1
$\begingroup$

I have a simple question. Can one look for correlations between percentage and average ? For example, percentage of unemployment and average salary. Or shoul I use total numbers instead ? If I look for a relation between percentages and average I get significant correlation, but between totals data is not significant.

$\endgroup$
3
  • $\begingroup$ Usually, when looking for correlations, it is an attempt to define cause and effect relationships. I can find correlations between the date of the year and my check number, but the fact both increase sequentially is rather expected. There is also a significant correlation between pirates and global warming, but there is (to my knowledge) no cause and effect relationship between the data. $\endgroup$
    – Tavrock
    Mar 6, 2017 at 8:14
  • $\begingroup$ @subhashc.davar My variables are percentages of unemployed people in state every month and average salary of employed people in state every month. So I have unemployment data from bureau of statistics USA and average salary is calculated from randomly selected 10K people in that state. $\endgroup$
    – Extria
    Mar 6, 2017 at 9:24
  • $\begingroup$ @subhashc.davar Correct $\endgroup$
    – Extria
    Mar 6, 2017 at 9:44

2 Answers 2

2
$\begingroup$

The biggest risk of looking at statistics of the aggregates is ecological fallacy (example here). Group-level aggregates do not necessarily are suitable for inferring individual-level characteristics. The most vivid example is the Simpson's paradox, which shows how you can get completely different results when looking at group-level vs individual-level data. If you only have the aggregated data, your results do not necessarily translate to the individuals.

$\endgroup$
0
$\begingroup$

I'd be really uncomfortable with a correlation between a(n average of) continuous variable and a percentage, because the percentage is bounded by $[0,100]$. So if the percentage is near these boundaries, the relation (i.e. correlation) between the two is bound to be non-linear. A more suitable way of associating one continuous and a categorical variable is, for example, through regression.

$\endgroup$
3
  • $\begingroup$ Thank you for your answer. I didn't thought that percentages are categorical variable so I used Pearson method to calculate correlation. But I do not understand why percentages are categorical variable ? $\endgroup$
    – Extria
    Mar 6, 2017 at 8:58
  • $\begingroup$ Most often a percentage is calculated by grouping in some way (e.g. the proportion of men out of the total sample). I assumed this is what you meant. But I must admit, sometimes a percentage is used to show change from a certain baseline point (e.g. my income has changed by 150%). This second case is slightly different because in this case the percentage is bounded only on the lower side ([$0,∞]$), but still the problem of non-linearity remains, especially if you include '100%' as the neutral (no-change) point. $\endgroup$
    – IWS
    Mar 6, 2017 at 9:05
  • $\begingroup$ Now I understand why my data is categorical, that you for explanation. $\endgroup$
    – Extria
    Mar 6, 2017 at 9:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.