I have a table with 1 000 projects from past. Each has approxiametly 150 attributes (e.g. project manager, initial year of project, team members, budget, investors in project (yes or no) and a lots of other, place of project, ..). Some attributes for some projects are missing and there is just "no available" info. Each project has assigned value, whether it was successful or not. That's the key indicator.

I have to perform analysis to find relevant factors for these projects, based on their potential to become successful. What's the best way to do that?

Here's an example with the attribute "initial year of project". There is 1 000 project in the table. 600 projects were successful finished and 400 were not. From all 1 000 projects, 900 has available information about "initial year of the project" (so, 100 projects have "no available" info in this attribute). These are values:

  • year 2010 - 100 projects were started - 50 of them were successful, 50 unsuccessful
  • year 2011 - 200 projects were started - 150 of them were successful,50 unsuccessful
  • year 2012 - 300 projects were started - 100 of them were successful,200 unsuccessful
  • year 2013 - 100 projects were started - 20 of them were successful, 80 unsuccessful
  • year 2014 - 200 projects were started - 170 of them were successful, 30 unsuccessful

I thought, that the correct way to find the most relevant factors is to quantify the dependency ratio of each attribute. So, for attribute "initial year of project" with the data above, I made a computation like this: 0.85*0.22*0.9 (because 170/200 = 0.85; 200/900=0.22;900/1000 = 0.9)

Is it correct? How would you solve this?

  • $\begingroup$ About what is the proportion of missing? Can you try to do imputation? $\endgroup$ Feb 26 '17 at 12:36

I see what you're trying to do here, i.e. calculate the success rate for year 2014 (0.85), the proportion of projects attributed to that year (0.22), total proportion of projects that have a initial year (0.9) and multiplying them to get a KPI. But I don't understand how you would interpret this number that you get for each year as a measure of importance for the "initial year" variable. Also, this is a bi-variate way of looking at variable importance, as you are looking at the "initial year" variable's impact on project success, whereas any real relation between variables and project success could be a combination of 2 or more variables.

I'd suggest you do a quick Random Forest run to get an estimate of variable importance for your problem. And you don't have to necessarily be proficient in stats or R to do so. Just download R (https://cran.r-project.org/bin/windows/base/), and then a package called "rattle" (https://cran.r-project.org/web/packages/rattle/index.html) that provides a GUI interface to do RF (and many other correlation tests that you can try out!).

And about the missing values in your dataset- you can impute them (probably make rules looking at project end year, manager start date, end date, etc. In my past experience I have found that usually a few simple ones help in covering a majority of missing values. For ex, if a project end year was 2011 then start year will be 2010 or 2011; then if the manager who was assigned to work on the project started in your firm in 2011, then the project may also have started in 2011), but if there is no reliable way of doing that you can flag them (say code all values with missing initial year as 9999) to still incorporate them in your training data.


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.