2
$\begingroup$

I'm writing an economics thesis and I've found many econometric papers in which I found regression tables without the p-value. They only show the coefficient of the independent variable and its t-value, like in the example image.

enter image description here

My econometrics knowledge is very elementary, but I thought that the most important thing to evaluate the significance of a variable was the p-value. Searching on the web, I've found that sometimes $t-value>2$ is considered as a good value for significance. Is that right? Can I say that when $t-value>2$ the coefficient is statistically significant?

Edit: after Rob's suggestion, here's a link to the paper

$\endgroup$
  • 1
    $\begingroup$ Why not give both? At the 5% level, a $t$ is strictly only significant for $|t|>2$ if the degrees of freedom is $>60$. $\endgroup$ – Glen_b Aug 21 '13 at 9:35
  • $\begingroup$ That's why I don't understand these tables. Why do they give only the $t$ value and say that it's significant, even if they don't show the p-value $\endgroup$ – Luigi Aug 21 '13 at 9:47
  • 2
    $\begingroup$ They give the number of observations, and the number of explanatory variables is (hopefully!) made clear, so you could work out the degrees of freedom, and hence the p-value. The critical value for 13 d.f. (the smallest one there if the listed variables are the only ones) is 2.16, so all but one of the coefficients that exceed 2 is actually significant. $\endgroup$ – Glen_b Aug 21 '13 at 9:53
5
$\begingroup$

I think that, in general, there is far too much concentration on p-values and significance and far too little on effect sizes, so I applaud the idea of showing tables without p-values.

There are a number of problems with over-reliance on p values, but one of them is alluded to (almost in reverse) in @Rob 's answer, above: People who don't understand the model may rely on the p-value. This may give them a false sense of understanding.

The particular case that you show is a sensitivity analysis. Here, the use of p-values seems even less appealing than typically: The goal of a sensitivity analysis ought to be to see how much the parameter estimates change, not whether they change from "significant" to "non-significant"; the sensitivity analysis shows that democracies and non-democracies have quite different values on some parameters. That's important. The p-value of those parameters is much much less important.

| cite | improve this answer | |
$\endgroup$
0
$\begingroup$

With more complex models, it may be difficult or impossible to calculate p-values. In these cases, t-values are often given instead (if you can provide a link to the paper, we can have a look if p-values could have been calculated). An example are mixed models, where the absence of p-values is a topic of constant debate (see http://glmm.wikidot.com/faq for more information). However, if the model is simple enough to calculate reliable p-values, I personally would want to provide them to the readers - especially in applied areas where readers may have trouble judging the models without them.

I do agree with @peter-flom that ideally, the focus should be on effect sizes. However, in my experience the only way to convince an applied scientist that a result should not be considered significant is by simply saying "my calculations yielded that it is not statistically significant". When I suggested that the effect size was not e.g. biologically relevant, they usually were not very responsive ("but it's significant!"). So I would use p-values to convince an applied reader not to pursue a line of thought further without more evidence, rather than to use p-values to try to prove something meaningful is happening. And of course, the biggest effect size still is not enough if it might very well be pure chance.

| cite | improve this answer | |
$\endgroup$

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.