{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Summary"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Throughout this chapter we based our development of the theory behind inference and prediction on the urn model. The urn induced a probability distribution on the estimator, such as the sample mean and the least squares regression coefficients. We end this chapter with some cautions about these statistical procedures. "
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"We saw how the SD of an estimator has a factor of the square root of the sample size in the denominator. When samples are large, the SD can be quite small, and can lead to rejecting a hypothesis or very narrow confidence intervals. When this happens it's good to consider the following.\n",
"\n",
"+ Is the difference that you have detected an important difference? That is, a $p$-value may be quite small, indicating a surprising result, but the actual effect observed may be unimportant. This distinction is coined as *statistical significance* does not imply *practical significance*.\n",
"\n",
"+ Keep in mind that these calculations do not incorporate bias, such as non-response bias, measurement bias. The bias might well be larger than any difference due to chance variation in the sampling distribution."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"At times, we know the sample is not from a chance mechanism, but it can still be useful to carry out a hypothesis test. In this case, the null model would test whether the sample (and estimator) are as if they were at random. When this test is rejected, then we confirm that something non-random has led to the observed data. This can be a useful conclusion--that the difference between what we expect and what we observed is not explained by chance. \n",
"\n",
"At other times, the sample consists of the complete population. When this happens, we might not need to make confidence intervals or hypothesis tests because we have observed all values in the population. That is, inference is not required.\n",
"However, we can, instead, place a different interpretation on hypothesis tests: we can suppose that any relation observed between two features was randomly distributed without relation to one another."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"We have also seen how the bootstrap can be used when we don't have enough information about the population. The bootstrap is a powerful technique, but it does have limitations.\n",
"\n",
"+ Make sure that the original sample is large and random so that the sample resembles the population\n",
"+ Repeat the bootstrap process many times. Typically 10000 replications is a reasonable number\n",
"+ The bootstrap tends to have difficulties when\n",
" + The estimator is influenced by outliers\n",
" + The parameter is based on extreme values of the distribution\n",
" + The sampling distribution of the statistic is far from bell-shaped"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Alternatively, we rely on the sampling distribution being approximately normal in shape. At times, the sampling distribution looks roughly normal but has thicker tails. In these situations, the family of $t$-distributions might be appropriate to use instead of the normal. "
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"A model is usually only an approximation of underlying reality, and the precision of the statement that $\\theta^*$ 0 exactly equals 0 is at odds with this notion of a model. \n",
"The inference depends on the correctness of our model. We can partially check the model assumptions, but some amount of doubt goes with any model. In fact, it often happens that the data suggest more than one possible model, and these models may even be contradictory."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Furthermore, at times the number of hypothesis tests or confidence intervals we are carrying out is quite large. For example, this can occur with multiple linear regression, when we have a large number of features in the model and we separately test whether each coefficient is $0$. This situation can arise when we are trying to select a model from among many possibilities. This is the topic of the next section."
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.9.4"
}
},
"nbformat": 4,
"nbformat_minor": 4
}