Stute Test
Let $Y$ and $D$ be two random variables. Let $m(D) = E[Y|D]$. The null hypothesis of the test is that $m(D) = \alpha_0 + \alpha_1 D$ for two real numbers $\alpha_0$ and $\alpha_1$. This means that, under the null, $m(.)$ is linear in $D$. This hypothesis can be tested in a sample with $N$ i.i.d. realizations of $(Y, D)$ using the following test statistic from Stute (1997):
$$
S = \dfrac{1}{N^2}\sum_{i = 1}^N \left(\sum_{j = 1}^i \varepsilon_{(j)} \right)^2
$$
where $\varepsilon_{(j)}$ is the residual from a linear regression of $Y$ on $D$ and a constant of the $j$-th observation after sorting by $D$. In other words, $S$ is obtained by sorting the data from the smallest to the largest value of $D$ and summing the squares of the total cumulative sums of the linear regression residuals.
Stute et al. (1998) show that, under the null, $S$ is finite. Conversely, under the alternative, at least one of the inner sums tends to infinity, hence $S$ diverges. Inference is performed using wild bootstrap. Specifically, $S$ is re-computed replacing $Y$ with $Y^*$, i.e. the predicted value of $Y$ from the linear regression of $Y$ on $D$ and a constant, plus the residuals multiplied by a two-point random variable, denoted as $V_{(j)}$, such that:
$$
P\left(V_{(j)} = \frac{1 + \sqrt{5}}{2}\right) = \frac{\sqrt{5}-1}{2}, P\left(V_{(j)} = \frac{1 - \sqrt{5}}{2}\right) = \frac{3 - \sqrt{5}}{2}.
$$
Denote with $S^*_b$ the $S^*$ statistic computed at the $b$-th bootstrap replication. The p-value from $B$ bootstrap replications is computed as
$$
\frac{1}{B}\sum_{b = 1}^B \mathbb{1}\{S < S^*_b\}
$$
Intuitively, under the alternative, the p-value should be zero, due to the fact that S diverges.
This test also works with panel data. In that case, the $S$ statistic is computed for each value of the time variable. Moreover, $V_{(j)}$ remains constant at the group level across the computation of the period-specific test statistics. Hence, the residual of group $g$ from a linear regression of $Y_{g,t}$ on $D_{g,t}$ and a constant are multiplied by $V_g$, regardless of $t$. Lastly, the individual test results can be summed into a joint test statistic. In this case, inference is performed using the distribution of the sum of the bootstrap statistics. Denote with $S_\ell$ the period-$\ell$ test statistic and with $S^*_{\ell,b}$ its $b$-th bootstrap estimate. In a dataset with $L$ periods, the p-value of the joint test is computed as follows:
$$
\frac{1}{B}\sum_{b = 1}^B \mathbb{1}\left\{\sum_{\ell = 1}^L S_\ell < \sum_{\ell = 1}^L S^*_{\ell, b}\right\}.
$$
References
Stute, W. (1997). Nonparametric model checks for regression. The Annals of Statistics, 613–641.
Stute, W., W. G. Manteiga, and M. P. Quindimil (1998). Bootstrap approximations in model checks for regression. Journal of the American Statistical Association 93 (441), 141–149.