3
$\begingroup$

enter image description here

I want coefficients for ‘Isolate’, ‘Temperature’ and ‘Isolate*Temperature’. I want to use these coefficients to plot probabilities of death (for each isolate) as a function of temperature (25-50C). What would be the best way to analyze this data? I've read that Firth's logistic regression is probably best but if the model has "blown up" then I'm not sure how to proceed. Any thoughts would be greatly appreciated. Thanks.

Here is my output from firth's logistic regression analysis: Firths logistic regression output

Can anyone explain why the chi square values are equal to infinity for the intercept and temp?

$\endgroup$

closed as unclear what you're asking by Xi'an, Andy, gung, kjetil b halvorsen, Nick Cox Jan 11 '15 at 10:33

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.

  • $\begingroup$ Both Isolate and Temp are numeric variables? Maybe scaling them (ie, by their standard deviation, or by 100) would help? $\endgroup$ – Andrew M Jan 11 '15 at 7:18
  • 1
    $\begingroup$ Suggesting this be reopened because Firths method is still susceptible the the ordinary or typical pitfalls in logistic regression. So those should be avaialble to explain such results. A logistic regression coefficient of 22 is not entirely meaningless, but it is certainly "informative" ... informative of a problem. $\endgroup$ – DWin Jul 30 at 1:34
3
$\begingroup$

I'm not sure why your particular dataset got that particular result since the data is not available, but there's a reason logistic regression calculations "blow up": complete separation. If your interaction term creates a region of values where there are no expected events then the denominator of the chi-square calculation will go to zero and the statistic will go to infinity. Chi-square for tables is the sum of squared (observed-expected) divided by "expected". The fitted odds ratio would also be some number divided by zero or something very small and that might drive the odds ratio to very large or infinite values. Firth's penalized method is supposed to reduce the chances of this happening but you should also notice that some of your coefficients and standard errors have also "blown up". If you calculate odds ratios by exponentiation of values of 10 or 20, you get effective representations of numerical infinity.

So it's not Firth's method that is at fault, but rather idiosyncrasies in your particular model applied to your data. You can often see teh cause of this by using tabular displays where the row is the event in question and the columns are the predictors. Supposedly the likelihood ratio tests are still valid in this case. You might also to consider exact methods.

$\endgroup$

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