topic Re: PROC LOGISTIC for a categorical variable, one ML estimate is significant, but Type 3 is not. in Statistical Procedures

PROC LOGISTIC for a categorical variable, one ML estimate is significant, but Type 3 is not.

Orange1984 — Mon, 04 Mar 2019 21:12:57 GMT

I applied PROC LOGISTIC to predict a binary outcome variable, coded as 0 and 1. I have eleven predictors. Seven predictors are binary (VAR1 - VAR7), three predictors are continuous variable (VAR9 - VAR 11), and one predictor is categorical with four levels (VAR8: EDU4).

Regarding the four-level categorical variable,

(1) I found the "type 3 analysis of effects" was not significant (df = 3, chi-sq = 6.1259, p =0.1055).

However, one of the three estimates in "Analysis of Maximum Likelihood Estimates" was significant (df = 1, estimate = 0.9442, chi-sq = 5.2407, p = 0.0221) when"one level compared to the reference level" (Bachelor's degree or above vs. high school and some post-secondary).

In this case, should I interpret this variable has a significant effect on my binary outcome?

(3) In the model, I applied WEIGHT statement in PROC LOGISTIC. Should I use PROC SURVEYLOGISTIC to get a better estimate?

Here is my SAS code using PROC LOGISTIC:

PROC LOGISTIC data=SDE descending;
CLASS VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 / desc order = formatted param = ref;
model DV = VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8 VAR9 VAR10 VAR11 / lackfit rsquare;
weight RESPWT;
run;

Thank you for your insights.

Re: PROC LOGISTIC for a categorical variable, one ML estimate is significant, but Type 3 is not.

ballardw — Mon, 04 Mar 2019 23:57:11 GMT

Without any of the output or input data the only question I will attempt to answer at this point is about surveylogistic. You should use proc surveylogistic if your data comes from a complex sample design such as stratified or clustered or just about anything other than a simple random sample. The options and requirements for analysis would depend on the sample design.

Re: PROC LOGISTIC for a categorical variable, one ML estimate is significant, but Type 3 is not.

PaigeMiller — Tue, 05 Mar 2019 01:40:19 GMT

The two tests are not testing the same thing.

When you get a not significant (p=0.1055) p-value for the Type 3 test, this means that the slopes (regression coefficients) of the four different levels of EDU4 are not statistically different from each other — i.e. they are all the same.

When one of the estimates for the levels of EDU4 is significant (p=0.0221), this means the slope for this level is significantly different than zero.

These are not testing the same thing. One does not imply the other.