Undergrad Calculating a standard deviation involving weighted samples

[Buzz Bloom]

TL;DR

I used a spreadsheet (PNG image in main body section) to analyze 12 values with +/- error ranges which I assume for the purpose of this calculation to be standard deviation values. I am seeking conformation or corrections regarding the validity of the method I used.

Below is the spreadsheet image.

The source of the data in the two columns, Ho and +/-, is the following article

http://planck.caltech.edu/pub/2015results/Planck_2015_Results_XIII_Cosmological_Parameters.pdf .​

​

Let AHo be the conjoined weighted average values for Ho and the corresponding error ranges.

[1] AHo = AH +/- AE = 67.65 +/- 0.22.​

The Ho column contains the 12 values, H_i, for the Ho variable, i being the index (of a variable value) ranging from 1 to 12.

The +/- column contains for each H_i value the corresponding value of an error range, E_i, which I assume to be of 1 standard deviation σ_i.

The Source column specifies where in the article the particular 12 values for H_i and E_i are found.

The Wt=... column contains values, W_i, for the weights for calculating weighted sums. I assume that the appropriate choice of weight are the squares of the reciprocals of the E_i values.

[2] W_i = 1/E_i²​

The sum of this column is

[3] ΣW = Σ[i=1 to 12] W_i = 25.8982 .​

The Wt*Ho column contains weighted Ho values

[4] WH_i = W_i H_i.​

The sum of this column is

[5] ΣWH = Σ[i=1 to 12] WH_i = 1751.9631 .​

​

The weighted average, AH, is

[6] AH = ΣWH / ΣW = 67.65.​

I feel confident that this result is OK for AH. I am less confident that my calculation below of AE is OK. I will present my analysis, and explain my main concern about AE, in what follows.

The D2^2=... column contains values, D2_i, of the squares of the 12 differences between AH and a value of Ho.

[7] D2_i = (Ho-H_i)²​

Althought not part of the calculations, it will later be useful to also have defined the sum of the differences squared.

[8] ΣD2 = Σ[i=1 to 12] (D2_i)​

The Wt*D2 column contains values, WD2_i, of the product of a weight and a squared difference.​

[9] WD2_i = W_i D2_i​

[10] ΣWD2 = Σ[i=1 to 12] (WD2_i) = 1.2992​

I now define Aσ as square root of a quotient: the weighted sum of squared differences and the sum of weights. Using [3] and [10]

[11] Aσ = (WD2/ΣW)^1/2 = 1.2992/25.8982 = 0.22​

​

The main reason I am uncertain is that if all of the E_i values were the same, say 1/s, then then the sum of the weights, ΣW, would be

[12] ΣW = (12/s)².​

Then the sum of the weighted squares of differences, ΣWD, would be

[`13] ΣWD2 = (1/s²) Σ[i=1 to 12] (D2_i)​

= (1/s)²)ΣD2​

and Aσ would be

[14] Aσ = [(1/12)ΣD2]^1/21/2.​

Ordinarily, since there are 12 samples I would expect

[15] Aσ = [(1/11) ΣD2]^1/21/2.​

Equation [15] is based on the article

https://en.wikipedia.org/wiki/Standard_deviation​

I do not know how to adjust this for weighted samples.

Another minor concern is that my assumtion to use 1/E_i² as weights may not be the right approach. However, I think I have a good rational for doing this, but it is rather complicated, so I will not include it in this post. If anyone would like to see this rationale, I will post it later.

 

[mathman]

You can directly calculate the second moments using the weights. Then use the usual formula for the variance as an expression involving the first two moments.