1
$\begingroup$

I am writing a master's thesis on crowdfunding. For this research I set N independent variables and 3 DVs. The reason for doing so is that I want to explore the phenomenon from the aspects of all 3 DVs, as they are giving me different insights, and to crosscheck the results from all 3 models. Here is a bit more details on the three DVs and the models:

  1. Success [dummy] - I will use logistic regression for that.

  2. Success Ratio - Ratio variable [0-N] - I use Linear Regression and to linearize the model I use the following transformations:
    $\ln(x+1)$ to keep the zeroes, as the distributions are heavily skewed. ==> $\ln(y+1)=b_0+\ln(x_1+1)b_1$..., etc.

  3. N_Backers - Count variable [0-N] - I plan to use a Poisson GLM with log function, but it might be reasonable as well to be inline with DV2 and to use Linear Regression with transformed variables $\ln(y+1)=b_0+\ln(x_1+1)b_1$..., etc.

Question: What is your advice on these methods, do they sound reasonable? Further, should I use linear or Poisson for the 3rd DV? My concerns are that I am not sure how the results will be interpreted using all different models. Probably the use of consistent models is the right approach.

$\endgroup$
1
$\begingroup$

Transformation on the dependent variables is rarely advisable, its much better to use a GLM. So I would consider the poisson model for the 3rd DV, but be sure to check for overdispersion if you use it.

According to your clarifications, the second DV has over 50% of 0 values. I've modeled symilar response variables using a conditional strategy: You could fit a binomial model to classify as zero/non zero and then fit a gamma model to the observations classified as non zero by the first model.

$\endgroup$
5
  • $\begingroup$ Thanks for the advice Aghila. The second DV is a ratio variable, combined of 2 other variables. It might take positive values from 0-N, also decimals, where 0>=X<1 indicates unsucecssful project (corespoinding to the 1st DV) and 1 and above indicates successful project. So I checked one similar research and they guy was using ln transformation for dv and iv, thats why my idea was to use ln tranform variables. $\endgroup$ – Delyan Peyankov Jul 21 '13 at 13:55
  • $\begingroup$ Its hard to make a recomendation without knowing more about the process generating this variable and its actual distribution. Can you provide more details? Also, what is the reason for transforming all the independent variables? $\endgroup$ – Aghila Jul 21 '13 at 16:39
  • $\begingroup$ The DV indicates the success ratio - Funded Amount/Funding Goal. Values below 1 indicate unsuccessful funding, whereas 1 and above, successful funding. The sample is quite ample 6000+ The distribution is rougly like this: Range [0-600+]. About 50%+ are 0 values , 40%+ values [0-5], and a very small percentage of the values are higher than 5, so the picture is quite skewed. There is a higly distinctive 2 peak distribution - projects with 0 success ratio, and projects with 1-1.5 with sharp decaying tails after the 2 peaks. $\endgroup$ – Delyan Peyankov Jul 21 '13 at 17:23
  • $\begingroup$ Reasons for Transformation of the IV: Interpret the model as elasticities, Not sure about: Linerilize the IV; deal with outliers $\endgroup$ – Delyan Peyankov Jul 21 '13 at 17:23
  • $\begingroup$ Edited my answer with your new information $\endgroup$ – Aghila Jul 21 '13 at 17:50

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.