# Linear Model or Logistic Model: Can Someone Recommend a Book?

I have a huge data set that looks roughly like this:

x = [x1, x2, x3, ..., x800]
x is how much percent the real project cost is different from the planned project cost

y1 = [y1.1, y1.2, y1.3, ..., y1.800]
y1 is the estimated project cost in dollar (or another currency)

y2 = [y2.1, y2.2, y2.3, ..., y2.800]
y2 is the real projet duration

y3 = [y3.1, y3.2, y3.3, ..., y3.800]
y3 is the project manager (in coded form like "mk" and "op")

....

y50 = [y50.1, y50.2, y50.3, ..., y50.800]
y50 is the estimated project duration


Now the aim is to predict the x (cost difference in percent) in dependency of other y variables. It is a priori not clear which variables influence it. And some of the y are not interdependent like for example the difference in percent x, the real costs and the estimated costs since x is calculated from the real and estimated costs.

I am now wondering how is the best practice way to analyse the data or how to build a model (linear or maybe logistic). The binary output for the logistic model could be difference bigger than 10 %.

Since I have a lot of data sets (over 800) almost everything could get significant on a e. g. 5 % level. And It's not easy to handle/control the power if you do not have a simple t-test.

What would you recommend? Is there a good book for "normal people" about that kind of statistics?

• Unless you're looking for a run-of-the-mill Statistics textbook, it will depend a lot on the software that you plan on using. You might really like this article--it's geared towards a general audience (though still substantial). In addition, you can practice with the code/data the author provided on his website (and described throughout). Aug 6 '14 at 12:35
• Also, you probably don't want to use the logistic regression here since the restriction that results be between zero and 1 is most likely an unreasonable assumption on your part. As an aside, your data will be more readable if you follow the convention of using x to signify your predictors and y to be the response. Aug 6 '14 at 12:39
• But again, that article will cover some of the big issues that you'll have to consider (such as incorporating cross-validation into your analysis as well as regularization), giving you a better idea of where to go from there. Aug 6 '14 at 12:41
• Thanks for the comments. And yes - I should switch x and y ;) Aug 6 '14 at 16:00