Customer marketing and analytics SUMMARY
Knowledge clip 1
Before you can actually do data analysis you need to do data preparation to ensure
your data is good enough for analysis.
Taste evaluations of beer brands: Beer drinkers (unit of observation) are unable to
discriminate brand of beers with only taste and aroma (unbranded condition). But
when they know the brand, the brands they liked were rated higher (branded
condition).
Colums (horizontal): variables: e.g. age, gender, region
Observations (vertical): rows: respondent number
Numeric coding for non-metric variables: e.g. Amsterdam (1), Rotterdam (2),
Utrecht (3) or gender: female (1), male (2). Although numerically coded, the numbers
have no meaning.
Screening the dataset:
- Check for duplicates and missing data
- Find “strange” codes and errors (visually)
Consistency checks (out-of-range, extreme values, reverse-coding)
E.g. Check the ages (200 is not possible
Reverse coding: 5 punts scale in one question ranging from + to - and in
other question from - to +
Logical issues: when the answers of a respondent contradict each other: you
notice a customer that says they do not have a credit card but later on mention
they use it to buy stuff.
Respondent fatigue: people sometimes get tired and keep answering the
same way as their previous answers.
Missingness
- Survey respondents miss out (accidental or deliberate, people don’t answer the
questions)
- Responses other than original scale (make “don’t know” or “not applicable”
missing in your data set since the number they are assigned has no value)
- Survey flow logic: if some of the questions are not asked to all respondents, you
have to make them missing
- Soms is missing data niet random, denk bijvoorbeeld aan mensen die in een
lagere inkomensklasse zitten die minder snel geneigd zijn om transparant over
hun jaarlijkse inkomen te antwoorden, of überhaupt niet antwoorden.
What to do with bad data
- Go back, collect more data
- Assign missing values (mean/mode substitution: BEWARE: reduces variability)
- Delete bad responses/ missing value casewise vs. pairwise
Casewise: if there is one missing data for any variable you take the whole
observation out of the dataset, hierdoor verlies je ook informatie van de scores op
andere variabelen.
,Pairwise: remove observation when if there is missing data related to the analysis
only remove the observation if the response is necessary for the analysis. BEWARE:
this might lead to different samples when conducting different analysis, because you
keep taking other observations/ repondents out of your data.
Knowledge clip 2
Univariate descriptive analysis: exploring the data
Summarize data (numerically and visually)
4 Levels of measurement:
1. Nominal/ categorical (NON-METRIC)
- Multiple choice: e.g. brand chosen, education etc.
- Dichotomy: gender, yes vs. no, user vs. no user
2. Ordinal: rank order favorite brand on Likertscale (NON-METRIC)
The numbers represent relative positions. There is a fixed order.
3. Interval: attitudes, liking, (METRIC): the numbers are ordered: The numbers
are not equally distant, they have known equal intervals: the intervals can vary
but they have meaning, the distance between the different values are not set
apart by a specific boundary, think of Celsius: it can be 2 degrees but also 2,4
or 27,2 degrees. Also: the zero degree does not mean no temperature. There
is no true zero.
4. Ratio (METRIC): age, income, market share, sales (fixed origin). The numbers
are equally distant and there is a true zero. Zero has a meaning.
Ratio gives us the most information, more precision. Nominal the least.
,Univariate descriptive analysis
Measures of:
- Central tendency: mean, median, mode
- Variability/dispersion: (interquartile) range, variance, std. deviation
- Distribution: skewness, kurtosis
Mean: sum of all readings/ total number
Median: middle reading when data is sorted in size order
Mode: most frequent reading
Mean -> you need metric data (interval and ratio)
Median -> you need at least ordinal level, can be: ordinal, interval and ratio
Mode -> you can use all levels of data
Measures of central tendency:
Deviation= actual reading - sample average: how much this person deviates
Variance = The sum of the squared2 deviation divided by the number of observations
Standard deviation = the root of the variance
Range = largest – smallest reading
Interquartile range IGR= upper quartile – lower quartile
Why don’t we only rely on measures of central tendency?
Because it gives us limited information.
Skewness: indicator of distribution symmetry
normal bell-shaped distribution
In a perfectly symmetrical the dataset: the mean = mode = median.
, right skewed (positive): longer tail is to the right but highest point is at the left.
When you find a positive skewness (+) your distribution is right skewed.
Mean > median > mode
left skewed (negative): longer tail is to the left but highest point is at the right.
When you find a negative skewness (-) your distribution is left skewed.
Mean < median < mode
Kortosis: indicator of flatness or peakness of the distribution
Neutral kurtosis (0) if kurtosis is zero: we have a normal distribution.
Positive kurtosis (+) indicates heavier tails and a more peaked distribution. There are
probably also many cases with extreme scores.
Negative kurtosis (-) indicates suggests lighter tails and a flatter distribution
With larger samples, skewness and kurtosis is not a big problem.
Knowledge clip 3.1
Inferential analysis: tried to generalize the sample results to a population.
Parameter: number describing the population E.g. average brand perception.
We need a random sample to make sure our sample parameter is representative of
the population.
Sampling error (standard error): standard deviation of sample means.
Central Limit Theorem CLT: the more your sample size increases, the distribution of
means of a large number of random samples from any population will lead to a
normal distribution. SO: Randomly sampling large enough from a population and
taking means of each sample will lead to a normal distribution. The sample
distribution of sample means. A rule of thumb = more than 30 observations.
Knowledge clip 1
Before you can actually do data analysis you need to do data preparation to ensure
your data is good enough for analysis.
Taste evaluations of beer brands: Beer drinkers (unit of observation) are unable to
discriminate brand of beers with only taste and aroma (unbranded condition). But
when they know the brand, the brands they liked were rated higher (branded
condition).
Colums (horizontal): variables: e.g. age, gender, region
Observations (vertical): rows: respondent number
Numeric coding for non-metric variables: e.g. Amsterdam (1), Rotterdam (2),
Utrecht (3) or gender: female (1), male (2). Although numerically coded, the numbers
have no meaning.
Screening the dataset:
- Check for duplicates and missing data
- Find “strange” codes and errors (visually)
Consistency checks (out-of-range, extreme values, reverse-coding)
E.g. Check the ages (200 is not possible
Reverse coding: 5 punts scale in one question ranging from + to - and in
other question from - to +
Logical issues: when the answers of a respondent contradict each other: you
notice a customer that says they do not have a credit card but later on mention
they use it to buy stuff.
Respondent fatigue: people sometimes get tired and keep answering the
same way as their previous answers.
Missingness
- Survey respondents miss out (accidental or deliberate, people don’t answer the
questions)
- Responses other than original scale (make “don’t know” or “not applicable”
missing in your data set since the number they are assigned has no value)
- Survey flow logic: if some of the questions are not asked to all respondents, you
have to make them missing
- Soms is missing data niet random, denk bijvoorbeeld aan mensen die in een
lagere inkomensklasse zitten die minder snel geneigd zijn om transparant over
hun jaarlijkse inkomen te antwoorden, of überhaupt niet antwoorden.
What to do with bad data
- Go back, collect more data
- Assign missing values (mean/mode substitution: BEWARE: reduces variability)
- Delete bad responses/ missing value casewise vs. pairwise
Casewise: if there is one missing data for any variable you take the whole
observation out of the dataset, hierdoor verlies je ook informatie van de scores op
andere variabelen.
,Pairwise: remove observation when if there is missing data related to the analysis
only remove the observation if the response is necessary for the analysis. BEWARE:
this might lead to different samples when conducting different analysis, because you
keep taking other observations/ repondents out of your data.
Knowledge clip 2
Univariate descriptive analysis: exploring the data
Summarize data (numerically and visually)
4 Levels of measurement:
1. Nominal/ categorical (NON-METRIC)
- Multiple choice: e.g. brand chosen, education etc.
- Dichotomy: gender, yes vs. no, user vs. no user
2. Ordinal: rank order favorite brand on Likertscale (NON-METRIC)
The numbers represent relative positions. There is a fixed order.
3. Interval: attitudes, liking, (METRIC): the numbers are ordered: The numbers
are not equally distant, they have known equal intervals: the intervals can vary
but they have meaning, the distance between the different values are not set
apart by a specific boundary, think of Celsius: it can be 2 degrees but also 2,4
or 27,2 degrees. Also: the zero degree does not mean no temperature. There
is no true zero.
4. Ratio (METRIC): age, income, market share, sales (fixed origin). The numbers
are equally distant and there is a true zero. Zero has a meaning.
Ratio gives us the most information, more precision. Nominal the least.
,Univariate descriptive analysis
Measures of:
- Central tendency: mean, median, mode
- Variability/dispersion: (interquartile) range, variance, std. deviation
- Distribution: skewness, kurtosis
Mean: sum of all readings/ total number
Median: middle reading when data is sorted in size order
Mode: most frequent reading
Mean -> you need metric data (interval and ratio)
Median -> you need at least ordinal level, can be: ordinal, interval and ratio
Mode -> you can use all levels of data
Measures of central tendency:
Deviation= actual reading - sample average: how much this person deviates
Variance = The sum of the squared2 deviation divided by the number of observations
Standard deviation = the root of the variance
Range = largest – smallest reading
Interquartile range IGR= upper quartile – lower quartile
Why don’t we only rely on measures of central tendency?
Because it gives us limited information.
Skewness: indicator of distribution symmetry
normal bell-shaped distribution
In a perfectly symmetrical the dataset: the mean = mode = median.
, right skewed (positive): longer tail is to the right but highest point is at the left.
When you find a positive skewness (+) your distribution is right skewed.
Mean > median > mode
left skewed (negative): longer tail is to the left but highest point is at the right.
When you find a negative skewness (-) your distribution is left skewed.
Mean < median < mode
Kortosis: indicator of flatness or peakness of the distribution
Neutral kurtosis (0) if kurtosis is zero: we have a normal distribution.
Positive kurtosis (+) indicates heavier tails and a more peaked distribution. There are
probably also many cases with extreme scores.
Negative kurtosis (-) indicates suggests lighter tails and a flatter distribution
With larger samples, skewness and kurtosis is not a big problem.
Knowledge clip 3.1
Inferential analysis: tried to generalize the sample results to a population.
Parameter: number describing the population E.g. average brand perception.
We need a random sample to make sure our sample parameter is representative of
the population.
Sampling error (standard error): standard deviation of sample means.
Central Limit Theorem CLT: the more your sample size increases, the distribution of
means of a large number of random samples from any population will lead to a
normal distribution. SO: Randomly sampling large enough from a population and
taking means of each sample will lead to a normal distribution. The sample
distribution of sample means. A rule of thumb = more than 30 observations.
