# Can tools like SPSS find out which columns correspond to a certain data range

This is a beginner's questions about statistical tools.

We have a spreadsheet with many columns from a survey.

The first few columns describe each individual:

1. Gender (discrete values Male, Female)
2. City (discrete values X,Y,Z)
3. College Year (discrete values 1,2,3,4)
4. BMI (continuous values in the range 10-50)

Then there is a column "Score" for each person, which is continuous from 1 to 5.

We want to find out what interesting correlations between sub-populations and Scoring ranges. For example, "Males in City Y have scores between 1.5 and 2.5" or "Females with BMI between 24.8 and 28.2 have scores between 2.3 and 3.7"

I can crunch this by trial and error using Excel, Access, or C++ code, but then I have to think up of each query and write the code or formulas for it, or try various ranges and combinations methodically.

But then I may miss some pattern. Can the major tools like SPSS or R extract sub-populations that correspond most strongly to a given scoring range? In other words I want it to look at the data and output a fact like the example above. Ideally I don't even want to provide the Score range, the tool should infer and extract the interesting patterns.

Thank you.

• I would usually use CHAID analysis for this kind of problem - I use it in SPSS but you can find it in R too (for free). – eli-k Aug 19 '16 at 5:59

## 1 Answer

Rather than tell you how to do this, I'm going to tell you that you shouldn't do it, and why.

Your proposal is an example of data-dredging. Rather than trying to answer a question about the data, you're looking for questions the data could answer. The problem with data-dredging is that given enough things to check, you're going to find spurious relationships, which won't generalize beyond the sample you happen to be looking at. It's not useful to point out that "Females with BMI between 24.8 and 28.2 have scores between 2.3 and 3.7" in your sample if this says nothing about the population, assuming you want to do some kind of inference.

• Fair enough. What if I say that we are really interested in specific outcomes of score. For example, "high" scores between 3.5 and 5, or "low" scores between 1.0 and 2.5? If we fix those score ranges, is it still "data-dredging"? – royappa Aug 18 '16 at 23:35
• That's not data-dredging, but if your real goal is just to see how score can be predicted from the other variables (and you're looking for stochastic rather than deterministic relationships), there's no need for setting cutoffs like that. Just use your favorite sort of predictive model (e.g., linear regression) to try predicting score with the other four variables. – Kodiologist Aug 18 '16 at 23:40
• An alternative take on this exercise is that it's simple data exploration. Yes, it's good to warn that many patterns that appear in an EDA might not hold up when tested--but that's no reason not to try to find things out! – whuber Aug 18 '16 at 23:41
• @whuber I guess I don't see the point in doing an analysis that's quite that exploratory unless you already know you're going to collect more data in the future. – Kodiologist Aug 18 '16 at 23:43
• Also, this IS an initial survey. More precisely, it's the first time an established survey has been done in a new population; we took a well-known American survey and ran it in India. We did some easy analysis but don't see the usual (obvious) relationships between the four variables and "high" or "low" score, as the US data does. This was surprising. So I want to see if we're missing something, or if (unlike the US) there is truly no meaningful relationship between those four variables and the "score". Then that would open up new questions. Thanks again for the great comments. – royappa Aug 19 '16 at 0:13