ARTICLE IN PRESS

0308-5961/$ - se

doi:10.1016/j.te

�CorrespondE-mail addr

r.fildes@lancas

Telecommunications Policy 31 (2007) 265–275

www.elsevierbusinessandmanagement.com/locate/telpol

Segmental new-product diffusion of residentialbroadband services

Alastair Robertson�, Didier Soopramanien, Robert Fildes

Department of Management Science, Lancaster University Management School, Lancaster LA1 4YX, UK

Abstract

Aggregate models of innovation diffusion do not capitalise on valuable consumer adoption dynamics that may be useful

to policy makers and market planners. The non-diffusion choice literature shows quite clearly that these dynamics may

indeed be very important factors in the diffusion process. The authors present a segmental broadband diffusion model that

is estimated from consumer survey data that measure the effect that household income has on its propensity to adopt this

technological product. The results suggest that early broadband adopters are mostly made up from wealthy households

and only as time progresses do less well off households adopt. The findings presented in this paper will be important to

market planners and policy makers requiring a relatively simple technique that forecasts segmental innovation diffusion.

r 2007 Elsevier Ltd. All rights reserved.

Keywords: Internet; Broadband; Diffusion; Gompertz; Forecasting; Innovation

1. Introduction

The established route to the modelling of innovative new-product penetration throughout the early stages ofa product lifecycle relies on the use of diffusion models. Some examples of where diffusion models have beenapplied to forecast innovative product lifecycles are IBM mainframe computers (Mahajan & Muller, 1996),residential high-speed internet services technology (Vanston, 2002) and mobile phones (Botelho & Ligia CostaPinto, 2004). These papers show that diffusion models can form an important part of the armoury of thepolicy maker and market planner.

Diffusion models provide an estimate of the hazard rate, defined as the probability that an innovation willbe adopted at a particular time by a particular individual within a given social system, providing that theindividual has not yet adopted the innovation. This definition suggests that each potential adopter has anequal probability of adopting the product as it launches (Roberts & Lattin, 2000). This must be viewed as anunrealistic assumption, however, as for many product types early adopters can be characteristically differentfrom late adopters (see Robertson, Soopramanien, & Fildes, forthcoming). This implies that the hazard rate ispotentially non-homogeneous. Evidence for this hypothesis in technology markets is abundant. Analyses of

e front matter r 2007 Elsevier Ltd. All rights reserved.

lpol.2007.03.006

ing author. Tel.: +44 1524 593867; fax: +44 1524 844885.

esses: a.w.robertson@lancaster.ac.uk (A. Robertson), d.soopramanien@lancaster.ac.uk (D. Soopramanien),

ter.ac.uk (R. Fildes).

ARTICLE IN PRESSA. Robertson et al. / Telecommunications Policy 31 (2007) 265–275266

the digital divide show that there are individuals in society who use information and communicationtechnologies (ICTs) effectively when others do not, whether voluntarily or involuntarily (Cabinet Office,2005). Although academic models that measure residential ICT markets are relatively hard to find, those thathave been published highlight strongly that the characteristics of adopters and non-adopters of ICTs varysystematically. Determined via the application of probabilistic internet choice models (see Chaudhuri, Flamm,& Horrigan, 2005; Kridel, Rappoport, & Taylor, 2002; Robertson et al., forthcoming), households excludedby the digital divide tend to be poorer, older and less well educated. This is a firm evidence that the productlifecycle is not homogeneous in hazard rate as aggregate diffusion models assume. Given this compellingevidence, this research introduces a segmental broadband diffusion model that is estimated using householdpenetration rates, to take into account income differences.

One of the major constraints faced by researchers exploring new-product penetration dynamics is the lack ofdata from which time-dependent segmental analyses can be made. Recent publications (Chaudhuri et al., 2005;Kridel et al., 2002; Robertson et al., forthcoming) show how multivariate probabilistic choice models (e.g.logit) can be applied to predict broadband penetration. Although these techniques are useful to anunderstanding of the interior mechanisms of the market, they cannot easily be applied to forecast futuretechnology penetration, nor explain past temporal patterns. To overcome the problem of predicting futurebroadband penetration the authors use survey data to measure the time scale to adoption. The advantage ofthis process is that they are also able to estimate Gompertz broadband adoption curves for differenthousehold income levels that characterise the reality of income-related digital division. The advantage of thismethod over aggregated approaches is that the planner can observe and estimate consumer activities thatconform well to existing understanding of the underlying economic processes of demand; it is therefore easilydescribed to non-technical policy makers. A second advantage to this method is that the choice ofsegmentation variable is left open to the analyst to decide. Although this technique is exemplified by applyingit to forecast segmental broadband penetration, the technique that is described could be re-applied to mosttypes of technology where survey data are available, from which consumer segments may be derived (e.g.residential VoIP, online shopping).

This paper is organised as follows. The literature review introduces the research on diffusion that is used todevelop the model that is presented in Section 3. A description of how the data are collected and managed isthen provided under Section 5. This leads to Section 6 that provides a detailed description of the segmentalGompertz results. Sections 7 and 8 offer a discussion of the findings, limitations and avenues for furtherresearch with final conclusions.

2. Literature review

The process of diffusion of innovative products can be defined as the cumulative sales or distributionpathways that some new products take from the point of market entry to the time that the market becomessaturated. Observation suggests that diffusion processes tend to follow a sigmoid shape (Bass, 1969; Mahajan,Muller, & Wind, 2000). Based on this Rogers (1962, 1995) selected the normal density to model markets as thisprovided a close fit to product penetration patterns. Further developments in diffusion were provided by Bass(1969) who built on the Rogers’ conceptual framework by developing a mathematical model that captured thenon-linear structure of S-shaped curves. One of the more complex algorithms to estimate the product lifecycle,in its original form, produces a model that at 50% market saturation also accounts for 50% of the time tocomplete the product lifecycle (i.e. it is symmetric like the normal distribution). A problematic issue ariseshowever. When the model is estimated prior to the 50% saturation point it tends to over-estimate marketdemand (Meade & Islam, 2001), a serious limitation for policy makers and market planners needing reliabledemand estimates. Only when the point of inflexion is reached, i.e. the maximum rate of adoption, do theparameter estimates of the model converge. Of course, this usually happens too late in the product lifecycle tobe of use to practitioners and policymakers. Although the Bass model has been amended in a number of waysthat relax the normal curve assumption, this has led to models of high complexity that carry forecast accuracypenalties (Meade & Islam, 2001).

Meade and Islam (2001) describe a number of algorithmic models that may be employed to forecastinnovative product demand. A fundamental problem they warn of is that forecast accuracy generally falls as

ARTICLE IN PRESSA. Robertson et al. / Telecommunications Policy 31 (2007) 265–275 267

model complexity increases. This has important ramifications for researchers and practitioners who need toestimate demand using limited quantities of estimation data. As Fildes (2002) points out this is a strongcriticism of many of the existing research studies. Usually it is the case that novel and increasingly complexmodels are tested using data from already saturated markets (e.g. TV, IBM mainframe computers). This, ofcourse, does not reflect the needs of the practitioner with perhaps only a few years, sometimes less, of demanddata from which models may be estimated (Bottomley & Fildes, 1998). An alternative algorithm that is moreeasily estimated when limited quantities of data are available was provided by Gregg, Hassel, and Richardson(1964). This forecasting tool, the Gompertz curve, is shown below:

NðtÞ ¼ e�a e�bi t

, (1)

where N(t) is the cumulative proportion of adopters within the social system at time t. The Gompertz modelhas two parameters to be estimated, a and b, using an optimisation process such as non-linear least squares(NLLS). The larger the a coefficient becomes the longer it takes for the market to reach saturation. For b thereverse is the case where larger coefficients imply a faster adoption cycle. This model, like the more complexBass model, also estimates an S-shaped diffusion path although its shape is asymmetric, a property oftenfound in the diffusion of household technological products (Vanston, 2002). This would imply a non-normaldistribution to the diffusion process that may be more correctly applied to the market place under certaincircumstances. Using the Gompertz model Vanston (2002) provided diffusion forecasts of high-speed internetadoption for US households aggregately. At the time of the study little data were available from which thedemand for high bandwidth could be estimated as the product had only recently launched. Vanston overcamethis issue by applying historical analogies to estimate his model. Vanston’s need in late 2000 was to forecastthe demand for high-speed digital access at the household level in the USA. He inferred possible a and bestimates by estimating a Gompertz model for each of the following household technologies: radio, television,colour TV, CD player, VCR, pay cable, PC, cellular and narrowband. He then averaged the estimatedcoefficients and applied these to the high-speed access model. Using this process he predicted that by 2005,65% of all US households would be online and that some 33% of online households would have adoptedhigh-speed internet services. Ex post, these estimates have proved quite realistic given the length of theforecasting horizon (5 years).

Vanston’s study shows the value of the Gompertz curve when the availability of relevant existing marketdata were limited. The consensus opinion in the forecasting community, however, is that when actual demanddata are available then it should be used to estimate the model (Meade & Islam, 2001). In the UK, broadbandlaunched in late 2000 but never really gathered any critical mass until 2002 when the sole supplier of wholesalebroadband services, British Telecom PLC, began substantive marketing campaigns that successfullyhighlighted the attributes of the broadband product. Today, however, the product is mass market and99.6% of households can now receive this product to at least 0.5 kbp. Since this product has been available forsome time it is now possible to study the characteristics of the adopters of this product in detail. The questionthat Robertson et al. (forthcoming) raised is whether the socio-demographic profile of broadband users variedsystematically from the population at large and particularly current dial users. Specifically, using a logit choicemodel on a subset of consumers they estimated that the characteristics of broadband users varied fromnarrowband users using data collected from UK households during 2003. Kridel et al. (2002), using US data,also find that the demographic profile of high-speed cable users systematically varied between dial-access usersand high-speed cable users. Both studies highlight the importance of household income as strong contributorsto this divide. Additional evidence comes from British Government reports which suggest that householdincome is the most correlated variable with the digital divide (Cabinet Report, 2005) and also shows thatinternet penetration rates are slowest for low-waged households.

Given the importance of household income in driving the digital divide, the question is how this variablemay be exploited within the diffusion context to understand and forecast diffusion patterns. To date, fewacademic diffusion studies have applied economic variables successfully despite the evidence from alternativemodelling strategies (e.g. choice modelling) that suggest that they are important in the adoption process.Although Horsky (1990) provided an aggregate model that incorporated wages and price into an aggregatediffusion process successfully, more generally there has been little agreement as to the positioning of covariates(like price) within the equation structure (Bottomley & Fildes, 1998; Roberts & Lattin, 2000).

ARTICLE IN PRESSA. Robertson et al. / Telecommunications Policy 31 (2007) 265–275268

Segmentation offers an alternative route to the application of consumer-specific covariates into the diffusionprocess. All models of innovation diffusion, estimated on aggregate data, lose the richness of known consumerbehaviours. Segmental diffusion models have tried to address this issue by assuming that different consumerscome to market at different points in time, a process that naturally implies a consumer-specific hazard rate(Roberts & Lattin, 2000). Very often the limiting factor in estimating these types of model is data availability.Aggregate data sources are usually readily available to the planner but individual level data are much scarcer.The application of segmental new-product diffusion has been tested using consumer risk aversion as asegmentation variable (Chatterjee & Eliashberg, 1990). Their model assumes that Bayesian updating leads tochanging evaluations of the product due to changing information levels (e.g. increasing word of mouth effectsas the adopting cohort increases) which in turn leads to a shifted hazard rate for those not yet adopted. Lattinand Roberts (1999) provided a model that they estimated in a pre-launch setting using survey data. There areseveral advantages to their process. First, the use of individual level data gave the researcher the opportunityto look at adoption behaviour at different stages in the product lifecycle (i.e. how will adoption behaviour foreach segment evolve over time?). Secondly, the researcher is able to look at specific consumer segments todiagnose problematic adopting cohorts (e.g. why is the digital divide so persistent and what remedial action, ifany, should be taken?).

Overall the key point to highlight from this discussion is that as markets evolve, so may the characteristicsof the adopters. If this is the case for the specific markets studied then it may be possible to observe changeproviding data are available. In summary, Rogers (1962, 1995) provided an initial framework ofunderstanding that suggested that different types of consumer enter the market at different stages of theproduct lifecycle. Later research increased the richness of this hypothesis by deriving diffusion models basedon microeconomic fundamentals (Horsky, 1990). This was taken further still by applying consumercharacteristics to produce segmental diffusion processes. The choice analyses described suggest strongly thatthe characteristics of the adopters of the latest technological products vary from those of earlier generations(e.g. higher income households tend to adopt first). The important question is whether the differences inpenetration that seem to exist may be capitalised upon by policy planners using diffusion techniques whenonly limited amounts of segmental product penetration data are available? There may be limited quantities ofdata available for several reasons.

(i)

Either the product is very young but sufficient data are available from a consumer survey that measuresactual or potential product adoption paths.

or,

(ii)

1T

poli

Segmental consumer product adoption data have only been captured very recently on a more matureproduct, and that consumers have been asked to state when they adopted the product. This is similar to thecase that is presented in this paper.1

The remainder of this paper now looks at how household survey data can be applied in understanding theeffect that household disposable income has on Gompertz innovation diffusion.

3. Segmental new-product diffusion

If the Gompertz curve represented in Eq. (1) is the entire market, then a model for segments is representedby

NiðtÞ ¼ e�a e�bi t

, (2)

where Ni(t) is the cumulative proportion of adopters from segment i at time t. It implies that differentconsumer segments may yield exploitable patterns such that the coefficient sizes (ai and bi), and therefore rates

he second point may be applicable to developing markets in countries where consumer level databases are scarce but where recent

cy developments (through aid agencies like the UN and ITU) are encouraging their development (see ITU, 2005).

ARTICLE IN PRESSA. Robertson et al. / Telecommunications Policy 31 (2007) 265–275 269

of diffusion, will vary systematically. To recover diffusion for the entire market, i.e. N(t) from Eq. (1), it isassumed that each segment remains a constant proportion, ai, of the full social system. This leads to Eqs. (3)and (4) where Dit is defined as the weighted proportion of cumulative adoption for segment i at t; relative tothe entire social system:

NðtÞ ¼Xn

i¼1

Dit, (3)

where

Dit ¼ aiNiðtÞ (4)

and

Xn

i¼1

ai ¼ 1. (5)

This is the model employed to estimate broadband adoption for the national market. Although the choice ofsegmentation variable is left open to the researcher to choose (i.e. depending on the nature of the study),income is applied as this is one of the strongest influencing factors in the digital divide. Providing that apattern of income sensitive diffusion curves exists in the broadband market then wealthier households shouldhave smaller a coefficients and larger b coefficients than poorer households.

Two national market models are estimated in this paper. The first estimates national broadband diffusionusing aggregate survey data (i.e. using Eq. (1)) whilst the second estimates using the segmental approachdescribed in this section, again using data collected from the survey. It should be noted that the segmentalcumulative proportion of adopters may also be derived from national time series, like the Ofcom dataintroduced in Fig. 2, but applying the survey data to derive the segmental penetration patterns. This wouldrepresent the ideal case as the national estimates are likely to be more accurate than the survey implyingminimal uncertainty about the final forecast estimates. In this paper, however, the case is presented thatnormally faces practitioners and researchers i.e. where no national penetration estimates are available tovalidate survey data (e.g. residential VoIP penetration). How this is accomplished is described in greater detailin the next section.

4. Estimation data

The data used to estimate the models described in the previous section were collected from two sources.The first source was collected using postal surveys that were sent to households across the UK betweenJuly and August 2005. A response rate of approximately 20% (1221 cases) was achieved across a widespectrum of social groups. The survey focussed on ICT adoption and usage at the household and individuallevel. Non-response bias was encountered and the data were weighted to national averages, supplied by theOffice of National Statistics and the Department for Education, to compensate on key socio-demographicvariables (age, educational attainment and gender) and also on computer penetration (source Ofcom). Theprocess of rim weighting was used to correct the non-response bias that was encountered (Barnett, 1991;Elliot, 1995).

To derive cumulative broadband demand over time, respondents were asked to select from the following sixoptions the length of time that they had access to broadband in their homes: less than 3 months, 3–6 months,6–12 months, 1–2 years, 2–3 years and finally more than 3 years. Respondents were also asked to providedetails of household disposable income as this is used to segment respondents into six income categories, eachsegment becoming a broadband demand segment. Broadband adoption was also estimated for thosehouseholds who declined to provide income data (n ¼ 128). As it turns out they have a very low propensity toadopt this technology. The final income categories used in the remainder of this study are therefore MissingIncome, under £10,000, £10,000–14,999, £15,000–24,999, £25,000–39,999 and £40,000 and above.

ARTICLE IN PRESSA. Robertson et al. / Telecommunications Policy 31 (2007) 265–275270

5. Descriptive results

The evidence provided by households confirms that broadband penetration levels are income sensitive asexpected. Fig. 1 shows penetration levels for each income segment.

A consistent pattern emerges with the highest income households reporting a much faster adoption curvethan the lowest. For instance, 54% of the highest income earning households (income greater than £40,000 peryear) had adopted broadband by the middle part of 2005. For the lowest adopting segment (Missing Incomeand less than £10,000), barely 12% are connected to the net via a high-speed net connection. As income risesthere is a systematic increase in the adoption curve. More importantly for the model that has been selected, theunderlying slopes also vary systematically (i.e. they are steeper for higher income categories as expected). Thisis a considerable evidence that the ICT behavioural characteristics of wealthier households vary from thosefrom the lowest. Furthermore, this finding strongly suggests that for the broadband market the ‘innovators’tended to belong to wealthier households than poorer.

An essential element of any survey analysis is to ask how well the data that are used represent thepopulation at large. Although the authors do not have an independent source of segment-related broadbandpenetration data that we can be relied on for this purpose, the aggregated survey estimate can be comparedagainst national statistics provided by Ofcom. This is considered to be the most precise estimate of broadbanddemand for the UK as the information is provided under regulatory commitment from British Telecom PLC.Therefore, this data source can be used to validate the models that are estimated later in this paper. Note,however, that the validation is only conducted to confirm to the reader the validity of applying aggregatedconsumer reports of broadband adoption to estimate future broadband penetration. Fig. 2 compares theOfcom data to the weighted survey data. This figure also highlights by how much each income segment adds tototal broadband penetration.

The highest solid line represents the survey estimate of national penetration and is labelled Whole Market.Notice how closely this aligns to the data provided by Ofcom. These are the data that are used to estimate theaggregate model. Beneath the aggregate curves are the reported estimates for the penetration rates for eachsegment, scaled to the proportion of each segment that is measured in the survey (i.e. Eq. (4)). For instance,those earning £40,000 or more represent 18% of the survey sample of 1218. Survey estimates show that 55% ofthese households had broadband by the middle of 2005. In terms of the nation at large, this representsapproximately 10% (0:18 � 55%) of all households in the UK or approximately one-third of all broadbandhouseholds. The segmental Gompertz models are estimated on the data provided in Fig. 1; aggregation onlytakes place on the estimated curves once each model has been scaled to the proportions that its income class isfound in the survey.

0

0.1

0.2

0.3

0.4

0.5

0.6

2002 2003 2004 2005

Pro

po

rtio

n o

f H

ou

seh

old

s w

ith

in

Seg

men

t th

at

have B

road

ban

d

Missing Inc Broad <£10 Broad £0-£14.9

Broad £15-£24.9 Broad £25-£39.9 Broad >=£40

Fig. 1. Survey estimates of broadband adoption within income segment.

ARTICLE IN PRESS

Table 1

Results of segmental and aggregate Gompertz models

Model Coefficient Adjusted R2 values MAPE

Missing income a 12.96* 0.97 NA

b 0.36*

Less than £10,000 a 9.21** 0.99 NA

b 0.29***

£10,000–14,999 a 8.27** 0.99 NA

b 0.31***

£15,000–24,999 a 8.86* 0.99 NA

b 0.40*

£25,000–40,000 a 6.78** 0.99 NA

b 0.44**

More than £40,000 a 6.89** 0.99 NA

b 0.48*

Combined segmental model a NA 0.99 6.15%

bAggregate model a 6.67*** 0.99 9.27%

b 0.36***

***1% significance; **5% significance; *10% significance.

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

2002 2003 2004 2005

Pro

po

rtio

n o

f H

ou

seh

old

s

that

have B

road

ban

d

Missing Inc Broad <£10

Broad £0-£14.9 Broad £15-£24.9

Broad £25-£39.9 Broad >=£40

Ofcom Whole Market

Survey Estimate of thePenetration Rate

Fig. 2. Segmental and aggregate broadband adoption: national comparison.

A. Robertson et al. / Telecommunications Policy 31 (2007) 265–275 271

6. Estimation methodology and modelling results

The Gompertz curve is inherently non-linear and therefore requires estimation using non-linear estimationtechniques. Although a number of methods can be applied to solve this type of problem (e.g. log-likelihood)the recommendation of Srinivasan and Mason (1986) and Putsis and Srinivasan (2000) is followed; that is, touse NLLS as this method should provide least bias in the parameter estimates. The Gompertz models wereestimated in STATA 9, using its own NLLS estimator. Some attention is also required in the assessment of thestatistics that are provided. Since each model is estimated on only four points of data, the reliability ofstandard errors that are provided should be viewed with caution. Table 1 provides the discussed statistics foreach segmental Gompertz model and for the estimated aggregate model.

It was discussed earlier that the size of the a’s and b’s relate directly to the speed of adoption that is forecast(wealthier households should have smaller a’s and larger b’s). Recall also that Fig. 2 highlighted the generality

ARTICLE IN PRESS

Fig. 3. Predicted segmental adoption rates.

A. Robertson et al. / Telecommunications Policy 31 (2007) 265–275272

that higher income households have faster adoption curves than lower income households. The coefficienttable suggests this to be broadly correct. In terms of the a coefficients in the segmental Gompertz models, asincome rises the speed of adoption is ramped upwards by smaller coefficients. The Missing Income segmenthas the largest coefficient in this category. The b coefficients gain in magnitude consistently across the incomegroups as expected with the Missing Income segment lying between the £10,000–14,999 and £15,000–24,999segments. All coefficients are statistically significant at 10% or better. The a coefficient in the aggregate modelis measured to be 6.67 whereas the b coefficient lies in the middle of the segmental range at 0.36.

The adjusted R2 values, a measure of in-sample fit, are in excess of 0.97 in the segmental models and 0.99 forthe aggregate. It is important to note, however, that R2 values are typically large when estimation values aretemporally correlated.

The Mean Absolute Percentage Error (MAPE) statistics are provided for both the aggregated estimatedmodel and for the combined segmental model. They give the reader a sense of accuracy of the forecastagainst the Ofcom estimates of broadband penetration, it is therefore an effective hold-out fit measure.The MAPE statistics measure how far, on average, each aggregate forecast lies away from the Ofcomestimates. For instance, for the aggregately estimated model the MAPE value of 9.27 suggests that,on average, the forecast is 9.27% away from Ofcoms stated values. This value improves for the segmentalmodel to 6.15%.

So far the modelling strategy demonstrates two important points. First, that the use of consumer surveydata can yield accurate penetration rates for recently launched technologies. Secondly, that the segmentalapproach provided here may improve forecasting accuracy over aggregate approaches, as evidenced byMAPE. Fig. 3 shows the un-scaled adoption rate curves that occur for each income segment using theestimated coefficients provided in Table 1. Three groups of curves are characteristically similar. The twohighest income earning households show similar adoption curves with the peak rate of adoption likely tooccur during 2005. These income segments account for 40% of the respondents in the survey. The slowestthree adopting segments also have similar predicted peak adoption rates that are due to occur during 2008.Between the extremes lie a distinct income segment, £15,000–24,999 that is due to reach its peak adoption rateduring 2006.

Fig. 4 shows how the national penetration forecast that is derived from the segmental model is derivedfrom the income segments. This model suggests that early penetration will derive mostly from higherincome households. Fig. 5 shows the adoption rate curves for the aggregately estimated model and thecombined segment model. Both curves are remarkably similar in shape and size and conform well toOfcom penetration rates. The combined segmental curve has a slightly lower maximum rate of adoptionthan the aggregate model, although peak rate times are identical and predicted for 2006. Fig. 6 introduces thefinal forecasts for both models. These curves, as would be expected having looked at Fig. 5, are almost

ARTICLE IN PRESS

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

Pro

po

rtio

n o

f H

ou

seh

old

s

that

have B

road

ban

d

Missing Income Less than £10.0 £10.0 to £14.9 £15.0 to £24.9

£25-£39.9 More than £40.0 Ofcom

Fig. 4. Predicted segmental penetration curves.

0.00

0.02

0.04

0.06

0.08

0.10

0.12

0.14

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

Seg

men

tal A

do

pti

on

Rate

s

Combined Segmental Model Aggregate Mode Ofcom

Fig. 5. Predicted national adoption rate: segmental vs. aggregate.

A. Robertson et al. / Telecommunications Policy 31 (2007) 265–275 273

identical. They both suggest that this market should provide internet services to approximately 95% ofhouseholds by 2015.

7. Discussion

This paper provides a methodology for estimating diffusion models from households’ stated adoption pathsat the segment level. The results suggest that the information provided by the respondents reflect accuratelynational penetration levels. The advantage this process has over aggregate national data is that the researchercan design the survey to capture key drivers that are known to affect the diffusion/adoption process. In thiscase household income was chosen as a key driver of the digital divide as prior research indicated this causalfactor to have one of the strongest impacts. It would be possible, however, to estimate the demand forbroadband using alternative measures such as social class, educational attainment or geographic regionproviding these data are available to the researcher.

ARTICLE IN PRESS

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

Pro

po

rtio

n o

f H

ou

seh

old

s

that

have B

road

ban

d

Combined Segmental Model Aggregate Model Ofcom

Fig. 6. Predicted national adoption: segmental vs. aggregate.

A. Robertson et al. / Telecommunications Policy 31 (2007) 265–275274

On a technical note, the survey data used in this research successfully measured the adoption curves ofhouseholds to yield a time series for each segment that is four data points in length. Ideally one would estimateany diffusion model using more data but when data are limited, more complex models tend not to converge onreliable parameter estimates. The Bass model was estimated on each income segment but would not convergesuccessfully on three of them, even when coefficient boundaries were set. The Gompertz presented no difficultywhen estimating the parameters, however.

An issue that arises with the incorporation of consumer segments into the diffusion process is thatmovements between income segments over time, a natural life progression, is not accounted for bythe modelling strategy provided by this paper. Recent evidence from the Joseph Rowntree Foundation2 inthe UK suggests that socially excluded groups tend to fall below the poverty line but that this cohorthad decreased in size since 2000. This would imply that the hazard rate of poorer households of 2000that presently earn more money will have changed. Additional evidence of this issue became available duringthe data collection stage. Some respondents indicated in the survey that they wished to own a computer butcould not afford to do so. This would suggest that a proportion of the non-internet using households wouldbecome digitally included should their financial situation improve. As may be expected, the model predictsthat if lower income groups migrate to higher income categories the overall household adoption rate willincrease.

Although the key note of this paper is the test of a new segmental forecasting methodology, it alsocontains information for policy. Policy on the digital divide has difficulty in being effective without firsthaving some forecast of how the division may evolve over time. Such forecasts can provide more investmentinto areas that will ensure its effect will diminish. An issue for telecommunication firms wishing toexpand into new territories is that they will not invest where underlying demand forecasts are non-existentor very limited in scope as is more normally the case in developing countries. It has been shown herehow simple survey methods could be applied in difficult areas of the world that would give moreuseable and functional estimates of final ICT demand. It is noted that the International TelecommunicationUnion (ITU), through its Partnership on Measuring ICT for Development (ITU, 2005), provides guidanceto data collection agencies in developing countries as to the type of survey data it should collect.The authors show that asking respondents to provide a timeline for the adoption of ICT products thatreliable time series estimates of penetration can be made. We also show that the inclusion of segmentationdata into the surveys can yield a clear pattern of adoption that will be beneficial to policy makers in thesecountries and beyond.

2www.jrf.org.uk—‘‘Monitoring poverty and social exclusion in the UK 2005’’.

ARTICLE IN PRESSA. Robertson et al. / Telecommunications Policy 31 (2007) 265–275 275

8. Conclusion

This paper provides a methodology for understanding better the underlying processes involved ininnovative product diffusion. We find that by collecting data from householders during a single survey wavethat accurate time-based adoption paths could be estimated. Moreover, we find that well-defined householdadoption segments exist that explain known processes in the digital divide, indeed, these patterns aresufficiently clear that they are exploitable by diffusion models like the Gompertz. This will be useful to thepolicy maker and practitioner seeking guidance in such matters.

Acknowledgements

The authors are grateful to the ESRC (award number R42200124401), the Pump Priming Research FundingScheme at Lancaster University and the Small Research Grant Scheme at the Lancaster UniversityManagement School for funding this research.

References

Barnett, V. (1991). Sample survey principles and methods. London: Edward Arnold.

Bass, F. M. (1969). A new product growth for model consumer durables. Management Science, 15(5), 215–227.

Botelho, A., & Ligia Costa Pinto, L. C. (2004). The diffusion of cellular phones in Portugal. Telecommunications Policy, 28(5–6), 427–437.

Bottomley, P. A., & Fildes, R. (1998). The role of prices in models of innovation diffusion. Journal of Forecasting, 17, 539–555.

Chatterjee, R., & Eliashberg, J. (1990). The invasion diffusion process in a heterogeneous: A micro modelling approach. Management

Science, 36, 1057–1079.

Chaudhuri, A., Flamm, K. S., & Horrigan, J. (2005). An analysis of the determinants of internet access. Telecommunications Policy,

29(9–10), 731–755.

Elliot, D. (1995). Weighting for non-response: A survey researcher’s guide. OPCS Social Survey Division.

Fildes, R. (2002). Telecommunications demand forecasting—A review. International Journal of Forecasting, 18(4), 489–522.

Gregg, J. V., Hassel, C. H., & Richardson, J. T. (1964). Mathematical trend curves: An aid to forecasting. Edinburgh: Oliver & Boyd.

Horsky, D. (1990). A diffusion model incorporating product benefits, price, income and information. Marketing Science, 9, 342–365.

ITU. (2005). Core ICT indicators. Partnership on measuring ICT for development, United Nations, found at the international

telecommunication website /http://www.itu.intS.

Kridel, D., Rappoport, P. N., & Taylor, L. D. (2002). The demand for high-speed access to the internet: The case of cable modems.

In D. G. Loomis, & L. D. Taylor (Eds.), Forecasting the internet: Understanding the explosive growth of data communications

(pp. 11–22). Dordrecht: Kluwer Academic Publishers.

Lattin, J. A., & Roberts, J. H. (1999). Calibrating a restricted parameter individual level diffusion model prior to launch. CA: Working Paper,

Graduate School of Business, Stanford University.

Mahajan, V., & Muller, E. (1996). Timing, diffusion, and substitution of successive generations of technological generations: The IBM

mainframe case. Technological Forecasting and Social Change, 51, 109–132.

Mahajan, V., Muller, E., & Wind, Y. (2000). New-product diffusion models: From theory to practice. In V. Mahajan, E. Muller, &

Y. Wind (Eds.), New product diffusion models (pp. 3–24). Kluwer Academic Publishers.

Meade, N., & Islam, T. (2001). Forecasting the diffusion of innovations. In J. S. Armstrong (Ed.), Principles of forecasting. Boston:

Kluwer Academic Publishers.

Putsis, W. P., & Srinivasan, V. (2000). Estimation techniques for macro diffusion models. In V. Mahajan, E. Muller, & Y. Wind (Eds.),

New-product diffusion models (pp. 263–291). Dordrecht: Kluwer Academic Publishers.

Roberts, J. H., & Lattin, J. M. (2000). Disaggregate-level diffusion models. In V. Mahajan, E. Muller, & Y. Wind (Eds.), New-product

diffusion models (pp. 207–236). Dordrecht: Kluwer Academic Publishers.

Robertson, A. W., Soopramanien, D. & Fildes, R. (forthcoming). A segment-based analysis of internet service adoption among UK

households. In Technology in society.

Rogers, E. M. (1962). Diffusion of innovations. New York: The Free Press.

Rogers, E. M. (1995). Diffusion of innovations (4th ed.). New York: The Free Press.

Srinivasan, V., & Mason, C. H. (1986). Nonlinear least squares estimation of new-product diffusion models. Marketing Science, 5,

169–178.

Vanston, L. (2002). Forecasts for internet/online access. In D. G. Loomis, & L. D. Taylor (Eds.), Forecasting the internet: Understanding

the explosive growth of data communications (pp. 45–58). Dordrecht: Kluwer Academic Publishers.