Odd Results when generating predictions with nnet function

Why would you expect raising y_pred to the 0.5 to "backtransform" a model
sqrt(y)~x? Wouldn't you raise to the 2?

Why would you "backtransform" x in such a model if it were never
transformed in the first place? Dr Maechler did not suggest that.

And why are you mentioning some random unspecified video on Youtube? That
does not enlighten anyone here, apparently including you. Please reference
package documentation, and/or reproduce the analysis discussed in that
video to provide a contrasting (or supporting) point with the example you
gave.


On September 2, 2020 7:21:58 AM PDT, Paul Bernal <paulbernal07 at gmail.com>
wrote:

Dear Dr. Martin and Dr. Peter,

Hope you are doing well. Thank you for your kind feedback. I also tried
fitting the nnet using y ~ x, but the model kept on generating odd
predictions. If I understand correctly, from what Dr. Martin said, it
would
be a good idea to try modeling sqrt(y) ~ x and then backtransform
raising
both y and x to 0.5?

I was looking at a video where the guy modeled count data without doing
any
kind of transformation and didn't get odd results, which is rather
extrange.

Cheers,

Paul



El mi?., 2 sept. 2020 a las 2:37, Martin Maechler (<
maechler at stat.math.ethz.ch>) escribi?:

peter dalgaard
    on Wed, 2 Sep 2020 08:41:09 +0200 writes:

    > Generically, nnet(a$y ~ a$x, a ...) should be nnet(y ~ x,
    > data=a, ...) otherwise predict will go looking for a$x, no
    > matter what is in xnew.

    > But more importantly, nnet() is a _classifier_,
    > so the LHS should be a class, not a numeric variable.

    > -pd

Well, nnet() can be used for both classification *and* regression,
which is quite clear from the MASS book, but indeed, not from
its help page, which indeed mentions one formula  'class ~ ...'
and then only has classification examples.

So, indeed, the  ?nnet  help page could improved.

In his case, y are counts,  so  John Tukey's good old
"first aid transformation" principle would suggest to model

sqrt(y) ~ ..   in a *regression* model which nnet() can do.

Martin Maechler
ETH Zurich  and  R Core team

    >> On 1 Sep 2020, at 22:19 , Paul Bernal
    >> <paulbernal07 at gmail.com> wrote:
    >>
    >> Dear friends,
    >>
    >> Hope you are all doing well. I am currently using R
    >> version 4.0.2 and working with the nnet package.
    >>
    >> My dataframe consists of three columns, FECHA which is
    >> the date, x, which is a sequence from 1 to 159, and y,
    >> which is the number of covid cases (I am also providing
    >> the dput for this data frame below).
    >>
    >> I tried fitting a neural net model using the following
    >> code:
    >>
    >> xnew = 1:159 Fit <- nnet(a$y ~ a$x, a, size = 5, maxit =
    >> 1000, lineout = T, decay = 0.001)
    >>
    >> Finally, I attempted to generate predictions with the
    >> following code:
    >>
    >> predictions <- predict(Fit, newdata = list(x = xnew),
    >> type = "raw")
    >>
    >> But obtained extremely odd results: As you can see,
    >> instead of obtaining numbers, more or less in the range
    >> of the last observations of a$y, I end up getting a bunch
    >> of 1s, which doesn?t make any sense (if anyone could help
    >> me understand what could be causing this):
    >> dput(predictions) structure(c(1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    >> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1), .Dim
    >> = c(159L, 1L), .Dimnames = list(c("1", "2", "3", "4",
    >> "5", "6", "7", "8", "9", "10", "11", "12", "13", "14",
    >> "15", "16", "17", "18", "19", "20", "21", "22", "23",
    >> "24", "25", "26", "27", "28", "29", "30", "31", "32",
    >> "33", "34", "35", "36", "37", "38", "39", "40", "41",
    >> "42", "43", "44", "45", "46", "47", "48", "49", "50",
    >> "51", "52", "53", "54", "55", "56", "57", "58", "59",
    >> "60", "61", "62", "63", "64", "65", "66", "67", "68",
    >> "69", "70", "71", "72", "73", "74", "75", "76", "77",
    >> "78", "79", "80", "81", "82", "83", "84", "85", "86",
    >> "87", "88", "89", "90", "91", "92", "93", "94", "95",
    >> "96", "97", "98", "99", "100", "101", "102", "103",
    >> "104", "105", "106", "107", "108", "109", "110", "111",
    >> "112", "113", "114", "115", "116", "117", "118", "119",
    >> "120", "121", "122", "123", "124", "125", "126", "127",
    >> "128", "129", "130", "131", "132", "133", "134", "135",
    >> "136", "137", "138", "139", "140", "141", "142", "143",
    >> "144", "145", "146", "147", "148", "149", "150", "151",
    >> "152", "153", "154", "155", "156", "157", "158", "159"),
    >> NULL))
    >>
    >> head(a) FECHA x y 1 2020-03-09 1 1 2 2020-03-10 2 8 3
    >> 2020-03-11 3 14 4 2020-03-12 4 27 5 2020-03-13 5 36 6
    >> 2020-03-14 6 43
    >>
    >> dput(a) structure(list(FECHA = structure(c(18330, 18331,
    >> 18332, 18333, 18334, 18335, 18336, 18337, 18338, 18339,
    >> 18340, 18341, 18342, 18343, 18344, 18345, 18346, 18347,
    >> 18348, 18349, 18350, 18351, 18352, 18353, 18354, 18355,
    >> 18356, 18357, 18358, 18359, 18360, 18361, 18362, 18363,
    >> 18364, 18365, 18366, 18367, 18368, 18369, 18370, 18371,
    >> 18372, 18373, 18374, 18375, 18376, 18377, 18378, 18379,
    >> 18380, 18381, 18382, 18383, 18384, 18385, 18386, 18387,
    >> 18388, 18389, 18390, 18391, 18392, 18393, 18394, 18395,
    >> 18396, 18397, 18398, 18399, 18400, 18401, 18402, 18403,
    >> 18404, 18405, 18406, 18407, 18408, 18409, 18410, 18411,
    >> 18412, 18413, 18414, 18415, 18416, 18417, 18418, 18419,
    >> 18420, 18421, 18422, 18423, 18424, 18425, 18426, 18427,
    >> 18428, 18429, 18430, 18431, 18432, 18433, 18434, 18435,
    >> 18436, 18437, 18438, 18439, 18440, 18441, 18442, 18443,
    >> 18444, 18445, 18446, 18447, 18448, 18449, 18450, 18451,
    >> 18452, 18453, 18454, 18455, 18456, 18457, 18458, 18459,
    >> 18460, 18461, 18462, 18463, 18464, 18465, 18466, 18467,
    >> 18468, 18469, 18470, 18471, 18472, 18473, 18474, 18475,
    >> 18476, 18477, 18478, 18479, 18480, 18481, 18482, 18483,
    >> 18484, 18485, 18486, 18487, 18488), class = "Date"), x =
    >> 1:159, y = c(1, 8, 14, 27, 36, 43, 55, 69, 86, 109, 137,
    >> 200, 245, 313, 345, 443, 558, 674, 786, 901, 989, 1075,
    >> 1181, 1317, 1475, 1673, 1801, 1988, 2100, 2249, 2528,
    >> 2752, 2974, 3234, 3400, 3472, 3574, 3751, 4016, 4210,
    >> 4273, 4467, 4658, 4821, 4992, 5166, 5338, 5538, 5779,
    >> 6021, 6200, 6378, 6532, 6720, 7090, 7197, 7387, 7523,
    >> 7731, 7868, 8070, 8282, 8448, 8616, 8783, 8944, 9118,
    >> 9268, 9449, 9606, 9726, 9867, 9977, 10116, 10267, 10577,
    >> 10926, 11183, 11447, 11728, 12131, 12531, 13015, 13463,
    >> 13837, 14095, 14609, 15044, 15463, 16004, 16425, 16854,
    >> 17233, 17889, 18586, 19211, 20059, 20686, 21422, 21962,
    >> 22597, 23351, 24274, 25222, 26030, 26752, 27314, 28030,
    >> 29037, 29905, 30658, 31686, 32785, 33550, 34463, 35237,
    >> 35995, 36983, 38149, 39334, 40291, 41251, 42216, 43257,
    >> 44352, 45633, 47177, 48096, 49243, 50373, 51408, 52261,
    >> 53468, 54426, 55153, 55906, 56817, 57993, 58864, 60296,
    >> 61442, 62223, 63269, 64191, 65256, 66383, 67453, 68456,
    >> 69424, 70231, 71418, 72560, 73651, 74492, 75394, 76464,
    >> 77377, 78446, 79402)), row.names = c(NA, 159L), class =
    >> "data.frame") Any help and/or guidance will be greatly
    >> appreciated,
    >>
    >> Cheers,
    >>
    >> Paul
    >>
    >> [[alternative HTML version deleted]]
    >>
    >> ______________________________________________

    > --
    > Peter Dalgaard, Professor, Center for Statistics,
    > Copenhagen Business School Solbjerg Plads 3, 2000
    > Frederiksberg, Denmark Phone: (+45)38153501 Office: A 4.23
    > Email: pd.mes at cbs.dk Priv: PDalgd at gmail.com

    > ______________________________________________

      [[alternative HTML version deleted]]

--
Sent from my phone. Please excuse my brevity.