CODEWITHSUNDEEP
rerror-handlingregressionwarnings
XM_Z
XM_Z

Reputation: 41

Error in prettyNum(.Internal(format(x, trim, digits, nsmall, width, 3L,: invalid 'nsmall' argument

I am doing negative binomial analysis for some count data in the following link:https://www.dropbox.com/s/q7fwqicw3ebvwlg/stackquestion.csv?dl=0

I had some problems (error messages) when I tried to fit all the independent variables into the model, which makes me want to look at each independent variables one by one to find out which variable caused the problem. Here is what I found:

For all the other variables, when I fit the variables to the Y which is column A looks normal:

m2 <- glm.nb(A~K, data=d)
summary(m2)

Call:
glm.nb(formula = A ~ K, data = d, init.theta = 0.5569971932, 
    link = log)

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-2.5070  -1.2538  -0.4360   0.1796   1.9588  

Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept) -0.66185    0.84980  -0.779    0.436    
K            0.25628    0.03016   8.498   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for Negative Binomial(0.557) family taken to be 1)

    Null deviance: 113.202  on 56  degrees of freedom
Residual deviance:  70.092  on 55  degrees of freedom
AIC: 834.86

Number of Fisher Scoring iterations: 1


              Theta:  0.5570 
          Std. Err.:  0.0923 

 2 x log-likelihood:  -828.8570

However, I found this variable L, when I fit L to the Y, I got this:

m1 <- glm.nb(A~L, data=d)
There were 50 or more warnings (use warnings() to see the first 50)
summary(m1)

Call:
glm.nb(formula = A ~ L, data = d, init.theta = 5136324.722, link = log)

Deviance Residuals: 
   Min      1Q  Median      3Q     Max  
-67.19  -18.93  -12.07   13.25   64.00  

Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept)  3.45341    0.01796   192.3   <2e-16 ***
L            0.24254    0.00103   235.5   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for Negative Binomial(5136325) family taken to be 1)

    Null deviance: 97084  on 56  degrees of freedom
Residual deviance: 28529  on 55  degrees of freedom
AIC: 28941

Number of Fisher Scoring iterations: 1

Error in prettyNum(.Internal(format(x, trim, digits, nsmall, width, 3L,  : 
  invalid 'nsmall' argument

You can see that the init.theta and AIC is too large, and there are 50 warning and an error message.

The warning message is this

 In theta.ml(Y, mu, sum(w), w, limit = control$maxit, trace = control$trace >  ... :
  iteration limit reached

Actually, variables M and L are two observations of one thing. I did not find anything abnormal with variable L. For all the data, only column L has this problem.

So I am wondering what exactly does this error message mean: Error in prettyNum(.Internal(format(x, trim, digits, nsmall, width, 3L,: invalid 'nsmall' argument. Since I just observed these data, how should I fix this error? Thank you!

Upvotes: 0

Views: 5473

Answers (1)

andrew_reece
andrew_reece

Reputation: 21274

The important message is in the warnings(): when L is the independent variable, the default number of iterations in the GLM convergence procedure is not high enough to converge on a model fit.

If you manually set the maxit parameter to a higher value, you can fit A ~ L without error:

glm.nb(A ~ L, data = d, control = glm.control(maxit = 500))

See the glm.control documentation for more. Note that you can also set a reasonable value for init.theta - and this will prevent both theta and AIC from fitting to unreasonable values:

m1 <- glm.nb(A ~ L, data = df, control = glm.control(maxit = 500), init.theta = 1.0)

Output:

    Call:
glm.nb(formula = A ~ L, data = df, control = glm.control(maxit = 500), 
    init.theta = 0.8016681349, link = log)

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-3.3020  -0.9347  -0.3578   0.1435   2.5420  

Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept)  1.25962    0.40094   3.142  0.00168 ** 
L            0.38823    0.02994  12.967  < 2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for Negative Binomial(0.8017) family taken to be 1)

    Null deviance: 160.693  on 56  degrees of freedom
Residual deviance:  67.976  on 55  degrees of freedom
AIC: 809.41

Number of Fisher Scoring iterations: 1


              Theta:  0.802 
          Std. Err.:  0.140 

 2 x log-likelihood:  -803.405

Upvotes: 1

Related Questions