I have a problem similar to the one brought up here: Can't post question on freelance. Says 'this looks like spam.'
I am trying to post my first question to Cross Validated. I have tried to include the text of the question here, but I get the spam message here (on Meta) as well.
So, trying to post a much shorter question here.
OK, trying to post the draft of the question I was trying to put in on Cross-Validated:
Prediction model for binary dependent variable, numerical and categorical predictors when data consists of matched pairs of observations
SUMMARY I am looking for advice on what would be the best type of analysis that would allow me to figure out the nature of a relationship between the propensity of customers to complain about the service and the metric that quantitatively measures the quality of service.
Any advice on techniques and/or examples of studies with a similar structure would be welcome.
I have a data set that consists of matched pairs of users of a particular service. Each matched pair consists of: a) one person who filed a complaint about the service during the period of the study. Let’s call that person a “complainer”. b) one person who did not complain about the service during the same period, but who is similar (matched) to the complainer on a couple characteristics. Let’s call that person a “matched non-complainer”.
For each person the data set has the value of the quality of service metric, as well as the values of a few other variables that may help predict the propensity of a person to complain.
If the data set did not have the matched pairs structure, then I would build a prediction model for propensity to complain using classification tree. (I might also try logistic regression.) I would treat the status of each user as complainer or non-complainer as a binary dependent variable. I would treat the quality of service metric as a numeric predictor. I would also include as potential predictors various other quantitative and qualitative characteristics of the users that may help predict their propensity to complain.
But given the matched pairs data set, I am wondering if I should either modify the aforementioned approach in some way that takes into account this pairing, or use some more specialized technique.
DETAILS OF THE PROBLEM
We started by getting the set of people who filed a complaint about the service at any point during the period of the study. For each “complainer” a single “matched non-complainer” was selected at random from the pool of users of the service who 1) did not complain about the service during the period of the study and 2) were identical to the complainer in terms of two characteristics: class of service they subscribed to and geographical location where they received the service.
For each person in the data set, we know the value of the quality of service metric (a quantitative variable) that objectively measures how good was the service for that person during the period of the study. The goal of the study is to a) find out whether there is a quantitative relationship between the quality of service and the propensity to complain b) figure out where the inflection point is in that relationship and c) see if there are interactions between the relationship of the propensity to complain vs quality of service and other “predictors” of the propensity to complain.
The relationship between the quality of service and the propensity to complain is expected to be non-linear: most likely there is a threshold (inflection point) such that the propensity to complain rises roughly linearly with a small slope as service quality deteriorates from “perfect service” down to that threshold. Then, as the service quality deteriorates beyond that threshold, the propensity to complain continues to rise (again, roughly linearly) much more steeply.
In addition to the quality of service metric that is of interest, there are a couple other potential predictors that may affect the propensity to complain: 1) user’s tenure with the service and 2) the type of electronic gadget that the user relies on to interact with the service.
So, the structure of the data set is as follows. Each row of the data set corresponds to one complainer + matched non-complainer pair and has the following fields:
- Unique ID of complainer
- Class of service of complainer
- Geographical region of complainer
- Tenure of complainer with service (this may be treated as a quantitative variable, or may be treated as an ordinal categorical variable with a small number of categories, e.g., tenure may be “bucketed” into under 1 year vs. 1-3 years vs. over 3 years)
- Type of gadget that the complainer used to interact with the service (this is a categorical variable with a small number of categories)
- Quality of service metric value for complainer (a quantitative score between 0 and 100 that objectively measures how much the service that the complainer received deviated from “perfection”)
7 to 12: fields that describe the “matched non-complainer” who is paired with the given complainer. These fields are: class of service, region, tenure, gadget and quality of service metric. Of course, in each row the values in the class of service and the region fields for the matched non-complainers are going to equal the corresponding values for the complainer because the non-complainer was “matched” to complainer precisely on these 2 variables.
I have a large data set (more than a hundred thousand rows)