My use of a mixed effects model has been queried for a repeated measures analysis. I am working in SAS 9.4. In my experimental set up I have a randomised block design, with 5 blocks each containing 1 replicate of 3 levels of a treatment (5 * 3 , n= 15)- variable name "TREAT". Measurements of my reponse variable (LOG_N2O) were taken on 75 occasions (CYCLETIME), hence the need for a repeated measures analysis, but since there are missing data from nearly every measurement time, ANOVA is not suitable. The code I have used is below
PROC MIXED DATA= AUTOSTATS; CLASS BLOCK TREAT CYCLETIME; MODEL LOG_N2O= TREAT CYCLETIME BLOCK TREAT*CYCLETIME TREAT*BLOCK CYCLETIME*BLOCK; REPEATED CYCLETIME / SUBJECT= CHANNEL TYPE= AR(1); RANDOM CHANNEL BLOCK; LSMEANS TREAT CYCLETIME BLOCK /* POST HOC TEST FOR IDENTIFYING THE DIFFERENCES */ TREAT*CYCLETIME TREAT*BLOCK CYCLETIME*BLOCK / ADJUST= TUKEY ; ODS EXCLUDE CONVERGENCESTATUS; /*SUPPRESS THE PRINTING AS IT TAKES UP MEMORY*/ ODS EXCLUDE COVPARMS; ODS EXCLUDE DIMENSIONS; ODS EXCLUDE FITSTATISTICS; ODS EXCLUDE ITERHISTORY; ODS EXCLUDE LRT; ODS EXCLUDE MODELINFO; ODS EXCLUDE NOOBS; ODS OUTPUT DIFFS= MEANTREAT TESTS3= N2O_FSTATS; RUN;
This gives me output that is understandable and makes sense to me...
SAS Output
Class Level Information
Class
Levels
Values
BLOCK
5
1 2 3 4 5
TREAT
3
FER NH4 NO3
CYCLETIME
75
01APR2014:03:41:50 01APR2014:09:34:55 01APR2014:23:12:41 02APR2014:12:34:39 02APR2014:16:42:48 02APR2014:20:51:37 03APR2014:00:59:42 03APR2014:05:08:41 03APR2014:09:17:52 03APR2014:13:26:31 03APR2014:17:36:16 03APR2014:21:44:46 04APR2014:01:52:23 04APR2014:05:59:54 04APR2014:10:09:08 04APR2014:16:07:16 04APR2014:20:14:46 05APR2014:00:23:31 05APR2014:04:33:54 05APR2014:08:42:11 05APR2014:12:49:13 05APR2014:16:55:41 05APR2014:21:02:37 06APR2014:01:10:53 06APR2014:05:20:49 06APR2014:12:56:13 07APR2014:15:17:47 09APR2014:12:27:32 09APR2014:16:36:09 09APR2014:20:47:03 10APR2014:00:59:10 10APR2014:05:10:27 10APR2014:09:19:44 10APR2014:20:50:00 11APR2014:01:02:00 11APR2014:18:46:00 11APR2014:22:57:00 12APR2014:03:08:00 12APR2014:07:17:00 12APR2014:11:26:00 12APR2014:15:34:00 13APR2014:16:09:00 13APR2014:20:19:00 14APR2014:00:28:00 24MAR2014:17:41:36 24MAR2014:22:03:01 25MAR2014:02:38:52 25MAR2014:07:14:45 25MAR2014:15:52:05 25MAR2014:20:13:41 26MAR2014:00:49:28 26MAR2014:05:25:17 26MAR2014:10:01:16 26MAR2014:16:41:00 26MAR2014:21:16:41 27MAR2014:01:52:21 27MAR2014:06:27:58 27MAR2014:15:40:36 27MAR2014:20:02:04 28MAR2014:00:37:53 28MAR2014:05:13:38 28MAR2014:09:49:36 28MAR2014:14:25:30 29MAR2014:19:23:01 29MAR2014:23:57:56 30MAR2014:04:32:23 30MAR2014:09:07:49 30MAR2014:13:43:31 30MAR2014:18:18:59 30MAR2014:22:53:40 31MAR2014:03:28:38 31MAR2014:08:04:00 31MAR2014:13:32:48 31MAR2014:18:26:10 31MAR2014:23:04:18
Number of Observations
Number of Observations Read
1110
Number of Observations Used
794
Number of Observations Not Used
316
Type 3 Tests of Fixed Effects
Effect
Num DF
Den DF
F Value
Pr > F
TREAT
2
356
9.76
<.0001
CYCLETIME
61
356
27.18
<.0001
BLOCK
4
0
29.90
.
TREAT*CYCLETIME
122
356
1.35
0.0176
BLOCK*TREAT
8
356
18.85
<.0001
BLOCK*CYCLETIME
240
356
1.03
0.3987
However, the degrees of freedom provided in the SAS output have been questioned and I need to be certain of my analysis as it is for publication. To solve the issue, I then went to SPSS to run the analysis there.... and opened another can of worms, since the output of the model is so different, in that the F value for the effect of variable "TREAT" is much much lower in SPSS, the degrees of freedom are also much much lower and hence the p value now non-sig. It's a long time since I used SPSS so I am less inclined to trust my coding in that platform. My question is whether the model I have coded above is a valid way to analyse my data, and whether the degrees of freedom are correct (why they are what they are, so that I can justify my analysis)
... View more