Working In Uncertainty

Learning more from experience


How much difference does it make?

We all learn from our experiences, but with systematic application of simple methods we can learn even more. Consider the case of David Ogilvy, the advertising genius who founded the Ogilvy & Mather advertising agency in 1949 and by the 1980s had turned it into the 4th largest advertising agency in the world almost entirely by organic growth. In one period of seven years Ogilvy & Mather won every new account for which it competed. One day someone from IBM arrived unexpectedly and simply gave them the IBM account; he knew their work.

David Ogilvy had many gifts, but two things in particular distinguished his advertising. First, he could write interesting and charming copy. Second, he was totally focused on the effectiveness of his advertising (rather than its creative qualities) and used experimentation, not just to guide individual campaigns as others did, but also to build up a body of empirically supported rules of thumb about what works and what doesn't in advertisements of different kinds in different media. For example, he found that, in print advertising, pictures of human faces at larger than life size tended to repel people.

His favourite form of advertising was direct response, where consumers respond directly to advertising. It allowed him to conduct experiments with different ideas, such as headlines, and get easily measurable results.

In his book ‘Ogilvy on advertising’ he relates how the great Stanley Resor, who by then had been head of J Walter Thompson for 45 years, told him he had begun looking at research to find factors which usually work. In two years they had already found a dozen. Ogilvy said he was ‘too polite’ to mention that he already had 96.

Ogilvy exploited the opportunities for experimentation inherent in his business. If you are not in advertising your opportunities will be different, but there will probably be more than you are currently using.

Learning more from experience

As we come to a more realistic understanding of what we understand, can predict, and can control it becomes clear that we will often benefit from learning more about how our organisation and its environment really work, and what we can do that produces results we value.

From the organisation's perspective this learning is essential to improving performance.

For an individual within an organisation, being able to draw convincing lessons from experience as well as by reasoning is useful as a means of finding good ideas and winning support for them. If you can back up your claims with results there is less chance of your discovery being overlooked.

In some industries (e.g. design of chemical manufacturing plants) it is possible to carry out proper scientific experiments, with control conditions, or systematically vary independent variables and record the effects on dependent variables. It is even possible to build a detailed model of the connections between variables and find optimum settings for independent variables. A lot of work has been done on how to design and interpret these experiments efficiently to screen out unimportant variables and build a model.

Once such a process is in live operation it is still possible to use rigorous quantitive methods to adjust the optimum settings to meet slowly changing conditions. EVOP (Evolutionary Operation of Processes), for example, involves varying the independent variables by very small increments, so that the output of the live process is still acceptable but the slight changes in output can guide future trials towards new optimum levels.

Unfortunately, most business situations don't allow these methods to be effective. An advertising agency, hotel, car dealership, or travel agent, for example, is very different from a chemical plant. It can be difficult to vary independent variables in a controlled way. The volume of data available is often small, with many potentially confounding variables. Conditions change in many ways and often quickly. Finally, we are almost always in a situation of running a live process and it is difficult to defend not applying what appears to be the best approach wherever possible.

Despite all these difficulties we have to do something. Most often we ‘experiment’ by simply doing something to see what happens. There's no control group or other comparison. This falls far short of standards for scientific experimentation but it's all we have. What can stop this being a futile exercise in self-deception is that we can use our knowledge of the world and the conditions surrounding our ‘experiment’ do two things:

  • Make allowances: In the Design of Experiments terminology this is a covariate design, where you record the values of potentially confounding variables and try to adjust for their effects. We can bring lifetimes of experience of bear on this task though quantification is often difficult. (We make allowances most often when the results of a trial surprise us but we should do it every time.)

  • Observe steps of causality: If you think that A causes D, because A causes B, which causes C, which then causes D you can sometimes observe changes in B and C as well as the final result, D. This is particularly important if D is a delayed response, such as future purchases by customers. A change at B or C ‘stores’ the effects for later and it is usually easier and quicker to observe the intermediate effect than wait for the final result.

For example, suppose you think that offering slightly different payment terms will improve sales. You try it on the next sales lead but don't get the sale. However, the buyer tells you that the payment terms are more attractive than the usual terms but explains that a corporate decision has been taken to purchase only from another ‘strategic’ supplier. Overall, this is slightly encouraging for your idea. The confounding factor is the corporate decision and you make an allowance for it. The intermediate effect is that the buyer finds the payment terms more attractive, even though this did not lead to the ultimate result of a sale.

If we persist in trying to quantify effects it becomes possible to make quantitive allowances for more and more factors that could not be controlled in our experiments. The benefits of this approach increase over time.

The power of experimentation can be increased by the following:

  • Try more things: There is a tremendous advantage in simply trying more things. Don't just plough on with exhaustive trials of one idea. Screen lots. Self-made millionaires are more often energetic than smart. They try a lot of different things and so are more likely to find a hit.

    Typically, an idea will be tried on a small scale first. If the signs are that it is worth a further look it may be tried on a wider scale, and so on in stages. This way the results may give increasing certainty.

  • Try ideas in favourable conditions first: Whether your idea is a theory about a causal factor or an idea for a better way of doing things the chances are that it won't apply equally well to all the situations where it might apply. At an early stage you will learn more if you try your idea in conditions where its impact is likely to be easy to see.

    For example, if you have thought of a new way to do something and think it is a better way, at least in some situations, think about what might characterise those situations. If possible, try your new method first in situations that suit it. Use what you learn from this to repeatedly revise your ideas about what works and what conditions affect its usefulness. Gradually expand the range of situations where you have tried it out.

    Do this consciously or you run the risk of concluding that your idea has universal application when it does not. If you stratify the population on the basis of suitability and you are clear about your criteria you can select sample items from the most suitable sub-group and use statistical inference to generalise about the effectiveness of your idea in this group.

    The approach of experimenting in the most favourable conditions first means that it is easier to see the effects you are interested in and you spend most of your time doing things you believe to be the best approach.

  • Go for data volume if you can: The more trials you can do the better. If there's a way to cheaply and quickly do lots then do so and plot graphs to help you see directly what is going on. It helps you separate the effects you are interested in from all others. Design of experiments literature tends to concentrate on what are called ‘two level’ designs i.e. each factor is tested at just two levels. If you can vary factors over more levels cheaply and do a lot of trials it is possible to plot graphs that give a clear message without complicated statistics. Unfortunately, high volume isn't often an option.

  • Observe and record more conditions and make allowances: Think as widely as possible to identify conditions that might have been important but perhaps you had not noticed. These are often conditions that are normally true in your business and which tend to be assumed. Perhaps in future that assumption won't hold. Keeping a diary can be useful.

  • Observe more steps of causality: Don't forget to check for steps in the causal chain that might lead to effects you don't want, as well as effects you do want. Many apparently sensible strategies have delayed adverse effects.

  • Design experiments with comparisons: You can get a stronger indication of effects by varying factors systematically and over a wider range (within sensible limits). There are many alternative factorial designs, not all requiring every combination of levels of factors. You can also reduce the confounding effect of other variables by 'blocked' designs. This is where you separate your total population into relatively homogeneous sub-groups and then split each group between the experimental conditions. In particular, as business conditions tend to change a lot over time it is very helpful to run the comparisons in parallel rather than one after another. Even when there is pressure to apply what is thought to be the best strategy wherever possible there is still room for small variations and these may be enough to guide optimisation (following the principles of EVOP).

  • Start with what you are most certain of: Most outcomes are driven by many factors. If you need to eliminate the effects of factors you can't control it makes sense to start with the factors whose effects you are most certain of. This in turn suggests that it is good to use early experiments to try to understand factors whose effect is pretty obvious and easy to model, even if it is not very interesting. Once you have eliminated the effects you understand, take a look at what is left.

  • Quantify effects: It is more useful to quantify effects than just say ‘more’ or ‘less’ because you can make better allowances for factors you understand and reveal more clearly what is left. Mathematical and spreadsheet models are useful.

  • Start with very simple models: Model building can be time consuming and tiring, so start with embarrassingly simply models and use them to see what parts of the model produce most of the uncertainty in results. Iteratively refine your models, concentrating on the areas that are responsible for most of the uncertainty.

  • Use more than one model: Often it makes sense to have more than one model in use at the same time. First, it may be that there are alternative models for the same purpose. (This is discussed further below.) Second, it is often helpful to develop models to support specific decisions (one offs or regular decisions). For example, the models used to plan stock levels may be different from those used to plan staff levels. There may be no link or consistency between them and this is not necessarily a bad thing.

  • Keep the detail: Gather experience in the smallest units possible and combine your data with averages only as a last resort. For example, if you want to try that idea for more attractive payment terms you could divide your customers into two groups with the same level of sales in the previous year. One group is offered the new terms and the other is offered the usual terms. After a period of time you compare the total sales in each group. This gives you two numbers to compare, but it is hard to make allowances for potentially confounding factors. Alternatively, you could try to make allowances at the level of each customer account, or even at the level of each potential sale. This will give you more information.

  • Evolve your experimental designs: Many experiments give weak or unusable conclusions when first attempted and nearly all experiments could be better designed with hindsight. Be prepared to change your experimental approach and try again, and again, and again.

Learning from financial and management accounts

At the aggregated level of company financials it is usually difficult to see the effects of trying different approaches until they are widely used. Even then the gradual roll out and combination of many initiatives and other trends makes direct analysis of the summarised financials very difficult.

Nevertheless, we can and should try to learn from management and financial accounts. Here are some ideas:

  • Compare actuals with detailed forecasts: What are the differences between summarised actuals and the various forecasts made on the basis of analysing much more detailed data? Why have they arisen? Are individual predictions poor? Are there factors that are not being understood and allowed for at all? Have familiar causal links changed so that previously reliable predictions are now failing?

    This level of comparison provides assurance that lower level forecasting and learning is working, and a warning if it is not.

  • Model crudely at a high level: Treat the financials as a set of time series. Get as many past periods as possible into view and use graphs to show how things have changed over time. Look for the trends and quantify the variability between months. Even this simple view may give more realistic expectations than complicated and detailed analyses using more detail if the detailed modelling is not going well or hasn't started. The crude model is useful on its own, but can also be used to challenge more detailed models. If someone has been experimenting at a detailed level and now predicts future results that would be very different from past results without suggesting a radically different plan of action or pointing to a powerful external trend then you should doubt the prediction.

Some pitfalls

Some things that can go wrong are:

  • repeating an experiment until you get the results you want, either through chance or by unwittingly introducing factors that bias results in the direction you want/expect;

  • not noticing important confounding factors;

  • determinedly pursuing an idea that's not a good one and distorting contrary evidence by wrongly concluding that confounding factors are responsible for contrary results;

  • not anticipating indirect adverse effects, and not wanting to know about them; and

  • assuming an idea with immediate face value is a good one and not testing.

Don't be paralysed by analysis

Paralysis by analysis is another pitfall.

It's one thing to recognise the value of more information but quite another to be unable to act at the right time because you aren't sure what to do. So what can you do when the hard evidence is inconclusive? Here are four possibilities, each more sophisticated and rational than the last:

  1. Choose a ‘null hypothesis’ until you're confident it is wrong: For example, you might decide to assume that schemes for motivating employees have no effect on motivation unless there is strong evidence to the contrary. Or you might assume they always increase motivation until proven otherwise. Your null hypothesis can be anything you like; there is no magic to it. If your null hypothesis happens to be the wrong place to start from this can mean you are making bad decisions and ignoring helpful evidence for longer than you really need to.

  2. Choose whatever the data point to: Suppose you've tried a scheme to improve motivation and motivation increased by a certain amount. You could assume that this is the true effect and if you use the scheme again you will get the same result. This is very responsive to the data, but can lead to extreme results if the evidence is weak. If your data base is small you are more likely to get values that are away from the true value.

  3. Go for the most likely hypothesis taking into consideration data and prior expectations: In this strategy you need to recognise that you had some expectations before you even tried the scheme, and try to clarify what they are. The evidence of an experiment is then combined with this prior view to produce a revised view. You can do this by judgement or computation.

    Doing this involves setting out all the potentially true hypotheses and attaching a probability to each (a probability density in the continuous case) that it is the true one. Once you combine evidence from the experiment the result is a revised set of probabilities. You then choose the hypothesis that is most likely to be true on your revised view. However, it could be that the most likely hypothesis is barely more likely than others.

  4. Combine all your hypotheses when making forecasts and decisions: This is the same as the previous approach except that when making forecasts and decisions you do not take the most likely hypothesis. Instead, you average the predictions/decision values of all the hypotheses, weighting each by its probability of being true. This explicitly shows the uncertainty you have about what model to use and tends to produce more widely spread distributions for future predictions. Other modelling approaches described above tend to understate uncertainty.

    This approach is called ‘Bayesian model averaging’ but it doesn't have to be complicated to do.

This list of possible approaches is not complete. For example, sometimes it is possible to take a decision without needing to know much about prior beliefs. It may be that one strategy is very attractive across a wide range of hypotheses so it can be chosen without having a clear idea of how likely each hypothesis is.

The approaches that take into consideration prior beliefs as well as new evidence are usually more suited to business situations because so often the data available are far from conclusive. I personally find it helpful to think about my prior beliefs, especially when the evidence of experience is weak, as it so often is.


Science has made a huge impact on the human race but there are times in business when it seems to have no relevance. This is because the circumstances in which we work often do not suit the experimental designs most of us learned at school or university. But if we adapt the principles to our circumstances we can learn more from experience and build more powerful cases for good business ideas.

Readable, informative, and charming. I love this book and highly recommend it.

Made in England


Words © 2004 Matthew Leitch. First published 19 February 2004 (updated 25 February 2004)