Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

How to interpret the t-ratio?

Hey guys,

I have a question: Is the t-ratio the same as the t-value? Or how should the t-ratio be interpreted?

I have 5 attributes with 2 levels each (present/absent)  and a price attribute with 4 levels.
The logit output gives me the effects, standart error and t-ratio for each.

But I need t to know which attributes are significant in my model and on which level and which are not significant. How can I see this from the output?

Thank you.
asked Nov 3, 2013 by anonymous
edited Nov 3, 2013 by Walter Williams
Be careful about trying to interpret t-ratios for conjoint part-worths.  In conjoint analysis, we typically scale the utilities so that within each attribute (factor) the utilities are zero-centered.  That means that the average utility within each factor is zero.  So, even for a very important attribute that has 3 or more levels in it, it is quite likely that one of the levels will have a utility around zero (because it is a zero-sum game within each attribute).  Since the t-test we report is relative to whether one level's utility is significantly different from zero, a level from a very important attribute (that is in middle of the pack in terms of preference for that attribute's levels) could appear to have a t-value that would indicate less than 95% confidence.  But, that DOESN'T mean that this level doesn't impact choice significantly!  It just means that when the utility levels are zero-centered within attribute, this level is not significantly different from zero.
Thank you very much for hinting me at this point. I understand what you mean, but what exactly does this mean for any statements about significances?
Isn't there any possibility to say anything about the significance of the different attributes with the logit output?

What conclusions can I then draw from the outputs? As I said above. I have 5 2-level attributes (present/absent) so their effects are always the same beside one being the negative value of the other. And a 4 level price attribute.

 thanks a lot, Brian!
Dear Brian,

Thank you very much for this information. I also have the same question regarding the t ratios.

I am also looking for the significance of the effects yet can see how this poses a problem if the attribute levels are zero-centered.

Is there any other way to find out the significance level? For example, could the effects be estimated in their raw value rather than zero-centered?

Thank you in advance for your help!

1 Answer

+1 vote
The t-ratio is the estimate divided by the standard error. With a large enough sample, t-ratios greater than 1.96 (in absolute value) suggest that your coefficient is statistically significantly different from 0 at the 95% confidence level. A threshold of 1.645 is used for 90% confidence.
answered Nov 3, 2013 by anonymous
Thank you very much! Than it obviously seems to be what is known to me as t-value. I was just irritated by the word ratio.

Thanks a lot.
Very high t-ratio values - is that correct?
Utilities estimation with multinominal logit
Threshold for 99% confidence level
What is the meaning of a t value of 1.96?