I am using GLMM by SPSS to test different predictors of recognition memory performance for affective pictures. The outcome variable is binary (Yes/No); so I use binary logistic regression. Predictors include Gender, Age, etc. I notice that the p values I get for the predictors in the table "Fixed coefficients" are different from the p values I get with the pairwise contrasts. For instance, in the fixed coefficient table the gender effect has t = 3.185 p = .001 but in the pairwise contrast t = 2.727 and p = .006. In the table for the pairwise contrasts it says "adj. sig.". I cannot find what the p value is adjusted for. In general, for all predictors the p value of the pairwise contrasts is always lower than in the fixed coefficients table.
Has anyone an explanation?
Ad-hoc comparisons or pairwise comparisons need to account for testing multiple hypotheses, which increase the type I error. Thus, some kind of adjustment is done - Bonferonni, Hochberg, Holm, Tukey method to name a few. If you don't adjust, you just increase the likelihood of finding a significant result by chance.
thanks for your answer.
I chose /EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD.
Thus, logically the pairwise contrasts are not supposed to be adjusted for multiple comparisons (Bonferroni, etc). The adj. sig. in the table pairwise contrasts must refer to something else. Also, for gender it is anyway only 1 pairwise contrast = 1 hypothesis.
I also tested a simple model with Gender as the only predictor. I get the same difference in p values. This means that this difference has Nothing to do with the presence/absence of other predictors.
Patrick, I'm not sure I understand your question completely. But if you are hoping to see some of the pairwise contrasts match the values of coefficients in your table of coefficients, you'll need to change your setting to SCALE=TRANSFORMED. Here's an example using one of the data sets that comes with SPSS.
NEW FILE.
DATASET CLOSE ALL.
* Change path as needed.
GET FILE = "C:\SPSSdata\bankloan.sav".
* Try using Patrick's settings for EMMEANS_OPTIONS.
GENLINMIXED
/FIELDS TARGET=default
/TARGET_OPTIONS DISTRIBUTION=BINOMIAL LINK=LOGIT
/EMMEANS_OPTIONS SCALE=ORIGINAL PADJUST=LSD
/EMMEANS TABLES=ed COMPARE = ed CONTRAST = PAIRWISE
/FIXED EFFECTS=age ed employ address income debtinc creddebt othdebt.
* To get pairwise contrasts that show values in the table of coefficients,
* use SCALE=TRANSFORMED.
GENLINMIXED
/FIELDS TARGET=default
/TARGET_OPTIONS DISTRIBUTION=BINOMIAL LINK=LOGIT
/EMMEANS_OPTIONS SCALE=TRANSFORMED PADJUST=LSD
/EMMEANS TABLES=ed COMPARE = ed CONTRAST = PAIRWISE
/FIXED EFFECTS=age ed employ address income debtinc creddebt othdebt.
Look for pairwise contrasts that have one of the lower education levels minus post-undergraduate degree. Each of those differences corresponds to one of the coefficients for variable ed in the table of coefficients. And the p-values match too. HTH.
Dear Bruce,
Thank you for your answer. Indeed, with SCALE=TRANSFORMED I get the same results as in the table with the fixed coefficients!
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS