[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: Missing observations
Nick Cox <firstname.lastname@example.org>
Re: st: Missing observations
Fri, 21 Jun 2013 08:34:55 +0100
You started out with what looked like a data management question about
-drop-, a topic I think I understand. Now this is a question about
analysing your data. I have never worked with returns -- indeed I can
not even remember the formula for a return.
But your problem is now, if I understand you correctly, comparing time
series of returns calculated over different time scales.
Given the serial and scale dependence here, _none_ of the standard
machinery of t tests, Mann-Whitney U tests, bootstrapping etc. carries
over. Whoever is telling you to do otherwise should be able to
explain to you why I am wrong and it is legitimate to treat returns as
independent. Why anyone would study returns if they thought that is
I find it difficult to believe that no literature exists, but you
should be able to understand why I don't know what it is.
On 20 June 2013 19:44, Csaba Kertai <email@example.com> wrote:
> Thank you Nick. Could you let me know what is not clear about this, please? Let me explain what I want to do in another way. I have 9 variables each having different number of values. These 9 variables are return variables (e.g. 1-year raw return, 2-year raw return etc.) and I need to compare the means/medians/25th/75th/90th percentiles and the percentage of positive values (within one 'group') of these variables to see whether, say, the median difference between the 1-yr raw return 'group' and the 2-yr raw return 'group' is significant. For this, I have to use traditional parametric tests (i.e. the t-test) and non-parametric bootstrapping.
> Could you help me with this, please? I've been scouring the Internet for a solution to testing percentile differences but it seems that there's not much on this particular issue.
> There are basically three things I cannot get my head round: how to test the median difference of 2 'groups' (tried 'signrank' and 'signtest' but these tests are paired tests), the percentiles difference of two 'groups', and the difference of the percentage of positive values between 2 'groups'.
> So you say that one solution could be to stack the 9 variables on top of each other and then group them by, say, inserting a second column (grouping variable) with numbers that will identify the 9 groups?
> Thank you
>> Subject: Re: st: Missing observations
>> From: firstname.lastname@example.org
>> Date: Thu, 20 Jun 2013 18:29:32 +0100
>> To: email@example.com
>> This is really isn't clear to me, but it may be that -var1- and -var2- should be stacked on top of each other.
>> On 20 Jun 2013, at 15:41, Csaba <firstname.lastname@example.org> wrote:
>>> Thank you for your reply. Yes you are right I muddled up observations with values. I meant to write values not observations. My problem is that if I use 'drop if missing(var2)' that will drop values for each variable in my data set.
>>> I need to compare the means/medians of 2 variables. Var1 has 1125 non-missing values, var2 has 169 non-missing values. I might be doing sth wrong but when I try using bootstrapping I get a message saying that I should drop any missing values as bootstrapping cannot distinguish between missing and non-missing values. That's why I want to drop missing values for Var2. Basically, I want to achieve the same result as with the unpaired two-sample mean comparison test but with bootstrapping.
>>> Thanks a lot!
>>> On 20 Jun 2013, at 12:32, Nick Cox <email@example.com> wrote:
>>>> -drop- as used here drops entire observations (outside Stata
>>>> observations are known as rows, cases, records). You seem to be under
>>>> the impression that there is an operation
>>>> drop missing values
>>>> that is somehow different from
>>>> -drop- observations
>>>> but I don't know what that would look like.
>>>> In your example if -var2- has only 169 non-missing values (_not_
>>>> observations) then
>>>> drop if missing(var2)
>>>> will leave precisely 169 observations. I don't understand how that is
>>>> a surprise or what else you want.
>>>> On 20 June 2013 11:17, Csaba Kertai <firstname.lastname@example.org> wrote:
>>>>> I need a bit of help with dropping missing observations. If I use 'drop if missing(var)' or drop if 'var'==. etc. many other observations are dropped as well. More precisely, var1 has 1125 observations and var2 has 169 observations. I want to drop missing observations for var2 but if I use drop if var2==. then this will keep only 169 observations for each variable. I only want to drop values that are missing.
* For searches and help try: