<!-- markdown css tag --><div class="pinggu_markdown">
<div class="pinggu_markdown__html"><h1 id="stata绘图（二）--多期did的平行趋势检验">Stata绘图（二） | 多期DID的平行趋势检验</h1>
<blockquote>
<p>作者：石器时代的大菠萝</p>
</blockquote>
<p>多期DID的平行趋势检验有两种等价的展示方法，一是回归法，二是绘图法，前者相对容易，而后者的操作过程稍复杂。不少人借鉴了Beck et al.（2&#48;1&#48;）的做法，但该文实际用的模型是渐进DID，即最终所有个体均实施了政策，因此相关命令需要经过一定修改才能用于一般的多期DID。本文的前半部分将会简单介绍多期DID，后半部分将会参考Beck et al.（2&#48;1&#48;）对图片的设定提供一个绘图过程。</p>
<h2 id="多期did简介">多期DID简介</h2>
<p>两期DID：Yit=常数+Dt+Gi+ Dt* Gi+ eit<br>
——Stata命令：reg yDG D_G 其他控制变量<br>
注：D_G是D与G的交互项</p>
<p>多期DID：Yit=常数+Xit+时间虚拟变量+ ui+ eit<br>
——Stata命令：xtreg y x i.time 其他控制变量,fe</p>
<p>其中，Yit是被解释变量；Dt表示政策后虚拟变量（取1表示政策之后，取&#48;则表示政策之前）；Gi表示处理变量（取1表示处理组，取&#48;则表示控制组）；Xit表示“个体i为处理组”且“时间t在政策之后”则取值为1，其他情况取值为&#48;，有一种等价说法是——Xit表示个体i在t时间是否实施了政策。<br>
请注意：不要把Xit理解成是交互项，因为在多期DID中，控制组样本的Dt无法给出合适的定义。简单来说，控制组样本根本不存在“政策年度”一说，更谈不上样本是发生在政策“之前”还是“之后”了。</p>
<p>Dt表示政策实施前后的虚拟变量，在两期DID中，因为只存在两期数据，因此其等价于时间虚拟变量。两期DID的Dt到了多期DID中，就转化为时间虚拟变量；Gi表示处理变量，由于多期DID中的个体效应ui包含了Gi的信息（Gi是ui的子集），因此同时在模型中放入Gi与ui将导致严重的多重共线性问题，应该只放入信息含量更多的ui。多期DID中的Xit来自两期DID中的Dt* Gi，尽管本文一再强调不应该把Xit理解成Dt与Gi的乘积，但是不少初学者依然会习惯性认为Xit等价于两个变量的乘积。</p>
<p>综上，两期DID推导至多期DID的变化过程是：Dt→时间虚拟变量，Gi→ui，Dt*Gi→Xit。<strong>多期DID没有对政策时点是否一致性提出要求，因此多期DID还适用于政策时点不一致情形。</strong></p>
<p>生成Xit的Stata过程对初学者来说可能稍有难度，有的人习惯在Excel中整理数据，有的人喜欢用merge命令把数据全部匹配到一起。由于不同人有不同的习惯，下面介绍如何在Excel和Stata中应该怎么制作出这个变量。</p>
<p>第一种做法：使用Excel来制作Xit是非常直观的。</p>
<p>第一步，你需要为数据添加一列policy_year，对于控制组样本应该设定为空白值，X就是最终要生成的变量。例子中一共有16个样本。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/21/QiRDOyIqglrUu1c.png" alt="1.png" width="5&#48;&#48;" height="35&#48;"></p>
<p>第二步，从16个样本中，筛选出policy_year为空白值的样本，然后对这些样本的X全部赋值为&#48;。满足条件的一共有8个。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/21/TPUAcwX2lsbMIa6.png" alt="3.png" width="5&#48;&#48;" height="2&#48;&#48;"></p>
<p>第三步，从16个样本中，筛选出policy_year有取值的样本（一共8个样本），新变量dyear是用year减去policy_year。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/KyfwI1iLn6YRMWU.png" alt="4.png" width="5&#48;&#48;" height="19&#48;"></p>
<p>第四步，从16个样本中，筛选出dyear≥&#48;的样本（一共5个样本），然后对这些样本的X赋值为1。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/21/gvLpb6TEKDYxwc9.png" alt="5.png" width="5&#48;&#48;" height="3&#48;&#48;"><br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/Qf6HvXYKIq5gLSx.png" alt="6.png" width="5&#48;&#48;" height="135"></p>
<p>第五步，取消筛选功能后，你可以在X中发现仍有空白值（3个），用&#48;填充他们，最后再把dyear删除，X就生成完毕了。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/7anp3BRouOsbiJ9.png" alt="7.png" width="5&#48;&#48;" height="34&#48;">)<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/DIRcotMXnBy8PzT.png" alt="8.png"></p>
<p>第二种做法：使用Stata来生成x，过程相对简单，如果不想一步步操作Excel可以考虑这种做法。</p>
<p>第一步，数据导入Stata。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/HwrLgEjGY8fMayu.png" alt="9.png"></p>
<p>第二步，输入命令：</p>
<pre class=" language-python"><code class="prismlanguage-python">gen x<span class="token operator">=</span><span class="token number">&#48;</span>
replace x<span class="token operator">=</span><span class="token number">1</span> <span class="token keyword">if</span> year<span class="token operator">&gt;=</span> policy_year
<span class="token operator">*</span>平行趋势检验还需要生成处理变量treat（处理组取<span class="token number">1</span>，控制组取<span class="token number">&#48;</span>）。
<span class="token operator">*</span>这个变量在Excel中非常容易生成，因此方法一不详细介绍生成过程。
gen treat<span class="token operator">=</span><span class="token number">&#48;</span>
replace treat<span class="token operator">=</span><span class="token number">1</span> <span class="token keyword">if</span> policy_year<span class="token operator">!=</span><span class="token punctuation">.</span>
</code></pre>
<h2 id="多期did的stata命令">多期DID的Stata命令</h2>
<pre class=" language-python"><code class="prismlanguage-python">xtreg y x i<span class="token punctuation">.</span>time 其他控制变量<span class="token punctuation">,</span>fe r
</code></pre>
<p>提示：推荐使用聚类稳健标准误进行回归，也就是加上“r”，但是这么做或许会降低系数的显著性。为什么会这样呢？这个问题与t检验的自由度有关，本帖不展开讨论这个技术细节。如果你在多期DID的回归使用了聚类稳健标准误，那么在平行趋势检验中，你应该继续使用聚类稳健标准误，从而做到前后一致。<mark>请不要低估考虑“是否使用聚类稳健标准误”的意义，它不但影响系数的显著性，还可能影响绘图策略。</mark></p>
<h2 id="平行趋势检验的stata命令——回归法">平行趋势检验的Stata命令——回归法</h2>
<pre class=" language-python"><code class="prismlanguage-python"><span class="token operator">*</span>变量说明：y表示被解释变量，<span class="token builtin">id</span>表示样本个体；year表示样本年份；policy_year表示政策发生年份；
<span class="token operator">*</span>treat取<span class="token number">1</span>表示处理组，取<span class="token number">&#48;</span>表示控制组。

<span class="token builtin">set</span> more off
xtset <span class="token builtin">id</span> year
gen distance <span class="token operator">=</span> year <span class="token operator">-</span> policy_year

<span class="token operator">*</span>了解数据情况。
tab distance<span class="token punctuation">,</span> missing
<span class="token operator">*</span>请确认distance变量是否存在以下两类问题：
<span class="token operator">*</span><span class="token number">1</span><span class="token punctuation">.</span>样本稀疏的问题，即样本个数在某些年份非常少。<span class="token number">2.</span>distance的取值范围太宽，检验太多期的平行趋势可能是没有必要的。
<span class="token operator">*</span>你可以采用“缩尾处理策略”以应对上面两种问题：
<span class="token operator">*</span>replace distance <span class="token operator">=</span> <span class="token operator">-</span><span class="token number">4</span> <span class="token keyword">if</span> distance <span class="token operator">&lt;</span> <span class="token operator">-</span><span class="token number">4</span>
<span class="token operator">*</span>replace distance <span class="token operator">=</span> <span class="token number">5</span> <span class="token keyword">if</span> distance<span class="token operator">&gt;</span> <span class="token number">5</span>

<span class="token operator">*</span>生成一系列的变量：
<span class="token operator">*</span>d_j的数学含义是：若样本是”处理组“且为”政策实施前的第j期“则取值为<span class="token number">1</span>，其他情况取值为<span class="token number">&#48;</span>。
<span class="token operator">*</span>dj的数学含义是：若样本是”处理组“且为”政策实施后的第j期“则取值为<span class="token number">1</span>，其他情况取值为<span class="token number">&#48;</span>。
<span class="token operator">*</span>current的数学含义是：若样本是”处理组“且为”政策实施当期“则取值为<span class="token number">1</span>，其他情况取值为<span class="token number">&#48;</span>。
<span class="token operator">*</span>例如，某个个体的政策实施于<span class="token number">2&#48;13</span>年，那么该个体在<span class="token number">2&#48;12</span>年的变量D_1取值为<span class="token number">1</span>，其余均为<span class="token number">&#48;</span>。
</code></pre>
<p>尽管上面给出的数学定义是十分清晰的，但为了照顾初学者，下面给出一个直观的数据描述。以d_1为例，若样本是”处理组“且为”政策实施前的第1期“（distance=-1）则取值为1（橘色区域所示），其余情况取值为&#48;。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;6/&#48;1/MDj9dAh67Cz2w8v.jpg" alt="11.jpg"></p>
<pre class=" language-python"><code class="prismlanguage-python"><span class="token operator">*</span>第一步，生成变量d_j、dj、current。

<span class="token operator">*</span>（<span class="token number">1</span>）生成d_j，假设你在“tab distance<span class="token punctuation">,</span> missing”中发现，distance最小值是<span class="token operator">-</span><span class="token number">4</span>，那么生成过程如下：
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">4</span> <span class="token punctuation">{</span>
gen d_`i'<span class="token operator">=</span> <span class="token number">&#48;</span>
replace d_`i<span class="token string">'= 1 if treat== 1 &amp; distance== -`i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>（<span class="token number">2</span>）生成dj，假设你在“tab distance<span class="token punctuation">,</span> missing”中发现，distance最大值是<span class="token number">5</span>，那么生成过程如下：
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">5</span> <span class="token punctuation">{</span>
gen d`i'<span class="token operator">=</span> <span class="token number">&#48;</span>
replace d`i<span class="token string">'= 1 if treat== 1 &amp; distance== `i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>（<span class="token number">3</span>）生成current。
gen current<span class="token operator">=</span> <span class="token number">&#48;</span>
replace current <span class="token operator">=</span> <span class="token number">1</span> <span class="token keyword">if</span> treat<span class="token operator">==</span> <span class="token number">1</span> <span class="token operator">&amp;</span> distance<span class="token operator">==</span> <span class="token number">&#48;</span>

<span class="token operator">*</span>回归法进行平行趋势检验：
xtreg y d_4 d_3 d_2 d_1 d1 d2 d3 d4 d5 i<span class="token punctuation">.</span>year 控制变量<span class="token punctuation">,</span> fe r

<span class="token operator">*</span>判别方法：若d_4 d_3 d_2 d_1均不显著，则表明平行趋势假设成立。
</code></pre>
<p>你可能已经注意到了，current不被包含在回归模型中，尽管我们生成了它。原因是d_j、dj、current无法同时被放进模型，否则会产生严格多重共线性问题，Stata会自动在“d_4 d_3 d_2 d_1 current d1 d2 d3 d4 d5”中随机drop掉一个变量（哪个被drop掉与顺序有关）。为了进行平行趋势检验，我们应该在d_j与current中选择一个变量并手动去掉。如果你发现检验结果不理想，可以尝试调整drop对象。<mark>被drop掉的那个变量，我们称之为基期。</mark> 若d_j的回归系数是不显著的，说明d_j的系数与基期没有显著差异，从而支持了平行趋势假设。本文的例子是以current为基期，如果你希望改变基期的位置，我在49楼写了一个以d_1为基期的代码示例。</p>
<p>尽管有些人还认为，可以在dj中选择一个变量去掉，但严格意义上这是不合适的。若所有d_j系数均不显著倒也无妨，同样可以说明平行趋势假定成立，但如果所有d_j的系数均显著为正（或负），那么我们无从判断d_j中任意两个回归系数是否有显著差异。另外，偶尔也会见到一些不规范的做法，比如在不采劝缩尾处理策略”情况下（前文对此策略已经用例子介绍了），只对current附近几期进行平行趋势检验，这种情况下，如果你得到了一些显著的d_j，你可能会误以为平行趋势检验没有通过。总之，基期的选择对平行趋势检验的结果是有影响的，请不要忽略这个问题。基于上述观点，tvdiff这个专门用来进行平行趋势检验的命令，由于不允许指定基期，因此价值可能有限。</p>
<h2 id="平行趋势检验的stata命令——绘图法">平行趋势检验的Stata命令——绘图法</h2>
<p>平行趋势检验的绘图法需要你先完成回归法的所有步骤，也就是在执行下面这条命令之后，才可以进行绘图法。下面这个回归的结果，已经可以用来判断检验是否通过了。若检验没有通过，绘图法也就没必要做了。</p>
<pre class=" language-python"><code class="prismlanguage-python">xtreg y d_4 d_3 d_2 d_1 d1 d2 d3 d4 d5 i<span class="token punctuation">.</span>year 控制变量<span class="token punctuation">,</span> fe r
</code></pre>
<p>如果你认为以下内容有一定操作难度，那么你可以考虑放弃用绘图法来展示平行趋势检验的结果。正如前文所述，绘图法与回归法是等价的，绘图法对论文的意义只是“锦上添花”。</p>
<p>方法1：采用coefplot绘制简易图形</p>
<pre class=" language-python"><code class="prismlanguage-python">ssc install coefplot
coefplot<span class="token punctuation">,</span>keep<span class="token punctuation">(</span> d_4 d_3 d_2 d_1 d1 d2 d3 d4 d5 <span class="token punctuation">)</span> levels<span class="token punctuation">(</span><span class="token number">9&#48;</span><span class="token punctuation">)</span> vertical lcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span> mcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span> msymbol<span class="token punctuation">(</span>circle_hollow<span class="token punctuation">)</span> ytitle<span class="token punctuation">(</span>回归系数<span class="token punctuation">,</span> size<span class="token punctuation">(</span>small<span class="token punctuation">)</span><span class="token punctuation">)</span>ylabel<span class="token punctuation">(</span><span class="token punctuation">,</span> labsize<span class="token punctuation">(</span>small<span class="token punctuation">)</span> angle<span class="token punctuation">(</span>horizontal<span class="token punctuation">)</span> nogrid<span class="token punctuation">)</span> yline<span class="token punctuation">(</span><span class="token number">&#48;</span><span class="token punctuation">,</span> lwidth<span class="token punctuation">(</span>vthin<span class="token punctuation">)</span>lpattern<span class="token punctuation">(</span>solid<span class="token punctuation">)</span> lcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span><span class="token punctuation">)</span> xtitle<span class="token punctuation">(</span>政策实施相对时间<span class="token punctuation">,</span> size<span class="token punctuation">(</span>small<span class="token punctuation">)</span><span class="token punctuation">)</span> xlabel<span class="token punctuation">(</span><span class="token punctuation">,</span>labsize<span class="token punctuation">(</span>small<span class="token punctuation">)</span><span class="token punctuation">)</span>graphregion<span class="token punctuation">(</span>fcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span> lcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span> ifcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span> ilcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span><span class="token punctuation">)</span> ciopts<span class="token punctuation">(</span>recast<span class="token punctuation">(</span>rcap<span class="token punctuation">)</span><span class="token punctuation">)</span> xline<span class="token punctuation">(</span><span class="token number">1&#48;.5</span><span class="token punctuation">,</span> lwidth<span class="token punctuation">(</span>vthin<span class="token punctuation">)</span> lpattern<span class="token punctuation">(</span>solid<span class="token punctuation">)</span>lcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span><span class="token punctuation">)</span>
</code></pre>
<p>Stata15会出现字体问题；如果你是Stata14，字体会默认为宋体，因此推荐用Stata14来作图。</p>
<p>本文使用Beck et al.（2&#48;1&#48;）提供的数据，参照上面的命令可以得到的图片大致如下，部分地方仍需要进一步手工调整。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/DvxzspKforHnLqe.png" alt="Graph1.png" width="8&#48;&#48;" height="56&#48;"></p>
<p>例如，调整x轴和y轴的标签。<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/KnH3z6g2BdAw8U1.png" alt="12.png" width="8&#48;&#48;" height="45&#48;"></p>
<p>此处展示一个简单调整以后的图：<br>
<img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/SVD7tWh13fQEcH9.png" alt="Graph11.png" width="8&#48;&#48;" height="57&#48;"></p>
<p>上面这个图有个缺陷，就是无法在基期（current）的位置形成一个断点，优点是过程简单方便。一个可行的替代方案是，你可以将基期设定在最开头（即d_4），即回归模型设置为：</p>
<pre class=" language-python"><code class="prismlanguage-python">xtreg yd_3 d_2 d_1 current d1 d2 d3 d4 d5 i<span class="token punctuation">.</span>year 控制变量<span class="token punctuation">,</span> fe r
</code></pre>
<p><strong>方法2：命令直接绘图（非常推荐）</strong></p>
<p><strong>第一步：手工计算置信区间</strong></p>
<p>首先考虑置信区间是如何被计算出来的。</p>
<p>为计算9&#48;%置信区间，需要计算t统计量，最精确的命令是：gen t = invttail ( d2 , &#48;.&#48;5 )</p>
<p>其中，参数d2是F统计量的第二个自由度，可以从前文的xtreg回归中获得这个数字。</p>
<p><img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/3oPzAVS1H7QCeZy.png" alt="13.png"></p>
<p><img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/qETHAD3XzQ9uI7B.png" alt="14.png"></p>
<pre class=" language-python"><code class="prismlanguage-python">gen t <span class="token operator">=</span> invttail <span class="token punctuation">(</span> <span class="token number">48</span> <span class="token punctuation">,</span> <span class="token number">&#48;.&#48;5</span> <span class="token punctuation">)</span>

<span class="token operator">*</span>生成b_j的系数与置信区间（假设有<span class="token number">4</span>期，即distance最小值是<span class="token operator">-</span><span class="token number">4</span>；你可以手动调整下面命令的参数）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">4</span> <span class="token punctuation">{</span>
gen b_`i<span class="token string">' = _b[d_`i'</span><span class="token punctuation">]</span>
gen se_b_`i<span class="token string">' = _se[d_`i'</span><span class="token punctuation">]</span>
gen b_`i<span class="token string">'LB = b_`i'</span> <span class="token operator">-</span> t <span class="token operator">*</span> se_b_`i'
gen b_`i<span class="token string">'UB = b_`i'</span> <span class="token operator">+</span> t <span class="token operator">*</span> se_b_`i'
<span class="token punctuation">}</span>

<span class="token operator">*</span>生成bj的系数与置信区间（假设有<span class="token number">5</span>期，即distance最大值是<span class="token number">5</span>）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">5</span> <span class="token punctuation">{</span>
gen b`i<span class="token string">' = _b[d`i'</span><span class="token punctuation">]</span>
gen se_b`i<span class="token string">' = _se[d`i'</span><span class="token punctuation">]</span>
gen b`i<span class="token string">'LB = b`i'</span> <span class="token operator">-</span> t <span class="token operator">*</span> se_b`i'
gen b`i<span class="token string">'UB = b`i'</span> <span class="token operator">+</span> t <span class="token operator">*</span> se_b`i'
<span class="token punctuation">}</span>
</code></pre>
<hr>
<p><strong>第二步：将上述计算结果集中反映在三个变量中</strong></p>
<pre class=" language-python"><code class="prismlanguage-python">gen b <span class="token operator">=</span> <span class="token punctuation">.</span>
gen LB <span class="token operator">=</span> <span class="token punctuation">.</span>
gen UB <span class="token operator">=</span> <span class="token punctuation">.</span>

<span class="token operator">*</span>生成系数（政策前）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">4</span> <span class="token punctuation">{</span>
replace b <span class="token operator">=</span> b_`i<span class="token string">'if distance == -`i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>生成系数（政策后）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">5</span><span class="token punctuation">{</span>
replace b <span class="token operator">=</span> b`i<span class="token string">'if distance == `i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>生成系数置信区间下限（政策前）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">4</span> <span class="token punctuation">{</span>
replace LB <span class="token operator">=</span> b_`i<span class="token string">'LB if distance == -`i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>生成系数置信区间下限（政策后）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">5</span> <span class="token punctuation">{</span>
replace LB <span class="token operator">=</span> b`i<span class="token string">'LB if distance == `i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>生成系数置信区间上限（政策前）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">4</span> <span class="token punctuation">{</span>
replace UB <span class="token operator">=</span> b_`i<span class="token string">'UB if distance == -`i'</span>
<span class="token punctuation">}</span>

<span class="token operator">*</span>生成系数置信区间上限（政策后）
forvalues i<span class="token operator">=</span><span class="token number">1</span><span class="token operator">/</span><span class="token number">5</span> <span class="token punctuation">{</span>
replace UB <span class="token operator">=</span> b`i<span class="token string">'UB if distance == `i'</span>
<span class="token punctuation">}</span>
</code></pre>
<p><strong>第三步：删除重复值与无用的变量</strong></p>
<p>绘图不会用到那么多变量与样本，只需要保留关键数据即可。</p>
<pre class=" language-python"><code class="prismlanguage-python">keep distance b LB UB
duplicates drop distance<span class="token punctuation">,</span>force
sort distance
</code></pre>
<p><strong>第四步：绘制图片</strong></p>
<p>下面是绘图命令，与coefplot一样，即使有了绘图设定，很多地方还是需要你手动调整。如果你了解graph editor，非常推荐直接在editor中对图片进行加工。</p>
<pre class=" language-python"><code class="prismlanguage-python">twoway <span class="token punctuation">(</span>connected b distance<span class="token punctuation">,</span> sort lcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span> mcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span> msymbol<span class="token punctuation">(</span>circle_hollow<span class="token punctuation">)</span> cmissing<span class="token punctuation">(</span>n<span class="token punctuation">)</span><span class="token punctuation">)</span><span class="token punctuation">(</span>rcap LB UB distance<span class="token punctuation">,</span> lcolor<span class="token punctuation">(</span>black<span class="token punctuation">)</span>lpattern<span class="token punctuation">(</span>dash<span class="token punctuation">)</span> msize<span class="token punctuation">(</span>medium<span class="token punctuation">)</span><span class="token punctuation">)</span><span class="token punctuation">,</span>ytitle<span class="token punctuation">(</span>Percentage change<span class="token punctuation">)</span> ytitle<span class="token punctuation">(</span><span class="token punctuation">,</span> size<span class="token punctuation">(</span>small<span class="token punctuation">)</span><span class="token punctuation">)</span>yline<span class="token punctuation">(</span><span class="token number">&#48;</span><span class="token punctuation">,</span> lwidth<span class="token punctuation">(</span>vthin<span class="token punctuation">)</span> lpattern<span class="token punctuation">(</span>dash<span class="token punctuation">)</span> lcolor<span class="token punctuation">(</span>teal<span class="token punctuation">)</span><span class="token punctuation">)</span> ylabel<span class="token punctuation">(</span><span class="token punctuation">,</span> labsize<span class="token punctuation">(</span>small<span class="token punctuation">)</span> angle<span class="token punctuation">(</span>horizontal<span class="token punctuation">)</span> nogrid<span class="token punctuation">)</span> xtitle<span class="token punctuation">(</span>Years relative to branch deregulation<span class="token punctuation">)</span> xtitle<span class="token punctuation">(</span><span class="token punctuation">,</span> size<span class="token punctuation">(</span>small<span class="token punctuation">)</span><span class="token punctuation">)</span> xline<span class="token punctuation">(</span><span class="token number">&#48;</span><span class="token punctuation">,</span> lwidth<span class="token punctuation">(</span>vthin<span class="token punctuation">)</span> lpattern<span class="token punctuation">(</span>dash<span class="token punctuation">)</span> lcolor<span class="token punctuation">(</span>teal<span class="token punctuation">)</span><span class="token punctuation">)</span>xlabel<span class="token punctuation">(</span><span class="token operator">-</span><span class="token number">2</span><span class="token punctuation">(</span><span class="token number">1</span><span class="token punctuation">)</span><span class="token number">2</span><span class="token punctuation">,</span> labsize<span class="token punctuation">(</span>small<span class="token punctuation">)</span><span class="token punctuation">)</span>xmtick<span class="token punctuation">(</span><span class="token operator">-</span><span class="token number">2</span><span class="token punctuation">(</span><span class="token number">1</span><span class="token punctuation">)</span><span class="token number">2</span><span class="token punctuation">,</span> nolabels ticks<span class="token punctuation">)</span>legend<span class="token punctuation">(</span>off<span class="token punctuation">)</span>graphregion<span class="token punctuation">(</span>fcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span> lcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span> ifcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span> ilcolor<span class="token punctuation">(</span>white<span class="token punctuation">)</span><span class="token punctuation">)</span>
</code></pre>
<p><img src="https://i.loli.net/2&#48;2&#48;/&#48;3/22/a8iRqbp7TJLnXl6.png" alt="Graph2.png"><br>
优点是图片美观，基期位置生成了一个断点，这是coefplot无法做到的。最重要的是，这种方法允许你以任意d_j为基期，并且也允许你相应位置形成断点（本帖没有给如何在d_j位置形成断点的方法）。缺点是整个绘图过程比较麻烦。</p>
<p>当政策时点一致时，不少论文选择用两期DID的设定（即Dt、Gi、Dt<em>Gi）来回归多期数据，这是一种不合适的处理方法。如果你用的命令是reg，那么信息多的ui就会被信息少的Gi取代，内生性问题可能很严重；如果你用的命令是xtreg-,fe，那么Gi的系数将会无法被固定效应模型估计出系数，因为Gi通常不随时间改变，所以不少求助帖子会问分组变量Gi的系数为什么ommited了。一种不好的做法可能是这样的，以reg Dt、Gi、Dt</em>Gi i.year对多期数据进行回归，模型中还额外控制了时间虚拟变量。正如前文所述，Dt与时间虚拟变量存在很强的联系，因此如果同时在模型中放入两者，容易导致Dt、Dt<em>Gi的方差膨胀因子增大，进而使Dt和Dt</em>Gi系数的显著性下降。如果你的政策实施时点是一致的，你也希望用多期DID进行回归（实际上也本应如此），那么你的实际操作过程会更简单，但是本文没有把这种特例单独列出来讨论。你可以按照本文的思路照搬照做，就当它是时点不一致，只需要在最后一步绘图中做出修改，将x轴可以直接写某某年，而不是“-1”“-2”这样子。最后需要指出的是，两期DID只是多期DID的特例，切勿用两期DID的思想套到多期DID身上，这句话很重要，再怎么强调也不过分。但是，你可以拿多期DID的思想直接套在两期DID上，如果你能“反套”成功，那就真的理解了多期DID了。</p>
<pre><code></code></pre>
<p><strong>参考文献</strong><br>
Beck T, Levine R, Levkov A. Big bad banks? The winners and losers from bank deregulation in the United States[J]. The Journal of Finance, 2&#48;1&#48;, 65(5): 1637-1667.</p>
</div>
</div>