What can Bass’ mathematical approach to innovation diffusion tell/show us? Well quite a lot. We can:
- explore impact of innovators (those in the social system that are external influences) and imitators (those internally influenced)
- predict diffusion of our innovation – allowing us to understand when to ramp up service provision capability (or product supply lines)
- get a more accurate sizing of adopter types to better target our, influencing budgets
- understand the best timing to launch the next generation of an innovation
Exciting stuff, right? Even more exciting, we get to explore the model hands-on in Excel.
Key Take Aways
- Bass models “The probability of adoption at time t given that adoption has not yet occurred” based on
- a coefficient of innovation called p (representing people that adopt from external influence) with average value of 0.003 and usually less than 0.01
- a coefficient of imitation called q (representing people that adopt through word of mouth..and so there have to have been some adoptions alread) with average value of 0.3 and usually between 0.38 and 0.5
- the amount of adopters already.
- It resembles the classic adoption curve from Rogers
- Coefficients p and q can be adjusted so the output maps remarkably close to real sales data
- so is beneficial in predicting sales for new product innovations
- …and, in some instances, new services)
- Using real examples, we can see more accurate values for the size of adopter types beyond Rogers use of standard deviations
- For example, Rogers has Innovators as 2.5% of market size; empirical use of Bass shows it is more likely 0.2-2.8%
- We can model multi-generations of innovations (Bass looks at computers)
- This opens up for us to determine the best time to launch the next generation to get quickest adoption
(You can jump to this article if you want a quick refresher on innovation diffusion).
Ready? Let’s get mathematical!
Bass’ diffusion model – the background
Bass’ A New Product Growth Model for Consumer Durables paper introduced the formula that allowed us to mathematically model
diffusion of innovation. You can find out a lot about the model and its background at the Bass Basement Research Institute.
Before we look at the formula, let’s see if the output is familiar (Figure 1).
I find it quite reassuring that we can mathematically create the same curve that Rogers’ described from his literature based studies. Of course, you may have spotted that Rogers’ curve is essentially a normal distribution with category sizes very close to standard deviations. That hardly needs a formula revelation to recreate….
However, Bass’ model is not the formula for normal distribution. It just happens to closely resembles that for certain choices of coefficients. We can best see this, by looking at how the model fits some real life examples.
Bass’ Model and Real life
Not only can we fit Bass’ model to Rogers’ work, we can reassuringly fit it against known sales data for innovations. You can see some examples in Figure 3.
Let’s take the bottom three graphs first. These come from Bass’ own work. And they show the sales (solid line) of clothes dryers, black and white tvs, and power lawnmowers over a number of years (his work is from the 1969). On the same graphs, in dots, is a fitted curve of the Bass model. OK, it’s not a perfect match, but it is remarkably close.
Now, let’s get a little bit more up to date. The two graphs at the top of Figure 3 look at diffusion of iPhone and Samung Galaxy devices. It comes from a paper that looks extends Bass’ model to deal with competitive price variations. However, I have just copied the graph of the standard Bass model. Taking figures from different papers has the joy of showing inconsistencies in how those papers draw graphs! Sales are now the dots, and the fitted Bass model is the solid line.
Again the Bass model fits quite well the actual sales. For the iPhone it is pretty much spot on. The Galaxy graph shows the shape, but there are many points above and below. Why is that? Well, the sales data is simply reflecting how price changes affect the decision to adopt an innovation. Whereas iPhone prices for a generation remain pretty stable, Samsung wildly vary their prices resulting in peaks and troughs of adoption. The underlying curve tend though is still aligning to Bass’ model.
Now we have a feeling for Bass’ model, let’s move towards the maths.
So what is the magic formula? It is a differential equation which Bass summarises as:
The portion of the potential market that adopts at t given that they have not yet adopted is equal to a linear function of previous adopters.From the Bass Basement
And Majaham, Muller and Bass make the definition a little more accessible:
The probability of adoption at time t given that adoption has not yet occurred is equal to: p + (q * cumulative fraction of adopters at time T).Majahan, Muller, Bass
This p and q , relate to coefficients of innovation and imitation.
The coefficients, p and q , relate back to our discussion in a previous article that a social system has two types of people and influences. Innovator types are influenced from outside the social system (often by PR, or by actively searching, or by knowing the innovator). They are represented by p, the coefficient of innovation.
Imitators, on the other hand, are influenced inside the social system. This is often by word of mouth, or observing the innovation in use. They are represented by q, the coefficient of imitation. To be influenced internally, there needs to have been people who have already adopted. Which is why q is multiplied by the cumulative fraction of adopters at time T.
We can see the impact of these coefficients in Figure 4. It is a little exaggeration to show the effect of innovators and imitators.
The intuition behind the model
At time 0, for a market size of M, the number of adopters is M*p. That is to say, Bass’ model assumes innovators adopt early (which aligns with othe diffusion theories). As time progresses the number of new innovators adopting, who have not previously adopted, diminishes.
However, the number of imitators adopting starts to increase to a peak. This also makes sense as imitator types need to be influenced internallly. As more and more people adopt there is greater chance of internal influence. More people are talking, more people can be observed using the innovation etc. At some point though, we start reaching saturation. And so the number of new adopters starts decreasing.
For most innovations the value of p is found to be very small (an average of 0.003 as will see shortly). This means the big step we see at time 0 in the exaggeration in Figure 4 (or the missing part of the curve if we compare to Rogers curve) is in reality quite small.
Bass Model Formula
So, here is the big reveal. I’ve waited on presenting the actual formula to minimse heart attacks. But I can wait no longer! Here in Figure 5 is the formula that describes the probability of an adoption taking place.
We’ve already come across the coefficients p and q. And the remaining terms are:
- f(t) — the portion of M that adopts at time t
- F(t) — the portion of M that have adopted by time t
What if we want to find, more usefully, the sales at any time t? Well, let’s say we have a market size of m. Then the number of sales S(t) at time t is equal to that market size multiplied by the portion of m adopting at time t. Or, m.f(t). And, if we juggle the formula in Figure 6 to get it in terms of f(t), then we get S(t) as defined in Figure 6.
Easy, right! Or intimidating? Well, either way, there’s nothing better than playing with the model to get a better understanding. And that’s what we do next.
Time to get hands-on
OK. Time to get hands on and play with Bass’ model. Over at the Bass Basement Research Institute site you can download an Excel spreadsheet with the Bass model in it.Its fun to play with Bass' Diffusion Model for innovations. Now you can predict how many new products you need per year – and its remarkably grounded in reality! Find out how and where to get the Excel model in this article (or… Click To Tweet
Once you download and open up, you’ll see something similar to Figure 7.
Up in the top left you can find changeable cells for M (market size, initially set to 1,000,000), p (coefficient of innovation, initially set to 0.003) and q (coefficient of imitation, initially set to 0.5). Over on the right you can see the diffusion curve generated at the top, and the cumulative adoption at the bottom.
Have a play with the values and see what it does to the diffusion curve.
Why is this important? Well, altering values of M, p and q allows us to fit the model to previous actual sales. But more interestingly, it allows us, with some caveats, to start predicting diffusion.
How many people do you need to staff your new service delivery? What stocking levels do you need in your supply chain for your new product? These are questions when launching an innovation. And, Bass’ model can help us understand.
If we can determine M, p, and q. then we can simulate the adoption at any time. You’ve already been playing at that in the last section. And earlier we saw that Bass’ model aligns pretty well with real life data.
However, how do we get these three values? First, your marketing team can run, say, market surveys to determine a realistic market size. Let’s say that they come back with the value of M=10 million people.
Second, we need to find values for the coefficients p and q. If you have launched similar innovations before, you can determine these coefficient values by fitting Bass’ model to sales data. If you haven’t, then perhaps you can still find a comparable from competitors data. Or you have to go back to good old assumptions and guestimates (which of course you can refine as you start to get sales data). But, for are examples sake, let’s say you determine p=0.008 and q=0.7.
Plug the data into the model, and look at the graphs, such as I’ve done in Figure 8.
Let’s assume our example is for a product, and that each adopter buys 1 product. Looking at the graph / data, we can see in
- year 1 there are 80K adoptions
- year 2 we need to produce 134K products
- rising to 1.78M in year 9
After year 9 the there is reducing volume – since we are saturating the market.
Lessons from predicting Colour TV sales
I just want to highlight a little anecdote from Bass’ “Comments on ‘A New Product Growth for ModelConsumer Durables The Bass Model’” paper
I decided to try my luck at forecasting color television sales, which had just begun to take off in the early 1960s….The result was a forecast that color television sales would peak in 1968at 6.7 million units…Industry forecasts were much more optimistic than mine and it was perhaps to be expected that my forecast would not be well received. As it turned out, color televi-sion sales did peak in 1968 and at a slightly lower level than my forecast. The industry had built capacity for 14 million color picture tubes and there was substantial economic dislocation following the sharp downturn in sales following the 1968 peak.
Predicting for Services (to be developed)
You may already have guessed that services have the potential to be slightly different. I find limited literature on diffusion of services. Most likely it is assumed services diffuse the same as products (i.e. Rogers’ and Bass’ models). This might not be correct. Service innovation:
- spreading through service companies
- (a new service) spreading through customers
- forced onto customers
The model is based on “first purchase” and not repeat purchase. Services are hopefully repeat – so the cumulative values of adoption are probably more relevant than the curve. But of course the model similarly does not account for people that stop using the service (unless you assume that people only use the service once and then abandon -in which case the curve fits…)
Of course, if you estimate the market size, M, wrong or misjudge values of p and q then predictions are going to be wrong. You can revisit your predictions once you have sales sizes. We can also compare the coefficient values to knowledge gained elsewhere.
Typical sizes of Bass’ coefficients
If, in our prediction, we are reduced to guestimates, then it is useful to know typical values of the coefficients. That way, we know if our guesses are not realistic.
The paper “Diffusion of new products: empirical generalizations and managerial uses” helps us. It collects together results of various other papers. When it comes to the coefficients, p and q, this is what it has to say:
- p + q lies between 0.3 and 0.7
- average value of p = 0.03
- average value of q = 0.38
- p is often quite small, 0.01 or less
- q is rarely greater than 0.5 and rarely less than 0.3.
The paper also helps us refine our view of Rogers’ adopter type sizes.
Refining Adopter Type Sizes
Rogers’ work on the size of adopter types and sizes is a fantastic insight and a great rule of thumb. His curve is one of normal distribution, and adopter sizes are essentially standard deviations (give or take a few decimal places). That is to say, the early and late majorities are 1 standard deviation, 34%.
But you may have noticed when playing around with Bass’ model in Excel that whilst the shape is similar to Rogers, it peak shifts left or right. It can also be higher or lower. It depends upon the values of the innovation and imitator coefficients. And this reflects real life.
Again, Mahajan, Muller and Bass’ “Diffusion of new products: empirical generalizations and managerial uses” gives an insight into what real-life values people have found.
It turns out that real-life is not far from Rogers’ thinking. For example, the early and late majority are 29.1-32.1% of market size compared to Rogers’ near standard deviation value of 34%. However,the size of laggards is higher than Roger’s 16% at 21.4-23.5%. Looking at innovators, they are found in practice to be 0.2-2.8% rather than 2.5%. And early adopters between 9.5-20% compared to Rogers fixed value of 13.5%.
Rather than the innovators and early majority making up 16% of market size they could be between 9.7% and 22.8%. The implications of this is on the shift from external to internal influences, where the chasm to be crossed might be found and whether Maloney’s 16% rule needs to be cheekily renamed.
What about generations of innovations?
Bass extended his model to look at several generations of innovations. We can interpret “generations” as incremental innovation. In Figure 11 we see sales and fitted curves for several different generations of computers. This comes from the working paper “Diffusion Of Technology Generations: A Model Of Adoption And Repeat Sales “.
We start with kit computers back in the 1970’s and move through to powerful setups built for the internet generation at te start of the century. Each generation being a sizeable incremental innovation from the last.
For each generation, a familiar diffusion curve appears. Enough for us to be able to make predictions on production volumes, or service size, for new innovations. Assuming we are accurate with our expected market size.
Additionally, we can see that the next incremental innovation were launching into the early adopters around the time the existing generation was drifting to the end of the late majority. This gives us an insight into timing of innovation launching. Useful to keep in mind.
Do price and other attributes affect the model?
We don’t take account of price changes, advertising effort, or so on in the standard Bass Model. These variables can be called decision variables. And in Bass et al’s paper “Why the Bass Model Fits without Decision Variables” a Generalised Bass Model is derived that takes account of decision variables.
I won’t go into this paper in this article. But, suffice to say, you can take account of decision variables in Bass’ model (the generalised version). And, if there are no decision variables to take account of, then the generalised model reduces to the standard Bass Model.
A wider discussion of the marketing mix influence can be found in “Modelling The Marketing-Mix Influence In New-Product Diffusion“.
So, I think this is a really useful model. And the set of papers lifted up in this article are truly interesting. But perhaps do not get the visibility deserved in the innovation community. For example, many can highlight Rogers’ 5 adopter types; fewer can draw the link to Bass’ model.
We have seen that under certain parameters, Bass’ model mimics Rogers adoption curve. Further, we saw that Bass’ model can be fit to a range of products sales data. This give credence to its validity.
That also gives us the confidence to use the model as a prediction tool. And we saw that Bass did just that for colour TVs in the 60’s. Much to the anger of industry experts. His predictions were not far off; the industry’s predictions led to over supply. It has also been used for predictions for “satellite television/telephone/radio, a new LCD projector, wireless telephones, and wireless Internet phones (3G)” .
To get predictions requires estimating the target market size and values for the coefficients p and q. Studies have given us average values and typical bounds on those coefficients. These can be used to check our guesses have some relevance.
Another useful empirical insight is the refinement of Rogers’ adopter type sizes from being standard deviations. We saw that innovators rather than being a standard 2.5% are somewhere between 0.2 and 2.8%. Other adopter types similarly had ranges.
As a final note, Bass paper was noted in the top 10 influential papers published in Management Science’s 50 year anniversary edition.