CODEWITHSUNDEEP
rregression
Eshmel
Eshmel

Reputation: 49

number of items to replace is not a multiple of replacement length in weighting

I am using multinomial regression to get the probability of belonging to four sub-groups for 500,000 regions. The data.frame looks like this:

Regions        groupadmit        mid-pop
1                 2               1764
2                 3               1254
25                1               1452
674               4               2665
3001              2               1097
56                3               9864
98                1               2675
500,000          ....             .....

I wrote the following code:

library (nnet)
mlogit<- multinom(groupadmit~mid_pop, data = admissionLSOA1)
probs <- predict(mlogit, type="probs")

The codes work fine till this point, giving the probability of belonging to each group (1, 2, 3, 4) for each observation (region).

Probabilities:

Regions     groupadmit1       groupadmit2     groupadmit3    groupadmit4         
52            0.2484091        0.2494408       0.2505393      0.2516109
97            0.2483949        0.2494358       0.2505441      0.2516252
1300          0.2483253        0.2494112       0.2505676      0.251695
287           0.2483623        0.2494242       0.2505551      0.2516584
500,000         ....            .....            ....          ....

But, when I go to weight the sample (regions) according to their probability, it brings back the following error:

Warning message:
In wts[groupadmit == 1] <- probs[groupadmit == 1, 1]/probs[groupadmit ==  :
number of items to replace is not a multiple of replacement length

What I am doing is weighting the regions according to their probability of belonging to each groupadmit proportional to the probability of belonging to groupadmit one in order to balance any chance for selection bias. It is very similar to inverse probability weighting. The codes are:

wts[groupadmit==1] <- probs[groupadmit==1,1]/probs[groupadmit==1,1]
wts[groupadmit==2] <- probs[groupadmit==2,1]/probs[groupadmit==2,2]
wts[groupadmit==3] <- probs[groupadmit==3,1]/probs[groupadmit==3,3]
wts[groupadmit==4] <- probs[groupadmit==4,1]/probs[groupadmit==4,4]

But, the above error comes up whenever I do the the analysis. May someone please help me to understand why I get this error and how can I solve it?

Many thanks in advance

Upvotes: 0

Views: 174

Answers (1)

Sal-laS
Sal-laS

Reputation: 11649

Why R complains?

Warning message:
In wts[groupadmit == 1] <- probs[groupadmit == 1, 1]/probs[groupadmit ==  :
number of items to replace is not a multiple of replacement length

it means that, the right handside of you assign (<-) is bigger than, what you have on the left handside which is wts[groupadmit==1]

Therefore, i suggest you to do:

length(probs[groupadmit==1,1]/probs[groupadmit==1,1])

and then

length(wts[groupadmit==1])

Then i suppose, it shows the lefthand side is smaller.

Then simply run 

wts[groupadmit==1] <- probs[groupadmit==1,1]/probs[groupadmit==1,1]

and finnally print

wts[groupadmit==1]

Solution:

A quick fix is to use rbind to build your wts:

wts<-rbind(probs[groupadmit==1,1]/probs[groupadmit==1,1], 
probs[groupadmit==2,1]/probs[groupadmit==2,2],
probs[groupadmit==3,1]/probs[groupadmit==3,3],
probs[groupadmit==4,1]/probs[groupadmit==4,4])

Upvotes: 0

Related Questions