# Normality and tests

#### leejones15

##### New Member
Ok, awesome. This has been super helpful, thank you!

#### obh

##### Active Member
Unlike the linear regression, the logistic regression is a bit confusing to interpret... the result is the probability to get a specific value

Test for significance is not everything, it is important to know how does it influence.

In your case you may get a result like: "The fact that a person knows a science (KS=yes) will increase the odds of "pos emb"=1 in comparison to "pos emb"=0 by 23%. You may play with http://www.statskingdom.com/430logistic_regression.html for interpretation in parallel to your regular stat software.

#### leejones15

##### New Member
Test for significance is not everything, it is important to know how does it influence.
When you say that significance is not everything, what does that mean for my results? Using your example, if the results show that odds increase by 23%, but p > 0.05, does that mean that knowing a scientist does not have an impact on positive embodiment?

#### noetsi

##### Fortran must die
What test would I use to compare a binary predictor to a binary DV, though? It wouldn't still be logistic regression, would it?
It does not matter what the predictor is. Only the DV. There are no predictor distribution assumptions formally in regression. If the DV is categorical and has two levels than you use logistic regression or probit (in practice these are different ways to do the same thing, they use a different link function to do it). If your DV has 3 levels (or 4 or 5) you would use multinomial logistic regression if they are nominal or ordinal logistic regression if they can be logically ordered.

#### noetsi

##### Fortran must die
Unlike the linear regression, the logistic regression is a bit confusing to interpret... the result is the probability to get a specific value

Test for significance is not everything, it is important to know how does it influence.

In your case you may get a result like: "The fact that a person knows a science (KS=yes) will increase the odds of "pos emb"=1 in comparison to "pos emb"=0 by 23%. You may play with http://www.statskingdom.com/430logistic_regression.html for interpretation in parallel to your regular stat software.
Odds ratios or proportional risk are probably the best way to interpret them. You use something like negative 2 log likelihood to test the model and Wald tests to test individual parameters.

#### obh

##### Active Member
When you say that significance is not everything, what does that mean for my results? Using your example, if the results show that odds increase by 23%, but p > 0.05, does that mean that knowing a scientist does not have an impact on positive embodiment?
I meant that you need to look at both.

If we ignore for a moment the multiple tests problem, and under the assumption that you took the proper sample size:
P value=0.23 says only that there is a probability of 0.23 that there is no effect.
or there may be a meaningless effect, per your definition, that is smaller than the required effect you used to calculate the sample size.

But if P value=0.00001, you know there is an effect, but it doesn't say anything about the effect size.

For example, you may find that medication improves the disease healing process with P value=0.00001, but the effect is very small, like healing time will be improved from 90 days to 89.8 days. you probably won't use this medication ...

#### obh

##### Active Member
Odds ratios or proportional risk are probably the best way to interpret them.
Hi Noetsi ,

The best way I found to understand logistic regression: how a change in one unit of the predictor will increase/decrease the odds of an event (1).

#### noetsi

##### Fortran must die
Hi Noetsi ,

The best way I found to understand logistic regression: how a change in one unit of the predictor will increase/decrease the odds of an event (1).
I think that is an odds ratio This is what I am talking about.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2938757/

Relative risk may be even better, but most software does not calculate that I think (I can't speak to R it may).

#### leejones15

##### New Member
Here is a sample from a Stata run. How would you interpret this? Knowing a scientist (1 = knows) is the only one that is [nearly] significant. Would the other values not be reported?

#### Attachments

• 69 KB Views: 4

#### noetsi

##### Fortran must die
_cons is significant at the .05 level. It is the only variable that is. 1=knows has a p value of .055 so it is not significant at the .05 level (one way to know this is the CI contains 1 - if it does the variable is not statistically significant in logistic regression).

Is _cons interval or a categorical predictor (and how is it coded. For example 1/0 if it has two levels).

I did not see an intercept. Did you suppress it? If you did it changes the interpretation (which is why I don't suppress it). #### leejones15

##### New Member
_cons is significant at the .05 level. It is the only variable that is. 1=knows has a p value of .055 so it is not significant at the .05 level (one way to know this is the CI contains 1 - if it does the variable is not statistically significant in logistic regression).

Is _cons interval or a categorical predictor (and how is it coded. For example 1/0 if it has two levels).

I did not see an intercept. Did you suppress it? If you did it changes the interpretation (which is why I don't suppress it). _cons is the constant/intercept

#### obh

##### Active Member
I think that is an odds ratio This is what I am talking about.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2938757/

The relative risk maybe even better, but most software does not calculate that I think (I can't speak to R it may).
I tried to traslate the coefficient to a meaningful sentence, as you may see.
It seems that R has the "relative risk".
https://cran.r-project.org/web/packages/logisticRR/vignettes/logisticRR.html