choosing and executing an analysis

Hi All,

I am having a bit of trouble in choosing and executing an analysis, as the data is quite simple, but seems a bit too complex for a simple chi-square or a paired t-test. I have 8 observations and 4 variables.
I had some difficulty conduction a chi-square on a subset of the data, so I did it in matrix form, and it seems to work fine:

unmarked_dataset <- matrix(c(29,53,1538,470),2)
chisq.test(unmarked_dataset)

Pearson's Chi-squared test with Yates' continuity correction

data: unmarked_dataset
X-squared = 69.192, df = 1, p-value < 2.2e-16

The result is significant, but I cannot know where that significance is coming from.

With a paired T-test, I did not find a way to insert all the variables that I need, as I need at least three levels, when following this general code.

t.test(unmarked$Occurrences, unmarked$Speaker_type, unmarked$Structure, unmarked$Condition, correct == FALSE)

Something like lme4 on the other hand seems too complicated, as I basically don't know what to count as random effects.

So, considering that my full dataset is the following, what tests do I need to apply to see if, for example, Children use significantly more PreN structures than Adults in the UnMarked conditions?

Thank you!

Screenshot 2020-05-19 11.28.35

Hello MarVel,

Welcome to the forum! So your table provided is not a true 2 x 2 contingency table. As you actually have the interaction here of Speaker_type x Structure x Condition. I think what you're looking for is Log - linear analysis as it is a technique used in statistics to examine the relationship between more than two categorical variables.

This link shows an in depth run in R:
https://data.library.virginia.edu/an-introduction-to-loglinear-models/

Let me know if this was helpful :slight_smile:

Hi Grey,

Thank you for the advice. I will definitely try this as it seems to give a very clear output.
I have tried running an analysis with the lme4 package, which I often used for other datasets, but in the current dataset I do not know which my random effects would be, and thus the model did not seem appropriate. SO maybe that test was too detailed for my current dataset.

Marta

Hi Grey,

The tutorial is great and I am getting appropriate results. I have a question about poisson as part of glm.
It says to add family = poisson to the model to assume that the variables are independent form one another. In my case they are not as UnMarked context should yield more PreN structures and vice versa. So do I use it or not?

I ran models with both and I get more realistic results if I exclude the poisson. Like so:

Call:
glm(formula = Freq ~ Speaker.type + Structure + Context, data = ItAll.df)

Deviance Residuals:
1 2 3 4 5 6 7 8
145.00 -99.25 -9.75 -36.00 -90.25 44.50 -45.00 90.75

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 226.00 81.13 2.786 0.0495 *
Speaker.typeChild -53.25 81.13 -0.656 0.5474
StructurePostN -123.75 81.13 -1.525 0.2019
ContextMarked -124.75 81.13 -1.538 0.1989

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

(Dispersion parameter for gaussian family taken to be 13163.12)

Null deviance: 120077  on 7  degrees of freedom

Residual deviance: 52652 on 4 degrees of freedom
AIC: 103.04

Number of Fisher Scoring iterations: 2

It shows how the differences are not significant, but it is a plausible explanation.

However, when it is ran with family=poisson all of the comparisons are significant at the highest level, which I think is unlikely.

Call:
glm(formula = Freq ~ Speaker.type + Structure + Context, family = poisson,
data = ItAll.df)

Deviance Residuals:
1 2 3 4 5 6 7 8
1.6777 -6.7330 0.0878 -0.6918 -4.2241 7.0066 -3.7895 6.5545

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 5.82782 0.05265 110.686 <2e-16 ***
Speaker.typeChild -0.74096 0.08724 -8.493 <2e-16 ***
StructurePostN -2.33598 0.14385 -16.239 <2e-16 ***
ContextMarked -2.37809 0.14638 -16.246 <2e-16 ***

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

(Dispersion parameter for poisson family taken to be 1)

Null deviance: 1208.64  on 7  degrees of freedom

Residual deviance: 172.89 on 4 degrees of freedom
AIC: 219.72

Number of Fisher Scoring iterations: 6

So, I ahve decided to go with the model without poisson. However, as I went further eith the anlysis and when the degrees of freedom need to be extracted. When I do that with the mod0 (without poisson), the result is 0:

pchisq(deviance(mod0), df = df.residual(mod0), lower.tail = F)
[1] 0

Which makes me have some doubts on the model. Then I tried running it with the model in which I included family=poisson (modPoiss) and there is a more realistic number (I guess):

pchisq(deviance(modPoiss), df = df.residual(modPoiss), lower.tail = F)
[1] 2.50456e-36

Thus, i am really unsure what is the correct way to procede with these data. Do you have any insight on that?

Thank you!
Marta

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.