[R-sig-ME] Intervention analysis with mixed models?

[Edgar Torres]

Dear all,


I wouldlike to know if it is possible to perform an intervention analysis with linearmixed models. I have three response variable (two are counts, and one is continuous).Each dependent variable was measure for fishing vessels (23 vessels) over time(I can estimate them either monthly or annually). The time series is from 1992to 2008, in 2002 there was a fishing restriction so I would like to know ifthis restriction impacted my dependent variables. Because I do not have acontrol group of vessels I cannot carry out a BACI design, so I just canevaluate the Before-After effect. Moreover, I would like to know if there is aneffect from sea surface temperatures (sst) anomalies in my dependent variables.An example of my data is as follows:


          Year    Boat   BA   LMAMb       MAMb  ANOM1_2      ANOM3  ANOM4    Time

1        1992      AP     B               25         579.8             0.59           0.44          0.49            1

2        1992      AW    B               27         638.0             0.59           0.44          0.49            1

3        1992      AY     B               25        229.7             0.59           0.44         0.49            1

325    2008      QQ     A               24         496.0             0.35          -0.26        -0.89           17
326    2008     WB     A              11         135.6             0.35          -0.26        -0.89           17

327    2008      XY     A              33         501.3             0.35          -0.26        -0.89           17

 I found a similarpost trying to answer this type of analysis (https://stat.ethz.ch/pipermail/r-sig-mixed-models/2010q3/004399.html).In the post the suggestion is:


mod <-lmer(Attack ~ Group*Year + Group*BA + (1|Subject), data=dat, family=poisson)


"This testswhether the difference in Attack before and after differs by Group, and whetherthe trajectory of Attack over time (Year) differs by Group. Note that year isnumeric and re-centered at 1994 (the year the intervention started)."



Followingthe example I formulated my model as follows (without a “Group” effect):


mam.glmm1 <- glmer(LMAMb ~ BA +Time + ANOM3 + (1 | Vessel), data=o, family=poisson(), na.action=na.omit,verbose=T)


I performthree additional models, modifying ANOM3, which refers to sst anomalies, then Icarry out an anova() to selecta model, and I get the following ANOVA table:


                          Df          AIC          BIC       logLik    deviance        Chisq   Chi Df     Pr(>Chisq)    

lmam.glmm1   5    2598.9    2617.9    -1294.5        2588.9                             

lmam.glmm2    5    2601.1    2620.0    -1295.5        2591.1     0.0000             0                     1   

lmam.glmm3    5    2595.7    2614.6    -1292.8        2585.7     5.4112            0           <2e-16 ***

lmam.glmm4    5    2601.2    2620.2    -1295.6        2591.2     0.0000            0                     1 



>From these resultsI select mam.glmm3 model and I perform an Anova() in this model to test forfixed effects, looking for the effect of BA:

                            Chisq    Df          Pr(>Chisq)  
BA                 10.3832      1              0.001272 **

Time                4.9562      1              0.025997 * 

ANOM3_4       9.1747      1              0.002454 **



>From these resultsI was wondering:

1)      If it is enough to estimate theBefore-After effect on response variables? 

2)      Is there a correlation between Time (continuous)and BA (factor) effects? Is it necessary to make an interaction between theseeffects?

3)      Am I missing something, is it normalto have p-values of 1 results and Chi Df of 0 from model comparison withanova()? 

4)      Can I conclude that effectivelythere was an effect Before-After on my dependent variable?

Thanks inadvance, I am beginning in the use of mixed models.
EDGAR
	[[alternative HTML version deleted]]