Dear colleagues, (especially those who are in love with statistics)
I tried to study the influence of a dichotomus independent variable (X) on a continuous dependant one (Y) {e.g.lets say the influence of sex on the body content of a heavy metal}. To do so I first; conducted T test and found a significant difference in Y between the categories of X, second; I used stepwise multiple linear regression analysis with X and some other factors as independent variables and Y as dependant one and the result was insignificant influence for X on Y.
My questions are:
1. Which result should I go with, T-test's result or that of the regression?
2. What about vise versa (if the regression result was significant and that of T-test wasn't), which result should I depend on?
I will appreciate every single contribution. Thank you in advance.
Hi Mohammad. I think that you should rely on the t-test results, since your independent variable is not continuous, so you have two populations (two sex) that may behave differently. Regression analysis should be better if your have both independent and dependent continuous variables, which is not the case.
Dear Mohammad,
your question is interesting. I suggest you check if your data has a normal distribution because the t-test is based on the assumption that the data follows a normal distribution.
Dear Antonio and Alessandro,
Thank you for your answers. However,
Dear Antonio, someone may think that the significant influence of X appeared in T-test is not real (i.e., may have resulted from X's association to another dichotomus factor really significantly affecting Y) and the regression analysis adjusted the independent variables on eachother and thus showed X insignificant. What do think about this claim?
Dear Mohammed Jamil
A t-test is part of univariate analyses. Here you examine the variation in mean content of heavy metals in relation to other variables which are thought as independent or explanatory variables. For significance testing you use t-test. If you have more than one explanatory variable then you may find some have significant effect and other do not. To control for the confounding effect of these variables you then use regression analysis. Step-wise multiple regression analysis is good here as your dependent variable is continuous (mean content of heavy metals). The explanatory variables need to be carefully used to fit such analysis. You may present both results if your work is towards a dissertation or a thesis. If your work is publish a paper then regression results might be preferred and more economic.
Thank you very much dear Dr. Habib. Your contribution is highly useful.
"The explanatory variables need to be carefully used to fit such analysis", would you please give me more on this statement?
Dear Mohammed
Please try this link, it might be useful
https://statistics.laerd.com/spss-tutorials/multiple-regression-using-spss-statistics.php
الدرس 49 من سلسلة دروس د / أسماء الميرغني لشرح البرنامج الإحصائي
https://www.youtube.com/watch?v=4nU8ux9bsTI
You should look at your data, and see if the assumptions made by both approaches are reasonable. Assumptions of normality, independence, and equality of variance.
Assuming that both the t-test and the regression model are appropriate for your data, then I would believe the regression model over the t-test. However, I can arrange a collection of numbers in such a way as to make either or both tests give incorrect results.
1) There could be an unmeasured variable that is the real cause-effect relationship. The observed effect is through correlation with the unmeasured variable.
2) One or more variables currently in the regression model are the real cause-effect relationship. X is correlated with one of these other variables.
3) The independent variables are not independent of one another.
4) The error term in the regression model is not normally distributed.
It is bad policy to run a bunch of statistical tests and select the outcome that you like. It is good policy to run a bunch of statistical tests and eliminate those approaches that (in hindsight) you should not have tried because your data does not satisfy the assumptions of the model. You may be left with "these tests" give one answer, and "those tests" give another. I would be perfectly happy in this case for you to select the outcome that you like, BUT include one of the following:
1) Note that doing the analysis in a different way leads to different conclusions and here are the data.
2) Include a supplement where the alternative analysis outcomes are presented along with the original data.
I would say that #2 is the better option because it leaves nothing to the imagination, and it takes advantage of work that you have already done. Also, with a supplement you can cite it, and discuss reasons why you did not choose that possibility.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression