0
$\begingroup$

All my models for my 7 response variables are not significant but upon checking the VIF, the independent variables do not show multicollinearity. I have also checked the plot residual plot of my models but it shows that the residuals follows a horizontal line. Should I still go with the models? Or should I transform my data (as some researchers have suggested). If you recommend to transform my data, the how? What functions should I use?

           Df Deviance Resid. Df    Resid. Dev Pr(>Chi)  
NULL                         44     18.800           
Treatment  2  2.13333        42     16.667   0.0765 .
Site       1  0.02977        41     16.637   0.7888  
Pole       1  0.12078        40     16.516   0.5895  
Clump      1  0.74395        39     15.772   0.1806  
Internode  1  0.00316        38     15.769   0.9305  
---

Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

> summ(Model1.SN, vifs = TRUE)
MODEL INFO:
Observations: 45
Dependent Variable: No.Shoots
Type: Linear regression 

MODEL FIT:
χ²(6) = 3.03, p = 0.29
Pseudo-R² (Cragg-Uhler) = 0.19
Pseudo-R² (McFadden) = 0.09
AIC = 96.52, BIC = 110.97 

Standard errors: MLE
------------------------------------------------------------
                          Est.   S.E.   t val.      p    VIF
---------------------- ------- ------ -------- ------ ------
(Intercept)               0.19   0.52     0.36   0.72       
Treatment200             -0.26   0.25    -1.02   0.32   1.77
TreatmentControl         -0.64   0.29    -2.21   0.03   1.77
Site                      0.03   0.05     0.53   0.60   1.49
Pole                      0.02   0.02     0.94   0.35   1.33
Clump                     0.16   0.12     1.34   0.19   1.17
Internode                -0.01   0.12    -0.09   0.93   1.18
------------------------------------------------------------

Call:
lm(formula = No.Shoots ~ Treatment + Site + Pole + Clump + Internode, 
    data = zdmc_branch, family = gaussian)

Residuals:
    Min      1Q  Median      3Q     Max 
-0.9288 -0.3569 -0.1244  0.3040  2.2411 

Coefficients:
                 Estimate Std. Error t value Pr(>|t|)  
(Intercept)       0.18832    0.51765   0.364   0.7180  
Treatment200     -0.25615    0.25211  -1.016   0.3161  
TreatmentControl -0.63823    0.28937  -2.206   0.0335 *
Site              0.02902    0.05487   0.529   0.5999  
Pole              0.02225    0.02356   0.944   0.3509  
Clump             0.15938    0.11880   1.342   0.1877  
Internode        -0.01058    0.12133  -0.087   0.9310  
---

Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.6442 on 38 degrees of freedom
Multiple R-squared:  0.1612,    Adjusted R-squared:  0.02878 
F-statistic: 1.217 on 6 and 38 DF,  p-value: 0.3189
$\endgroup$
6
  • 3
    $\begingroup$ Welcome to Cross Validated! This question is very hard to read. Try using the code {} tool on the editing toolbar to include your models and summaries in a way that's easier to read. Also, it would help if you could edit the question to say more about the nature of your data and the hypotheses you want to test. We often find that there are better ways to analyze data than a questioner is aware of. For example, if you have a count value as an outcome, you probably should use a model other than simple linear regression. $\endgroup$
    – EdM
    Apr 2 at 18:10
  • 5
    $\begingroup$ It sounds as if you are under pressure to produce certain results and want to engineer your analyses so they can give you those types of results. Is that a fair assessment? $\endgroup$
    – rolando2
    Apr 2 at 18:21
  • 4
    $\begingroup$ I formatted your output for you. I agree with @EdM about probably not using linear reg and about his request for you to give details. I do note that it looks like N = 44, which would mean that your model could be overfit. But .... after all that, you may have to conclude that your theory was wrong. Not only are your p values not < 0.05, but your $R^2$ is pretty low. But add details so we can suggest things. $\endgroup$
    – Peter Flom
    Apr 2 at 18:39
  • 1
    $\begingroup$ Are site, pole, clump and internode all numerical values(1, 2, 3)? They all have 1 degree of freedom in the model output. I'd expect that these would be categories and have multiple levels (site 1, site 2, site 3) your dependent variable no. shoots does sound like a count variable and you may need something like poison or negative binomial regression. Log+1 transformation appears in older biology Stats books but had many downsides. $\endgroup$
    – N Brouwer
    Apr 2 at 23:55
  • $\begingroup$ I would just like to confirm that All explanatory variables are numerical values. However the dependent variable is continuous although some just produced 0 shoots. Thank you so much for all your comments. Really appreciate it. $\endgroup$
    – user409313
    Apr 4 at 3:17

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.