14 Hypothesis Testing
Error types:

Type I Error (False Positive):
 Reality: nope
 Diagnosis/Analysis: yes

Type II Error (False Negative):
 Reality: yes
 Diagnosis/Analysis: nope
Power: The probability of rejecting the null hypothesis when it is actually false
Note:
Always written in terms of the population parameter (\(\beta\)) not the estimator/estimate (\(\hat{\beta}\))

Sometimes, different disciplines prefer to use \(\beta\) (i.e., standardized coefficient), or \(\mathbf{b}\) (i.e., unstandardized coefficient)
\(\beta\) and \(\mathbf{b}\) are similar in interpretation; however, \(\beta\) is scale free. Hence, you can see the relative contribution of \(\beta\) to the dependent variable. On the other hand, \(\mathbf{b}\) can be more easily used in policy decisions.
\[ \beta_j = \mathbf{b} \frac{s_{x_j}}{s_y} \]
Assuming the null hypothesis is true, what is the (asymptotic) distribution of the estimator
Twosided
\[ \begin{aligned} &H_0: \beta_j = 0 \\ &H_1: \beta_j \neq 0 \end{aligned} \]
then under the null, the OLS estimator has the following distribution
\[ A1A3a, A5: \sqrt{n} \hat{\beta_j} \sim N(0,Avar(\sqrt{n}\hat{\beta}_j)) \]
 For the onesided test, the null is a set of values, so now you choose the worst case single value that is hardest to prove and derive the distribution under the null
 Onesided
\[ \begin{aligned} &H_0: \beta_j\ge 0 \\ &H_1: \beta_j < 0 \end{aligned} \]
then the hardest null value to prove is \(H_0: \beta_j=0\). Then under this specific null, the OLS estimator has the following asymptotic distribution
\[ A1A3a, A5: \sqrt{n}\hat{\beta_j} \sim N(0,Avar(\sqrt{n}\hat{\beta}_j)) \]
14.1 Types of hypothesis testing
\(H_0 : \theta = \theta_0\)
\(H_1 : \theta \neq \theta_0\)
How far away / extreme \(\theta\) can be if our null hypothesis is true
Assume that our likelihood function for q is \(L(q) = q^{30}(1q)^{70}\) Likelihood function
q = seq(0, 1, length = 100)
L = function(q) {
q ^ 30 * (1  q) ^ 70
}
plot(q,
L(q),
ylab = "L(q)",
xlab = "q",
type = "l")
LogLikelihood function
q = seq(0, 1, length = 100)
l = function(q) {
30 * log(q) + 70 * log(1  q)
}
plot(q,
l(q)  l(0.3),
ylab = "l(q)  l(qhat)",
xlab = "q",
type = "l")
abline(v = 0.2)
Figure from(Fox 1997)
typically, The likelihood ratio test (and Lagrange Multiplier (Score)) performs better with small to moderate sample sizes, but the Wald test only requires one maximization (under the full model).
14.2 Wald test
\[ \begin{aligned} W &= (\hat{\theta}\theta_0)'[cov(\hat{\theta})]^{1}(\hat{\theta}\theta_0) \\ W &\sim \chi_q^2 \end{aligned} \]
where \(cov(\hat{\theta})\) is given by the inverse Fisher Information matrix evaluated at \(\hat{\theta}\) and q is the rank of \(cov(\hat{\theta})\), which is the number of nonredundant parameters in \(\theta\)
Alternatively,
\[ t_W=\frac{(\hat{\theta}\theta_0)^2}{I(\theta_0)^{1}} \sim \chi^2_{(v)} \]
where v is the degree of freedom.
Equivalently,
\[ s_W= \frac{\hat{\theta}\theta_0}{\sqrt{I(\hat{\theta})^{1}}} \sim Z \]
How far away in the distribution your sample estimate is from the hypothesized population parameter.
For a null value, what is the probability you would have obtained a realization “more extreme” or “worse” than the estimate you actually obtained?
Significance Level (\(\alpha\)) and Confidence Level (\(1\alpha\))
 The significance level is the benchmark in which the probability is so low that we would have to reject the null
 The confidence level is the probability that sets the bounds on how far away the realization of the estimator would have to be to reject the null.
Test Statistics
 Standardized (transform) the estimator and null value to a test statistic that always has the same distribution
 Test Statistic for the OLS estimator for a single hypothesis
\[ T = \frac{\sqrt{n}(\hat{\beta}_j\beta_{j0})}{\sqrt{n}SE(\hat{\beta_j})} \sim^a N(0,1) \]
Equivalently,
\[ T = \frac{(\hat{\beta}_j\beta_{j0})}{SE(\hat{\beta_j})} \sim^a N(0,1) \]
the test statistic is another random variable that is a function of the data and null hypothesis.
 T denotes the random variable test statistic
 t denotes the single realization of the test statistic
Evaluating Test Statistic: determine whether or not we reject or fail to reject the null hypothesis at a given significance / confidence level
Three equivalent ways
Critical Value
Pvalue
Confidence Interval
Critical Value
For a given significance level, will determine the critical value \((c)\)
 Onesided: \(H_0: \beta_j \ge \beta_{j0}\)
\[ P(T<cH_0)=\alpha \]
Reject the null if \(t<c\)
 Onesided: \(H_0: \beta_j \le \beta_{j0}\)
\[ P(T>cH_0)=\alpha \]
Reject the null if \(t>c\)
 Twosided: \(H_0: \beta_j \neq \beta_{j0}\)
\[ P(T>cH_0)=\alpha \]
Reject the null if \(t>c\)
 pvalue
Calculate the probability that the test statistic was worse than the realization you have
 Onesided: \(H_0: \beta_j \ge \beta_{j0}\)
\[ \text{pvalue} = P(T<tH_0) \]
 Onesided: \(H_0: \beta_j \le \beta_{j0}\)
\[ \text{pvalue} = P(T>tH_0) \]
 Twosided: \(H_0: \beta_j \neq \beta_{j0}\)
\[ \text{pvalue} = P(T<tH_0) \]
reject the null if pvalue \(< \alpha\)
 Confidence Interval
Using the critical value associated with a null hypothesis and significance level, create an interval
\[ CI(\hat{\beta}_j)_{\alpha} = [\hat{\beta}_j(c \times SE(\hat{\beta}_j)),\hat{\beta}_j+(c \times SE(\hat{\beta}_j))] \]
If the null set lies outside the interval then we reject the null.
 We are not testing whether the true population value is close to the estimate, we are testing that given a field true population value of the parameter, how like it is that we observed this estimate.
 Can be interpreted as we believe with \((1\alpha)\times 100 \%\) probability that the confidence interval captures the true parameter value.
With stronger assumption (A1A6), we could consider Finite Sample Properties
\[ T = \frac{\hat{\beta}_j\beta_{j0}}{SE(\hat{\beta}_j)} \sim T(nk) \]
 This above distributional derivation is strongly dependent on A4 and A5
 T has a student tdistribution because the numerator is normal and the denominator is \(\chi^2\).
 Critical value and pvalues will be calculated from the student tdistribution rather than the standard normal distribution.
 \(n \to \infty\), \(T(nk)\) is asymptotically standard normal.
Rule of thumb
if \(nk>120\): the critical values and pvalues from the tdistribution are (almost) the same as the critical values and pvalues from the standard normal distribution.

if \(nk<120\)
 if (A1A6) hold then the ttest is an exact finite distribution test
 if (A1A3a, A5) hold, because the tdistribution is asymptotically normal, computing the critical values from a tdistribution is still a valid asymptotic test (i.e., not quite the right critical values and p0values, the difference goes away as \(n \to \infty\))
14.2.1 Multiple Hypothesis

test multiple parameters as the same time
 \(H_0: \beta_1 = 0\ \& \ \beta_2 = 0\)
 \(H_0: \beta_1 = 1\ \& \ \beta_2 = 0\)
perform a series of simply hypothesis does not answer the question (joint distribution vs. two marginal distributions).
The test statistic is based on a restriction written in matrix form.
\[ y=\beta_0+x_1\beta_1 + x_2\beta_2 + x_3\beta_3 + \epsilon \]
Null hypothesis is \(H_0: \beta_1 = 0\) & \(\beta_2=0\) can be rewritten as \(H_0: \mathbf{R}\beta \mathbf{q}=0\) where
 \(\mathbf{R}\) is a \(m \times k\) matrix where m is the number of restrictions and \(k\) is the number of parameters. \(\mathbf{q}\) is a \(k \times 1\) vector
 \(\mathbf{R}\) “picks up” the relevant parameters while \(\mathbf{q}\) is a the null value of the parameter
\[ \mathbf{R}= \left( \begin{array}{cccc} 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ \end{array} \right), \mathbf{q} = \left( \begin{array}{c} 0 \\ 0 \\ \end{array} \right) \]
Test Statistic for OLS estimator for a multiple hypothesis
\[ F = \frac{(\mathbf{R\hat{\beta}q})\hat{\Sigma}^{1}(\mathbf{R\hat{\beta}q})}{m} \sim^a F(m,nk) \]

\(\hat{\Sigma}^{1}\) is the estimator for the asymptotic variancecovariance matrix
When \(m = 1\), there is only a single restriction, then the \(F\)statistic is the \(t\)statistic squared.
\(F\) distribution is strictly positive, check FDistribution for more details.
14.2.2 Linear Combination
Testing multiple parameters as the same time
\[ \begin{aligned} H_0&: \beta_1 \beta_2 = 0 \\ H_0&: \beta_1  \beta_2 > 0 \\ H_0&: \beta_1  2\times\beta_2 =0 \end{aligned} \]
Each is a single restriction on a function of the parameters.
Null hypothesis:
\[ H_0: \beta_1 \beta_2 = 0 \]
can be rewritten as
\[ H_0: \mathbf{R}\beta \mathbf{q}=0 \]
where \(\mathbf{R}\)=(0 1 1 0 0) and \(\mathbf{q}=0\)
14.2.3 Estimate Difference in Coefficients
There is no package to estimate for the difference between two coefficients and its CI, but a simple function created by Katherine Zee can be used to calculate this difference. Some modifications might be needed if you don’t use standard lm
model in R.
difftest_lm < function(x1, x2, model) {
diffest <
summary(model)$coef[x1, "Estimate"]  summary(model)$coef[x2, "Estimate"]
vardiff < (summary(model)$coef[x1, "Std. Error"] ^ 2 +
summary(model)$coef[x2, "Std. Error"] ^ 2)  (2 * (vcov(model)[x1, x2]))
# variance of x1 + variance of x2  2*covariance of x1 and x2
diffse < sqrt(vardiff)
tdiff < (diffest) / (diffse)
ptdiff < 2 * (1  pt(abs(tdiff), model$df, lower.tail = T))
upr <
# will usually be very close to 1.96
diffest + qt(.975, df = model$df) * diffse
lwr < diffest + qt(.025, df = model$df) * diffse
df < model$df
return(list(
est = round(diffest, digits = 2),
t = round(tdiff, digits = 2),
p = round(ptdiff, digits = 4),
lwr = round(lwr, digits = 2),
upr = round(upr, digits = 2),
df = df
))
}
14.2.4 Application
library("car")
# Multiple hypothesis
mod.davis < lm(weight ~ repwt, data=Davis)
linearHypothesis(mod.davis, c("(Intercept) = 0", "repwt = 1"),white.adjust = TRUE)
#> Linear hypothesis test
#>
#> Hypothesis:
#> (Intercept) = 0
#> repwt = 1
#>
#> Model 1: restricted model
#> Model 2: weight ~ repwt
#>
#> Note: Coefficient covariance matrix supplied.
#>
#> Res.Df Df F Pr(>F)
#> 1 183
#> 2 181 2 3.3896 0.03588 *
#> 
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
# Linear Combination
mod.duncan < lm(prestige ~ income + education, data=Duncan)
linearHypothesis(mod.duncan, "1*income  1*education = 0")
#> Linear hypothesis test
#>
#> Hypothesis:
#> income  education = 0
#>
#> Model 1: restricted model
#> Model 2: prestige ~ income + education
#>
#> Res.Df RSS Df Sum of Sq F Pr(>F)
#> 1 43 7518.9
#> 2 42 7506.7 1 12.195 0.0682 0.7952
14.2.5 Nonlinear
Suppose that we have q nonlinear functions of the parameters
\[
\mathbf{h}(\theta) = \{ h_1 (\theta), ..., h_q (\theta)\}'
\]
The,n, the Jacobian matrix (\(\mathbf{H}(\theta)\)), of rank q is
\[ \mathbf{H}_{q \times p}(\theta) = \left( \begin{array} {ccc} \frac{\partial h_1(\theta)}{\partial \theta_1} & ... & \frac{\partial h_1(\theta)}{\partial \theta_p} \\ . & . & . \\ \frac{\partial h_q(\theta)}{\partial \theta_1} & ... & \frac{\partial h_q(\theta)}{\partial \theta_p} \end{array} \right) \]
where the null hypothesis \(H_0: \mathbf{h} (\theta) = 0\) can be tested against the 2sided alternative with the Wald statistic
\[ W = \frac{\mathbf{h(\hat{\theta})'\{H(\hat{\theta})[F(\hat{\theta})'F(\hat{\theta})]^{1}H(\hat{\theta})'\}^{1}h(\hat{\theta})}}{s^2q} \sim F_{q,np} \]
14.3 The likelihood ratio test
\[ t_{LR} = 2[l(\hat{\theta})l(\theta_0)] \sim \chi^2_v \]
where v is the degree of freedom.
Compare the height of the loglikelihood of the sample estimate in relation to the height of loglikelihood of the hypothesized population parameter
Alternatively,
This test considers a ratio of two maximizations,
\[ \begin{aligned} L_r &= \text{maximized value of the likelihood under $H_0$ (the reduced model)} \\ L_f &= \text{maximized value of the likelihood under $H_0 \cup H_a$ (the full model)} \end{aligned} \]
Then, the likelihood ratio is:
\[ \Lambda = \frac{L_r}{L_f} \]
which can’t exceed 1 (since \(L_f\) is always at least as large as \(Lr\) because \(L_r\) is the result of a maximization under a restricted set of the parameter values).
The likelihood ratio statistic is:
\[ \begin{aligned} 2ln(\Lambda) &= 2ln(L_r/L_f) = 2(l_r  l_f) \\ \lim_{n \to \infty}(2ln(\Lambda)) &\sim \chi^2_v \end{aligned} \]
where \(v\) is the number of parameters in the full model minus the number of parameters in the reduced model.
If \(L_r\) is much smaller than \(L_f\) (the likelihood ratio exceeds \(\chi_{\alpha,v}^2\)), then we reject he reduced model and accept the full model at \(\alpha \times 100 \%\) significance level