There's not really enough information in your question - and I don't use JMP.

However, there's at least three ways to get a Chi-squared statistic from a logistic regression.

1) The classic LRT which uses the deviance of a simpler model compared to the deviance of a more complex model. The deviance has an approximate Chi-square statistic and the differences between two deviances has a Chi-square distribution that is usually an even closer approximation. This statistic has a Chi-square distribution with df equal to the number of parameters that change between the model. The deviance is simply -2 * the maximised log likelihood of the model.

2) You can use Pearson Chi-square to get a test statistic. This is less common but some computer packages report it. This is asymptotically identical to 1) but a better approximation I seem to recall for small samples.

3) You can use a Wald test. This is a normal approximation of a sort. You divide each parameter by its SE to get a test statistic for each parameter (with an asymptotic z distribution). If you square this and sum the squared values for a number of parameters this gives you a test statistic for the joint significance of those parameters with an approximate Chi-square distribution (with df equal to the number of parameters being tested). If its a single parameter the Wald test is often reported as a z statistic (the estimate divided by its SE).

Note that the three p values from these tests need not agree ... except asymptotically. Generally 1) and 2) are superior to 3) for logistic regression. If I had to guess (I stress this is a guess) I'd say the Chi-square test in JMP is 1) and the test of a single parameter a Wald z test. That's quite common in stats packages. CIs are often analogous to the Wald test (though profiling the likelihood is more accurate and is the default in some packages - notably R. This basically uses an approach based on the deviance used in 1) and should agree with that method).

If the parameter and SE are reported you can check whether a Wald test is being used .... you can get the exact p value easily from Excel or R or doubtless any number of online calculators. If it doesn't match it's probably the classic LRT testing a model that drops that parameter vs. one including it.

Dear William. Some information is missing in your question. However, from what I intuitively gather, I could tell you that those likelihood ratio tests act as "global tests" for significance of each predictor, while parameter estimates' significance is (as you may notice) parameter-specific. This might seem redundant but it is not in some cases. For example, in the case of categortical variables with multiple levels, the interpretation of coefficient is trickier because they represent a 'change from the baseline' (usually). Therefore you'll get k-1 coefficients (k = number of categories), and only some (or perhaps even only 1) might be significant. In such a case, the LRT (global) might result in non-significance of this categorical variable 'overall' but one, or some significant coefficients for the different baseline-category comparisons.

If you share more details (or even the output) it will be easier to help.

Hope this helps.

Best,

J.

Hello Josue and Thom.

Thanks for your messages.

I have attached an output from the JMP statistical program for the dependent variable "two or more SA for full sample", with two independent variables "right average constriction rate" and " GAD". You will see in the report of the parameter estimates that only the constriction rate is significant, but in the Effect Likelihood Ratio tests that both independent variables are significant. What are the implications of the differences between the parameter report and the Effect Likelihood Ratio test report, and how to best describe this in the results section of a scientific paper?

Thank you

Thank you Jochen - when the parameter estimates and the LR-test offer different perspectives on the significance of an independent variable, is it best to report both the parameter tests and the LR-tests?