It’s standard practice in social science journals to report only point estimates and hypothesis tests for the coefficients. Most statisticians, on the other hand, hold that confidence intervals give a better picture of the sampling variability in the estimates.
Conventional confidence intervals for logit regression coefficients are easily computed by hand. For an approximate 95% confidence interval around a coefficient, simply add and subtract the standard error multiplied by 1.96 (2.00 is close enough for most purposes). But you can save yourself the trouble by asking LOGISTIC or GENMOD to do it for you. In LOGISTIC, the option in the MODEL statement for conventional (Wald) confidence intervals is CLPARM=WALD. In GENMOD, the corresponding MODEL option is WALDCI. The default is a 95% confidence interval. To change that to a 90% interval, put the option ALPHA=.10 in the MODEL statement.
There’s another method, called profile likelihood confidence intervals, that may produce better approximations, especially in smaller samples. This method involves an iterative evaluation of the likelihood function and produces intervals that are not generally symmetric around the coefficient estimate. GENMOD computes profile likelihood confidence intervals with the MODEL option LRCI (likelihood ratio confidence intervals). In LOGISTIC, the model option is CLPARM=PL (for profile likelihood). The profile likelihood method is computationally intensive so you may want to use it sparingly for large samples.
Here’s an example of both kinds of confidence intervals using LOGISTIC with the death-penalty data set:
PROC LOGISTIC DATA=my.penalty DES; MODEL death = blackd whitvic culp / CLPARM=BOTH; RUN;
Besides the usual results, we get the numbers shown in Output 3.1. The intervals produced by the two methods are very similar but not identical.
Parameter Estimates and 95% Confidence Intervals Wald Confidence Limits Parameter Variable Estimate Lower Upper INTERCPT -5.2182 -7.0407 -3.3957 BLACKD 1.6360 0.4689 2.8031 WHITVIC 0.8477 -0.2427 1.9381 CULP 1.2709 0.8863 1.6555 Profile Likelihood Confidence Limits Parameter Variable Estimate Lower Upper INTERCPT -5.2182 -7.2485 -3.5710 BLACKD 1.6360 0.5289 2.8890 WHITVIC 0.8477 -0.2190 1.9811 CULP 1.2709 0.9187 1.6953 |
The output from GENMOD is sufficiently different that it’s worth examining as well. The program is:
PROC GENMOD DATA=my.penalty; MODEL death = blackd whitvic culp / D=B WALDCI LRCI; RUN;
In the top half of Output 3.2, we see the Wald confidence intervals, labeled “Normal” because they are based on the normal distribution. Unfortunately, variable names aren’t given so you have to remember that PRM1 (parameter 1) is the intercept, PRM2 is the coefficient for BLACKD, and so on.
In the lower portion of the output, we find the profile likelihood results. The first column of numbers is the one to pay attention to. The remaining columns can usually be ignored, but here’s a brief explanation. When GENMOD is evaluating the profile likelihood function, it varies not only the parameter of interest but also the other parameters in the model. The additional columns tell us what values the other parameters took on when the parameter of interest was at its lower (or upper) limit.
Normal Confidence Intervals For Parameters Two-Sided Confidence Coefficient: 0.9500 Parameter Confidence Limits PRM1 Lower -7.0408 PRM1 Upper -3.3956 PRM2 Lower 0.4689 PRM2 Upper 2.8031 PRM3 Lower -0.2427 PRM3 Upper 1.9381 PRM4 Lower 0.8863 PRM4 Upper 1.6555 Likelihood Ratio Based Confidence Intervals For Parameters Two-Sided Confidence Coefficient: 0.9500 Param Confidence Limits Parameter Values PRM1 PRM2 PRM3 PRM4 PRM1 Lower -7.2485 -7.2485 2.6226 1.5941 1.6140 PRM1 Upper -3.5710 -3.5710 0.8248 0.1857 1.0038 PRM2 Lower 0.5289 -4.0304 0.5289 0.4032 1.1420 PRM2 Upper 2.8890 -6.8385 2.8890 1.3096 1.5126 PRM3 Lower -0.2190 -4.2409 1.1922 -0.2190 1.2389 PRM3 Upper 1.9811 -6.6011 2.1640 1.9811 1.4047 PRM4 Lower 0.9187 -3.9489 1.1767 0.6457 0.9187 PRM4 Upper 1.6953 -6.8361 2.3064 1.1261 1.6953 |
We’ve seen how to get confidence intervals for the β parameters in a logistic regression. What about confidence intervals for the odds ratios? LOGISTIC can compute them for you. In the MODEL statement, the CLODDS=WALD option requests the conventional Wald confidence intervals, and the CLODDS=PL option requests the profile likelihood intervals. Output 3.3 shows the results for the model just estimated. The “Unit” column indicates how much each independent variable is incremented to produce the estimated odds ratio. The default is 1 unit. For the variable CULP, each 1-point increase on the culpability scale multiplies the odds of a death sentence by 3.564.
Conditional Odds Ratios and 95% Confidence Intervals Profile Likelihood Confidence Limits Odds Variable Unit Ratio Lower Upper BLACKD 1.0000 5.135 1.697 17.976 WHITVIC 1.0000 2.334 0.803 7.251 CULP 1.0000 3.564 2.506 5.448 Conditional Odds Ratios and 95% Confidence Intervals Wald Confidence Limits Odds Variable Unit Ratio Lower Upper BLACKD 1.0000 5.135 1.598 16.495 WHITVIC 1.0000 2.334 0.784 6.945 CULP 1.0000 3.564 2.426 5.236 |
If you want odds ratios for different increments, you can easily calculate them by hand. If O is the odds ratio for a 1-unit increment, Ok is the odds ratio for a k-unit increment. If that’s too much trouble, you can use the UNITS statement to produce “customized” odds ratios. For example, to get the odds ratio for a 2-unit increase in CULP, include the following statement in the LOGISTIC procedure:
UNITS culp=2 / DEFAULT=1;
The DEFAULT option tells SAS to print odds ratios and their confidence intervals for a one-unit increase in each of the other variables in the model.
Because GENMOD doesn’t compute odds ratios, it won’t produce confidence intervals for them. However, they’re easy to get with a hand calculator. The first step is to get confidence intervals for the original parameters. Let β be a parameter estimate, and let U and L be the upper and lower confidence limits for this parameter. The odds ratio estimate is eβ. The upper and lower odds ratio limits are eU and eL. This works for either conventional confidence intervals or profile likelihood confidence intervals. If you want confidence intervals for the transformation 100(eβ–1), discussed in Section 2.9, just substitute the upper and lower limits for β in this formula.
18.226.187.24