z****h
发帖数: 203 | 1 关于捐款的，有历史数据，包含捐款人的各种信息的predictor variables， 还包括这
个人捐款数量（可能很多，也可能很少， 更可能是0）， 我开始回答说用linear
regression model来model 捐款数量和predictors的关系，但是interviewer说如果那
些捐款人里面99%的都是捐款为0，只有1%的捐款了，数量还不齐。那这样model会很有
问题，请问什么办法可以解决。
我已经最近被两个公司问了这种类似问题，一个是APPLE，一个是今天面试的一个银行。 |
k****n
发帖数: 165 | 2 For regression, usually a transformation for the response variable is
required if its distribution is skewed. If 99% are 0, a common approach is
to split the model into 1) predict whether an observation will donate, 2) if
yes to 1), how much will he donate. |
z****h
发帖数: 203 | 3 Thanks. Are you saying I should use logistic regression model first, then
use linear model on the non 0% part? Fitting two models separately?
if
【在 k****n 的大作中提到】
: For regression, usually a transformation for the response variable is
: required if its distribution is skewed. If 99% are 0, a common approach is
: to split the model into 1) predict whether an observation will donate, 2) if
: yes to 1), how much will he donate.
|
k****n
发帖数: 165 | 4 Yes. Google zero inflated regression for more details.
BTW, I'm wondering what kind of position were you interviewing for APPLE.
This is a very statistical question. |
s********0
发帖数: 2625 | 5 你说的是two-stage method，比较常用，因为简单，易懂。
复杂一点，可以google censored regression model and tobit model.我知道stata里
有现成的package，SAS/R不太清楚。
【在 z****h 的大作中提到】
: Thanks. Are you saying I should use logistic regression model first, then
: use linear model on the non 0% part? Fitting two models separately?
:
: if
|
z****h
发帖数: 203 | 6 the Apple position is Statistician/Algorithm Analys
【在 k****n 的大作中提到】
: Yes. Google zero inflated regression for more details.
: BTW, I'm wondering what kind of position were you interviewing for APPLE.
: This is a very statistical question.
|
