Corrado (2017a) price index for telephones for the 1987-2014 period as extended and published by the Federal Reserve Board through 2017. 38

secirP dna gnidnepS latipaC remusnoC rof sdohteM dna secruoS :21 erugiF tnempiuqE snoitacinummoC dna lausiV-oiduA )a( 39

secirP dna gnidnepS latipaC remusnoC rof sdohteM dna secruoS :21 erugiF tnempiuqE gnissecorP noitamrofnI )b( 40

Figure 13: Price Indexes for ICT Durable Equipment Categories 41

A4 IT Durable Use Intensity, Service Prices and Consumption We construct measures of use intensity for each type of capital employed to connect to the access services discussed in the paper. These include personal computers and related capital (monitors, software, and data storage equipment), televisions, and cell phones. These use intensity measures allow us to identify the effect on IT capital services from users spending a greater share of their time on digital access services and consequently the imprint that free and purchased services have on consumption. Mechanically, constructing use intensity for a particular type of capital requires allocating time spent on accessingeachdigitalservicetothecapitalusedfortheaccess. Forexample,useintensityforpersonalcomputers isproportionaltotheshareofhouseholdtimespentaccessingfixedinternetservicesplustheportionoftimespent using SVOD when viewing programming through the computer. Likewise, television use intensity is affected by cable access and by a portion of SVOD viewing time as well. Using the ratio of aggregate time spent on each access service to the number of each type of capital held by households, we construct intensity measures as the share of the working day a given PC, TV, or cell phone is in use. As shown in 15, we measure the share of households with at least one of each device, the number of devices in use conditional on the household having such a device, the total number of hours hoursholds spend on each devices, and finally, the share of the day each device is in use. Service adoption The adoption of access services by household is derived from the household figures calculated in the previous section and is shown in figure 14. Subscription video on demand penetration has also risen briskly since appearing in 2007. The share of households with at least one of the major services reached 60 percent in 2017. Time spent on each service is allocated by device as discussed below. Computers Estimates of households with a personal computer are provided by U.S. Census Bureau for 1984, 1989,1993,androughlyannuallyfrom1989forwardincollaborationwiththesupplementalsurveypublishedby Current Population Survey. ThenumberofPCsperhouseholdisbasedonperiodicreportsfromtheResidential Electricity Consumption Survey published by the Energy Information Agency. As shown by the black line in figure15,internetaccessamongcomputerhouseholdswasroughly20percentasof1990andwasover90percent by2007. ThenumberofPCspercomputer-holdinghouseholdsrosefrom1toover2. Dividingthetotalnumber of hours on the computer by the number of devices, we find that the share of the working day the average PC was in use for accessing the internet or SVOD rose from 18 percent in 1987 to roughly 29 percent in 2017. 21 Televisions EstimatesofhouseholdswithatelevisionareprovidedbyStatisticalAbstractsoftheUnitedStates, citing figures from Census of Housing. As shown by the grey solid line in figure 15, nearly all households had a television at the beginning of our period of study and this share remained above 90 percent as of 2017. The number of televisions per household is based on periodic reports from the Residential Electricity Consumption Survey publishedbytheEnergyInformationAgency. TelevisionsperTV-usinghouseholdmovedupfromroughly 2 to roughly 2-1/2 by 2005 and has eased down a touch since then. Dividing the total number of hours by the number of TVs in use yields a share of the day that peaked in 2013 at roughly a third and has moved down noticeably since then. The use intensity of PCs and TVs was roughly equal in 2017. Cell phones Mobile phones (whether feature phones or smartphones) are assumed to be present whenever individuals have service, so the issue of adoption of the service conditional on the presence of the equipment, does not arise. However, figure 15 shows that mobile phone adoption rose rapidly from 2007 to 2015 and has stabilized since then, and cell phone adoption overall has stabilized at 90 percent. As noted above, we use individual adoption rates as proxies for the household adoption rate in the case of cell phones. The share of households with mobile phone service rose rapidly from essentially zero at the beginning of our period of study to over 90 percent as of 2013 and was stable through 2017. The number of hours of use shot up with the advent of widespread smartphone use, and the share of the working day phones are in use shot up as well and stood at 18 percent as of 2018. 21Note that time spent using of the computer for other purposes, which averaged about 2-1/2 hours per day, is not included in this figure. 42

Figure 14: Conditional Access Service Use Intensity Shareof day 0.4 0.35 0.3 0.25 Cable 0.2 Internet 0.15 Mobile SVOD 0.1 0.05 0 1987 1992 1997 2002 2007 2012 2017 Source. Authors' calculations. Note. Day excludes 10 hours for personal care. 43

Figure 15: Equipment Use (a) Penetration (b) Multiplicity (c) Intensity 44