Research Marketing Exam 2

experiment

deliberately changes one thing (the independent variable) and measures its effect on another (the dependent variable), under controlled conditions. We run experiments because we want to know what causes what (causal research).

Independent Variable (IV)

• The variable the researcher deliberately changes; the treatment variable

• ex. price of the Hershey’s Kiss ($1 vs free)

Dependent Variable (DV)

• The outcome the researcher measures

• ex. Which chocolate each student chose

Test Units

• The people, stores, etc. whose response is measured

• ex. Students in the class

Treatment

• The specific level of the IV given to a group

• ex. Free offer: 0¢ Kiss / 14¢ Lindt

Treatment Group

• Receives the treatment

• ex. Students who saw the free offer

Control Group

• Does not receive the treatment; the benchmark

• Students who saw the paid offer: 1¢ / 15¢

extraneous variables

other factors (weather, a competitor’s sale, exam week) that can also affect the DV. A good experiment controls for them so the change in the DV can be traced to the IV.

association (covariation)

two variables move together

causality

a change in X leads to a change in Y (“X causes Y ”).

association vs. causality

• association is a necessary but insufficient condition for causality: no association ⇒ no causality, but association̸ ⇒ causality

• ex. Margarine and divorce: Maine’s divorce rate and margarine consumption move together al- most perfectly. Should Maine ban margarine? No.

• Ice cream and crime: both rise together because hot weather drives both. A third factor creates the association.

When do you look for a third factor/timing problem?

When two things move together and someone claims one causes the other

what are the 3 necessary conditions for causality

1. temporal precedence

2. assocation

3. no alternative explanations

(all three must be held)

temporal precednece

• The cause (X) happens before the effect (Y )

• Students saw the price before choosing → met

Association

• X and Y move together (covariation)

• Hershey’s share rose from 44% to 90% when free → met

no alternative explanations

• Nothing else could explain the change in Y

• “Maybe the free group just likes Hershey’s more” → not yet met; random assignment fixes it

Random Assignment

• each test unit is placed in the treatment or control group by chance (a coin flip, or Qualtrics’ randomizer)

• chocolate lovers, dieters, and hungry students land in both groups, so the groups are comparable. The treatment becomes the only systematic difference between groups, which rules out alternative explanations based on who is in each group

Internal Validity

did the treatment really cause the result? Have we ruled out alternative explanations?

External Validity

would the result hold outside the study, with other people, places, real

money and real decisions?

Lab Experiments

• Setting: Artifical, controlled

• Other factors: Few, well controlled

• Strength: High internal validity

• Cost and Time: Lower, Faster

• Participants Aware: Usually Yes

• ex. our class, a hypothetical Qualtrics survey with random assignment, no real money

Field Experiment

• Setting: Real World

• Other Factors: many, little control

• Strength: High external validity

• Cost and TIme: Higher, Slower

• Participants Aware: Usually no

• ex. cafeteria: real customers adding chocolate to a real bill

What does X mean?

exposure to the treatment

What does O mean

an observation (measurement of the DV) numbered O1, O2, …

What does Random mean

random assignment to the group

X, O, R notation

• Read left to right as time. Each row is a group. A row without X is the control/comparison group.

• (1) how many groups? one row each; (2) formed randomly? add R; (3) DV measured before? add O before X; (4) DV measured after? add O after X.

One shot case study (Pre experimental)

• notation: X O

• ex. Display up; 500 sold

• remaining weakness: no comparison; 500 compared to what?

One group pretest posttest (pre experimental)

• notation: O1 X O2

• Smoothie example: Week 1 = 400; week 2 with display = 500

• Remaining weakness: no control group; weather or eam week could explain +100

Pretest Posttest with comparison group (quasi)

• Notation: O1 X O2 —> O3 X O4

• Smoothie Example: Cafe A (gym) get display; Cafe B (library) doesn’t

• Remaining Weakness: Not randomly assigned: the cafes may differ

Posttest only with comparison group (quasi)

• Notation: X O1 —> O2

• Smoothie Exapmle: Compare cafés after only

• Remaining Weakness: Can’t tell if Café A was already selling more

Pretest Posttet with control group (true)

• Notation: R O1 X O2 —> R O3 O4

• Smoothie Exapmle: Coin flip for each of 20 cafés

• Remaining Weakness: Few internal threats

Posttest only with control group (true)

• Notation: R X O2 —> R O1

• Example: Chocolate study, macaron demo

• Remaining Weakness: Few internal threats

(___) rules out events in time

control group

(____) rules out differences between groups

random assignment

What do you need to have a true experiment?

both a control group and a random assignment

Treatment effect of posttest only with control group

TE = O2 - O1

• how much of the treatement group changed

Treatment effect of Pretest Posttest with control group

TE = (O2 - O1) - (O4 - O3)

• how much it owuld have changed anyway

Nominal Data Format

Property: label only no natural order

Comes from: categorical questions

examples: gender, brand chosesn, ZIP code, Student ID, jerssey number

You can compute: frequency/mode %

Ordinal Data Format

Property: + Order and gaps unknown

Comes From: numbers in brackets

Exapmles: never/rarely/often;letter grade; age 18-24/25-34

You can Compute: + median

Interval Data Format

Property: +Equal gaps; zero is arbitrary

Comes From: likert, N-Point, semantic differential, stapel

Examples: 1-7 ratings; temp F

You can Compute: +Mean, Range, Stand. Dev.

Ratio Data Format

Property: +true zero (None at all)

Comes From: open ended number questions

Examples: age in years, income, $ spent, # visits

You can compute: everything, including “twice as“

What are the 4 questions to classify any answer

1. can i label it? Yes —> at least nominal

2. can i rank it? Yes—> at least ordinal

3. Are the gaps equal? Yes —> at least interval

4. Is zero real? Does “twice as much“ make sense? Yes —> ratio

hypothesis

a statement about population information (the value of a population parameter, such as a mean or a proportion).

Hypothesis testing

uses sample data to decide whether the hypothesis is true.

Why do we need hypothesis testing?

• we use it to infer population information, asking how likely the sample in hand would be if the null hypothesis were true

• it’s decision making under incertainty

“Presumption of innocence: innocent unless proven guilty”

H0 presumed true until there is enough evidence to reject it

“The defendant is innocent“

Null hypothesis (H0)

“The defendant is guilty”

Alternative Hypothesis (H1)

“Evidence presented by the prosecutor”

Sample data and the test statistic

“convicted only if all jurors believe the defendant is guilty“

level of significance (a) sets how strong the evidence must be

“Number of jurors who still believe the defendant is innocent“

p-value

What are the 4 steps in Hypothesis Testing

1. state the hypotheses: one null, one alternative (always a pair)

2. set the criterion for a decision a = 5% (0.05)

3. Compute the test statistic from the sample data

4. make a decision: reject H0 and accept H1; or retain H0 and reject H1

Null Hypothesis (H0)

always comes with an “=” sign (=, ≤, or ≥); the status quo, presumed true; usually what we hope to reject; can be rejected or retained, never accepted.

Alternative Hypothesis (H1)

contradicts H0: “totally opposite” (̸=, two-tailed) or “one-way opposite” (> or <, one-tailed); the statement useful for marketing decisions; can be accepted or rejected.

Level of significance (a)

the decision criterion; always 5% in this course. 1 − α = confidence level (95%).

Test Statistic

computed from the sample data and H0 only; measures the distance between H0 and the sample data.

p-value

the likelihood of obtaining the sample data if H0 is true

Larger test statistic —> smaller p-value —> (___)

• stronger evidence to reject H0

• A smaller test statistic means a larger p-value and weaker evidence

Problems with raw survey data

1. one concept, many questions: Q5a–Q5d all measure service quality, but a regression needs one variable per concept. Fix: composite score.

2. Categories stored as numbers. “Where do you live?” coded 1/2/3 are labels, not amounts: a commuter (3) is not “3×” an on-campus student (1).

   1. Fix: dummy variables

Multi item scales

• Example (service quality): SQ1 “Staff are friendly,” SQ2 “My order is prepared correctly,” SQ3 “Service is slow” (reverse-worded), SQ4 “Staff know the menu.”

• Any single question carries random noise (a bad day, a misread word). Averaging several items cancels much of that noise, so the score is more reliable.

• In a regression, the construct (not each question) is the predictor.

reverse worded item

• is phrased in the opposite direction

• xr = (min + max) - x

• 1-5 scale: 6-x; 1-7 scale: 8-x

• on a 1-7 scale: 1 - 7, 2 - 6, 3 - 5, 4 - 4

• always create a new variable never overwrite the raw number

Cronchbach’s alpha (reliability)

• Measures internal consistency: do the items “hang together” (measure the same thing)?

• Rule of thumb: α ≥ .70 is the usual bar for combining items and reporting a scale.

• Campus Grind: service quality α = .84 with SQ3_R, but only 0.17 if SQ3 is not reversed.

• Negative item-total r ⇒ almost always a reverse-worded item you forgot to flip. Fix the coding; don’t drop the item.

• Low positive r (< .30) ⇒ the item may measure something else. Drop it only if α-if-deleted clearly rises and the content still makes sense.

Rule for Dummy Coding

• k categories → k − 1 dummies, each coded 1 = in the group, 0 = not.

• The group with all zeros is the reference group; every dummy coefficient compares its group with the reference group.

Choosing the reference group

1. a meaningful baseline (current customers, control condition, status quo);

2. a reasonably large group;

3. never “Other”

read regression output in three checks

1. is the model significant in ANOVA?

2. how much does it explain? (Adjusted R²)

3. which predictiors matter? (Coefficients table)

Beta

• What it tells you: change in Y per 1-unit change in X, in original units

• Rule: use for the equation and interpretation

Standard Beta

• what it tells you: same, in standard deviation units (unit free)

• Rule: biggest |B| = strongest driver

Significance

• what it tells you: is the predictor’s effect real?

• Rule: < 0.05 = significant

VIF

• what it tells you: are predictors too similar to each other?

• Rule: < 10 = OK

continuous predictor

Each 1-point increase in X goes with Y being B points higher (lower), holding the other predictors constant. ServQual B = 0.48.

Dummy Predictor

Group = 1 scores B points higher/lower than the reference group, holding the others constant. Commuter B = −0.74 vs. on-campus

Not Significant (p > 0.05)

do not interpret it, but keep it in the model as a control. FitFuel Female B = 0.09, p = .58.

Strongest Driver

compare Std. Beta, not B (B values are in different units).