Mean Difference of 4-Level Categorical Variable

mekono — Sun, 07 Mar 2021 04:47:53 GMT

Hi!

I'm performing an analysis of 2 treatment groups (group a: ~700; group b: ~1400) and need to calculate the standardized mean difference between the groups. Propensity score adjusted was performed rather than propensity score matching to maintain higher sample size. The majority of category variables are binary but I have a handful that are 3-4 levels, like patient location; patient race; patient marital status.

I used a macro (http://www.lerner.ccf.org/qhs/software/lib/stddiff.sas ) to get the unadjusted SMD, and am using proc glimmix to calculate the adjusted SMD of the categorical variables. I found another macro but it was more geared to propensity score matching.

What is the best way to interpret the proc glimmix output to come up with a single mean difference value between the two treatment groups?

The code I'm using is below- I was advised to not include treatment group as random effect- and outcome.

<code>

    proc glimmix data=work.clinical_data;
      class trx_group region ;
      model region (descending)=  trx_group propensity_weight
           /  link=cumlogit  dist=multi  solution;
      output out=region predicted(blup ilink)=predProbs lcl(blup ilink)=lower ucl(blup ilink)=upper;;
      run;

Class Level Information
Class	Levels	Values
TRX_group	2	0 1
REGION	4	WEST SOUTH NORTHEAST MIDWEST

Number of Observations Read	2156
Number of Observations Used	2156

Response Profile
Ordered Value	REGION	Total Frequency
1	WEST	316
2	SOUTH	820
3	NORTHEAST	541
4	MIDWEST	479
The GLIMMIX procedure is modeling the probabilities of levels of REGION having lower Ordered Values in the Response Profile table.

Dimensions
Columns in X	6
Columns in Z	0
Subjects (Blocks in V)	1
Max Obs per Subject	2156

Optimization Information
Optimization Technique	Newton-Raphson
Parameters in Optimization	5
Lower Boundaries	0
Upper Boundaries	0
Fixed Effects	Not Profiled

Iteration History
Iteration	Restarts	Evaluations	Objective Function	Change	Max Gradient
0	0	4	2868.0488989	.	48.79174
1	0	3	2840.3830991	27.66579987	4.798868
2	0	3	2840.3402403	0.04285876	0.024003
3	0	3	2840.3402388	0.00000149	8.377E-7

Convergence criterion (GCONV=1E-8) satisfied.

Fit Statistics
-2 Log Likelihood	5680.68
AIC (smaller is better)	5690.68
AICC (smaller is better)	5690.71
BIC (smaller is better)	5719.06
CAIC (smaller is better)	5724.06
HQIC (smaller is better)	5701.06

Parameter Estimates
Effect	REGION	TRX_group	Estimate	Standard Error	DF	t Value	Pr > \|t\|
Intercept	WEST		-1.4138	0.1248	2151	-11.33	<.0001
Intercept	SOUTH		0.4960	0.1203	2151	4.12	<.0001
Intercept	NORTHEAST		1.6671	0.1256	2151	13.27	<.0001
TRX_group		0	-0.00609	0.09912	2151	-0.06	0.9510
TRX_group		1	0	.	.	.	.
Propensity_score			-1.1431	0.1834	2151	-6.23	<.0001

Type III Tests of Fixed Effects
Effect	Num DF	Den DF	F Value	Pr > F
TRX_group	1	2151	0.00	0.9510
Propensity_score	1	2151	38.87	<.0001

Re: Mean Difference of 4-Level Categorical Variable

StatDave — Sun, 07 Mar 2021 19:27:45 GMT

I have to say that REGION as a response variable, instead of as a predictor, seems very odd and makes me wonder you really want a model that predicts REGION. Be that as it may, see this note which illustrates using the NLMeans macro with a multinomial model. While the specific example in the note models a nominal response (LINK=GLOGIT) and computes relative risk estimates, you can also use it for an ordinal (LINK=CUMLOGIT) response and obtain difference estimates (which is the default).

topic Re: Mean Difference of 4-Level Categorical Variable in Statistical Procedures

Mean Difference of 4-Level Categorical Variable

Re: Mean Difference of 4-Level Categorical Variable