One-sample \(t\)-test (\(n\) individuals) | One-factor ANOVA (\(m\) groups of \(n\) individuals) | |
---|---|---|
Null hypothesis | The mean \(\bar{x}\) equals \(\mu\) | Same mean in each of the \(m\) groups |
Statistic | \(t = (\bar{x} - \mu) / (s/\sqrt{n})\) | \(F = MSA/MSE\) |
Distribution | \(t\) with \(n-1\) degrees of freedom | \(F\) with \(m(n-1)\) and \((m - 1)\) degrees of freedom |
The significance level is the probability of rejecting the null hypothesis when it is true.
You must choose a test and a significance level before analyzing the results.
If several tests are performed in an experiment, the probability of mistakenly rejecting one of the null hypotheses increases (multiple comparison problem).
The power of a test is the probability of rejecting the null hypothesis if it is false. The smaller the effect to be detected compared to the variance of the response (low signal-to-noise ratio), the higher \(n\) must be to have the same power.
With a sufficiently large \(n\), even a very small effect will be judged statistically significant; it does not mean that the effect is important.
Bias: systematic difference between the estimate of a parameter and its exact value.
Standard error: standard deviation of the estimate of a parameter due to limited sampling; decreases when \(n\) increases.
Confidence interval: In X% of the possible samples, the X% confidence interval of a parameter estimate contains the true value of this parameter.
Relationship between confidence interval and hypothesis testing: the hypothesis \(\theta = \theta_0\) can be rejected at a threshold \(\alpha\) if the \(100 \% (1 - \ alpha)\) confidence interval of \(\hat{\theta}\) does not include \(\theta_0\).
ANOVA tolerates moderate deviations from normality, so that assumption is less critical than the other two.
With only 2 groups, the \(t\)-test allows for unequal variances.
Models in gray not seen in this course.
Model names in italics not seen in this course.
\[ y \sim w + x + z \]
\(y\) is the numerical response, \(w\) and \(x\) are numerical predictors, \(z\) is a factor with treatment coding (default in R) and three levels: A (reference), B, and C.
Estimated coefficients:
\[ y \sim x * z \]
Estimated coefficients:
In other words, to get the slope of \(y\) vs. \(x\) if \(z\) = B, we must add the coefficients x and x:zB.
\[ y \sim w * x \]
Estimated coefficients:
In these models, the mean of \(y\) is not equal to the linear combination of predictors \(\eta\), but to a transformation of \(\eta\) according to a link function.
The interpretation of the parameters above indicates the effect on \(\eta\). To obtain the effect on the mean of \(y\), you must apply the inverse of the link function.
\[ x_{norm} = \frac{x - \mu_x}{\sigma_x} \] - Since \(x_{norm} = 0\) corresponds to the mean of \(x\), it is easier to interpret the intercept in all cases, and the coefficients in the case of a model with interactions.
Choice between models of different complexities: compromise between underfitting and overfitting.
Underfitting: important effects not included in the model.
Overfitting: the model reproduces very well the data used to fit it, but performs worse on new data.
In the absence of independent data to evaluate the predictive power of different models, we can estimate it with the AIC (and its variants).
Compare models based on the same response variable and the same observations.
The best model may not be good: check the fit.
If several models are plausible, the weighted average of their predictions is often better than the predictions of the best model.
Problem where two or more predictors are strongly correlated.
Different options: