0

I'm doing a logistic regression concerning voluntary choice for an audit( 1 = yes, 0 = no)

Now first i want to look at the two groups (the ones who choose for a voluntary audit and the ones who dont) and compare the means and std deviation of both groups. For the categorical variables im going to use Chi square and for the contious variables i have the choice between independant sample t-test and man whitney test.

Now i've tested if my variables are normally distributed (needed for independant sample t-test). one variable was log transformated (lntotactief) but it doesn't seem to be normally distibuted according to both tests. although when i look at the histogram it seems to be normally distrubted.

Now my questions is can i use the man whitney test for all my continious variables to compare the means and test if they are statistically different between both groups? Or should i just use the independ sample t test for the log transformated variable?enter image description here

Thanks in advance

Michael Lew
  • 10,995
  • 2
  • 29
  • 47
Latriuz
  • 21
  • 6
  • On "Now i've tested if my variables are normally distributed (needed for independant sample t-test)" see [this answer](https://stats.stackexchange.com/questions/2492/is-normality-testing-essentially-useless/2501#2501) – Glen_b May 04 '18 at 05:13

1 Answers1

1

To start ... you really are not giving us enough information to give well-considered opinions. The t-test is reasonably robust to some departures from normality. The key questions are whether there is a suggestion of a bimodal distribution within groups, and whether the standard deviations within groups are similar (homoscedasticity).

Your samples sizes look pretty big so there is a large probability that a "test of normality" will get rejected even when the shape of the distribution is reasonably symmetric. So that's an argument against using the LS or the You have a negative skewness and negative values in the one variable we can see in that screenshot, so it would not be a good candidate for log transformation. If it were one of the ones that you already log transformed you might reconsider that decision. One possible approach is to perform both methods and see if the results are different. From what you have said, I suspect they will give the same verdict.

DWin
  • 7,005
  • 17
  • 32
  • I log transformated the second variable (lntotactief). In 2 other papers they transformated this variable in this manner. But it seems a little strange that i transformate it to be more like a normal distributed variable. and then i still use the man whitney test – Latriuz May 04 '18 at 12:28
  • There is no need for log transformation if you use Mann-Whitney. That test only takes into account ordering and log() does not affect order. Log transformation may improve some statistical properties but may distort other. If you do regression on logged variables you are making inferences about geometric means rather than arithmetic means. T-tests are equivalent to regression on one variable, so the same caveat applies there as well. – DWin May 04 '18 at 16:30