ANSWERED on Tue 30 Mar 2010 - 11:25 am UTC by leader

**Actions: **
Add Comment

I'm going to a job interview at a market research firm. They have people their who make hierarchical Bayesian models. I'd like to know, in short and simple terms, what they are useful for. I don't need to know the details. I certainly don't need to understand the maths.

In short I'm just looking for a paragraph or two on hierarchical Bayesian models and their use for market research. Specifically:

- what they're useful for

- why they're better than other techniques

- any flaws or limitation

I have a layman's working understanding of a few basic concepts (sampling, correlation, statistical significance) but that's about it.

A few sentences I can understand is fine, but links to 10 articles I can't understand is not helpful! Please note that I've already tried Googling this, and found plenty of articles but not been able to understand them. Terms like 'priors', 'Monte Carlo' etc are beyond me.

You're welcome to either answer by writing a helpful paragraph or two, or by finding an article that I can understand!

Finally, I need this by 10am GMT Wednesday (31st) morning.

Many thanks,

Daniel Kirk

HIERARCHICAL BAYES MODEL

"Hierarchical Bayes models free researchers from computational constraints that allows researchers and practitioners to develop more realistic models of buyer behavior and decision making. Moreover, this freedom enables exploration of marketing problems that have proven elusive over the years, such as models for advertising ROI, sales force effectiveness, and similarly complex problems that often involve simultaneity." (Reference 1)

Whenever there is a lack of data or different sets of data from multiple Reference s, scientist uses to predict future event with more confidence. Actually, these researchers take whatever data is available to them and carry out experiments to strengthen the results from original set of data. As these experiments continue to, evidence becomes stronger. As evidence becomes stronger, the lack of data effectively transform into a reliable data set. Loaded with new information, researchers predict future probabilities.

Here is the process:

1. Researcher have small sets of data from variety of different sources

2. Researchers conduct experiments to uncover more data from the original small set of data

3. Overtimes, data becomes large and is supported by evidence.

4. Based on new data set, researchers make future probabilities.

WHAT THEY’RE USEFUL FOR?

What is the Problem and How Bayesian Method helps solve the problem?

"Bayesian method is a useful tool for modeling multi-faceted, non-linear phenomena such as those encountered in marketing.

The problem with marketing data is that it is characterized by many "units" of analysis (e.g., many respondents, households or customers), each with just a few observations. The lack of data at the individual-level, coupled with the desire to account for individual differences and not treat all respondents alike, results in severe challenges to the analysis of marketing data. There often is not enough data to infer about a specific respondent's preferences, or their sensitivity to variables like prices, without making some educated guesses (e.g., people would rather spend less than more, holding all else constant) or building up a model that bridges the analysis of respondents to each other." (Reference 1)

WHY THEY’RE BETTER THAN OTHER TECHNIQUES?

"In application after application where respondents provide multiple-observation data, HIERARCHICAL BAYES MODEL estimation seems at least to match and usually to beat traditional models. Conjoint analysis is a prime example of an application that benefits from HIERARCHICAL BAYES MODEL estimation.

HIERARCHICAL BAYES MODEL permits estimation of individual-level models, which lets marketers more accurately target/model individuals. More specifically, HIERARCHICAL BAYES MODEL permits estimation of models too demanding for traditional methods: even when estimating more beta coefficients per individual than there are individual observations.

Aggregate estimation models confound heterogeneity and noise. By modeling individuals rather than the average, HIERARCHICAL BAYES MODEL can separate signal (heterogeneity) from noise. This leads to more stable, accurate models whether viewed in terms of individual- or aggregate-level performance.

The draws (replicates) for each respondent provide a rich Reference of information for more accurately conducting statistical tests and, for example, estimating nonlinear functions of parameters such as shares of preference." (Reference 2)

FLAWS OR LIMITATIONS?

There are very limited programs that can do complex HB calculations.

Most software will need to be customized for each new model.

Complex codes are required to program HB models.

Even after large calculations, variance in parameters may still be large.

-------------------------------------

Reference 1:

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=655541

Reference 2:

http://www.sawtoothsoftware.com/download/techpap/hbwhy.pdf

Thank You Dan:

I am happy that the answer was useful. Yes, I must acknowledge that Roger's answer was better than mine and I should have provided a more simple answer. Nevertheless, your constructive feedback will go a long way in honing my skills and providing a better answer, in future...Thanks again.

**Actions: **
Add Comment

Frequently Asked Questions | Terms & Conditions | Disclaimer | Privacy Policy | Contact Us | Spread the word!

Thu 5 May 2016 - 6:39 am UTC - © 2016 Uclue Ltd

Roger Browne

Researcher

30 Mar 2010 12:18 UTCTue 30 Mar 2010 - 12:18 pm UTC

Here's how I understand it, but I'm not a statistician so this may be a load of rubbish. Please treat it as nothing more than an extra freebie.

Statistics is often used to establish new facts. For example, suppose you have created an advertisement for a new product. You show the advertisement to 10,000 people, and you don't show it to another 10,000 people (the control group). Then you monitor their spending patterns, and you might be able to come up with a conclusion such as the following: "With 99% confidence, we can state that people who saw the ad are more likely to buy the product".

That's all very good if you're trying to discover new facts (e.g. "With 99% confidence, we can say that some new drug cures green warts"), but facts that are expressed in this way they are not particularly useful to market researchers.

Instead, market researchers want to gain insights like the following: "Given that 50% of teenagers and 18% of adults already buy this product, and given that 67% of this product's sales are currently to females, and given that a trial run in Arizona showed a 24% increase in sales following an advertising campaign with a reach of 83%, what increase in sales would we expect from a nationwide advertising campaign on shows that reach mostly male adults?"

Bayesian models provide a useful set of tools that can point towards an answer this kind of question.

Here's another example of the difference between the statistics you probably learned in high school, and Bayesian statistics. Suppose you have established that some test for illegal drugs in sports is 95% reliable. In other words, 95% of the time it gives the correct answer, and 5% of the time it gives the wrong answer. That's a useful figure for some purposes, but the drugs tester needs a different figure.

Suppose an athlete fails their drug test. You might be tempted to say that there is a 95% chance that the athlete really was taking the banned drug, but that's wrong. Bayesian analysis leads us to a useful answer here.

Suppose we also know or can establish that (overall) around 1% of competing athletes are taking the banned drug. Then, for a given athlete who failed the test, we know that the likelihood that the athlete took the drug and got caught is 95% times 1% (i.e. 0.95%), but the likelihood that the athlete didn't take the drug and is the victim of a false positive is 5% times 99% (i.e. 4.95%).

It is the Bayesian statistical analysis that lets us combine different pieces of knowledge to derive useful probabalistic information about an individual.