I will write only about those items to which less than 50% of the group answered correctly. And I will try to generalize my feedback, so it won't be just answers to exact questions.

I will write only about those items to which less than 50% of the group answered correctly. And I will try to generalize my feedback, so it won't be just answers to exact questions.

1. Significance level (P) is the probability of type I error. Type I error is about rejecting H0 hypothesis, when it is actually true. So if P=.05, then you have 5% chance that you are wrong when you reject H0 and conclude that there is an effect or a difference. For full interpretation of the effect size and significance level, you need to present the exact numbers for both (and not just p<.05), or (and) the sample size, which will allow the reader to calculate the p-value from the effect size (or vice versa) if needed.

2. The significance level sometimes cannot be trusted, if, for example, sample size is small. BUT the actual measure of reliability of your significance test is statistical power of your study. Statistical power is about type II error (1 – Beta). It means that if stat.power is high, then the chances that you are accepting H1 correctly are also high. So in studies with high statistical power, rejection of H0 means that probably the effect is truly not there.

3. This is mediation:

This is moderation:

4. Exploratory factor analysis does not allow testing any hypothesis. Confirmatory factor analysis does. And there are no other types of analysis except structural equation modeling that would allow testing hypothesis about the relationships between latent variables. Exploratory factor analysis can only show that there is a correlation between the factors, but you cannot test any hypothesis with it, as you don’t know in advance what will your factors be. The measure of the association between the variable and the factor is called factor loading. And, finally, Oblimin rotation takes into account the correlating factors.

5. Analysis of variance requires normally distributed dependent variable(s), and categorical (=nominal scale) independent variable(s).

6. Students T test does not allow to test for interaction effects between categorical variables. ANOVA does. And if N in the groups are equal, you can still proceed with ANOVA, even if the variances are not homogenous. Though, if the dependent variable is not normally distributed, you need to change to Kruskal-Wallis ANOVA. If results of ANOVA are significant (p-value <.05), it normally means that the groups (IV) do not come from the same population.

7. To compare the frequency of appearance of something (specific value) in your sample with the pre-known frequency, you use binomial test. To test the association (the effect) of some variable (usually categorical) with the other variable, measured in frequencies, you use Chi-square test.

8. If the SD in two groups is the same, then the effect size id simply (X1-X2)/SD

9. Qual methods usually do not give you a concise description of any phenomena. But if you translate your analysis into numbers, it can give you a possibility to do some statistical analysis.

10. The rule of thumb to estimate the sample size for the regression analysis is the following: to test the whole model: N >= 50 + 8*IV (IV = number of independent variables in the regression); to test individual predictors: N >= 104 + IV

11. If the effect of IV-1 disappears when we control for IV-2, it means that most of the variance explained in dependent variable by IV-1 is shared by IV-2. So IV-2 “eats” all the variance that they can both explain, and there is nothing left for IV-1 to explain anymore J

12. If the mean is higher than the median, it usually means that we observe negative skewness (because there are more values at the beginning of the scale) or presence of outliers (drawing the distribution line here would really help).

Дата добавления: 2015-11-05