topic SAS build non-linear model to predict missing values in Statistical Procedures

SAS build non-linear model to predict missing values

GeorgeSAS — Wed, 19 Jul 2017 00:15:50 GMT

Hello everyone,

I have a dataset with two missing values.I hope I can predict these two value from other rows.(curved line)

so I want build a regression model (this is not a linear line,but a curved like line)to predict the missing values when var in (98,99).

please advise me.

Thank you very much!

data have;
input value var;
cards;
75.447     1
74.628     2
73.737     3
72.846     4
71.964     5
71.064     6
70.173     7
69.273     8
68.382     9
67.482    10
66.582    11
65.691    12
64.800    13
63.900    14
63.009    15
62.118    16
61.227    17
60.345    18
59.463    19
58.590    20
57.735    21
56.871    22
56.007    23
55.134    24
54.252    25
53.388    26
52.506    27
51.624    28
50.742    29
49.860    30
48.969    31
48.087    32
47.214    33
46.332    34
45.450    35
44.577    36
43.695    37
42.822    38
41.940    39
41.067    40
40.185    41
39.330    42
38.466    43
37.602    44
36.747    45
35.883    46
35.028    47
34.173    48
33.318    49
32.490    50
31.653    51
30.816    52
29.988    53
29.160    54
28.359    55
27.540    56
26.730    57
25.929    58
25.119    59
24.309    60
23.517    61
22.752    62
21.969    63
21.195    64
20.457    65
19.710    66
18.954    67
18.207    68
17.469    69
16.776    70
16.065    71
15.381    72
14.697    73
14.049    74
13.401    75
12.771     76
12.168     77
11.565     78
10.980     79
10.404     80
 9.846     81
 9.288     82
 8.784     83
 8.289     84
 7.794     85
 7.335     86
 6.903     87
 6.489     88
 6.102     89
 5.733     90
 5.346     91
 5.040     92
 4.734     93
 4.410     94
 4.077     95
 3.780     96
 3.483     97
 .         98
 .         99
 2.357      100
;
run;

let me get a start model;

proc sgscatter data=have;
  plot value*var/ 
        reg=(nogroup clm degree=2) grid ;
run;

I want use regression like this line to estimate values when var in(98,99). please help me find out how to do this.

Thanks!

Re: SAS build non-linear model to predict missing values

Ksharp — Wed, 19 Jul 2017 13:40:32 GMT

PROC ADAPTIVE

PROC GAML

PROC LOESS

Re: SAS build non-linear model to predict missing values

Ksharp — Wed, 19 Jul 2017 13:44:21 GMT

proc loess data=have;
model value=var;
output out=want;
run;
proc sgscatter data=want;;
  plot Predicted*var/ grid ;
run;

Re: SAS build non-linear model to predict missing values

GeorgeSAS — Wed, 19 Jul 2017 18:39:53 GMT

Thank you Ksharp,
I can't run the program code on SAS 9.3,can you tell me the predicted value when var in (98,99) from your method?

Thanks!

Re: SAS build non-linear model to predict missing values

Rick_SAS — Wed, 19 Jul 2017 19:05:14 GMT

Unfortunately, PROC LOESS drops observations for which ANY variable has a missing value. However, you can use any other nonparametric regression procedure, such as PROC TPSPLINE or PROC TRANSREG, which handle spline fits.

proc transreg data=have;
   model identity(value) = spline(var);
   output out=Want predicted;
run;

proc print data=want;
where Var > 96;
var Var Value PValue;
run;

Re: SAS build non-linear model to predict missing values

Ksharp — Thu, 20 Jul 2017 12:56:03 GMT

SAS Output

SAS 系统

value	var	SmoothingParameter	DepVar	Obs	Predicted
75.447	1	0.045918	value	1	75.4556
74.628	2	0.045918	value	2	74.6074
73.737	3	0.045918	value	3	73.7370
72.846	4	0.045918	value	4	72.8486
71.964	5	0.045918	value	5	71.9588
71.064	6	0.045918	value	6	71.0666
70.173	7	0.045918	value	7	70.1704
69.273	8	0.045918	value	8	69.2756
68.382	9	0.045918	value	9	68.3794
67.482	10	0.045918	value	10	67.4820
66.582	11	0.045918	value	11	66.5846
65.691	12	0.045918	value	12	65.6910
64.800	13	0.045918	value	13	64.7974
63.900	14	0.045918	value	14	63.9026
63.009	15	0.045918	value	15	63.0090
62.118	16	0.045918	value	16	62.1180
61.227	17	0.045918	value	17	61.2296
60.345	18	0.045918	value	18	60.3450
59.463	19	0.045918	value	19	59.4656
58.590	20	0.045918	value	20	58.5952
57.735	21	0.045918	value	21	57.7324
56.871	22	0.045918	value	22	56.8710
56.007	23	0.045918	value	23	56.0044
55.134	24	0.045918	value	24	55.1314
54.252	25	0.045918	value	25	54.2572
53.388	26	0.045918	value	26	53.3828
52.506	27	0.045918	value	27	52.5060
51.624	28	0.045918	value	28	51.6240
50.742	29	0.045918	value	29	50.7420
49.860	30	0.045918	value	30	49.8574
48.969	31	0.045918	value	31	48.9716
48.087	32	0.045918	value	32	48.0896
47.214	33	0.045918	value	33	47.2114
46.332	34	0.045918	value	34	46.3320
45.450	35	0.045918	value	35	45.4526
44.577	36	0.045918	value	36	44.5744
43.695	37	0.045918	value	37	43.6976
42.822	38	0.045918	value	38	42.8194
41.940	39	0.045918	value	39	41.9426
41.067	40	0.045918	value	40	41.0644
40.185	41	0.045918	value	41	40.1927
39.330	42	0.045918	value	42	39.3274
38.466	43	0.045918	value	43	38.4660
37.602	44	0.045918	value	44	37.6046
36.747	45	0.045918	value	45	36.7444
35.883	46	0.045918	value	46	35.8856
35.028	47	0.045918	value	47	35.0280
34.173	48	0.045918	value	48	34.1730
33.318	49	0.045918	value	49	33.3257
32.490	50	0.045918	value	50	32.4874
31.653	51	0.045918	value	51	31.6530
30.816	52	0.045918	value	52	30.8186
29.988	53	0.045918	value	53	29.9880
29.160	54	0.045918	value	54	29.1677
28.359	55	0.045918	value	55	28.3538
27.540	56	0.045918	value	56	27.5426
26.730	57	0.045918	value	57	26.7326
25.929	58	0.045918	value	58	25.9264
25.119	59	0.045918	value	59	25.1190
24.309	60	0.045918	value	60	24.3142
23.517	61	0.045918	value	61	23.5247
22.752	62	0.045918	value	62	22.7468
21.969	63	0.045918	value	63	21.9716
21.195	64	0.045918	value	64	21.2053
20.457	65	0.045918	value	65	20.4544
19.710	66	0.045918	value	66	19.7074
18.954	67	0.045918	value	67	18.9566
18.207	68	0.045918	value	68	18.2096
17.469	69	0.045918	value	69	17.4819
16.776	70	0.045918	value	70	16.7708
16.065	71	0.045918	value	71	16.0727
15.381	72	0.045918	value	72	15.3810
14.697	73	0.045918	value	73	14.7073
14.049	74	0.045918	value	74	14.0490
13.401	75	0.045918	value	75	13.4062
12.771	76	0.045918	value	76	12.7787
12.168	77	0.045918	value	77	12.1680
11.565	78	0.045918	value	78	11.5702
10.980	79	0.045918	value	79	10.9826
10.404	80	0.045918	value	80	10.4092
9.846	81	0.045918	value	81	9.8460
9.288	82	0.045918	value	82	9.3035
8.784	83	0.045918	value	83	8.7866
8.289	84	0.045918	value	84	8.2890
7.794	85	0.045918	value	85	7.8043
7.335	86	0.045918	value	86	7.3427
6.903	87	0.045918	value	87	6.9082
6.489	88	0.045918	value	88	6.4967
6.102	89	0.045918	value	89	6.1072
5.733	90	0.045918	value	90	5.7278
5.346	91	0.045918	value	91	5.3692
5.040	92	0.045918	value	92	5.0400
4.734	93	0.045918	value	93	4.7288
4.410	94	0.045918	value	94	4.4074
4.077	95	0.045918	value	95	4.0873
3.780	96	0.045918	value	96	3.7800
3.483	97	0.045918	value	97	3.4830
.	98	0.045918	value	98	3.1084
.	99	0.045918	value	99	2.7337
2.357	100	0.045918	value	100	2.3591

Re: SAS build non-linear model to predict missing values

GeorgeSAS — Thu, 03 Aug 2017 16:11:07 GMT

Thank you!
Rick, may I ask what is the model theory of the transreg procedure? like general linear model? or something else?

Re: SAS build non-linear model to predict missing values

Rick_SAS — Thu, 03 Aug 2017 16:34:17 GMT

It transforms data (nonlinearly) and then fits a linear model to the transformed data.

For an overview of the different models and transformations, see

http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_introreg_sect026.htm

Re: SAS build non-linear model to predict missing values

GeorgeSAS — Thu, 03 Aug 2017 18:31:49 GMT

Thank you Rick,
what is the mythology/theory behind this exact model? least square ? or something else?

Thank you!

Re: SAS build non-linear model to predict missing values

Rick_SAS — Thu, 03 Aug 2017 19:03:48 GMT

See the TRANSREG doc for specifics. For regular data analysis, it uses ordinary least squares (OLS). For optimal variable transformations, it iterates between OLS estimates of the parameters and OLS estimates of the transformation parameters (a method that is called the method of alternating least squares).