1
$\begingroup$

I apologize if this question will seem odd, but I am quite new to statistical analysis. I performed coinfection experiments with a total of 12 conditions, and I have 3 measurements per condition. I did a Shapiro-Wilk test to check for normality: the values of some conditions are normally distributed, while others are not. I know that in case of normal distribution I can perform a one-way ANOVA test, while if my data is not normally distributed I have to do a Kruskal-Wallis test. However, I am not sure how to proceed since my data are "both" normally and non-normally distributed. How do I proceed from this point? (I attached the results from the normality test) Thank you to anyone who will answer, I appreciate the help very much.

Kind regards, Elena

Shapiro-Wilk test, results Shapiro-Wilk test, results

$\endgroup$
1
  • 1
    $\begingroup$ It is the residuals from the ANOVA model that need to be normally distributed, not the original data for each level of the factor. // Doing Shapiro-Wilk tests on a dozen samples of size three, it would not be surprising to get one P-value below 0.05. $\endgroup$
    – BruceET
    May 10 at 15:50

1 Answer 1

1
$\begingroup$

Using today's date as seed for a simulation, I happened to get one sample (of size three) out of 12 that "failed" the Shapiro-Wilk normality test at the 5% level (specifically, the 7th).

set.seed(510)
pv = replicate(12, shapiro.test(rnorm(3))$p.val)
pv
 [1] 0.11687631 0.61252325 0.88430727 0.85727925 0.18020280 0.70942521
 [7] 0.04484017 0.50968817 0.51974599 0.35479127 0.91914364 0.53469289

Using these same 36 observations to make a fictitious dataset for a one-way ANOVA (in this case with no effects), we have the following:

set.seed(510)
x = rnorm(3 * 12)
g = as.factor(rep(1:12, each=3))
anova(lm(x~g))

Analysis of Variance Table

Response: x
          Df Sum Sq Mean Sq F value Pr(>F)
g         11  9.878 0.89798  0.6026 0.8082
Residuals 24 35.766 1.49025      

The 36 residuals can be obtained, displayed, and tested for normality as follows:

r = resid(lm(x~g))
stripchart(r, pch="|")

enter image description here

shapiro.test(r)

        Shapiro-Wilk normality test

data:  r
W = 0.97451, p-value = 0.5606

Moreover, a normal quantile-quantile plot of the residuals is nearly linear, consistent with normality.

qqnorm(r);  qqline(r, col="blue", lwd=2)

enter image description here

For this particular example, we have one level of the factor that fails the Shapiro-Wilk normality test, while the test on all $36$ residuals shows no departure from normality.

I'm not saying that an ANOVA is OK for your data; only saying that you have not yet correctly assessed normality of your data for doing a one-factor ANOVA.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.