MAR4832 - Exam 1

Purpose of Maximum Difference (MaxDiff):

Used for measuring the importance or preference within a list of items.

• Helps understand what people care about the MOST and the LEAST in a list of options)

• Best/Worst Item Scaling.

• Example: “Most Important vs. Least Important, Most Influential vs. Least Influential.”

Where Prioritization (MaxDiff) can be helpful:

Better understanding:

• Brand equity associated with competing brands.

• Which consumer needs need to be addressed ASAP.

• Relative effectiveness of different marketing messages.

  • To use in campaigns or planning.

• What product features to modify.

• Other issues in brand portfolio management.

Examples:

• Narrowing possible launch messages (marketing) to boost conversions.

• Identified which promotions resonated most with loyalty members.

Importance of using a data-based approach to prioritization (MaxDiff):

Reduces guesswork and personal bias.

Compared to direct questioning, MaxDiff tends to be more useful under what conditions?

The list has:

• More than 8 items.

• Works best with 20 to 30 items.

Why is MaxDiff, better than direct questioning?

Direct questioning can be cumbersome for consumers; MaxDiff breaks up the task and makes it easier for consumers.

Connectivity (As seen in Sawtooth)

Each item should be compared, directly or indirectly, with every other item in the study. This allows all items to be placed on a common scale.

• Who gets compared to whom.

• Example: A vs B.

• Ensures no overlap amongst questions.

Frequency Balance (As seen in Sawtooth)

Each item appears an equal number of times as every other item.

• Ensures all items are compared and appear the same number of times.

Orthogonality (As seen on Sawtooth)

Each item appears an equal number of times as every other item.

• Spreads pairings out

• Example ↓ wouldn’t work, as A & B are paired together too often.

  • Q1: A, B, C, D

  • Q2: A, B, E, F

  • Q3: A, B, G, H

Positional Balance (As seen on Sawtooth)

Where an item appears within the a question.

• No item gets an advantage from always being at the top or bottom. (Randomness is needed)

When you run a MaxDiff study on Sawtooth you get:

Item Scores (Utilities)

What are two methods for determining “Item Scores”?

1. Logit Model Method

2. Best-Worst Count Method

Under what conditions do you use the Best-Worst Count Method (Simpler)?

Use the Best-Worst Count Method when consumers are relatively similar and when customer segments are known beforehand (a priori segmentation).

Best-Worst Count Model

• Count the number of times an item is selected as “Best”.

• Count the number of times an item is selected as “Worst”.

• Calculate B-W (Substract) and rank items accordingly.

Under what conditions do you use the Logit Model Method (Simpler)?

Because it estimates choice probabilities that are constrained between 0% and 100% and sum to 100%, making it more appropriate for modeling consumer choices.

• The logit model also provides a more advanced analysis of MaxDiff choice data and works better when consumers have different preferences or when customer segments are not known beforehand.

Purpose of Conjoint Analysis (Conjoint)

Figures out what consumers value in a product and uses that to predict what they’d buy.

• Example: You break a product into features, and then show them different combinations and ask them to rate or choose.

• PREDICTIVE: Tells you what will happen if something changes, such as price or a feature.

  • You’re able to test this on Sawtooth; you can change features and see how many people switch to your product or others.

Business problems where Optimization (Conjoint) is important:

• Deciding whether to launch a high-end product, a budget product, or both.

• Identifying how to modify a product to reduce costs while maintaining share (profit).

• Set a price that is competitive but reasonable. (You do this by finding out the price at which consumers switch to other competitors.)

• Best combination of features for a new product.

Why is it important to use a data-based approach for Optimization (Conjoint)?

Because without data, decisions rely on gut feeling, which leads to:

• Over- or underpriced products.

• Bloated, confusing bundles.

• Costly trial-and-error launches.

BUT with it, Decisions are based on trade-offs:

• Clear view of share, revenue, and profit.

• Smarter bundles and pricing.

Examples of how Optimization (Conjoint), has had a positive impact on businesses:

• Retail baking tested account features to design a product that won a 30% share in pilot markets.

• CPG identified the best flavor and more… for a snack brand, boosting trial and repeat by 22%.

Why are simulations that compare “current state” and “future state” outcomes useful to managers (and how do these compare to alternatives such as concept testing or real-world test marketing)?

They turn abstract utility scores into something managers can use: predicted market share.

• Example: A manager can compare today's market (current state) to a "what-if" version (future state) by changing a price or feature and seeing the share change right away.

• Similar gives you dozens of combinations and is cheaper.

Compares to concept testing or real-world testing since:

• You can only typically test 2-3 ideas.

• Slow and expensive.

• No competitors in the test.

Conjoint Analysis vs. MaxDiff

MaxDiff scales items on a single dimension, such as brand, flavor, or ice cream. (The focus is on contrasting items.)

vs.

Conjoint Analysis, which handles multiple attributes at once. (The focus is on consumer preference for combination of attributes.)

Weakness of asking consumers directly about attribute importance:

• There are no trade-offs present.

• They tend to say everything is important.

Conditions under which conjoint analysis may not be appropriate:

• Key attributes are missing from the study

• Too many attributes or levels

• Buyers use screening rules instead of making trade-offs

• The market doesn't match its assumptions (equal availability of products at all stores, full awareness - everyone knows about every possible product, no products sold out - always in stock)

What are the three outputs from Conjoint Analysis?

• Utilities.

• Relative importance of attributes

• Simulations.

Utilities (Par Worths)

A score for each level of each attribute shows how desirable it is.

• Describes individual levels.

Importance Scores

Measure how much influence each attribute has on someone else’s choice.

• Formula: (Highest utility - lowest utility) / (Sum of all ranges)

• How much a whole feature matters (Example: Consumers value price more than flavor.)

Market Simulations

Competitive market scenarios to predict which products respondents would choose.

How are utilities useful in running simulations?

Using utilities, you’re able to add up all utilities of the product features (E.g., Flavor) to get its total score. Then you’re able to compare scores.

THE HIGHER THE UTILITY SCORE THE MORE APPEALING.

True

How are part-worth utilities different from the relative importance of attributes?

Part-worth utilities score each option within a feature. Relative importance scores the whole feature.

Three ways to conduct Conjoint Analysis simulations

1. First-Choice: The product with the highest score gets 100% of the choice.

2. Linear: Each product gets a share equal to its score divided by the total of all scores.

3. Logit: Like linear, but scores are first put through e^score so higher scores get a bigger boost.

How are ‘Linear’ and ‘Logit’-based simulation approaches different?

In LINEAR, shares are proportional to the item scores.

vs.

In LOGIT, exponentiating makes the product bigger, so the leader gets a larger share than in linear.

Choose the method whose predicted shares come closest to what really happens in the market.

Why Share or Preference (SOP) rather than Market Share when reviewing simulation results?

Because the simulation only measures interest not real sales.

Sales of Preference (SOP)

Percent of respondents who would prefer each product in the survey’s PRETEND market.

If SOP were to equal Market Shares, we’d be assuming…

• All features that affect real choices were included in the study.

• Every product is equally available.

• Everyone is aware of every product.

• Products have reached long-term equilibrium.

• No product is ever out of stock.

Dummy Variable

A variable that equals 1 if a level is present and 0 if it’s not.

How many dummy variables do we need to represent the levels of an attribute?

The number of levels MINUS 1.

• Example:

  • 2 Levels: 1 Dummy.

  • 3 Levels: 2 Dummies.

  • 4 Levels: 3 Dummies.

• THE LEVEL LEFT OUT IS THE REFERENCE LEVEL.

Full-Factorial vs. Partial Factorial- Conjoint Study Designs

In a full factorial, every possible combination of levels is shown.

vs.

In a partial factorial, only a carefully selected subset of levels is shown. (Less work for respondents)

Design Matrix

A table where each row is a product combination and each column is a dummy variable.

• 1 = Level is present.

• 0 = Not present.

• IN CONJOINT ANALYSIS: It is the X in the regression.

  • Tells the regression exactly which levels (features) each rating was based on.

What are the steps to conduct traditional CA using Excel for a sample of respondents?

Steps in Excel:

1. Build the Design Matrix using Dummy variables for the attributes and levels.

2. Collect the ratings. (0 to 10)

3. Put all the respondents’ ratings into one worksheet.

4. Name the data array and use INDEX() so you can pull out one respondent’s ratings at a time.

5. Run LINEST with the respondent’s ratings as Y and the design matrix as X; you have to include the constant. Press Ctrl + Shift - Enter to get the coefficients.

6. Use a data table. Data, then What if analysis, Data Table. Use the consumer ID as the column input cell.

How should one use the resulting Excel output— in particular, how does one interpret the regression coefficient of a dummy variable, and what is the role for a constant term in the regression equation?

The dummy’s coefficient shows how much more or less that level is worth compared to the reference level. A positive value means it’s preferred over the reference level, and a negative means it’s less preferred.

What is the role of the constant in CA in Excel?

It’s the total value of all three reference levels combined (ITS THE BASELINE - Since a starting point is needed).

Given the CA Excel output, how is the total value of a specific bundle of attributes (with specified levels) determined?

Total value = constant + the coefficients of the levels in the bundle.

How is the relative importance of different attributes obtained?

1. For each attribute, find the range (highest - lowest)

2. Add all ranges to get the total range.

3. (Individual range / Total range) x 100

Why is it important to consider competitors’ products in conducting simulations?

Real customers choose among available options, not one product in isolation. A product's share depends on how good it is relative to the others. Including competitors lets you see who you take share from, how competitors' changes affect you, and whether your own products cannibalize each other.

How do we generate coefficients analogous to the commercial software from the Excel output for traditional CA?

1. Give the reference level a missing value called K.

2. Write every other level as K plus its coefficient.

3. Add all the levels together and set the sum equal to 0.

4. Solve for K, then plug K back in to get each level’s new value.

Priori vs. Endogenous Segmentation

Priori: You decide which groups in advance using known characteristics, such as age, gender, or member vs. prospect. Then compare their preferences.

Endogenous: You let the conjoint coefficients create the groups. Respondents with similar utilities end up together, so the data decide the segments.

Way to use Priori and Endogenous Segmentation together

You find the endogenous segments first, then check which a priori traits (age, gender) describe each one. That tells you who is in each segment and how to reach them.

What are the typical steps for conducting endogenous segmentation with the conjoint coefficients obtained from multiple respondents?

1. Standardize each coefficient or importance score so no variable dominates (average, standard deviation, then standardize)

2. Pick a number or segments and give each a starting “center” (INDEX will pull the data)

3. Measure the distance of each respondent to every center (SUMXMY2)

4. Find each respondent's closest center (MIN) and assign the segment (MATCH).

5. Add up all the minimum distances to get a total, T.

6. Run Solver (Evolutionary method) to move the centers so T is as small as possible.

7. Repeat for however many segments are chosen.

8. Graph the minimum T against the number of segments.

A typical product may have many features (e.g., think of a product like a cellphone); are all those features explicitly used in a typical conjoint study?

No. Only the key attributes and levels are included.

• 4-6 attributes.

• 2-4 levels each.

Which attributes and levels are useful?

• Ones that actually drive consumers' choices.

• Ones the company can change (and that competitors differ on).

• Realistic levels, including price.

Role of exploratory research

It helps you discover which features and levels matter to consumers, so you pick the right ones before designing the study.

Funnel Approach

Start with broad topics (how they use and choose products in the category), then narrow to specific features, then to specific products or concepts.

How are the endogenous segments selected?

Try 1, 2, 3, 4, and 5 segments, then graph how well each fits. Pick the number where the curve bends (the "kink"), or just below it.