I have been tasked with creating a composite indicator for the purpose of comparing the demographic/socio-economic situation of communities within a city.

I'm using US Census data for three demographic/socio-economic indicators and have transformed them all to percentage measurements:

  • % People of Color
  • % Households Burdened by Housing Costs
  • % Linguistically Isolated Households

Percentage values provide a much more intuitive unit of comparison than the raw counts, and these indicators all share the same directionality (i.e. higher percentages indicate a more advantaged community, lower percentages indicate less advantage).

Now I need to find a method of combining the values of each indicator to form an index score, which itself will be used to divide the communities into quintile groups.

Here's a look at the data distributions:

Histogram comparison

While all three variables are percentage measurements (and therefore share the same scale), an important assumption of this exercise is that the indicators should have equal influence on the index score.

The index score should reflect a given community's relationship with the hypothetical "average community" (one that has the mean value for all three indicators). But what unit provides interpretable comparative value? It seems that many composite indicator methodologies use z-scores to standardize data before combining them (see here). This means that the unit of comparison between communities would be standard deviations from the "average community" score.


What can I do if the data are not normally distributed? A quick glance at the histograms above suggests that these are not normal distributions. Should I try further transforming the data (e.g., log, logit, arcsine) with the hope that the distribution will more closely resemble normality? Or should I leave the data as-is and abandon the idea that I will be able to interpret the final index's comparative unit (as seen here: "Standardizing a non normal dataset"; "Z-Score and Normal Distribution")?

Any general advice on the process I should go through to create a useful composite indicator would be much appreciated!


There are really two parts to this question: 1) How to standardize variables with different non-Gaussian distributions to make their values comparable, and 2) How to combine these standardized values into a single index?

For the second part, something like PCA would be recommended, to prevent "double counting" where the measures are correlated. Peter's answer addresses this issue.

In terms of the first part, a common approach is to essentially use quantiles/percentiles. These can be converted to equivalent "z-scores", in which case this technique is sometimes called the normal score transform. In the simplest version of this, for each variable you estimate quantiles from the Empirical distribution function. (More broadly, you could use any parametric or non-parametric approach to estimate the CDF.) The variables will then be in "comparable units", each approximately uniformly distributed.

In the second step, you transform these to approximately normal by putting through the inverse normal CDF function to get the z-scores corresponding to those quantiles. This second step is optional, but can be useful if you want to "treat the variables as Gaussian" for subsequent analyses (e.g. PCA or least-squares more generally is most well suited for Gaussian variables).

  • $\begingroup$ Very helpful! It's strange that I haven't come across this method being used in any of the case studies I've collected – it seems to be the right tool for the job. $\endgroup$ – Tiernan Aug 22 '16 at 15:46

One possibility in addition to those already mentioned is to use item-reponse-theory. You could say that the number $y_{ij}$ positives out of the total $n_{ij}$ in city $i$ for item $j$ has a $Y_{ij} \sim \text{Bin}(n_{ij}, \pi_{ij})$ distribution, where $y_{ij}/n_{ij}$ corresponds to your observed proportions. In the simplest form of this approach you would then say that $$\pi_{ij} = \frac{e^{\theta_j-b_j}}{1+e^{\theta_j-b_j}}$$ so that the $\text{logit}(\pi_{ij})=\theta_j-b_j$ (a more complicated model might have $\text{logit}(\pi_{ij})=a_j[\theta_j-b_j]$).

If you then fit this model, $\hat{\theta}_j$ is what you are interested in as an index of how the city is doing.

  • $\begingroup$ Am I correct in thinking that item response theory is best suited to scoring a set of indicators that have unequal levels of influence? $\endgroup$ – Tiernan Aug 22 '16 at 15:50
  • $\begingroup$ @Tiernan A Rasch model (the 1 parameter IRT model) assumes equal levels of influence. Other IRT models will estimate different discriminations (i.e. different levels of influence) $\endgroup$ – Ian_Fin Aug 23 '16 at 9:08

Statistically, you could do factor analysis to combine the percentages into a measure of a single latent variable. This will almost certainly result in one factor explaining a lot of the variance. The score on that factor will have mean 0 and sd 1, but may or may not be normally distributed.

Alternatively, you could try to figure out, substantively, which of the variables will have what influence on other measures. This might involve a lot of literature review, trying to find out how the 3 variables are related to some other variable that is of interest (such as SES or income or whatever).

  • $\begingroup$ I'm now realizing that my dataset presents a challenge to using PCA or FA: "The presence of spatial heterogeneity and spatial autocorrelation invalidates two basic assumptions of many standard statistical analyses: that data are independently generated and identically distributed." researchgate.net/publication/… Perhaps it's time to migrate this question to GIS Stackexchange... $\endgroup$ – Tiernan Aug 22 '16 at 20:52

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.