0
$\begingroup$

Psychology masters thesis: One predictor variable (3 total) shows a slope coefficient of 0 on a partial regression plot.

  1. Do I need to transform this
  2. If so, what transformation is recommended and
  3. do I then need to transform all the predictor variables and the outcome variable?

I am fairly new to this so would be very grateful for step by step suggestions. Many thanks

$\endgroup$

closed as unclear what you're asking by Glen_b Sep 27 '17 at 8:37

Please clarify your specific problem or add additional details to highlight exactly what you need. As it's currently written, it’s hard to tell exactly what you're asking. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

  • 2
    $\begingroup$ Why do you think that is a problem ? $\endgroup$ – user83346 Sep 27 '17 at 6:55
  • $\begingroup$ No suitable advice could reasonably be given with so little information $\endgroup$ – Glen_b Sep 27 '17 at 8:39
  • $\begingroup$ ok,- I am conducting a multiple regression to see if parent stress, parent screen time viewing and parent limit setting on child screen time viewing predict child screen time viewing. All are continuous variables. The model including all three predictor variables is significant (F(3, 71)= 6.542, p<.002). However, t-tests reveal that while parent restriction and parent screen time show a significant relationship, parent stress does not. Parent stress shows a scatterplot slope coefficient of 0 which is not linear and therefore does not meet the assumption of linearity necessary for the test. $\endgroup$ – Annie Sep 27 '17 at 9:05
1
$\begingroup$

First thing you look at is the variance of outcome and input. If any of them are 0 then the coefficients are zero. After that lets go on.

Do I need to transform this? Probably not. Ask your self if it make sense that the underlying relation are linear in the current form. For example imagine I want to investigate the effect of wealth on happiness a linear scalling does not make sense 1 dollar does not have the same effect for Bill Gates as for the musician on the street. Here a relative scaling are preferred, so you perform a log-transformation.

Transforming the outcome. Sometimes it can be relavant. The most obvious tranformation is to use logistic or multinomial regression if you have categorical variables. Not doing it is quite wrong.

The next thing you do is you ask your self can my y be independent of this variable? Use your commen sense and your domain knowledge. There can be reasons why a coeficient are different than it should be. These group into 2 groups:

  • Correlated variables mess the results up. If you have 2 variables that are highly correlated it is difficult for the regression to find out variable actually made the difference.

  • There are reverse causality. For example I want to investigate the effect of going to a psychiatrist. I am stupid so I look at the suiciderate of the persons going to a psychiatrist and the persons not going. I find higher suiciderate for the persons going to a psychiatrist. This is not because going to a psychiatrist makes me want to kill myself, but because wanting to kill my self makes me go to a psychiatrist. So the causality is reversed.

So in your case you should look at the variables highly correlated with the variable of interest. Ask your self are there reverse casality? And what variables would have been nice include to explain the problem better and how might they have an effect. Remember finding something is insignificant is also a result.

$\endgroup$
  • $\begingroup$ Thank you very much for your prompt reply. I am very sorry for my infancy in this field but what exactly do you mean by the variance of outcome and input,- in which output table might I find this? $\endgroup$ – Annie Sep 27 '17 at 8:27
  • $\begingroup$ Input variable(s) is a synonym predictor variable(s). The "var" function finds it. The primary purpose was just to check if there was variation in both outcome and input a.k.a if all the input or output values are the same then it does not work. Naturally. By the way how large is your dataset? $\endgroup$ – Peter Mølgaard Pallesen Sep 27 '17 at 9:02
  • $\begingroup$ input and output variables are not the same. Data set is fairly small, (n=81) including a few missing values $\endgroup$ – Annie Sep 27 '17 at 9:09
  • $\begingroup$ Make a plot of the points with the explanatory variable on the x-axis and the outcome on the y axis. If the data is categorical try to make histograms of the different classes and put into the original question. $\endgroup$ – Peter Mølgaard Pallesen Sep 27 '17 at 9:14
  • $\begingroup$ I imagine the best way forward is to not include the variable in the test and take care to discuss why parent stress is not related to child screen time in this instance, using this measure, in this population. $\endgroup$ – Annie Sep 27 '17 at 9:15

Not the answer you're looking for? Browse other questions tagged or ask your own question.