Stata中ARIMA模型观测值不足报错的解决方法咨询

问题说明 #

我要在Stata里构建月度自回归时间序列模型，分析avg_happy是否随output变化，用了以下代码：

arima avg_happy, ar(1) ma(1)

但即使删掉部分缺失观测，还是报错：

(note: insufficient memory or observations to estimate usual starting values [2]) insufficient observations

数据示例 #

* Example generated by -dataex-. For more info, type help dataex
clear
input double date float(output avg_happy) str6 country
1552           0         0 "France"
1595           0 .12514286 "France"
1626           0     .1898 "France"
1697           0 .06636364 "France"
1707           0 .07551724 "France"
1725           0 .15442307 "France"
1740           0 .15208332 "France"
1761           0 .11060607 "France"
1767           0 .07891892 "France"
1785           0 .10645833 "France"
1801           0 .14339286 "France"
1819           0 .07742424 "France"
1840           0 .16685714 "France"
1858           0 .20857143 "France"
1870           0 .07963637 "France"
1881           0  .1796923 "France"
1901           0 .09269842 "France"
1913           0     .1825 "France"
1962           0 .11526316 "France"
1972           0 .10245614 "France"
1979           0  .1946667 "France"
1989           0   .167541 "France"
2002           0  .1596875 "France"
2016           0      .146 "France"
2023           0 .11526316 "France"
2045           0  .0768421 "France"
2055           0 .22160715 "France"
2065           0 .15531915 "France"
2068           0  .0768421 "France"
2077           0 .02607143 "France"
2087           0  .0931915 "France"
2110           0 .14847457 "France"
2118           0 .06083333 "France"
2156           0 .07423729 "France"
2188           0 .10265625 "France"
2194           0 .12807018 "France"
2204 .0041370597 .10645833 "France"
2215           0 .28175437 "France"
2225           0  .1636207 "France"
2233           0 .15208332 "France"
2244           0 .26404256 "France"
2254           0  .1946667 "France"
2264           0 .28895834 "France"
2274  .007244105  .2019149 "France"
2285           0 .16946428 "France"
2294           0  .2053125 "France"
2307           0  .1932353 "France"
2316           0 .16387755 "France"
2324           0 .13978724 "France"
2367           0 .19367346 "France"
2372           0 .27872726 "France"
2381           0         . "France"
2382           0 .26756355 "France"
2389           0 .17085107 "France"
2396           0  .2830612 "France"
2400           0 .24333334 "France"
2406           0 .16649123 "France"
2415           0  .2174468 "France"
2428           0  .1798551 "France"
2433           0   .186383 "France"
2444  .005748244  .1855932 "France"
2454           0  .1361017 "France"
2463           0      .146 "France"
2468           0 .11108695 "France"
2477           0 .11918367 "France"
2508           0  .1361017 "France"
2531           0 .17929825 "France"
2541           0  .1871795 "France"
2554           0         . "France"
2555           0  .1970552 "France"
2561           0 .28565216 "France"
2577           0 .17456523 "France"
2589   .00506966    .27375 "France"
2602           0 .24333334 "France"
2616           0 .19770834 "France"
2622           0 .15869565 "France"
2633           0 .09605263 "France"
2640           0  .2329787 "France"
2651           0 .17756757 "France"
2657           0 .14129032 "France"
2672           0 .15927273 "France"
2700           0 .21291666 "France"
2728           0  .1883871 "France"
2735           0     .2555 "France"
2749           0         . "France"
2750           0 .22420147 "France"
2761           0 .31465515 "France"
2769           0     .1825 "France"
2775           0  .1795082 "France"
2782  .011836817 .13518518 "France"
2798           0 .25131148 "France"
2811           0   .167541 "France"
2815           0 .10814815 "France"
end

解决步骤 #

1. 修复时间序列格式 #

你的date变量不是Stata识别的标准时间格式，且观测存在大量时间断档（比如1913到1962间隔49年），这会导致模型无法识别连续时序。先转换并声明时间序列：

* 假设date是从1960年1月开始的月份数，根据实际含义调整基准
gen monthly_date = ym(1960 + floor((date-1)/12), mod(date-1,12)+1)
format monthly_date %tm
sort monthly_date
tsset monthly_date

2. 正确纳入自变量 #

你要分析avg_happy与output的关系，当前代码没加自变量，应该用ARIMAX模型：

arima avg_happy output, ar(1) ma(1)

3. 清理缺失值 #

彻底删除含缺失值的观测，避免模型自动排除过多样本：

drop if missing(avg_happy, output)

4. 简化模型或调整时序 #

如果清理后样本仍不足，先尝试更简单的模型验证：

arima avg_happy output, ar(1)

若数据断档无法修复，可考虑拆分时间段分析，或改用动态面板类模型（需补充其他假设）。

内容的提问来源于stack exchange，提问作者daviddupont